CROSS-REFERENCE TO RELATED APPLICATIONS

This application is a continuation application of U.S. patent application Ser. No. 14/167,709, entitled: METHODS, SYSTEMS, AND SOFTWARE FOR IDENTIFYING BIO-MOLECULES WITH INTERACTING COMPONENTS, filed Jan. 29, 2014, which claims benefit under 35 U.S.C. Â§119(e) to U.S. Provisional Patent Application No. 61/759,276, entitled: METHODS, SYSTEMS, AND SOFTWARE FOR IDENTIFYING BIO-MOLECULES WITH INTERACTING COMPONENTS, filed Jan. 31, 2013, and U.S. Provisional Patent Application No. 61/799,377, entitled: METHODS, SYSTEMS, AND SOFTWARE FOR IDENTIFYING BIO-MOLECULES USING MODELS OF MULTIPLICATIVE FORM, filed Mar. 15, 2013, which are herein incorporated by reference in their entireties for all purposes.

BACKGROUND

The present disclosure relates to the fields of molecular biology, molecular evolution, bioinformatics, and digital systems. More specifically, the disclosure relates to methods for computationally predicting the activity of a biomolecule and/or guiding directed evolution. Systems, including digital systems, and system software for performing these methods are also provided. Methods of the present disclosure have utility in the optimization of proteins for industrial and therapeutic use.

Protein design has long been known to be a difficult task if for no other reason than the combinatorial explosion of possible molecules that constitute searchable sequence space. The sequence space of proteins is immense and is impossible to explore exhaustively using methods currently known in the art. Because of this complexity, many approximate methods have been used to design better proteins; chief among them is the method of directed evolution. Today, directed evolution of proteins is dominated by various high throughput screening and recombination formats, often performed iteratively.

In parallel, various computational techniques have been proposed for exploring sequence-activity space. While each computational technique has advantages in certain contexts, new ways to efficiently search sequence space to identify functional proteins would be highly desirable.

SUMMARY

The present disclosure presents techniques for generating and using sequence-activity models that employ non-linear terms, particularly terms that account for interactions between two or more sub-units in a sequence. The sequence-activity models describe activities, characteristics, or properties of biological molecules as functions of various biological sequences. These non-linear terms may be â€œcross productâ€ terms that involve multiplication of two or more variables, each representing the presence (or absence) of the sub-units participating in the interaction. Some embodiments involve techniques for selecting the non-linear terms that best describe the activity of the sequence. Note that there are often far more possible non-linear interaction terms than there are true interactions between sub-units. Hence, to avoid overfitting, only a limited number of non-linear terms are typically considered and those employed should reflect interactions that appreciably affect activity.

One aspect of the disclosure provides a method of preparing a sequence-activity model that can assist in identifying biological molecules with enhanced desired activity, the method comprising: (a) receiving sequence and activity data for a plurality of biological molecules; (b) preparing a base model from the sequence and activity data, wherein the base model predicts activity as a function of the presence or absence of sub-units of the sequence; (c) preparing at least one new model by adding or subtracting at least one new interaction term to or from the base model, wherein the new interaction term represents the interaction between two or more interacting sub-units; (d) determining the ability of the at least one new model to predict activity as a function of the presence or absence of the subunits; and (e) determining whether to add or subtract the new interaction term to or from the base model based on the ability of the at least one new model to predict activity as determined in (d) and with a bias against adding the new interaction term. The model derived can then be used in various applications, such as in directed evolution of protein libraries to identify proteins with desired biological activities and properties.

In some embodiments, wherein the method determines that the new interaction term should be added to the base model to produce an updated model, the method further includes additional steps to search for additional interaction terms that can further improve the updated model. Specifically, the method includes: (f) repeating (c) using the updated model in place of the base model and adding or subtracting a different interaction term than the one added/subtracted in (c); and (g) repeating (d) and (e) using the updated model in place of the base model. In some embodiments, the method further includes (h) repeating (f) and (g) using a further updated model. In various embodiments, the sequence can be a whole genome, whole chromosome, chromosome segment, collection of gene sequences for interacting genes, gene, nucleic acid sequence, protein, polysaccharide, etc. In one or more embodiments, the sub-units of the sequences can be chromosomes, chromosome segments, haplotypes, genes, nucleotides, codons, mutations, amino acids, carbohydrates (mono, di, tri, or oligomeric), etc.

In one or more implementations consistent with the embodiments above, a method is provided for identifying amino acid residues to be modified in a protein variant library. In these embodiments, a plurality of biological molecules constitutes a training set of a protein variant library. The protein variant library may include proteins from various sources. In one example, the members include naturally occurring proteins such as those encoded by members of a single gene family. In another example, the sequences include proteins obtained by using a recombination-based diversity generation mechanism. For example, DNA fragmentation-mediated recombination, synthetic oligonucleotide-mediated recombination or a combination thereof may be performed on nucleic acids encoding all or part of one or more naturally occurring parent proteins for this purpose. In still another example, the members are obtained by implementing a design of experiment (DOE) protocol to identify the systematically varied sequences.

In some embodiments, at least one interaction term is a cross-product term containing a product of one variable representing the presence of one interacting residue and another variable representing the presence of another interacting residue. The form of the sequence-activity model may be a sum of at least one cross-product term and one or more linear terms, with each of the linear terms representing the effect of a variable residue in a training set of a protein variant library. The at least one cross-product term may be selected from a group of potential cross-product terms by various techniques including step-wise addition or subtraction of terms without replacement.

In one or more embodiments, a model including cross-product terms is fitted to given data using Bayesian regression techniques, wherein prior knowledge is used to determine posterior probability distributions of the model.

In one or more embodiments, two or more new models are created, each of which includes at least one different interaction term. In such embodiments, the method further comprises preparing an ensemble model based on the two or more new models. The ensemble model includes interaction terms from the two or more new models. The ensemble model weights the interaction terms according to the two or more new models' abilities to predict activity of interest.

The sequence-activity model may be produced from the training set by many different techniques. In certain embodiments, the model is a regression model such as a partial least squares model, a Bayesian regression model, or a principal component regression model. In another embodiment, the model is a neural network.

Using the sequence-activity model to identify residues for fixing or variation may involve any of many different possible analytical techniques. In some cases, a â€œreference sequenceâ€ is used to define the variations. Such sequence may be one predicted by the model to have the highest value (or one of the highest values) of the desired activity. In another case, the reference sequence may be that of a member of the original protein variant library. From the reference sequence, the method may select subsequences for effecting the variations. Additionally or alternatively, the sequence-activity model ranks residue positions (or specific residues at certain positions) in order of impact on the desired activity.

One goal of the method may be to generate a new protein variant library. As part of this process, the method may identify sequences that are to be used for generating this new library. Such sequences include variations on the residues identified in (e), (g), or (h) above or are precursors used to subsequently introduce such variations. The sequences may be modified by performing mutagenesis or a recombination-based diversity generation mechanism to generate the new library of protein variants. This may form part of a directed evolution procedure. The new library may also be used in developing a new sequence-activity model. The new protein variant library is analyzed to assess effects on a particular activity such as stability, catalytic activity, therapeutic activity, resistance to a pathogen or toxin, toxicity, etc.

In some embodiments, the method involves selecting one or more members of the new protein variant library for production. One or more of these may then be synthesized and/or expressed in an expression system. In a specific embodiment, the method continues in the following manner: (i) providing an expression system from which a selected member of the new protein variant library can be expressed; and (ii) expressing the selected member of the new protein variant library.

In some embodiments, rather than using amino acid sequences, the methods employ nucleotide sequences to generate the models and predict activity. Variations in groups of nucleotides, e.g., codons, affect the activity of peptides encoded by the nucleotide sequences. In some embodiments, the model may provide a bias for codons that are preferentially expressed (compared to other codons encoding the same amino acid) depending upon the host employed to express the peptide.

Another aspect of the disclosure pertains to apparatus and computer program products including machine-readable media on which are provided program instructions and/or arrangements of data for implementing the methods and software systems described above. Frequently, the program instructions are provided as code for performing certain method operations. Data, if employed to implement features of this disclosure, may be provided as data structures, database tables, data objects, or other appropriate arrangements of specified information. Any of the methods or systems described herein may be represented, in whole or in part, as such program instructions and/or data provided on any suitable machine-readable media.

These and other features are described in more detail below in the detailed description and in conjunction with the following figures.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 illustrates a general stepwise method for preparing a sequence-activity model.

FIG. 2 is a flow chart depicting a sequence of operations to generate one or more generations of protein variant libraries, wherein the operations use a sequence-activity model such as one of those obtained in FIG. 1 to guide the generation of protein variant libraries. The generated variant libraries can provide sequence and activity data to prepare one or more new sequence-activity models, forming a modeling-exploring loop of guided evolution.

FIGS. 3A-3H are graphs showing examples in which the predictive capabilities of certain linear and non-linear models are compared.

FIG. 4A-4B illustrates flow charts of processes implementing stepwise addition and subtraction methods for preparing a sequence-activity model. FIG. 4A illustrates a specific example of a stepwise addition method for preparing a model; and FIG. 4B illustrates a specific example of a stepwise subtraction method for preparing a model.

FIG. 5 illustrates a flow chart of a process implementing Bayesian regression in directed evolution of sequence variants in accordance with an embodiment.

FIG. 6 illustrates a flow chart of a process implementing ensemble regression in directed evolution of sequence variants in accordance with an embodiment.

FIG. 7 is a flow chart depicting a bootstrap p-value method of generating protein variant libraries in accordance with an embodiment.

FIG. 8 is a schematic of an exemplary digital device.

DETAILED DESCRIPTIONI. Definitions

Unless defined otherwise herein, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. Various scientific dictionaries that include the terms included herein are well known and available to those in the art. Any methods and materials similar or equivalent to those described herein find use in the practice of the embodiments disclosed herein.

The terms defined immediately below are more fully understood by reference to the specification as a whole. The definitions are for the purpose of describing particular embodiments only and aiding in understanding the complex concepts described in this specification. They are not intended to limit the full scope of the disclosure. Specifically, it is to be understood that this disclosure is not limited to the particular sequences, compositions, algorithms, systems, methodology, protocols, and reagents described, as these may vary, depending upon the context they are used by those of skill in the art.

As used in this specification and appended claims, the singular forms â€œaâ€ , â€œanâ€ , and â€œtheâ€ include plural referents unless the content and context clearly dictates otherwise. Thus, for example, reference to â€œa deviceâ€ includes a combination of two or more such devices, and the like.

Unless indicated otherwise, an â€œorâ€ conjunction is intended to be used in its correct sense as a Boolean logical operator, encompassing both the selection of features in the alternative (A or B, where the selection of A is mutually exclusive from B) and the selection of features in conjunction (A or B, where both A and B are selected). In some places in the text, the term â€œand/orâ€ is used for the same purpose, which shall not be construed to imply that â€œorâ€ is used with reference to mutually exclusive alternatives.

A â€œbio-moleculeâ€ or â€œbiological moleculeâ€ refers to a molecule that is generally found in a biological organism. In some embodiments, biological molecules comprise polymeric biological macromolecules having multiple subunits (i.e., â€œbiopolymersâ€ ). Typical bio-molecules include, but are not limited to, molecules that share some structural features with naturally occurring polymers such as RNAs (formed from nucleotide subunits), DNAs (formed from nucleotide subunits), and peptides or polypeptides (formed from amino acid subunits), including, e.g., RNAs, RNA analogues, DNAs, DNA analogues, polypeptides, polypeptide analogues, peptide nucleic acids (PNAs), combinations of RNA and DNA (e.g., chimeraplasts), or the like. It is not intended that bio-molecules be limited to any particular molecule, as any suitable biological molecule finds use in the present invention, including but not limited to, e.g., lipids, carbohydrates, or other organic molecules that are made by one or more genetically encodable molecules (e.g., one or more enzymes or enzyme pathways) or the like.

The terms â€œpolynucleotideâ€ and â€œnucleic acidâ€ refer to deoxyribonucleotides or ribonucleotides and polymers (e.g., oligonucleotides, polynucleotides, etc.) thereof in either single- or double-stranded form. These terms include, but are not limited to, single-, double- or triple-stranded DNA, genomic DNA, cDNA, RNA, DNA-RNA hybrid, polymers comprising purine and pyrimidine bases, and/or other natural, chemically or biochemically modified, non-natural or derivatized nucleotide bases. The following are non-limiting examples of polynucleotides: genes, gene fragments, chromosomal fragments, ESTs, exons, introns, mRNA, tRNA, rRNA, ribozymes, cDNA, recombinant polynucleotides, branched polynucleotides, plasmids, vectors, isolated DNA of any sequence, isolated RNA of any sequence, nucleic acid probes, and primers. In some embodiments, polynucleotides comprise modified nucleotides, such as methylated nucleotides and nucleotide analogs, uracyl, other sugars and linking groups such as fluororibose and thioate, and/or nucleotide branches. In some alternative embodiments, the sequence of nucleotides is interrupted by non-nucleotide components.

Unless specifically limited, the term encompasses nucleic acids containing known analogs of natural nucleotides that have similar binding properties as the reference nucleic acid and are metabolized in a manner similar to naturally occurring nucleotides. Unless otherwise indicated, a particular nucleic acid sequence also implicitly encompasses conservatively modified variants thereof (e.g., degenerate codon substitutions) and complementary sequences as well as the sequence explicitly indicated. Specifically, degenerate codon substitutions may be achieved by generating sequences in which the third position of one or more selected (or all) codons is substituted with mixed-base and/or deoxyinosine residues (Batzer et al. (1991) Nucleic Acid Res. 19:5081; Ohtsuka et al. (1985) J. Biol. Chem. 260:2605-2608; Rossolini et al. (1994) Mol. Cell. Probes 8:91-98). The term nucleic acid is used interchangeably with, e.g., oligonucleotide, polynucleotide, cDNA, and mRNA.

The terms â€œprotein,â€ â€œpolypeptideâ€ and â€œpeptideâ€ are used interchangeably to denote a polymer of at least two amino acids covalently linked by an amide bond, regardless of length or post-translational modification (e.g., glycosylation, phosphorylation, lipidation, myristilation, ubiquitination, etc.). In some cases, the polymer has at least about 30 amino acid residues, and usually at least about 50 amino acid residues. More typically, they contain at least about 100 amino acid residues. The terms include compositions conventionally considered to be fragments of full-length proteins or peptides. Included within this definition are D- and L-amino acids, and mixtures of D- and L-amino acids. The polypeptides described herein are not restricted to the genetically encoded amino acids. Indeed, in addition to the genetically encoded amino acids, the polypeptides described herein may be made up of, either in whole or in part, naturally-occurring and/or synthetic non-encoded amino acids. In some embodiments, a polypeptide is a portion of the full-length ancestral or parental polypeptide, containing amino acid additions or deletions (e.g., gaps) or substitutions as compared to the amino acid sequence of the full-length parental polypeptide, while still retaining functional activity (e.g., catalytic activity).

As used herein, the term â€œcellulaseâ€ refers to a category of enzymes capable of hydrolyzing cellulose (Î²-1,4-glucan or Î²-D-glucosidic linkages) to shorter cellulose chains, oligosaccharides, cellobiose and/or glucose. In some embodiments, the term â€œcellulaseâ€ encompasses beta-glucosidases, endoglucanases, cellobiohydrolases, cellobiose dehydrogenases, endoxylanases, beta-xylosidases, arabinofuranosidases, alpha-glucuronidases, acetylxylan esterases, feruloyl esterases, and/or alpha-glucuronyl esterases. In some embodiments, the term â€œcellulaseâ€ encompasses hemicellulose-hydrolyzing enzymes, including but not limited to endoxylanases, beta-xylosidases, arabinofuranosidases, alpha-glucuronidases, acetylxylan esterase, feruloyl esterase, and alpha-glucuronyl esterase. A â€œcellulase-producing fungal cellâ€ is a fungal cell that expresses and secretes at least one cellulose hydrolyzing enzyme. In some embodiments, the cellulase-producing fungal cells express and secrete a mixture of cellulose hydrolyzing enzymes. â€œCellulolytic,â€ â€œcellulose hydrolyzing,â€ â€œcellulose degrading,â€ and similar terms refer to enzymes such as endoglucanases and cellobiohydrolases (the latter are also referred to as â€œexoglucanasesâ€ ) that act synergistically to break down the cellulose to soluble di- or oligosaccharides such as cellobiose, which are then further hydrolyzed to glucose by beta-glucosidase. In some embodiments, the cellulase is a recombinant cellulase selected from Î²-glucosidases (BGLs), Type 1 cellobiohydrolases (CBH1s), Type 2 cellobiohydrolases (CBH2s), glycoside hydrolase 61s (GH61s), and/or endoglucanases (EGs). In some embodiments, the cellulase is a recombinant Myceliophthora cellulase selected from Î²-glucosidases (BGLs), Type 1 cellobiohydrolases (CBH1s), Type 2 cellobiohydrolases (CBH2s), glycoside hydrolase 61s (GH61s), and/or endoglucanases (EGs). In some additional embodiments, the cellulase is a recombinant cellulase selected from EG1b, EG2, EG3, EG4, EG5, EG6, CBH1a, CBH1b, CBH2a, CBH2b, GH61a, and/or BGL.

The term â€œsequenceâ€ is used herein to refer to the order and identity of any biological sequences including but not limited to a whole genome, whole chromosome, chromosome segment, collection of gene sequences for interacting genes, gene, nucleic acid sequence, protein, polysaccharide, etc. In some contexts, a sequence refers to the order and identity of amino acid residues in a protein (i.e., a protein sequence or protein character string) or to the order and identity of nucleotides in a nucleic acid (i.e., a nucleic acid sequence or nucleic acid character string). A sequence may be represented by a character string. A â€œnucleic acid sequenceâ€ refers to the order and identity of the nucleotides comprising a nucleic acid. A â€œprotein sequenceâ€ refers to the order and identity of the amino acids comprising a protein or peptide.

â€œCodonâ€ refers to a specific sequence of three consecutive nucleotides that is part of the genetic code and that specifies a particular amino acid in a protein or starts or stops protein synthesis.

â€œNative sequenceâ€ or â€œwild type sequenceâ€ refers to a polynucleotide or polypeptide isolated from a naturally occurring source. Included within â€œnative sequenceâ€ are recombinant forms of a native polypeptide or polynucleotide which have a sequence identical to the native form.

The term â€œgeneâ€ is used broadly to refer to any segment of DNA or other nucleic acid associated with a biological function. Thus, genes include coding sequences and optionally, the regulatory sequences required for their expression. Genes also optionally include unexpressed nucleic acid segments that, for example, form recognition sequences for other proteins. Genes can be obtained from a variety of sources, including cloning from a source of interest or synthesizing from known or predicted sequence information, and may include sequences designed to have desired parameters.

A â€œmotifâ€ refers to a pattern of subunits in or among biological molecules. For example, the term â€œmotifâ€ can be used in reference to a subunit pattern of the unencoded biological molecule or to a subunit pattern of an encoded representation of a biological molecule.

The term â€œchromosomeâ€ is used in reference to an organized structure of DNA and associated protein found cells, comprising a single piece of coiled DNA including many genes, regulatory elements, and other nucleotide sequences. The term is also used in reference to the DNA sequence of the structure.

â€œScreeningâ€ refers to the process in which one or more properties of one or more bio-molecules are determined. For example, typical screening processes include those in which one or more properties of one or more members of one or more libraries are determined. An â€œexpression systemâ€ is a system for expressing a protein or peptide encoded by a gene or other nucleic acid.

â€œHost cellâ€ or â€œrecombinant host cellâ€ refers to a cell that comprises at least one recombinant nucleic acid molecule. Thus, for example, in some embodiments, recombinant host cells express genes that are not found within the native (i.e., non-recombinant) form of the cell.

â€œDirected evolution,â€ â€œguided evolution,â€ or â€œartificial evolutionâ€ refers to in vitro or in vivo processes of artificially changing one or more biomolecule sequences (or a character string representing that sequence) by artificial selection, recombination, or other manipulation. In some embodiments, directed evolution occurs in a reproductive population in which there are (1) varieties of individuals, with some varieties being (2) heritable, of which some varieties (3) differ in fitness. Reproductive success is determined by outcome of selection for a predetermined property such as a beneficial property. The reproductive population can be, e.g., a physical population or a virtual population in a computer system.

In certain embodiments, directed evolution methods generate protein variant libraries by recombining genes encoding variants of a parent protein variant library. The methods may employ oligonucleotides containing sequences or subsequences for encoding the proteins of a parental variant library. Some of the oligonucleotides of the parental variant library may be closely related, differing only in the choice of codons for alternate amino acids selected to be varied by recombination with other variants. The method may be performed for one or multiple cycles until desired results are achieved. If multiple cycles are used, each involves a screening step to identify which variants having acceptable performance are to be used in a subsequent recombination cycle.

â€œShufflingâ€ and â€œgene shufflingâ€ refer to directed evolution methods for introducing diversity by recombining a collection of fragments of the parental polynucleotides through a series of chain extension cycles. In certain embodiments, one or more of the chain extension cycles is self-priming; i.e., performed without the addition of primers other than the fragments themselves. Each cycle involves annealing single stranded fragments through hybridization, subsequent elongation of annealed fragments through chain extension, and denaturing. Over the course of shuffling, a growing nucleic acid strand is typically exposed to multiple different annealing partners in a process sometimes referred to as â€œtemplate switching.â€ As used herein, â€œtemplate switchingâ€ refers to the ability to switch one nucleic acid domain from one nucleic acid with a second domain from a second nucleic acid (i.e., the first and second nucleic acids serve as templates in the shuffling procedure).

Template switching frequently produces chimeric sequences, which result from the introduction of crossovers between fragments of different origins. The crossovers are created through template switched recombinations during the multiple cycles of annealing, extension, and denaturing. Thus, shuffling typically leads to production of variant polynucleotide sequences. In some embodiments, the variant sequences comprise, a â€œlibraryâ€ of variants. In some embodiments of these libraries, the variants contain sequence segments from two or more of parent polynucleotides.

When two or more parental polynucleotides are employed, the individual parental polynucleotides are sufficiently homologous that fragments from different parents hybridize under the annealing conditions employed in the shuffling cycles. In some embodiments, the shuffling permits recombination of parent polynucleotides having relatively limited homology. Often, the individual parent polynucleotides have distinct and/or unique domains and/or other sequence characteristics of interest. When using parent polynucleotides having distinct sequence characteristics, shuffling can produce highly diverse variant polynucleotides.

Various shuffling techniques are known in the art. See e.g., U.S. Pat. Nos. 6,917,882, 7,776,598, 8,029,988, 7,024,312, and 7,795,030, all of which are incorporated herein by reference in their entireties.

A â€œfragmentâ€ is any portion of a sequence of nucleotides or amino acids. Fragments may be produced using any suitable method known in the art, including but not limited to cleaving a polypeptide or polynucleotide sequence. In some embodiments, fragments are produced by using nucleases that cleave polynucleotides. In some additional embodiments, fragments are generated using chemical and/or biological synthesis techniques. In some embodiments, fragments comprise subsequences of at least one parental sequence, generated using partial chain elongation of complementary nucleic acid(s).

â€œParental polypeptide,â€ â€œparental polynucleotide,â€ â€œparent nucleic acid,â€ and â€œparentâ€ are generally used to refer to the wild-type polypeptide, wild-type polynucleotide, or a variant used as a starting point in a diversity generation procedure such as a directed evolution. In some embodiments, the parent itself is produced via shuffling or other diversity generation procedure. In some embodiments, mutants used in directed evolution are directly related to a parent polypeptide. In some embodiments, the parent polypeptide is stable when exposed to extremes of temperature, pH and/or solvent conditions and can serve as the basis for generating variants for shuffling. In some embodiments, the parental polypeptide is not stable to extremes of temperature, pH and/or solvent conditions, and the parental polypeptide is evolved to make a robust variants.

A â€œparent nucleic acidâ€ encodes a parental polypeptide.

â€œMutant,â€ â€œvariant,â€ and â€œvariant sequenceâ€ as used herein, refer to a biological sequence that differs in some respect from a standard or reference sequence. The difference may be referred to as a â€œmutationâ€ . In some embodiments, a mutant is an amino acid (i.e., polypeptide) or polynucleotide sequence that has been altered by at least one substitution, insertion, cross-over, deletion, and/or other genetic operation. For purposes of the present disclosure, mutants and variants are not limited to a particular method by which they are generated. In some embodiments, a mutant or variant sequence has increased, decreased, or substantially similar activities or properties, in comparison to the parental sequence. In some embodiments, the variant polypeptide comprises one or more amino acid residues that have been mutated, as compared to the amino acid sequence of the wild-type polypeptide (e.g., a parent polypeptide). In some embodiments, one or more amino acid residues of the polypeptide are held constant, are invariant, or are not mutated as compared to a parent polypeptide in the variant polypeptides making up the plurality. In some embodiments, the parent polypeptide is used as the basis for generating variants with improved stability, activity, or other property.

â€œMutagenesisâ€ is the process of introducing a mutation into a standard or reference sequence such as a parent nucleic acid or parent polypeptide.

A â€œlibraryâ€ or â€œpopulationâ€ refers to a collection of at least two different molecules, character strings, and/or models, such as nucleic acid sequences (e.g., genes, oligonucleotides, etc.) or expression products (e.g., enzymes or other proteins) therefrom. A library or population generally includes a number of different molecules. For example, a library or population typically includes at least about 10 different molecules. Large libraries typically include at least about 100 different molecules, more typically at least about 1000 different molecules. For some applications, the library includes at least about 10000 or more different molecules. In certain embodiments, the library contains a number variant or chimeric nucleic acids or proteins produced by a directed evolution procedure.

Two nucleic acids are â€œrecombinedâ€ when sequences from each of the two nucleic acids are combined in a progeny nucleic acid. Two sequences are â€œdirectlyâ€ recombined when both of the nucleic acids are substrates for recombination.

â€œSelectionâ€ refers to the process in which one or more bio-molecules are identified as having one or more properties of interest. Thus, for example, one can screen a library to determine one or more properties of one or more library members. If one or more of the library members is/are identified as possessing a property of interest, it is selected. Selection can include the isolation of a library member, but this is not necessary. Further, selection and screening can be, and often are, simultaneous.

A â€œdependent variableâ€ represents an output or effect, or is tested to see if it is the effect. The â€œindependent variablesâ€ represent the inputs or causes, or are tested to see if they are the cause. A dependent variable may be studied to see if and how much it varies as the independent variables vary.

In the simple stochastic linear model

y_i=a+bx_i+e_i

where the term y_iis the i^thvalue of the dependent variable and x_iis i^thvalue of the independent variable. The term e_iis known as the â€œerrorâ€ and contains the variability of the dependent variable not explained by the independent variable.

An independent variable is also known as a â€œpredictor variableâ€ , â€œregressorâ€ , â€œcontrolled variableâ€ , â€œmanipulated variableâ€ , â€œexplanatory variableâ€ , or â€œinput variableâ€ .

â€œOrthogonal/orthogonalityâ€ refers to an independent variable that is uncorrelated with other independent variables in a model or other relationship.

The term â€œsequence-activity modelâ€ refers to any mathematical models that describe the relationship between activities, characteristics, or properties of biological molecules on the one hand, and various biological sequences on the other hand.

The term â€œencoded character stringâ€ refers to a representation of a biological molecule that preserves sequence/structural information regarding that molecule. In some embodiments, the encoded character string contains information about sequence mutations in a library of variants. Encoded character strings of bio-molecules along with activity information for the bio-molecules may be used as a training set for a sequence activity model. Non-sequence properties of bio-molecules can be stored or otherwise associated with encoded character strings for the bio-molecules.

â€œReference sequenceâ€ is a sequence from which variation of sequence is effected. In some cases, a â€œreference sequenceâ€ is used to define the variations. Such sequence may be one predicted by a model to have the highest value (or one of the highest values) of the desired activity. In another case, the reference sequence may be that of a member of an original protein variant library. It certain embodiments, a reference sequence is the sequence of a parent protein or nucleic acid.

â€œTraining setâ€ refers to a set of sequence-activity data or observations that one or more models are fitted to and built upon. For instance, for a protein sequence-activity model, a training set comprises residue sequences for an initial or improved protein variant library. Typically, these data include complete or partial residue sequence information, together with an activity value for each protein in the library. In some cases, multiple types of activities (e.g., rate constant data and thermal stability data) are provided together in the training set. The activity is sometimes a beneficial property.

The term â€œobservationâ€ is information about protein or other biological entity that may be used in a training set for generating a model such as a sequence activity model. The term â€œobservationâ€ may refer to any sequenced and assayed biological molecules, including protein variants. In certain embodiments, each observation is an activity value and an associated sequence for a variant in a library. Generally, the more observations employed to create a sequence-activity model, the better the predictive power of that sequence-activity model.

As used herein, the term â€œbeneficial propertyâ€ is intended to refer to a phenotypic or other identifiable feature that confers some benefit to a protein or a composition of matter or process associated with the protein. Examples of beneficial properties include an increase or decrease, when compared to a parent protein, in a variant protein's catalytic properties, binding properties, stability when exposed to extremes of temperature, pH, etc., sensitivity to stimuli, inhibition, and the like. Other beneficial properties may include an altered profile in response to a particular stimulus. Further examples of beneficial properties are set forth below. Values of beneficial properties may be used as activity values in the observations used in a training set for a sequence activity model.

â€œNext-generation sequencingâ€ or â€œhigh-throughput sequencingâ€ are sequencing techniques that parallelize the sequencing process, producing thousands or millions of sequences at once. Examples of suitable next-generation sequencing methods include, but are not limited to, single molecule real-time sequencing (e.g., Pacific Biosciences, Menlo Park, Calif.), ion semiconductor sequencing (e.g., Ion Torrent, South San Francisco, Calif.), pyrosequencing (e.g., 454, Branford, Conn.), sequencing by ligation (e.g., SOLid sequencing of Life Technologies, Carlsbad, Calif.), sequencing by synthesis and reversible terminator (e.g., Illumina, San Diego, Calif.), nucleic acid imaging technologies such as transmission electron microscopy, and the like. Further descriptions of exemplary techniques are described in the detailed description of this disclosure.

â€œPredictive powerâ€ refers to the ability of a model to correctly predict the values of a dependent variable for data under various conditions. For example, the predictive power of a sequence activity model refers to the ability of the model to predict activity from sequence information.

â€œCross validationâ€ refers to a method for testing the generalizability of a model's ability to predict a value of interest (i.e. the value of the dependent variable). The method prepares a model using one set of data, and tests the model error using a different set of data. The first set of data is viewed as a training set, and the second set of data is a validation set.

â€œSystematic varianceâ€ refers to different descriptors of an item or set of items being changed in different combinations.

â€œSystematically varied dataâ€ refers to data produced, derived, or resulting from different descriptors of an item or set of items being changed in different combinations. Many different descriptors can be changed at the same time, but in different combinations. For example, activity data gathered from polypeptides in which combinations of amino acids have been changed is systematically varied data.

The term â€œsystematically varied sequencesâ€ refers to a set of sequences in which each residue is seen in multiple contexts. In principle, the level of systematic variation can be quantified by the degree to which the sequences are orthogonal from one another (i.e., maximally different compared to the mean).

The term â€œtogglingâ€ refers to the introduction of multiple amino acid residue types into a specific position in the sequences of protein variants in the optimized library.

The terms â€œregressionâ€ and â€œregression analysisâ€ refer to techniques used to understand which among the independent variables are related to the dependent variable, and to explore the forms of these relationships. In restricted circumstances, regression analysis can be used to infer causal relationships between the independent and dependent variables. It is a statistical technique for estimating the relationships among variables. It includes many techniques for modeling and analyzing several variables, when the focus is on the relationship between a dependent variable and one or more independent variables. More specifically, regression analysis helps one understand how the typical value of the dependent variable changes when any one of the independent variables is varied, while the other independent variables are held fixed. Regression techniques may be used to generate sequence activity models from training sets comprising multiple observations, which may contain sequence and activity information.

Partial Least Squares or PLS is a family of methods that finds a linear regression model by projecting predicted variables (e.g., activities) and the observable variables (e.g., sequences) to a new space. PLS is also known as projection to latent structures. Both the X (independent variables) and Y (dependent variables) data are projected to new spaces. PLS is used to find the fundamental relations between two matrices (X and Y). A latent variable approach is used to model the covariance structures in the X and Y spaces. A PLS model will try to find the multidimensional direction in the X space that explains the maximum multidimensional variance direction in the Y space. PLS regression is particularly suited when the matrix of predictors has more variables than observations, and when there is multicollinearity among X values.

A â€œdescriptorâ€ refers to something that serves to describe or identify an item. For example, characters in a character string can be descriptors of amino acids in a polypeptide being represented by the character string.

In a regression model, the dependent variable is related to independent variables by a sum of terms. Each term includes a product of an independent variable and an associated regression coefficient. In the case of a purely linear regression model, the regression coefficients are given by Î² in the following form of expression:

y_i=Î²₁x_i1+ . . . +Î²_px_ip+Îµ_i=x_i^TÎ²+Îµ_i

where y_iis the dependent variable, the x_iare the independent variables, Îµ_iis the error variable, and T denotes the transpose, that is the inner product of the vectors x_iand Î².

â€œPrincipal component regressionâ€ (PCR) refers to a regression analysis that uses principal component analysis when estimating regression coefficients. In PCR instead of regressing the dependent variable on the independent variables directly, the principal components of the independent variables are used. PCR typically only uses a subset of the principal components in the regression.

â€œPrincipal component analysisâ€ (PCA) refers to a mathematical procedure that uses an orthogonal transformation to convert a set of observations of possibly correlated variables into a set of values of linearly uncorrelated variables called principal components. The number of principal components is less than or equal to the number of original variables. This transformation is defined in such a way that the first principal component has the largest possible variance (that is, accounts for as much of the variability in the data as possible), and each succeeding component in turn has the highest variance possible under the constraint that it be orthogonal to (i.e., uncorrelated with) the preceding components.

â€œNeural networkâ€ is a model containing an interconnected group of processing elements or â€œneuronsâ€ that process information using a connectionist approach to computation. Neural networks are used to model complex relationships between inputs and outputs or to find patterns in data. Most neural networks process data in a non-linear, distributed, parallel fashion. In most cases a neural network is an adaptive system that changes its structure during a learning phase.

Functions are performed collectively and in parallel by the processing elements, rather than there being a clear delineation of subtasks to which various units are assigned.

Generally, a neural network involves a network of simple processing elements that exhibit complex global behavior determined by the connections between the processing elements and element parameters. Neural networks are used with algorithms designed to alter the strength of the connections in the network to produce a desired signal flow. The strength is altered during training or learning.

â€œRandom forestâ€ refers to a combination of classification tree predictors such that each tree depends on the values of a random vector sampled independently and with the same distribution for all trees in the forest. A random forest is a learning ensemble consisting of a bagging of un-pruned decision tree learners with a randomized selection of features at each split of the decision tree. A random forest grows a large number of classification trees, each of which votes for the most popular class. The random forest then classifies a variable by taking the most popular voted class from all the tree predictors in the forest.

â€œPrior probability distributionâ€ , or â€œprior,â€ of an uncertain quantity p is the probability distribution that expresses the uncertainty about p before data of interest (e.g., a training set of protein sequences) are taken into account. The unknown quantity may be a parameter, coefficient, variable, latent variable, or the like (e.g., a coefficient in a multiple regression model).

â€œPosterior probability distribution,â€ or â€œposterior,â€ of an uncertain quantity p is the probability distribution that expresses the uncertainty about p after the data of interest are taken into account.

The term â€œBayesian linear regressionâ€ refers to an approach to linear regression in which the statistical analysis is undertaken within the context of Bayesian inference. The prior belief about the linear regression model, including the prior probability distribution function of the model's parameter, is combined with the data's likelihood function according to Bayes theorem to yield the posterior probability distribution about the parameters.

â€œOverfittingâ€ refers to a condition that occurs when a statistical model describes random error or noise instead of the underlying relationship. Overfitting generally occurs when a model is excessively complex, such as having too many parameters relative to the number of observations. A model which has been overfit will generally have poor predictive performance, as it can exaggerate minor fluctuations in the data. In some embodiments, a mathematical model is used to describe the relationship between one or more independent variables (IVs) and a dependent variable (DV). The model can be written as DV=Algebraic expression of (IVs). An â€œalgebraic expressionâ€ can include variables, coefficients, constants, and operating symbols, such as plus and minus signs. 4x²+3xy+7y+5 is a bivariate algebraic expression.

In some embodiments, â€œtermsâ€ of an algebraic expression or a mathematical model are the elements separated by the plus or minus signs. In this context, the example above has four terms, 4x², 3xy, 7y, and 5. Terms may consist of variables and coefficients (4x², 3xy, and 7y), or constants (5). In algebraic expressions, variables can take on various values to represent changing conditions of a system. For instance, it can be a continuous variable representing the speed of a traveling car or a discrete variable with multiple non-continuous values representing amino acid types. A variable can be a bit value variable representing the presence or absence of an entity, e.g., the presence or absence of a residue of a specific type at a specific position. In the algebraic expression above, the variables are x and y.

In some embodiments, â€œtermsâ€ of an expression may be elements of the expression that are demarcated by other signs, such as by multiplication.

â€œCoefficientâ€ refers to a scalar value multiplied by a dependent variable or an expression containing a dependent variable. In the example above, â€œcoefficientsâ€ are the number part of the terms in an algebraic expression. In 4x²+3xy+7y+5, the coefficient of the first term is 4. The coefficient of the second term is 3, and the coefficient of the third term is 7. If a term consists of only variables, its coefficient is 1.

â€œConstantsâ€ are the terms in the algebraic expression that contain only numbers. That is, they are the terms without variables. In the expression 4x²+3xy+7y+5, the constant term is

A â€œlinear termâ€ is a term with a degree of 1, or a single variable raised to the power of 1. In the example above, the term 7y is a linear term because its degree is 1 (y¹or simply y). In contrast, the term 4x²is a quadratic term because the x has a degree of 2, and 3xy is a bivariate quadratic term because x and y each has a degree of 1, the product leading to a degree of 2.

In some places of the text, â€œlinear termâ€ and â€œnon-interaction termâ€ are used interchangeably herein to refer to a term of a regression model comprising the product of a single independent variable and an associated coefficient, wherein the single IV represents the presence/absence of a single residue.

In some embodiments, â€œnon-linear term,â€ â€œcross-product term,â€ and â€œinteraction termâ€ are used interchangeably in this disclosure when they refer to a term of a regression model comprising the product of two or more independent variables and an associated coefficient. More generally, â€œnon-linear termsâ€ are used to indicate terms with a degree of more or less than 1, e.g., a power function or exponential function of the independent variable. Some examples of non-linear terms include xy, x², x^1/3, x^y, and e^x. Thus, in some places in the text, â€œnon-linear termâ€ refers to a broader sense than a term including the product of two independent variables.

In some embodiments, an interaction term may be implemented as a term including a non-linear function of two or more IVs, e.g., the product function, power function, or exponential function of two or more IVs, each IV representing the presence of a residue of a specific type at a specific position. For instance, in y=ax₁+bx₂+cx₁x₂, variables x₁and x₂may represent the presence/absence of two particular residues at one particular location, and the term cx₁x₂is an interaction term representing the effect of the two particular residues' interaction. In other embodiments, an interaction term may be implemented as a term including a single IV representing the interaction of two or more residues. For instance, in y=ax₁+bx₂+cz, variables x₁and x₂may represent the presence/absence of two particular residues at a particular location, and the term cz is an interaction term representing the effect of the two particular residues' interaction. In this last example, the interaction term cz is not a cross-product term. Although technically cz is a linear term, it is not so labeled herein to avoid confusion with the linear, non-interaction terms ax₁and bx₂. As used in the disclosure, the term â€œlinear modelâ€ refers to models including only linear terms. In contrast, the term â€œnon-linear modelâ€ refers to models including both linear and non-linear terms. In some embodiments, non-linear models include interaction terms implemented as cross product terms.

More generally, a linear model or a linear system satisfies the superposition principal and homogeneity of degree 1. The superposition principal states that, for all linear systems, the net response at a given place and time caused by two or more stimuli is the sum of the responses which would have been caused by each stimulus individually. This is also known as additivity. If input A produces response X and input B produces response Y then input (A+B) produces response (X+Y). Homogeneity of degree 1 refers to any model whose output or dependent variable (DV) changes proportionately as its input or independent variable. Conversely, a â€œnon-linear modelâ€ is a model that does not satisfy the superposition principal or homogeneity of degree 1.

â€œInteracting sub-unitsâ€ refers to two or more sub-units of a sequence that have a synergistic effect on the modeled activity of the sequence, the synergistic effect being separate and different from the sub-units' individual effects on the modeled activity.

The term â€œbase modelâ€ is used in reference to a sequence-activity model provided at the beginning of a process of improving a model.

The term â€œupdated modelâ€ is used in reference to a sequence-activity model that is derived directly or indirectly from a base model, which has improved predictive power compared to the base model and/or another model from which it is derived from.

A â€œlikelihood functionâ€ or â€œlikelihoodâ€ of a model is a function of the parameters of a statistical model. The likelihood of a set of parameter values given some observed outcomes equals to the probability of those observed outcomes given those parameter values, i.e., L(Î¸|x)=P(x|Î¸).

â€œMonte Carlo simulationsâ€ are simulations that rely on a large number of random sampling to obtain numerical results that simulate a real phenomenon. For instance, drawing a large number of pseudo-random uniform variables from the interval (0,1], and assigning values less than or equal to 0.50 as heads and greater than 0.50 as tails, is a Monte Carlo simulation of the behavior of repeatedly tossing a coin.

A â€œMetropolis algorithmâ€ or â€œMetropolis-Hastings algorithmâ€ is a Markov chain Monte Carlo (MCMC) method for obtaining a sequence of random samples from a probability distribution for which direct sampling is difficult. This sampling sequence can be used to approximate the distribution (i.e., to generate a histogram), or to compute an integral (such as an expected value). Metropolis-Hastings and other MCMC algorithms are generally used for sampling from multi-dimensional distributions, especially when the number of dimensions is high. The objective of the Metropolis-Hastings algorithm is to asymptotically generate states x according to a desired distribution P(x) and uses a stochastic process to fulfill it. The idea of the algorithm is to condition the stochastic process such that it asymptotically converges to the unique distribution P(x).

A â€œMarkov chainâ€ is a sequence of random variables X₁, X₂, X₃. . . with the Markov property. In other words, given the present state, the future and past states are independent. Formally,

Pr(X_n+1=x|X₁=x₁,X₂=x₂, . . . ,X_n=x_n)=Pr(X_n+1=x|X_n=x_n).

The possible values of X_iform a countable set S called the state space of the chain. A â€œMarkov chainâ€ system is a mathematical system that undergoes transitions from one state to another, between a finite or countable number of possible states. It is a random process usually characterized as memoryless: the next state depends only on the current state and not on the sequence of events that preceded it.

The â€œAkaike Information Criterionâ€ (AIC) is a measure of the relative goodness of fit of a statistical model, and it is often used as a criterion for model selection among a finite set of models. The AIC is grounded in the concept of information entropy, in effect offering a relative measure of the information lost when a given model is used to describe reality. It can be said to describe the tradeoff between bias and variance in model construction, or loosely speaking between accuracy and complexity of the model. The AIC can be calculated as: AIC=âˆ’2 log_eL+2k, wherein L is the maximum likelihood of the function and k is the number of free parameters of the model to be estimated.

â€œBayesian Information Criterionâ€ is a criterion for model selection among a finite set of models, and is closely related to AIC. The BIC can be calculated as: BIC=âˆ’2 log_eL+klog_e(n), wherein n is the number of data observations. As the number of observations increased, BIC often penalizes extra number of free parameters more heavily than AIC.

A â€œgenetic algorithmâ€ is a process that mimics evolutionary processes. Genetic algorithms (GAs) are used in a wide variety of fields to solve problems which are not fully characterized or too complex to allow full characterization, but for which some analytical evaluation is available. That is, GAs are used to solve problems which can be evaluated by some quantifiable measure for the relative value of a solution (or at least the relative value of one potential solution in comparison to another). In the context of the present disclosure, a genetic algorithm is a process for selecting or manipulating character strings in a computer, typically where the character string corresponds to one or more biological molecules (e.g., nucleic acids, proteins, or the like).

The term â€œgenetic operationâ€ (or â€œGOâ€ ) refer to biological and/or computational genetic operations, wherein all changes in any population of any type of character strings (and thus in any physical properties of physical objects encoded by such strings) can be described as a result of random and/or predetermined application of a finite set of logical algebraic functions. Examples of GO include but are not limited to multiplication, crossover, recombination, mutation, ligation, fragmentation, etc.

â€œEnsemble modelâ€ is a model whose terms include all the terms of a group of models, wherein the ensemble model's coefficients of the terms are based on the weighted coefficients of the corresponding terms of the individual models of the group. The weighting of coefficients is based on the predictive power and/or fitness of the individual models.

II. Generating Improved Protein Variant Libraries

In a guided evolution approach to exploring protein sequences, sequence-activity models are used to guide the generation of protein variants. One aspect of the disclosure provides various methods for preparing sequence-activity models that are based on protein libraries and can be used to search for new and improved protein libraries. This section first provides an overview of the process for searching for new and improved proteins, and then provides further details on issues related to selecting a starting library, building a sequence-activity model, and using the model to guide exploration of new proteins.

This disclosure provides illustrative examples involving amino acid residue sequences and protein activities, but it is understood that the approach described herein can also be implemented for other biological sequences and activities. For instance, in various embodiments, a sequence can be a whole genome, whole chromosome, chromosome segment, collection of gene sequences for interacting genes, gene, nucleic acid sequence, protein, polysaccharide, etc. In one or more embodiments, the sub-units of the sequences can be chromosomes, chromosome segments, haplotypes, genes, nucleotides, codons, mutations, amino acids, mono, di, tri, or oligomeric carbohydrates, etc.

Typically, at the beginning of a particular round of directed evolution of sequences, a training set of sequenced and assayed protein variants are obtained. A given round of directed evolution produces a number of variant proteins that vary by one or more mutations from the parent peptide or peptides used at the beginning of the round of directed evolution. The variant peptides produced during a round of directed evolution are assayed for activity. Those peptides having desired activity and/or improved activity compared to the parent peptide(s) are selected for use in at least one further round of directed evolution.

Sequenced and assayed protein variants may also be used to produce a sequence-activity model. Typically, they are used in a sequence-activity model if in fact they are sequenced. Each of the sequenced and assayed protein variants is referred to as an â€œobservation.â€ Generally, the more observations employed to create a sequence-activity model, the better the predictive power of that sequence-activity model.

Until the advent of next-generation massively parallel sequencing technology, it was difficult to economically sequence more than 10 to 30 variant peptides produced in any round of directed evolution. Now with the application of next-generation sequencing, many more variant proteins produced in a round of directed evolution can be sequenced. As a consequence, a much greater pool of training set data can be used to produce sequence-activity models. The sequence-activity models can now be generated using a training set that includes not only the top performing peptides from a round, but also some peptides which would not be of interest for further rounds of directed evolution, but whose sequence activity information could be applied to produce a more robust sequence-activity model.

In some embodiments, it is generally desirable to produce sequence-activity models having a good ability to predict the activity of an arbitrary sequence. The predictive power may be characterized by the accuracy of the prediction as well as the consistency with which the model accurately predicts activity. Further, a model may be characterized by its ability to accurately predict activity over a wide range of sequence space. For instance, the predictive power can be characterized in terms of residuals between the calculated and actual activities for a given test and/or validation set of peptides. A model with higher generalized predictive power tends to yield smaller and more consistent residuals across different sets of validation data. A model that is overfitted to a test set of data tends to yield larger and less consistent residuals for validation data, as shown by an example below. One aspect of the disclosure provides a method to efficiently find a model with high predictive power across different sets of data.

A. Overview of the Process to Search for Improved Protein Variants

Sequence-activity models as described herein may be used to assist in identifying one or more parent â€œgenesâ€ in an initial variant library to undergo directed evolution. After a round of evolution is performed, a new variant library is identified, providing a new set of observations, which can then be fed back as data to prepare a new or refined sequence-activity model. This process of alternating between preparing a sequence-activity model based on new observations and conducting directed evolution based on the sequence-activity model may form an iterative loop of modeling-exploring, which may be repeated until desired proteins and libraries are obtained.

Because of the feedback loop between sequence-activity models and variant libraries, better models and better variant libraries depend on each other in exploration of proteins with improved activities. Therefore, bottle necks and improvements in either modeling and/or sequencing domains may affect both domains. In some embodiments of the invention, improvements of modeling efficiencies due to better modeling techniques provide better models to guide sequence exploration. In some embodiments, next generation sequencing technologies are used to improve sequencing speed in vitro, as well as to provide cross validation data to improve in silico computational models.

In some embodiments of the invention, useful sequence-activity models require robust mathematical modeling techniques and a large number of â€œobservations.â€ These observations are data provided in a training set for a model. Specifically, each observation is an activity value and an associated sequence for a variant in a library. Historically, sequencing has been a limiting step in the development of large training sets and consequently, increasingly robust sequence-activity models. In methods commonly used currently, variant libraries having perhaps hundreds of variants are generated. However, only a small fraction of these variants is actually sequenced. In a typical round of directed evolution, only about 10 to 30 variants with the highest activity are actually sequenced. Ideally, a much larger fraction of the variants in the library would be sequenced, including some variants with relatively low activities. Next generation sequencing tools have greatly improved sequencing speed, making it possible to include the low activity and high activity variants in a training set. In some embodiments, inclusion of variants having a range of activity levels results in production of models that perform better and/or are better at predicting activity over a wider range of sequence and activity space.

Some linear sequence-activity regression models referred to herein include individual residues as independent variables to predict any activity of interest. The linear sequence-activity regression models do not include terms to account for the interactions between two or more residuals. If an interaction between two of the residuals has a synergistic effect on activity, a linear model may provide an artificially inflated value of the coefficients associated with the two interacting residues. As a consequence, someone working with the model may erroneously conclude that by simply making a residue substitution as proposed by the relatively high value of the coefficient, the activity of a resulting peptide would be higher than expected. This is because the researcher does not understand from using a linear model that the increased activity associated with the residue substitution is primarily a result of that substitution's interaction with another substitution. If the researcher understood the importance of this interaction, then he or she could make both substitutions concurrently and achieve the increase in activity suggested by the linear model.

If two residues interact to suppress activity in a nonlinear fashion, the linear model ascribes lower values to the coefficients associate with these residues than would be appropriate if the residues were considered purely in isolation from one another. In other words, making one of the substitutions but not the other for the interacting residues will produce a result in activity that is greater than would be suggested by the linear model.

As a linear model can be inadequate when residue-residue interactions have strong impact on activity, nonlinear models with nonlinear interaction terms accounting for the interactions among residues are often necessary for accurate predictions of activity. However, models that utilize nonlinear terms pose computational and empirical challenges. Most notably, there are a great many potential interaction terms to take into consideration in developing/utilizing a model, which requires a considerable amount of computation. A much bigger limitation is the potential number of observations necessary to produce a model with a significant number of residue-residue interaction terms. Additionally, there may be a tendency for the model creation technique to over fit the data, given a particular number of available observations. To address this challenge, carefully selecting and limiting the interaction terms provided in the sequence-activity model is an important consideration in the development of many models.

FIG. 1 presents a flow chart showing one implementation of a process of preparing a sequence-activity model. As depicted, a process 100 begins at a block 103 to provide sequence and activity data for variant genes (â€œobservationsâ€ ). The sequence data can be taken from, for instance, a training set comprising residue sequences for an initial or improved protein variant library. Typically, these data include complete or partial residue sequence information, together with an activity value for each protein in the library. In some cases, multiple types of activities (e.g., rate constant data and thermal stability data) are provided together in the training set. Other data sources may be considered as well, as determined by the desired results. Some suitable data sources include, but are not limited to, literature references that describe information on particular peptides of relevance to the sequence activity model under construction. Additional information sources include, but are not limited to, earlier or different rounds of directed evolution in the same project. Indeed, it is intended that information derived from previous rounds of directed evolution (using any suitable method, including but not limited to those provided herein) will find use in the development of later produced libraries, variants, etc.

In many embodiments, the individual members of the protein variant library represent a wide range of sequences and activities. This facilitates the generation of a sequence-activity model that is applicable over a broad region of sequence space. Techniques for generating such diverse libraries include, but are not limited to, systematic variation of protein sequences and directed evolution techniques, as described herein. However, in some alternative embodiments, it is desirable to generate models from gene sequences in a particular gene family (e.g., a particular kinase found in multiple species or organisms). As many residues will be identical across all members of the family, the model describes only those residues that vary. Thus, in some embodiments, statistical models based on such relatively small training sets, compared to the set of all possible variants, are valid in a local sense. Namely, the models are valid only for the given observations of the given variants. In some embodiments, the goal is not to find a global fitness function, as it is recognized that in some models, this is beyond the capacity and/or need of the model system(s) under consideration.

Activity data may be obtained using any suitable means known in the art, including but not limited to assays and/or screens appropriately designed to measure magnitudes of the activity/activities of interest. Such techniques are well known and are not essential to the current invention. The principles for designing appropriate assays or screens are widely understood and known in the art. Techniques for obtaining protein sequences are also well known and are not key to the current invention. As mentioned, next-generating sequencing technologies may be used. The activity used with the embodiments described herein may be protein stability (e.g., thermal stability). However, many important embodiments consider other activities such as catalytic activity, resistance to pathogens and/or toxins, therapeutic activity, toxicity, and the like. Indeed, it is not intended that the present invention be limited to any particular assay/screening method(s) and/or sequencing method(s), as any suitable method known in the art finds use in the present invention.

After the training set data have been generated or acquired, the process uses it to generate a base sequence-activity model that predicts activity as a function of sequence information. See block 105. This model is an expression, algorithm or other tool that predicts the relative activity of a particular protein when provided with sequence information for that protein. In other words, protein sequence information is input and an activity prediction is output. In some embodiments, the base model does not include any interaction terms. In such cases, the base model may be described as a â€œlinear model.â€ In other embodiments, the base model includes all available interaction terms, in which case the base model may be described as a non-linear model or an interaction model.

For many embodiments, the base model can rank the contribution of various residues to activity. Methods of generating such models, which all fall under the rubric of machine learning, (e.g., partial least squares regression (PLS), principal component regression (PCR), and multiple linear regression (MLR), Bayesian linear regression) are discussed below, along with the format of the independent variables (sequence information), the format of the dependent variable(s) (activity), and the form of the model itself (e.g., a linear first order expression).

After a base sequence activity model is generated, the process iteratively adds or subtracts interaction terms from a pool of available interaction terms to or from the base model and evaluates the resulting new models for improvement over the base model to produce a final model. See block 107. When the base model includes all available interaction terms, the process subtracts such terms in a step-wise manner. When the base model includes no interaction terms, the process adds such terms in a step-wise manner.

In evaluating a new model, the methods of the present disclosure not only takes the variance that a model accounts for given a set of data into consideration, but also the ability of the model to predict new data. In some embodiments, this model selection approach penalizes models having more coefficients/parameters than equivalent models having fewer coefficients/parameters to avoid over fitting the model to the given data set. Examples of selection methods include, but are not limited to, Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), and variations thereof.

In a series of nested models, as in regression models with progressively more interaction terms (and associated coefficients) than a base model, more complex models provide equally good or better fits than simpler ones even if the extra coefficients are spurious, because the more complex model enjoys extra degrees of freedom. Certain embodiments of the present disclosure employ model selection methods that penalize more complex models to the extent that the gain in goodness of fit is more than offset by the cost of spurious parameters.

Exemplary algorithms for generating sequence-activity models according to the operations in blocks 105 and 107 are presented below. Such techniques include, but are not limited to, step-wise techniques that bias against inclusion of additional interaction terms in a model. However, it is not intended that the present disclosure be limited to these specific examples.

In one aspect, the present disclosure provides methods of preparing a sequence-activity model that can assist in identifying biological molecules to affect a desired activity. In some embodiments, the method comprises: (a) receiving sequence and activity data for a plurality of biological molecules; (b) preparing a base model from the sequence and activity data, wherein the base model predicts activity as a function of the presence or absence of sub-units of the sequence; (c) preparing at least one new model by adding or subtracting at least one new interaction term to or from the base model, wherein the new interaction term represents the interaction between two or more interacting sub-units; (d) determining the ability of the at least one new model to predict activity as a function of the presence or absence of the sub-units; and (e) determining whether to add or subtract the new interaction term to or from the base model based on the ability of the at least one new model to predict activity as determined in (d) and with a bias against adding the new interaction term. The model derived can then be used in various applications, such as in directed evolution of protein libraries to identify proteins with desired biological activities and properties.

In some embodiments, wherein the method determines that the new interaction term should be added to the base model to produce an updated model, the method also includes: (f) repeating (c) using the updated model in place of the base model and adding or subtracting a different interaction term than the one added/subtracted in (c); and (g) repeating (d) and (e) using the updated model in place of the base model. In some embodiments, the method further includes: (h) repeating (f) and (g) using a further updated model.

After the observations for a training set are selected and a mathematical technique for producing the sequence-activity model is chosen, the base model is created. The base model is typically generated without regard for its predictive ability. It is simply produced in accordance with a defined procedure for producing a base model from the available observations (i.e., the observation set), as described herein. As stated above, the sequence models may describe various sequences, while in some embodiments, the models describe proteins. In the latter case, the base model is simply a linear model with a single term for each of the mutations present in the collection of peptides used to create the training set. In these embodiments, the base model does not include any terms representing interactions between residues in the peptides. In some embodiments, the base model does not include a separate term for each and every mutation present in the observation set.

In alternative approaches, the base model includes not only the terms describing each of the mutations in isolation, but additionally includes terms for all the potential interacting residues. In the extreme case, every conceivable interaction between the noted mutations is used in the base model. This includes a term for each and every pairwise interaction between mutations, as well as terms for each and every possible three residue interaction, as well as every possible four residue interaction, etc. Some embodiments include only the pairwise interactions or the pairwise interactions and the three-way interactions. A three-way interaction is an activity-affecting interaction between three distinct sub-units.

In one or more embodiments using a simple linear model as the base model, subsequent efforts to improve the model include adding new terms representing distinct interactions. In alternative embodiments where the base model includes all the linear and non-linear terms, subsequent efforts to improve the model involves selectively removing some of the non-linear, interaction terms.

In one or more embodiments of the invention, the process of improving the base model involves iteratively adding or subtracting interaction terms from the base model in determining whether the resulting model sufficiently improves the quality of the model. At each iteration, the predictive power of the current model is determined and compared to another model, e.g., the base model or the updated model.

In embodiments in which a measure of predictive power already takes into account the ability of a model to be generalized to other sets of data, that measure alone can determine whether a candidate model should be selected. For instance, a measure such as AIC or BIC takes both the model likelihood (or residual error) and the number of parameters into consideration. A â€œlikelihood functionâ€ or â€œlikelihoodâ€ of a model is a function of the parameters of a statistical model. The likelihood of a set of parameter values given some observed outcomes equals to the probability of those observed outcomes given those parameter values, i.e. L(Î¸|x)=P(x|Î¸). An exemplary calculation of model likelihood is described in a section below. Measures such as AIC and BIC are biased against a model having more parameters if the model having more parameters captures the same amount of data variance as does a model having fewer parameters. If a measure of predictive power only considers residual error, the magnitude of the improvement in residual error must be considered in order to determine whether or not to incorporate the change associated with the current iteration into the current best updated model. This may be accomplished by comparing the magnitude of the improvement against a threshold. If the magnitude is less than the threshold, the change under consideration in the current iteration is not accepted. If, alternatively, the magnitude of the improvement exceeds the threshold, then the change under consideration is incorporated into the updated model and the updated model serves as the new best model going forward for the remaining iterations.

In certain embodiments, each iteration considers the addition or subtraction of a single interaction term from the current best model under consideration. In the case of an additive model, i.e., the case for the base model contains only linear terms, a pool of all available interaction terms can be considered. Each of these interaction terms is considered in succession until the process is completed and a final best model is obtained.

In some cases, upon determining that the process has effectively converged and further improvement is unlikely, the model generation process is terminated before all of the available interaction terms in the pool have been considered.

FIG. 2 illustrates how a model can be iteratively used to guide the creation of new protein variant libraries for the purpose of exploring protein sequence and activity space, in a process (See, 200). After a final model is generated, the final model is employed to identify multiple residue positions (e.g., position 35) or specific residue values (e.g. glutamine at position 35) that are predicted to impact activity. See block 207. In addition to identifying such positions, the model may be used to â€œrankâ€ the residue positions or residue values based on their contributions to the desired activity (activities?). For example, the model may predict that glutamine at position 35 has the most pronounced, positive effect on activity; phenylalanine at position 208 has the second most pronounced, positive effect on activity; and so on. In one specific approach described below, PLS or PCR regression coefficients are employed to rank the importance of specific residues. In another specific approach, a PLS load matrix is employed to rank the importance of specific residue positions.

After the process has identified residues that impact activity, some of them are selected for variation as indicated at a block 209 (FIG. 2). This is done for the purpose of exploring sequence space. Residues are selected using any of a number of different selection protocols, some of which are described below. In one illustrative example, specific residues predicted to have the most beneficial impact on activity are preserved (i.e., not varied). A certain number of other residues predicted to have a lesser impact are, however, selected for variation. In another illustrative example, the residue positions found to have the biggest impact on activity are selected for variation, but only if they are found to vary in high performing members of the training set. For example, if the model predicts that residue position 197 has the biggest impact on activity, but all or most of the proteins with high activity have leucine at this position, position 197 would not be selected for variation in this approach. In other words, all or most proteins in a next generation library would have leucine at position 197. However, if some â€œgoodâ€ proteins had valine at this position and others had leucine, then the process would choose to vary the amino acid at this position. In some cases, it will be found that a combination of two or more interacting residues have the biggest impact on activity. Hence, in some strategies, these residues are co-varied.

After the residues for variation have been identified, the method next generates a new variant library having the specified residue variation. See block 211 (FIG. 2). Various methodologies are available for this purpose. In one example, an in vitro or in vivo recombination-based diversity generation mechanism is performed to generate the new variant library. Such procedures may employ oligonucleotides containing sequences or subsequences for encoding the proteins of the parental variant library. Some of the oligonucleotides will be closely related, differing only in the choice of codons for alternate amino acids selected for variation at 209. The recombination-based diversity generation mechanism may be performed for one or multiple cycles. If multiple cycles are used, each involves a screening step to identify which variants have acceptable performance to be used in a subsequent recombination cycle. This is a form of directed evolution. However, it is not intended that the present invention be limited to any specific method of recombination-based diversity generation method, as any suitable method/technique finds use in the present invention.

In an additional illustrative example, a â€œreferenceâ€ protein sequence is chosen and the residues selected at 209 of FIG. 2 are â€œtoggledâ€ to identify individual members of the variant library. The new proteins so identified are synthesized by an appropriate technique to generate the new library. In one example, the reference sequence may be a top-performing member of the training set or a â€œbestâ€ sequence predicted by a PLS or PCR model.

In another illustrative example, residues for variation in a round of directed evolution are selected in a single parent sequence. The parent may be identified using model results from a prior round of directed evolution or by using data that identify the library member having the best assay performance. Oligonucleotides for the next round of directed evolution may be defined to include portions of the selected parent's backbone with one or more mutations predicted algorithmically from a sequence activity model for the current round. These oligonucleotides may be produced using any suitable means, including but not limited to synthetic methods.

After the new library has been produced, it is screened for activity, as indicated in a block 213 (FIG. 2). Ideally, the new library provides one or more members with better activity than was observed in the previous library. However, even without such an advantage, the new library can provide beneficial information. Its members may be employed for generating improved models that account for the effects of the variations selected in 209 (FIG. 2), and thereby more accurately predict activity across wider regions of sequence space. Further, the library may represent a passage in sequence space from a local maximum toward a global maximum (e.g., in activity).

Depending on the goal of process 200 (FIG. 2), in some embodiments, it is desirable to generate a series of new protein variant libraries, with each one providing new members of a training set. The updated training set is then used to generate an improved model. To achieve the improved model, process 200 is shown with a decision operation as shown in block 215, which determines whether yet another protein variant library should be produced. Various criteria can be used to make this decision. Examples of decision criteria include but are not limited to the number of protein variant libraries generated so far, the activity of top proteins from the current library, the magnitude of activity desired, and the level of improvement observed in recent new libraries.

Assuming that the process is used to continue with a new library, the process returns to operation of block 100 (FIG. 2) where a new sequence-activity model is generated from sequence and activity data obtained for the current protein variant library. In other words, the sequence and activity data for the current protein variant library serves as part of the training set for the new model (or it may serve as the entire training set). Thereafter, operations shown in blocks 207, 209, 211, 213, and 215 (FIG. 2) are performed as described above, but with the new model.

When it is determined that the end-point of the method has been reached, the cycle illustrated in FIG. 2 is ended and no new library is generated. At that point, the process is either simply terminated or in some embodiments, one or more sequences from one or more of the libraries is/are selected for development and/or manufacture. See block 217.

B. Generating Observations

Protein variant libraries are groups of multiple proteins having one or more residues that vary from member to member in a library. These libraries may be generated using the methods described herein and/or any suitable means known in the art. These libraries find use in providing data for training sets used to generate sequence-activity models in accordance with various embodiments of the present invention. The number of proteins included in a protein variant library often depends on the application and the cost associated with their generation. It is not intended that the present invention be limited to any particular number of proteins in the protein libraries used in the methods of the present invention. It is further not intended that the present invention be limited to any particular protein variant library or libraries.

In one example, the protein variant library is generated from one or more naturally occurring proteins, which may be encoded by a single gene family. Other starting points including but not limited to recombinants of known proteins or novel synthetic proteins may be used. From these seed or starting proteins, the library may be generated by various techniques. In one case, the library is generated by DNA fragmentation-mediated recombination as described in Stemmer (1994) Proceedings of the National Academy of Sciences, USA, 10747-10751 and WO 95/22625 (both of which are incorporated herein by reference), synthetic oligonucleotide-mediated recombination as described in Ness et al. (2002) Nature Biotechnology 20:1251-1255 and WO 00/42561 (both of which are incorporated herein by reference), or nucleic acids encoding part or all of one or more parent proteins. Combinations of these methods may be used as well (e.g., recombination of DNA fragments and synthetic oligonucleotides) as well as other recombination-based methods described in, for example, WO97/20078 and WO98/27230, both of which are incorporated herein by reference. Any suitable methods used to generate protein variant libraries find use in the present invention. Indeed, it is not intended that the present invention be limited to any particular method for producing variant libraries.

In some embodiments, a single â€œstartingâ€ sequence (which may be an â€œancestorâ€ sequence) may be employed for purposes of defining a group of mutations used in the modeling process. In some embodiments, at least one of the starting sequence is a wild-type sequence.

In certain embodiments, the mutations are (a) identified in the literature as affecting substrate specificity, selectivity, stability, or other beneficial property and/or (b) computationally predicted to improve protein folding patterns (e.g., packing the interior residues of a protein), ligand binding, subunit interactions, family shuffling between multiple diverse homologs, etc. Alternatively, the mutations may be physically introduced into the starting sequence and the expression products screened for beneficial properties. Site directed mutagenesis is one example of a useful technique for introducing mutations, although any suitable method finds use. Thus, alternatively or in addition, the mutants may be provided by gene synthesis, saturating random mutagenesis, semi-synthetic combinatorial libraries of residues, directed evolution, recursive sequence recombination (â€œRSRâ€ ) (See e.g., US Patent Application No. 2006/0223143, incorporated by reference herein in its entirety), gene shuffling, error-prone PCR, and/or any other suitable method. One example of a suitable saturation mutagenesis procedure is described in US Published Patent Application No. 20100093560, which is incorporated herein by reference in its entirety.

The starting sequence need not be identical to the amino acid sequence of the wild type protein. However, in some embodiments, the starting sequence is the sequence of the wild type protein. In some embodiments, the starting sequence includes mutations not present in the wild type protein. In some embodiments, the starting sequence is a consensus sequence derived from a group of proteins having a common property, e.g., a family of proteins.

A non-limiting representative list of families or classes of enzymes which may serve as sources of parent sequences includes, but is not limited to the following: oxidoreducatses (E.C.1); transferases (E.C.2); hydrolyases (E.C.3); lyases (E.C.4); isomerases (E.C. 5) and ligases (E.C. 6). More specific but non-limiting subgroups of oxidoreducatses include dehydrogenases (e.g., alcohol dehydrogenases (carbonyl reductases), xylulose reductases, aldehyde reductases, farnesol dehydrogenase, lactate dehydrogenases, arabinose dehydrogenases, glucose dehyrodgenase, fructose dehydrogenases, xylose reductases and succinate dehyrogenases), oxidases (e.g., glucose oxidases, hexose oxidases, galactose oxidases and laccases), monoamine oxidases, lipoxygenases, peroxidases, aldehyde dehydrogenases, reductases, long-chain acyl-[acyl-carrier-protein] reductases, acyl-CoA dehydrogenases, ene-reductases, synthases (e.g., glutamate synthases), nitrate reductases, mono and di-oxygenases, and catalases. More specific but non-limiting subgroups of transferases include methyl, amidino, and carboxyl transferases, transketolases, transaldolases, acyltransferases, glycosyltransferases, transaminases, transglutaminases and polymerases. More specific but non-limiting subgroups of hydrolases include ester hydrolases, peptidases, glycosylases, amylases, cellulases, hemicellulases, xylanases, chitinases, glucosidases, glucanases, glucoamylases, acylases, galactosidases, pullulanases, phytases, lactases, arabinosidases, nucleosidases, nitrilases, phosphatases, lipases, phospholipases, proteases, ATPases, and dehalogenases. More specific but non-limiting subgroups of lyases include decarboxylases, aldolases, hydratases, dehydratases (e.g., carbonic anhydrases), synthases (e.g., isoprene, pinene and farnesene synthases), pectinases (e.g., pectin lyases) and halohydrin dehydrogenases. More specific, but non-limiting subgroups of isomerases include racemases, epimerases, isomerases (e.g., xylose, arabinose, ribose, glucose, galactose and mannose isomerases), tautomerases, and mutases (e.g. acyl transferring mutases, phosphomutases, and aminomutases. More specific but non-limiting subgroups of ligases include ester synthases. Other families or classes of enzymes which may be used as sources of parent sequences include transaminases, proteases, kinases, and synthases. This list, while illustrating certain specific aspects of the possible enzymes of the disclosure, is not considered exhaustive and does not portray the limitations or circumscribe the scope of the disclosure.

In some cases, the candidate enzymes useful in the methods described herein are capable of catalyzing an enantioselective reaction such as an enantioselective reduction reaction, for example. Such enzymes can be used to make intermediates useful in the synthesis of pharmaceutical compounds for example.

In some embodiments, the candidate enzymes are selected from endoxylanases (EC 3.2.1.8); Î²-xylosidases (EC 3.2.1.37); alpha-L-arabinofuranosidases (EC 3.2.1.55); alpha-glucuronidases (EC 3.2.1.139); acetylxylanesterases (EC 3.1.1.72); feruloyl esterases (EC 3.1.1.73); coumaroyl esterases (EC 3.1.1.73); alpha-galactosidases (EC 3.2.1.22); beta-galactosidases (EC 3.2.1.23); beta-mannanases (EC 3.2.1.78); beta-mannosidases (EC 3.2.1.25); endo-polygalacturonases (EC 3.2.1.15); pectin methyl esterases (EC 3.1.1.11); endo-galactanases (EC 3.2.1.89); pectin acetyl esterases (EC 3.1.1.6); endo-pectin lyases (EC 4.2.2.10); pectate lyases (EC 4.2.2.2); alpha rhamnosidases (EC 3.2.1.40); exo-poly-alpha-galacturonosidase (EC 3.2.1.82); 1,4-alpha-galacturonidase (EC 3.2.1.67); exopolygalacturonate lyases (EC 4.2.2.9); rhamnogalacturonan endolyases EC (4.2.2.B3); rhamnogalacturonan acetylesterases (EC 3.2.1.B11); rhamnogalacturonan galacturonohydrolases (EC 3.2.1.B11); endo-arabinanases (EC 3.2.1.99); laccases (EC 1.10.3.2); manganese-dependent peroxidases (EC 1.10.3.2); amylases (EC 3.2.1.1), glucoamylases (EC 3.2.1.3), proteases, lipases, and lignin peroxidases (EC 1.11.1.14). Any combination of one, two, three, four, five, or more than five enzymes find use in the compositions of the present invention.

In one or more embodiments of the invention, a single starting sequence is modified in various ways to generate the library. In some embodiments, the library is generated by systematically varying the individual residues of the starting sequence. In one illustrative example, a design of experiment (DOE) methodology is employed to identify the systematically varied sequences. In another example, a â€œwet labâ€ procedure such as oligonucleotide-mediated recombination is used to introduce some level of systematic variation. It is not intended that the present invention be limited to any particular method for generating systematically varied sequences, as any suitable method finds use.

As used herein, the term â€œsystematically varied sequencesâ€ refers to a set of sequences in which each residue is seen in multiple contexts. In principle, the level of systematic variation can be quantified by the degree to which the sequences are orthogonal from one another (i.e., maximally different compared to the mean). In some embodiments, the process does not depend on having maximally orthogonal sequences. However, the quality of the model will be improved in direct relation to the orthogonality of the sequence space tested. In a simple illustrative example, a peptide sequence is systematically varied by identifying two residue positions, each of which can have one of two different amino acids. A maximally diverse library includes all four possible sequences. Such maximal systematic variation increases exponentially with the number of variable positions; e.g., by 2^N, when there are 2 options at each of N residue positions. Those having ordinary skill in the art will readily recognize that maximal systematic variation, however, is not required. Systematic variation provides a mechanism for identifying a relatively small set of sequences for testing that provides a good sampling of sequence space.

Protein variants having systematically varied sequences can be obtained in a number of ways using techniques that are well known to those of ordinary skill in the art. As indicated, suitable methods include, but are not limited to recombination-based methods that generate variants based on one or more â€œparentalâ€ polynucleotide sequences. Polynucleotide sequences can be recombined using a variety of techniques, including, for example, DNAse digestion of polynucleotides to be recombined followed by ligation and/or PCR reassembly of the nucleic acids. These methods include, but are not limited to those described in, for example, Stemmer (1994) Proceedings of the National Academy of Sciences USA, 91:10747-10751, U.S. Pat. No. 5,605,793, â€œMethods for In Vitro Recombination,â€ U.S. Pat. No. 5,811,238, â€œMethods for Generating Polynucleotides having Desired Characteristics by Iterative Selection and Recombination,â€ U.S. Pat. No. 5,830,721, â€œDNA Mutagenesis by Random Fragmentation and Reassembly,â€ U.S. Pat. No. 5,834,252, â€œEnd Complementary Polymerase Reaction,â€ U.S. Pat. No. 5,837,458, â€œMethods and Compositions for Cellular and Metabolic Engineering,â€ WO98/42832, â€œRecombination of Polynucleotide Sequences Using Random or Defined Primers,â€ WO 98/27230, â€œMethods and Compositions for Polypeptide Engineering,â€ WO 99/29902, â€œMethod for Creating Polynucleotide and Polypeptide Sequences,â€ and the like, all of which are incorporated herein by reference.

Synthetic recombination methods are also particularly well suited for generating protein variant libraries with systematic variation. In synthetic recombination methods, a plurality of oligonucleotides are synthesized which collectively encode a plurality of the genes to be recombined. In some embodiments, the oligonucleotides collectively encode sequences derived from homologous parental genes. For example, homologous genes of interest are aligned using a sequence alignment program such as BLAST (See e.g., Atschul, et al., Journal of Molecular Biology, 215:403-410 (1990). Nucleotides corresponding to amino acid variations between the homologues are noted. These variations are optionally further restricted to a subset of the total possible variations based on covariation analysis of the parental sequences, functional information for the parental sequences, selection of conservative or non-conservative changes between the parental sequences, or other suitable criteria. Variations are optionally further increased to encode additional amino acid diversity at positions identified by, for example, covariation analysis of the parental sequences, functional information for the parental sequences, selection of conservative or non-conservative changes between the parental sequences, or apparent tolerance of a position for variation. The result is a degenerate gene sequence encoding a consensus amino acid sequence derived from the parental gene sequences, with degenerate nucleotides at positions encoding amino acid variations. Oligonucleotides are designed which contain the nucleotides required to assemble the diversity present in the degenerate gene. Details regarding such approaches can be found in, for example, Ness et al. (2002), Nature Biotechnology, 20:1251-1255, WO 00/42561, â€œOligonucleotide Mediated Nucleic Acid Recombination,â€ WO 00/42560, â€œMethods for Making Character Strings, Polynucleotides and Polypeptides having Desired Characteristics,â€ WO 01/75767, â€œIn Silico Cross-Over Site Selection,â€ and WO 01/64864, â€œSingle-Stranded Nucleic Acid Template-Mediated Recombination and Nucleic Acid Fragment Isolation,â€ each of which is incorporated herein by reference. The identified polynucleotide variant sequences may be transcribed and translated, either in vitro or in vivo, to create a set or library of protein variant sequences.

The set of systematically varied sequences can also be designed a priori using design of experiment (DOE) methods to define the sequences in the data set. A description of DOE methods can be found in Diamond, W. J. (2001) Practical Experiment Designs: for Engineers and Scientists, John Wiley & Sons and in â€œPractical Experimental Design for Engineers and Scientistsâ€ by William J Drummond (1981) Van Nostrand Reinhold Co New York, â€œStatistics for experimentersâ€ George E. P. Box, William G Hunter and J. Stuart Hunter (1978) John Wiley and Sons, New York, or, e.g., on the world wide web at itl.nist.gov/div898/handbook/. There are several computational packages available to perform the relevant mathematics, including Statistics Toolbox (MATLABÂ®), JMPÂ®, STATISTICAÂ®, and STAT-EASEÂ® DESIGN EXPERTÂ®. The result is a systematically varied and orthogonal dispersed data set of sequences that is suitable for building the sequence-activity model of the present invention. DOE-based data sets can also be readily generated using either Plackett-Burman or Fractional Factorial Designs, as known in the art. Diamond, W. J. (2001).

In engineering and chemical sciences, fractional factorial designs are used to define fewer experiments as compared to full factorial designs. In these methods, a factor is varied (i.e., â€œtoggledâ€ ) between two or more levels. Optimization techniques are used to ensure that the experiments chosen are maximally informative in accounting for factor space variance. The same design approaches (e.g., fractional factorial, D-optimal design) can be applied in protein engineering to construct fewer sequences where a given number of positions are toggled between two or more residues. In some embodiments, this set of sequences provides an optimal description of systematic variance present in the protein sequence space in question.

An illustrative example of the DOE approach applied to protein engineering includes the following operations:

- 1) Identify positions to toggle based on the principles described herein (e.g., present in parental sequences, level of conservation, etc.)
- 2) Create a DOE experiment using one of the commonly available statistical software packages by defining the number of factors (i.e., variable positions), the number of levels (i.e., choices at each position), and the number of experiments to run to provide an output matrix. The information content of the output matrix (typically consisting of 1s and 0s that represent residue choices at each position) depends directly on the number of experiments to run (typically, the more the better).
- 3) Use the output matrix to construct a protein alignment that codes the 1s and 0s back to specific residue choices at each position.
- 4) Synthesize the genes encoding the proteins represented in the protein alignment.
- 5) Test the proteins encoded by the synthesized genes in relevant assay(s).
- 6) Build a model based on the tested genes/proteins.
- 7) Follow the steps described herein to identify positions of importance and to build one or more subsequent libraries with improved fitness.

In an illustrative example, a protein in which the functionally best amino acid residues at 20 positions are to be determined (e.g., where there are 2 possible amino acids available at each position) is investigated. In this example, a resolution IV factorial design would be appropriate. A resolution IV design is defined as a design that is capable of elucidating the effects of all single variables, with no two-factor effects overlapping them. The design would then specify a set of 40 specific amino acid sequences covering the total diversity of 2²⁰(Ëœ1 million) possible sequences. These sequences are then generated using any standard gene synthesis protocol and the function and fitness of these clones is determined.

An alternative to the above approaches is to employ some or all available sequences (e.g., the GENBANKÂ® database and other public sources) to provide the protein variant library. This approach provides an indication of sequence space regions of interest.

C. Sequencing Methods

Historically, sequencing has been a limiting step in the development of large training sets and consequently, increasingly robust sequence-activity models. The high cost and long time required to sequence variants limited the number of observations to a few tens of variants. Next generation sequencing tools have greatly reduced cost and increased sequencing speed and volume, making it possible to include both low and high activity variants in a training set.

Next-generation sequencing tools can inexpensively sequence large numbers of base pairs (e.g., at least about 1,000,000,000 base pairs) in one run. This capacity can be utilized when sequencing variant proteins, which are typically only a few kilobase pairs in length, in a single run. Often next-generation sequencing tools are optimized for sequencing single large genomes (e.g., the human genome) rather than many smaller sequences in a single run. To realize the potential of next-generation sequencing tools for sequencing many observations in parallel, the origin of each of the observations being sequenced in a single run should be uniquely identified. In some embodiments, bar-coded sequences are used on each and every fragment fed to a next-generation sequencer for a single run. In one example, barcodes uniquely identify a particular well on a particular plate (e.g., 96 well plates). In some of these embodiments, each well of each plate contains a single unique variant. By barcoding each variant, or more specifically each fragment of each variant, the gene sequences of multiple different variants can be sequenced and identified in a single run. In the process, all fragment reads having the same barcode are identified and processed together by the algorithm identifying length sequences for the variants.

In some embodiments, the DNA from the cells of a variant in a given well is extracted and then fragmented. The fragments are then bar-coded to identify at least the well, and sometimes the well and plate associated with that variant. The resulting fragments are then size selected to produce sequences of appropriate length for the next-generation sequencer. In one illustrative example, the read lengths are about 200 base pairs. In some embodiments, the plate barcode is not applied until after the DNA fragments from the various wells of a plate are first pooled. The pooled DNA is then bar-coded to identify the plate. In some embodiments, each fragment, regardless of which well it is derived from, will have the same plate barcode. However, in some alternative embodiments, the fragments have different barcodes. In addition, the well and plate barcodes may be applied to identify the DNA extracted from a given well.

In one or more embodiments, sequence data can be obtained using bulk sequencing methods including, for example, Sanger sequencing or Maxam-Gilbert sequencing, which are considered the first generation sequencing methods. Sanger sequencing, which involves using labeled dideoxy chain terminators, is well known in the art; see, e.g., Sanger et al., Proceedings of the National Academy of Sciences of the United States of America 74, 5463-5467 (1997). Maxam-Gilbert sequencing, which involves performing multiple partial chemical degradation reactions on fractions of the nucleic acid sample followed by detection and analysis of the fragments to infer the sequence, is also well known in the art; see, e.g., Maxam et al., Proceedings of the National Academy of Sciences of the United States of America 74, 560-564 (1977). Another bulk sequencing method is sequencing by hybridization, in which the sequence of a sample is deduced based on its hybridization properties to a plurality of sequences, e.g., on a microarray or gene chip; see, e.g., Drmanac, et al., Nature Biotechnology 16, 54-58 (1998).

In one or more embodiments, sequence data is obtained using next-generation sequencing methods. Next-generation sequencing is also referred to as â€œhigh-throughput sequencingâ€ . The techniques parallelize the sequencing process, producing thousands or millions of sequences at once. Examples of suitable next-generation sequencing methods include, but are not limited to, single molecule real-time sequencing (e.g., Pacific Biosciences, Menlo Park, Calif.), ion semiconductor sequencing (e.g., Ion Torrent, South San Francisco, Calif.), pyrosequencing (e.g., 454, Branford, Conn.), sequencing by ligation (e.g., SOLid sequencing of Life Technologies, Carlsbad, Calif.), sequencing by synthesis and reversible terminator (e.g., Illumina, San Diego, Calif.), nucleic acid imaging technologies such as transmission electron microscopy, and the like.

In general, next-generation sequencing methods typically use an in vitro cloning step to amplify individual DNA molecules. Emulsion PCR (emPCR) isolates individual DNA molecules along with primer-coated beads in aqueous droplets within an oil phase. PCR produces copies of the DNA molecule, which bind to primers on the bead, followed by immobilization for later sequencing. emPCR is used in the methods by Marguilis et al. (commercialized by 454 Life Sciences, Branford, Conn.), Shendure and Porreca et al. (also known as â€œpolony sequencingâ€ ) and SOLiD sequencing, (Applied Biosystems Inc., Foster City, Calif.). See M. Margulies, et al. (2005) â€œGenome sequencing in microfabricated high-density picolitre reactorsâ€ Nature 437: 376-380; J. Shendure, et al. (2005) â€œAccurate Multiplex Polony Sequencing of an Evolved Bacterial Genomeâ€ Science 309 (5741): 1728-1732. In vitro clonal amplification can also be carried out by â€œbridge PCR,â€ where fragments are amplified upon primers attached to a solid surface. Braslaysky et al. developed a single-molecule method (commercialized by Helicos Biosciences Corp., Cambridge, Mass.) that omits this amplification step, directly fixing DNA molecules to a surface. I. Braslaysky, et al. (2003) â€œSequence information can be obtained from single DNA moleculesâ€ Proceedings of the National Academy of Sciences of the United States of America 100: 3960-3964.

DNA molecules that are physically bound to a surface can be sequenced in parallel. In â€œsequencing by synthesis,â€ a complementary strand is built based on the sequence of a template strand using a DNA polymerase like dye-termination electrophoretic sequencing, Reversible terminator methods (commercialized by Illumina, Inc., San Diego, Calif. and Helicos Biosciences Corp., Cambridge, Mass.) use reversible versions of dye-terminators, adding one nucleotide at a time, and detect fluorescence at each position in real time, by repeated removal of the blocking group to allow polymerization of another nucleotide. â€œPyrosequencingâ€ also uses DNA polymerization, adding one nucleotide at a time and detecting and quantifying the number of nucleotides added to a given location through the light emitted by the release of attached pyrophosphates (commercialized by 454 Life Sciences, Branford, Conn.). See M. Ronaghi, et al. (1996). â€œReal-time DNA sequencing using detection of pyrophosphate releaseâ€ Analytical Biochemistry 242: 84-89.

Specific examples of next-generation sequencing methods are described in further details below. One or more implementations of the current invention may use one or more of the following sequencing methods without deviating from the principles of the invention.

Single molecule real time sequencing (also known as SMRT) is a parallelized single molecule DNA sequencing by synthesis technology developed by Pacific Biosciences. Single molecule real time sequencing utilizes the zero-mode waveguide (ZMW). A single DNA polymerase enzyme is affixed at the bottom of a ZMW with a single molecule of DNA as a template. The ZMW is a structure that creates an illuminated observation volume that is small enough to observe only a single nucleotide of DNA (also known as a base) being incorporated by DNA polymerase. Each of the four DNA bases is attached to one of four different fluorescent dyes. When a nucleotide is incorporated by the DNA polymerase, the fluorescent tag is cleaved off and diffuses out of the observation area of the ZMW where its fluorescence is no longer observable. A detector detects the fluorescent signal of the nucleotide incorporation, and the base call is made according to the corresponding fluorescence of the dye.

Another single molecule sequencing technology applicable is the Helicos True Single Molecule Sequencing (tSMS) technology (e.g. as described in Harris T. D. et al., Science 320:106-109 [2008]). In the tSMS technique, a DNA sample is cleaved into strands of approximately 100 to 200 nucleotides, and a polyA sequence is added to the 3â€² end of each DNA strand. Each strand is labeled by the addition of a fluorescently labeled adenosine nucleotide. The DNA strands are then hybridized to a flow cell, which contains millions of oligo-T capture sites that are immobilized to the flow cell surface. In certain embodiments the templates can be at a density of about 100 million templates/cm². The flow cell is then loaded into an instrument, e.g., HeliScopeâ„¢ sequencer, and a laser illuminates the surface of the flow cell, revealing the position of each template. A CCD camera can map the position of the templates on the flow cell surface. The template fluorescent label is then cleaved and washed away. The sequencing reaction begins by introducing a DNA polymerase and a fluorescently labeled nucleotide. The oligo-T nucleic acid serves as a primer. The polymerase incorporates the labeled nucleotides to the primer in a template directed manner. The polymerase and unincorporated nucleotides are removed. The templates that have directed incorporation of the fluorescently labeled nucleotide are discerned by imaging the flow cell surface. After imaging, a cleavage step removes the fluorescent label, and the process is repeated with other fluorescently labeled nucleotides until the desired read length is achieved. Sequence information is collected with each nucleotide addition step. Whole genome sequencing by single molecule sequencing technologies excludes or typically obviates PCR-based amplification in the preparation of the sequencing libraries, and the methods allow for direct measurement of the sample, rather than measurement of copies of that sample.

Ion Semiconductor Sequencing is a method of DNA sequencing based on the detection of hydrogen ions that are released during the polymerization of DNA. This is a method of â€œsequencing by synthesis,â€ during which a complementary strand is built based on the sequence of a template strand. A microwell containing a template DNA strand to be sequenced is flooded with a single species of deoxyribonucleotide triphosphate (dNTP). If the introduced dNTP is complementary to the leading template nucleotide, it is incorporated into the growing complementary strand. This causes the release of a hydrogen ion that triggers an ISFET ion sensor, which indicates that a reaction has occurred. If homopolymer repeats are present in the template sequence, multiple dNTP molecules will be incorporated in a single cycle. This leads to a corresponding number of released hydrogens and a proportionally higher electronic signal. This technology differs from other sequencing technologies in that no modified nucleotides or optics are used. Ion semiconductor sequencing may also be referred to as ion torrent sequencing, pH-mediated sequencing, silicon sequencing, or semiconductor sequencing.

In pyrosequencing, the pyrophosphate ion released by the polymerization reaction is reacted with adenosine 5â€² phosphosulfate by ATP sulfurylase to produce ATP; the ATP then drives the conversion of luciferin to oxyluciferin plus light by luciferase. As the fluorescence is transient, no separate step to eliminate fluorescence is necessary in this method. One type of deoxyribonucleotide triphosphate (dNTP) is added at a time, and sequence information is discerned according to which dNTP generates significant signal at a reaction site. The commercially available Roche GS FLX instrument acquires sequence using this method. This technique and applications thereof are discussed in detail, for example, in Ronaghi et al., Analytical Biochemistry 242, 84-89 (1996) and Margulies et al., Nature 437, 376-380 (2005) (corrigendum at Nature 441, 120 (2006)). A commercially available pyrosequencing technology is 454 sequencing (Roche) (e.g. as described in Margulies, M. et al. Nature 437:376-380 [2005]).

In ligation sequencing, a ligase enzyme is used to join a partially double-stranded oligonucleotide with an overhang to the nucleic acid being sequenced, which has an overhang; in order for ligation to occur, the overhangs must be complementary. The bases in the overhang of the partially double-stranded oligonucleotide can be identified according to a fluorophore conjugated to the partially double-stranded oligonucleotide and/or to a secondary oligonucleotide that hybridizes to another part of the partially double-stranded oligonucleotide. After acquisition of fluorescence data, the ligated complex is cleaved upstream of the ligation site, such as by a type IIs restriction enzyme, for example, Bbvl, which cuts at a site a fixed distance from its recognition site (which was included in the partially double stranded oligonucleotide). This cleavage reaction exposes a new overhang just upstream of the previous overhang, and the process is repeated. This technique and applications thereof are discussed in detail, for example, in Brenner et al., Nature Biotechnology 18, 630-634 (2000). In some embodiments, ligation sequencing is adapted to the methods of the invention by obtaining a rolling circle amplification product of a circular nucleic acid molecule, and using the rolling circle amplification product as the template for ligation sequencing.

A commercially available example of ligation sequencing technology is the SOLiDâ„¢ technology (Applied Biosystems). In SOLiDâ„¢ sequencing-by-ligation, genomic DNA is sheared into fragments, and adaptors are attached to the 5â€² and 3â€² ends of the fragments to generate a fragment library. Alternatively, internal adaptors can be introduced by ligating adaptors to the 5â€² and 3â€² ends of the fragments, circularizing the fragments, digesting the circularized fragment to generate an internal adaptor, and attaching adaptors to the 5â€² and 3â€² ends of the resulting fragments to generate a mate-paired library. Next, clonal bead populations are prepared in microreactors containing beads, primers, template, and PCR components. Following PCR, the templates are denatured and beads are enriched to separate the beads with extended templates. Templates on the selected beads are subjected to a 3â€² modification that permits bonding to a glass slide. The sequence can be determined by sequential hybridization and ligation of partially random oligonucleotides with a central determined base (or pair of bases) that is identified by a specific fluorophore. After a color is recorded, the ligated oligonucleotide is cleaved and removed and the process is then repeated.

In reversible terminator sequencing, a fluorescent dye-labeled nucleotide analog that is a reversible chain terminator due to the presence of a blocking group is incorporated in a single-base extension reaction. The identity of the base is determined according to the fluorophore; in other words, each base is paired with a different fluorophore. After fluorescence/sequence data is acquired, the fluorophore and the blocking group are chemically removed, and the cycle is repeated to acquire the next base of sequence information. The Illumina GA instrument operates by this method. This technique and applications thereof are discussed in detail, for example, in Ruparel et al., Proceedings of the National Academy of Sciences of the United States of America 102, 5932-5937 (2005), and Harris et al., Science 320, 106-109 (2008).

A commercially available example of reversible terminator sequencing method is Illumina's sequencing-by-synthesis and reversible terminator-based sequencing (e.g. as described in Bentley et al., Nature 6:53-59 [2009]). Illumina's sequencing technology relies on the attachment of fragmented genomic DNA to a planar, optically transparent surface on which oligonucleotide anchors are bound. Template DNA is end-repaired to generate 5â€²-phosphorylated blunt ends, and the polymerase activity of Klenow fragment is used to add a single A base to the 3â€² end of the blunt phosphorylated DNA fragments. This addition prepares the DNA fragments for ligation to oligonucleotide adapters, which have an overhang of a single T base at their 3â€² end to increase ligation efficiency. The adapter oligonucleotides are complementary to the flow-cell anchors. Under limiting-dilution conditions, adapter-modified, single-stranded template DNA is added to the flow cell and immobilized by hybridization to the anchors. Attached DNA fragments are extended and bridge amplified to create an ultra-high density sequencing flow cell with hundreds of millions of clusters, each containing Ëœ1,000 copies of the same template. The templates are sequenced using a robust four-color DNA sequencing-by-synthesis technology that employs reversible terminators with removable fluorescent dyes. High-sensitivity fluorescence detection is achieved using laser excitation and total internal reflection optics. Short sequence reads of about 20-40 bp e.g. 36 bp, are aligned against a repeat-masked reference genome and unique mapping of the short sequence reads to the reference genome are identified using specially developed data analysis pipeline software. Non-repeat-masked reference genomes can also be used. Whether repeat-masked or non-repeat-masked reference genomes are used, only reads that map uniquely to the reference genome are counted. After completion of the first read, the templates can be regenerated in situ to enable a second read from the opposite end of the fragments. Thus, either single-end or paired end sequencing of the DNA fragments can be used. Partial sequencing of DNA fragments present in the sample is performed, and sequence tags comprising reads of predetermined length e.g. 36 bp, are mapped to a known reference genome are counted.

In nanopore sequencing, a single stranded nucleic acid molecule is threaded through a pore, e.g., using an electrophoretic driving force, and sequence is deduced by analyzing data obtained as the single stranded nucleic acid molecule passes through the pore. The data can be ion current data, wherein each base alters the current, e.g., by partially blocking the current passing through the pore to a different, distinguishable degree.

In another illustrative, but non-limiting, embodiment, the methods described herein comprises obtaining sequence information using transmission electron microscopy (TEM). The method comprises utilizing single atom resolution transmission electron microscope imaging of high-molecular weight (150 kb or greater) DNA selectively labeled with heavy atom markers and arranging these molecules on ultra-thin films in ultra-dense (3 nm strand-to-strand) parallel arrays with consistent base-to-base spacing. The electron microscope is used to image the molecules on the films to determine the position of the heavy atom markers and to extract base sequence information from the DNA. The method is further described in PCT patent publication WO 2009/046445.

In another illustrative, but non-limiting, embodiment, the methods described herein comprises obtaining sequence information using third-generation sequencing. In third-generation sequencing, a slide with an aluminum coating with many small (^Ëœ50 nm) holes is used as a zero mode waveguide (see, e.g., Levene et al., Science 299, 682-686 (2003)). The aluminum surface is protected from attachment of DNA polymerase by polyphosphonate chemistry, e.g., polyvinylphosphonate chemistry (see, e.g., Korlach et al., Proceedings of the National Academy of Sciences of the United States of America 105, 1176-1181 (2008)). This results in preferential attachment of the DNA polymerase molecules to the exposed silica in the holes of the aluminum coating. This setup allows evanescent wave phenomena to be used to reduce fluorescence background, allowing the use of higher concentrations of fluorescently labeled dNTPs. The fluorophore is attached to the terminal phosphate of the dNTPs, such that fluorescence is released upon incorporation of the dNTP, but the fluorophore does not remain attached to the newly incorporated nucleotide, meaning that the complex is immediately ready for another round of incorporation. By this method, incorporation of dNTPs into an individual primer-template complexes present in the holes of the aluminum coating can be detected. See, e.g., Eid et al., Science 323, 133-138 (2009).

D. Generating a Sequence-Activity Model

As indicated above, a sequence-activity model used with the embodiments herein relates protein sequence information to protein activity. The protein sequence information used by the model may take many forms. In some embodiments, it is a complete sequence of the amino acid residues in a protein (e.g., HGPVFSTGGA . . . ). However, in some embodiments, the complete amino acid sequence is unnecessary. For example, in some embodiments, it is sufficient to provide only those residues that are to be varied in a particular research effort. In some embodiments involving later research stages, many residues are fixed and only limited regions of sequence space remain to be explored. In some of such situations, it is convenient to provide sequence-activity models that require, as inputs, only the identification of those residues in the regions of the protein where the exploration continues. In some additional embodiments, the models do not require that the exact identities of residues at the residue positions be known. In some such embodiments, one or more physical or chemical properties that characterize the amino acid at a particular residue position is/are identified. In one illustrative example, the model requires specification of residue positions by bulk, hydrophobicity, acidity, etc. Furthermore, in some models, combinations of such properties are employed. Indeed, it is not intended that the present invention be limited to any particular approach, as the models find use in various configurations of sequence information, activity information and/or other physical properties (e.g., hydrophobicity, etc.).

Thus, the form of the sequence-activity model can vary widely, so long as it provides a vehicle for correctly approximating the relative activity of proteins based on sequence information, as desired. In some embodiments, the models generally treat activity as a dependent variable and sequence/residue values as independent variables. Examples of the mathematical/logical form of models include linear and non-linear mathematical expressions of various orders, neural networks, classification and regression trees/graphs, clustering approaches, recursive partitioning, support vector machines, and the like. In one embodiment, the model form is a linear additive model in which the products of coefficients and residue values are summed. In another embodiment, the model form is a non-linear product of various sequence/residue terms, including certain residue cross products (which represent interaction terms between residues). Indeed, it is not intended that the disclosed embodiments be limited to any specific format, as any suitable format finds use, as illustrated herein.

In some embodiments, models are developed from a training set of activity versus sequence information to provide the mathematical/logical relationship between activity and sequence. This relationship is typically validated prior to use for predicting the activity of new sequences or the impacts of residues on the activity of interest.

Various techniques for generating models are available and find use in the present invention. In some embodiments, the techniques involve optimization of models or minimization of model errors. Specific examples include partial least squares, ensemble regression, random forest, various other regression techniques, as well as neural network techniques, recursive partitioning, support vector machine techniques, CART (classification and regression trees), and/or the like. Generally, the technique should produce a model that can distinguish residues that have a significant impact on activity from those that do not. In some embodiments, the models also rank individual residues or residue positions based on their impact on activity. It is not intended that the present invention be limited to any specific technique for generating models, as any suitable method known in the art finds use in the present invention.

In some embodiments, models are generated by a regression technique that identifies covariation of independent and dependent variables in a training set. Various regression techniques are known and widely used. Examples include multiple linear regression (MLR), principal component regression (PCR) and partial least squares regression (PLS). In some embodiments, models are generated using techniques that involve multiple constituents, including but not limited to ensemble regression and random forest. These and any other suitable methods find use in the present invention. It is not intended that the present invention be limited to any particular technique.

MLR is the most basic of these techniques. It is used to simply solve a set of coefficient equations for members of a training set. Each equation relates to the activity of a training set member (i.e., dependent variables) with the presence or absence of a particular residue at a particular position (i.e., independent variables). Depending upon the number of residue options in the training set, the number of these equations can be quite large.

Like MLR, PLS and PCR generate models from equations relating sequence activity to residue values. However, these techniques do so in a different manner. They first perform a coordinate transformation to reduce the number of independent variables. They then perform the regression on the transformed variables. In MLR, there is a potentially very large number of independent variables: two or more for each residue position that varies within the training set. Given that proteins and peptides of interest are often quite large and the training set may provide many different sequences, the number of independent variables can quickly become very large. By reducing the number of variables to focus on those that provide the most variation in the data set, PLS and PCR generally require fewer samples and simplify the steps involved in generating models.

PCR is similar to PLS regression in that the actual regression is done on a relatively small number of latent variables obtained by coordinate transformation of the raw independent variables (i.e., residue values). The difference between PLS and PCR is that the latent variables in PCR are constructed by maximizing covariation between the independent variables (i.e., residue values). In PLS regression, the latent variables are constructed in such a way as to maximize the covariation between the independent variables and the dependent variables (i.e., activity values). Partial Least Squares regression is described in Hand, D. J., et al. (2001) Principles of Data Mining (Adaptive Computation and Machine Learning), Boston, Mass., MIT Press, and in Geladi, et al. (1986) â€œPartial Least-Squares Regression: a Tutorial,â€ Analytica Chimica Acta, 198:1-17. Both of these references are incorporated herein by reference for all purposes.

In PCR and PLS, the direct result of the regression analysis is an expression for activity that is a function of the weighted latent variables. This expression can be transformed to an expression for activity as a function of the original independent variables by performing a coordinate transformation that converts the latent variables back to the original independent variables.

In essence, both PCR and PLS first reduce the dimensionality of the information contained in the training set and then perform a regression analysis on a transformed data set, which has been transformed to produce new independent variables, but preserves the original dependent variable values. The transformed versions of the data sets may result in only a relatively few expressions for performing the regression analysis. In protocols in which no dimension reduction has been performed, each separate residue for which there can be a variation must be considered. This can be a very large set of coefficients (e.g., 2^Ncoefficients for two-way interactions, where N is the number of residue positions that may vary in the training set). In a typical principal component analysis, only 3, 4, 5, 6 principal components are employed.

The ability of machine learning techniques to fit the training data is often referred to as the â€œmodel fitâ€ and in regression techniques such as MLR, PCR and PLS, the model fit is typically measured by the sum squared difference between measured and predicted values. For a given training set, the optimal model fit will be achieved using MLR, with PCR and PLS often having a worse model fit (higher sum squared error between measurements and predictions). However, the chief advantage of using latent variable regression techniques such as PCR and PLS lies in the predictive ability of such models. Obtaining a model fit with very small sum squared error in no way guarantees the model will be able to accurately predicted new samples not seen in the training setâ€”in fact, it is often the opposite case, particularly when there are many variables and only a few observations (i.e., samples). Thus latent variable regression techniques (e.g., PCR, PLS), while often having worse model fits on the training data are usually more robust and are able to predict new samples outside the training set more accurately.

Another class of tools that can be used to generate models in accordance with this disclosure is the support vector machines (SVMs). These mathematical tools take training sets of sequences that have been classified into two or more groups based on activity as inputs. Support vector machines operate by weighting different members of a training set differently depending upon how close they are to a hyperplane interface separating â€œactiveâ€ and â€œinactiveâ€ members of the training set. This technique requires that the scientist first decide which training set members to place in the â€œactiveâ€ group and which training set members to place in the â€œinactiveâ€ group. In some embodiments, this is accomplished by choosing an appropriate numerical value for the activity level that serves as the boundary between â€œactiveâ€ and â€œinactiveâ€ members of the training set. From this classification, the support vector machine generates a vector, W, that can provide coefficient values for the individual independent variables defining the sequences of the active and inactive group members in the training set. These coefficients can be used to â€œrankâ€ individual residues as described elsewhere herein. The technique is used to identify a hyperplane that maximizes the distance between the closest training set members on opposite sides of that plane. In another embodiment, support vector regression modeling is carried out. In this case, the dependent variable is a vector of continuous activity values. The support vector regression model generates a coefficient vector, W, which can be used to rank individual residues.

SVMs have been used to look at large data sets in many studies and have found wide use with DNA microarrays. Their potential strengths include the ability to finely discriminate (by weighting) the factors that separate samples from each other. To the extent that an SVM can tease out precisely which residues contribute to function, it can be a particularly useful tool for ranking residues. SVMs are described in S. Gunn (1998) â€œSupport Vector Machines for Classification and Regressions,â€ Technical Report, Faculty of Engineering and Applied Science, Department of Electronics and Computer Science, University of Southampton, which is incorporated herein by reference for all purposes.

In some embodiments of the invention, another class of tools that can be used to generate models is classification and regression based on an ensemble of classification trees using random inputs, an example of which is random forest. See Breiman (2001). â€œRandom Forests,â€ Machine Learning 45 (1): 5-32. Random forests are a combination of tree predictors such that each tree depends on the values of a random vector sampled independently and with the same distribution for all trees in the forest. A random forest is a learning ensemble consisting of a bagging of un-pruned decision tree learners with a randomized selection of features at each split of the decision tree. The generalization error for forests converges to a limit as the number of trees in the forest becomes large.

Random forests may be constructed in the following manner:

1) If the number of cases in the training set is N, sample N cases at randomâ€”but with replacement, from the original data. This sample will be the training set for growing the tree.

2) If there are M input independent variables, a number m<<M is specified such that at each node in the tree, m variables are selected at random out of the M and the best split on these m is used to split the node. The value of m is held constant during the forest growing.

3) In some implementations, each tree is grown to the largest extent possible. There is no pruning.

4) A large number of trees are then generated, k=1, . . . , K (usually K>=100).

5) After a large number of trees have been generated, they all vote for the classification of variables of interest. For example, they may each contribute the final prediction of activity or the contribution of particular mutations.

6) The random forest then classifies x (e.g., a sequence of mutations or other independent variable) by taking the most popular voted class from all the tree predictors in the forest.

The forest error rate depends on the correlation between any two trees in the forest. Increasing the correlation increases the forest error rate. The forest error rate depends on the strength of each individual tree in the forest. A tree with a low error rate is a strong classifier. Increasing the strength of the individual trees decreases the forest error rate. Reducing m reduces both the correlation and the strength. Increasing it increases both. Somewhere in between is an â€œoptimalâ€ range of mâ€”usually quite wide.

Random forest techniques can be used for categorical variables as well as continuous variables in regression models. In some embodiments of the invention, random forest models have predictive power comparable to SVM and neural net models, but tend to have higher computational efficiency because, among other reasons, cross validation is built into the modeling process and a separate process for cross validation is not necessary.

i) Linear Models

While the present disclosure is directed to non-linear models, these may be more easily understood in the context of linear models of sequence versus activity. Additionally, in some embodiments, a linear model is used as a â€œbaseâ€ model in a step-wise process for generating a non-linear model. In general, a linear regression model of activity versus sequence has the following form:

$\begin{matrix} y = c_{0} + {âˆ‘}_{i = 1}^{N} î¢ž î¢ž {âˆ‘}_{j = 1}^{M} î¢ž î¢ž c_{ij} î¢ž x_{ij} & (1) \end{matrix}$

In this linear expression, y is predicted response, while c_ijand x_ijare the regression coefficient and bit value or dummy variable used to represent residue choice, respectively at position i in the sequence. There are N residue positions in the sequences of the protein variant library and each of these may be occupied by one or more residues. At any given position, there may be j=1 through M separate residue types. This model assumes a linear (additive) relationship between the residues at every position. An expanded version of equation 1 follows:

y=c₀+c₁₁x₁₁+c₁₂x₁₂+ . . . c_1Mx_1M+c₂₁x₂₁+c₂₂x₂₂+ . . . c_2Mx_2M+ . . . +c_NMx_NM

As indicated, data in the form of activity and sequence information is derived from the initial protein variant library and used to determine the regression coefficients of the model. The dummy variables are first identified from an alignment of the protein variant sequences. Amino acid residue positions are identified from among the protein variant sequences in which the amino acid residues in those positions differ between sequences. Amino acid residue information in some or all of these variable residue positions may be incorporated in the sequence-activity model.

Table I contains sequence information in the form of variable residue positions and residue types for 10 illustrative variant proteins, along with activity values corresponding to each variant protein. These are representative members of a larger set that is required to generate enough equations to solve for all of the coefficients. Thus, for example, for the illustrative protein variant sequences in Table I, positions 10, 166, 175, and 340 are variable residue positions and all other positions, i.e., those not indicated in the Table, contain residues that are identical between Variants 1-10.

In this example, the 10 variants may or may not include the wild-type backbone sequence.

In some embodiments, a model developed to account for data of all variants including the wild-type backbone sequence may introduce a problem of perfect multi-collinearity, or a dummy variable trap. This problem may be addressed by various techniques. Some embodiments may exclude the wild-type backbone data from developing the model. Some embodiments may drop out those coefficients representing the wild-type backbone. Some embodiments may use techniques such as PLS regression to address multi-collinearity.

TABLE IIllustrative Sequence and Activity DataVariable Residue Position10166175340y (activity)Variant 1 (SEQ ID NO: 1)AlaSerGlyPhey₁Variant 2 (SEQ ID NO: 2)AspPheValAlay₂Variant 3 (SEQ ID NO: 3)LysLeuGlyAlay₃Variant 4 (SEQ ID NO: 4)AspIleValPhey₄Variant 5 (SEQ ID NO: 5)AlaIleValAlay₅Variant 6 (SEQ ID NO: 6)AspSerGlyPhey₆Variant 7 (SEQ ID NO: 7)LysPheGlyPhey₇Variant 8 (SEQ ID NO: 8)AlaPheValAlay₈Variant 9 (SEQ ID NO: 9)LysSerGlyPhey₉Variant 10 (SEQ ID NO: 10)AspLeuValAlaâ€‚y₁₀

Thus, based on equation 1, a regression model can be derived from the systematically varied library in Table I, i.e.:

$\begin{matrix} y = c_{0} + c_{10 î¢ž î¢ž Ala} î¢ž x_{10 î¢ž Ala} + c_{10 î¢ž Asp} î¢ž x_{10 î¢ž Asp} + c_{10 î¢ž î¢ž Lys} î¢ž x_{10 î¢ž Lys} + c_{166 î¢ž Ser} î¢ž x_{166 î¢ž Ser} + c_{166 î¢ž î¢ž Phe} î¢ž x_{166 î¢ž Phe} + c_{166 î¢ž Leu} î¢ž x_{166 î¢ž Leu} + c_{166 î¢ž Ile} î¢ž x_{166 î¢ž Ile} + c_{175 î¢ž Gly} î¢ž x_{175 î¢ž Gly} + c_{175 î¢ž î¢ž Val} î¢ž x_{175 î¢ž Val} + c_{340 î¢ž î¢ž Phe} î¢ž x_{340 î¢ž Phe} + c_{340 î¢ž î¢ž Ala} î¢ž x_{340 î¢ž Ala} & (Eq . î¢ž 2) \end{matrix}$

The bit values (x dummy variables) can be represented as either 1 or 0 reflecting the presence or absence of the designated amino acid residue or alternatively, 1 or âˆ’1, or some other surrogate representation. For example, using the 1 or 0 designation, x_10Alawould be â€œ1â€ for Variant 1 and â€œ0â€ for Variant 2. Using the 1 or âˆ’1 designation, x_10Alawould be â€œ1â€ for Variant 1 and â€œâˆ’1â€ for Variant 2. The regression coefficients can thus be derived from regression equations based on the sequence activity information for all variants in library. Examples of such equations for Variants 1-10 (using the 1 or 0 designation for x) follow:

y₁=c₀+c_{10 Ala}(1)+c_10Asp(0)+c_{10 Lys}(0)+c_166Ser(1)+c_{166 Phe}(0)+c_166Leu(0)+c_166Ile(0)+c_175Gly(1)+c_{175 Val}(0)+c_{340 Phe}(1)c_{340 Ala}(0)

y₂=c₀+c_{10 Ala}(0)+c_10Asp(1)+c_{10 Lys}(0)+c_166Ser(0)+c_{166 Phe}(0)+c_166Leu(0)+c_166Ile(0)+c_175Gly(0)+c_{175 Val}(1)+c_{340 Phe}(0)c_{340 Ala}(1)

y₃=c₀+c_{10 Ala}(0)+c_10Asp(0)+c_{10 Lys}(1)+c_166Ser(0)+c_{166 Phe}(0)+c_166Leu(1)+c_166Ile(0)+c_175Gly(1)+c_{175 Val}(0)+c_{340 Phe}(0)c_{340 Ala}(1)

y₄=c₀+c_{10 Ala}(0)+c_10Asp(1)+c_{10 Lys}(0)+c_166Ser(0)+c_{166 Phe}(0)+c_166Leu(0)+c_166Ile(1)+c_175Gly(0)+c_{175 Val}(1)+c_{340 Phe}(1)c_{340 Ala}(0)

y₅=c₀+c_{10 Ala}(1)+c_10Asp(0)+c_{10 Lys}(0)+c_166Ser(0)+c_{166 Phe}(0)+c_166Leu(0)+c_166Ile(1)+c_175Gly(0)+c_{175 Val}(1)+c_{340 Phe}(0)c_{340 Ala}(1)

y₆=c₀+c_{10 Ala}(0)+c_10Asp(1)+c_{10 Lys}(0)+c_166Ser(1)+c_{166 Phe}(0)+c_166Leu(0)+c_166Ile(0)+c_175Gly(1)+c_{175 Val}(0)+c_{340 Phe}(1)c_{340 Ala}(0)

y₇=c₀+c_{10 Ala}(0)+c_10Asp(0)+c_{10 Lys}(1)+c_166Ser(0)+c_{166 Phe}(1)+c_166Leu(0)+c_166Ile(0)+c_175Gly(1)+c_{175 Val}(0)+c_{340 Phe}(1)c_{340 Ala}(0)

y₈=c₀+c_{10 Ala}(1)+c_10Asp(0)+c_{10 Lys}(0)+c_166Ser(0)+c_{166 Phe}(1)+c_166Leu(0)+c_166Ile(0)+c_175Gly(0)+c_{175 Val}(1)+c_{340 Phe}(0)c_{340 Ala}(1)

y₉=c₀+c_{10 Ala}(0)+c_10Asp(0)+c_{10 Lys}(1)+c_166Ser(1)+c_{166 Phe}(0)+c_166Leu(0)+c_166Ile(0)+c_175Gly(1)+c_{175 Val}(0)+c_{340 Phe}(1)c_{340 Ala}(0)

y₁₀=c₀+c_{10 Ala}(0)+c_10Asp(1)+c_{10 Lys}(0)+c_166Ser(0)+c_{166 Phe}(0)+c_166Leu(1)+c_166Ile(0)+c_175Gly(0)+c_{175 Val}(1)+c_{340 Phe}(0)c_{340 Ala}(1)

The complete set of equations can be readily solved using any suitable regression technique (e.g., PCR, PLS, or MLR) to determine the value for regression coefficients corresponding to each residue and position of interest. In this example, the relative magnitude of the regression coefficient correlates to the relative magnitude of contribution of that particular residue at the particular position to activity. The regression coefficients may then be ranked or otherwise categorized to determine which residues are more likely to favorably contribute to the desired activity. Table II provides illustrative regression coefficient values corresponding to the systematically varied library exemplified in Table I:

TABLE IIIllustrative Rank Ordering of Regression CoefficientsREGRESSIONCOEFFICIENTVALUEc_166Ile62.15c_175Gly61.89c_10Asp60.23c_{340 Ala}57.45c_{10 Ala}50.12c_{166 Phe}49.65c_166Leu49.42c_{340 Phe}47.16c_166Ser45.34c_{175 Val}43.65c_10Lys40.15

The rank ordered list of regression coefficients can be used to construct a new library of protein variants that is optimized with respect to a desired activity (i.e., improved fitness). This can be done in various ways. In one embodiment, it is accomplished by retaining the amino acid residues having coefficients with the highest observed values. These are the residues indicated by the regression model to contribute the most to desired activity. If negative descriptors are employed to identify residues (e.g., 1 for leucine and âˆ’1 for glycine), it becomes necessary to rank residue positions based on the absolute value of the coefficient. Note that in such situations, there is typically only a single coefficient for each residue. The absolute value of the coefficient magnitude gives the ranking of the corresponding residue position. Then, it becomes necessary to consider the signs of the individual residues to determine whether each of them is detrimental or beneficial in terms of the desired activity.

ii) Non-Linear Models

Non-linear modeling is employed to account for residue-residue interactions that contribute to activity in proteins. An N-K landscape describes this problem. The parameter N refers to the number of variable residues in a collection of related polypeptides sequences. The parameter K represents the interaction between individual residues within anyone of these polypeptides. Interaction is usually a result of close physical proximity between various residues whether in the primary, secondary, or tertiary structure of the polypeptide. The interaction may be due to direct interactions, indirect interactions, physicochemical interactions, interactions due to folding intermediates, translational effects, and the like. See Kauffman, S. and Levin, S. (1987), â€œTowards a general theory of adaptive walks on rugged landscapesâ€ , Journal of Theoretical Biology 128 (1) 11-45.

The parameter K is defined such that for value K=1, each variable residue (e.g., there are 20 of them) interacts with exactly one other residue in its sequence. In the case where all residues are physically and chemically separate from the effects of all other residues, the value of K is zero. Obviously, depending upon the structure of the polypeptide, K can have a wide range of different values. With a rigorously solved structure of the polypeptide in question, a value for K may be estimated. Often, however, this is not the case.

A purely linear, additive model of polypeptide activity (as described above) can be improved by including one or more non-linear, interaction terms representing specific interactions between 2 or more residues. In the context of the model form presented above, these terms are depicted as â€œcross-productsâ€ containing two or more dummy variables representing the two or more particular residues (each associated with a particular position in the sequence) that interact to have a significant positive or negative impact on activity. For example, a cross-product term may have the form c_abx_ax_b, where x_ais a dummy variable representing the presence of a particular residue at a particular position on the sequence and the variable x_brepresents the presence of a particular residue at a different position (that interacts with the first position) in the polypeptide sequence. A detailed example form of the model is shown below.

The presence of all residues represented in the cross-product term (i.e., each of two or more specific types of residue at specifically identified positions) impacts the overall activity of the polypeptide. The impact can be manifested in many ways. For example, each of the individual interacting residues when present alone in a polypeptide may have a negative impact on activity, but when they are present in the polypeptide, the overall effect is positive. The opposite may be true in other cases. In addition, there may be a synergistic effect produced, in which each of the individual residues alone has a relatively limited impact on activity, but when all of them are present, the effect on activity is greater than the cumulative effects of all the individual residues.

In some embodiments, non-linear models include a cross-product term for every possible combination of interacting variable residues in the sequence. However, this does not represent physical reality, as only a subset of the variable residues actually interact with one another. In addition, it would result in â€œoverfittingâ€ to produce a model that provides spurious results that are manifestations of the particular polypeptides used to create the model and do not represent real interactions within the polypeptide. The correct number of cross-product terms for a model that represents physical reality, and avoids overfitting, is dictated by the value of K. For example, if K=1, the number of cross-product interaction terms equals N.

In constructing a non-linear model, in some embodiments, it is important to identify those cross-product interaction terms representing true structural interactions that have a significant impact on activity. This can be accomplished in various ways, including but not limited to forward addition in which candidate cross-product terms are added to the initial linear term only model one at a time until the addition of terms is no longer statistically significant, and reverse subtraction, in which all possible cross product terms are provided in an initial model and removed one at a time. The illustrative examples presented below involve the use of stepwise addition and subtraction techniques to identify the useful non-linear interaction terms.

In some embodiments, the approach to generating a non-linear model containing such interaction terms is the same as the approach described above for generating a linear model. In other words, a training set is employed to â€œfitâ€ the data to a model. However, one or more non-linear terms, preferably the cross-product terms discussed above, are added to the model. Further, the resulting non-linear model, like the linear models described above, can be employed to rank the importance of various residues on the overall activity of a polypeptide. Various techniques can be used to identify the best combination of variable residues as predicted by the non-linear equation. Approaches to ranking the residues are described below. In some embodiments, very large numbers of possible cross-product terms for variable residues are used, even when limited to interactions caused by only two residues. As more interactions occur, the number of potential interactions to consider for a non-linear model grows in an exponential manner. If the model includes the possibility of interactions that include three or more residues, the number of potential terms grows even more rapidly.

In a simple illustrative example, in which there are 20 variable residues and K=1 (this assumes that each variable residue interacts with one other variable residue), there should be 20 interaction terms (cross-products) in the model. If there are any fewer interaction terms, the model will not fully describe the interactions (although some of the interactions may not have a significant impact on activity). In contrast, if there are any more interaction terms the model may overfit the data set. In this example, there are N*(Nâˆ’1)/2 or 190 possible pairs of interactions. Finding the combination of 20 unique pairs that describe the 20 interactions in the sequence is a significant computational problem, as there are approximately 5.48Ã—10²⁶possible combinations.

Numerous techniques can be employed to identify the relevant cross-product terms. Depending upon the size of the problem and the computational power available, it is possible to explore all possible combinations and thereby identify the one model that best fits the data. However, often the problem is computationally demanding. Thus, in some embodiments, an efficient search algorithm or an approximation it utilized. As indicated herein, one suitable search technique is a stepwise technique. However, it is not intended that the present invention be limited to any particular method for identification of the relevant cross-product terms.

An illustrative example is presented below in Table III to show the value of incorporating non-linear cross-product terms in a model predicting activity from sequence information. This example is a non-linear model in which it is assumed there are only two residue options at each variable position in the sequence. In this example, the protein sequence is cast into a coded sequence by using dummy variables that correspond to choice A or choice B, using +1 and âˆ’1 respectively. The model is immune to the arbitrary choice of which numerical value is used to assign each residue choice. The variable positions shown in the first row of Table III do not indicate the actual sequence positions of a protein sequence. Instead, they are arbitrary labels representing any 10 hypothetical positions in a protein sequence that can be varied with one of two options shown in the second and third rows of Table III for Residue Choice A and Residual Choice B.

TABLE IIIExample of coding residues at positions each having two optionsVariable Position Label12345678910Residue Choice AILLMGWKCSF(SEQ ID NO: 11)Residue Choice BVAIPHNRTAY(SEQ ID NO: 12)Protein Residue ChoiceVALPGWKTSF(SEQ ID NO: 13)Model Code Valueâˆ’1âˆ’11âˆ’1111âˆ’111

With this coding scheme, the linear model used to associate protein sequences with activity can be written as follows:

y=c₁x₁+c₂x₂+c₃x₃. . . +c_nx_n+ . . . +c_Nx_N+c₀â€ƒâ€ƒ(Eq. 3)

- where y is the response (activity), c_nthe regression coefficient for the residue choice at position n, x the dummy variable coding for the residue choice (+1/âˆ’1) at position n, and c₀the mean value of the response. This form of the model assumes there are no interactions between the variable residues (i.e., each residue choice contributes independently to the overall fitness of the protein).

The non-linear model includes a certain number of (as yet undetermined) cross-product terms to account for interactions between residues:

y=c₁x₁+c₂x₂+c₃x₃+ . . . +c_nx_n+c_1,2x₁x₂+c_1,2x₁x₃+c_2,3x₂x₃+ . . . +c₀â€ƒâ€ƒ(Eq. 4)

where the variables are the same as those in Eq. (3) but now there are non-linear terms, e.g., c_1,2is the regression coefficient for the interaction between variable positions 1 and 2.

In order to assess the performance of the linear and non-linear models, a synthetic data source known as the NK landscape was used (Kauffman and Levin, 1987). As mentioned above, N is the number of variable positions in a simulated protein and K is the epistatic coupling between residues. In addition, the synthetic data set was generated in silico.

This data set was used to generate an initial training set with S=40 synthetic samples, with N=20 variable positions and K=1 (to reiterate, for K=1 each variable position is functionally coupled to one other variable position). In generating the randomized proteins, each variable position had an equal probability of containing the dummy variable +1 or âˆ’1. The residue-residue interactions (represented by cross-products) and actual activities were known for each member of the synthetic training set. Another V=100 samples were generated for use in validation. Again, the residue-residue interactions and activities were known for each member of the validation set.

The training sets were used to construct both linear and non-linear models. Some non-linear models were generated with selection of the cross-product terms and other non-linear models were generated without selection of such terms. The models for FIG. 3A-F were generated using a genetic algorithm modeling method, while the models for Figure G-H were generated using stepwise modeling method. Although the quantitative advantage of models having both linear and non-linear terms relative to models having only linear terms differ between genetic algorithm and stepwise modeling methods, the results indicate the generalizable advantage of models with non-linear terms, regardless of the modeling methods. Indeed, it is not intended that the present invention be limited to any particular methods, as any suitable modeling methods find use in the present invention.

For the training set size of S=40 described above, the linear model was capable of correlating the measured and predicted values reasonably well, but demonstrated weaker correlation when validated against data not seen in the training set (see FIG. 3A). As shown, the dark data points represent the observed activity of 40 training data points vs. the predictions made by a linear model. The light data points represent the predictions made by the same model constructed from the 40 training samples and used to predict the validation samples V, none of which were seen in the original training set. The validation set provides a good measure of the true predictive capacity of the model, as opposed to the training set, which can suffer from the model overfit problem especially for the non-linear cases described below.

This result for the S=40 training set described above is notable, considering that a linear model was used to model a non-linear fitness landscape. In this case, the linear model could, at best, capture the average contribution to fitness for the choice of a given residue. Given a sufficient number of average contributions taken into consideration in combination, the linear model roughly predicts the actual measured response. The validation results for the linear model were marginally better when the training size was increased to S=100 (see FIG. 3B). The tendency of relatively simple models to underfit data is known as â€œbias.â€

When the non-linear model was trained using only S=40 samples, the correlation with the training set members was excellent (see FIG. 3C). Unfortunately, in this illustrative example, the model provided limited predictive power outside the training set, as evidenced by its limited correlation with measured values in the validation set. This non-linear model, with many potential variables (210 possible), and limited training data to facilitate identification of the proper cross-product terms, was able to essentially just memorize the data set it was trained on. This tendency of high complexity models to overfit the data is known as â€œvariance.â€ The bias-variance tradeoff represents a fundamental problem in machine learning and some form of validation is almost always required to address it when dealing with new or uncharacterized machine learning problems.

However, when the non-linear model was trained using a larger training set (S=100) as shown in FIG. 3D, the non-linear model performed exceedingly well for both the training prediction and, more importantly, the validation prediction. The validation predictions were sufficiently accurate that most of the data points are obscured by the dark circles used to plot the training set.

For comparison, FIGS. 3E and 3F show the performance of non-linear models prepared without careful selection of the cross-product terms. Unlike the models in FIGS. 3C and 3D, every possible cross-product term was chosen (i.e., 190 cross-product terms for N=20). As shown in these Figures, the ability to predict validation set activity is relatively poor compared to that of the non-linear models generated with careful selection of cross-product terms. This poor ability to predict validation data is a manifestation of overfitting.

FIGS. 3G and 3H respectively show the predictive power indicated by residuals of a linear model and a stepwise, non-linear model for data simulated in silico. The stepwise non-linear model was implemented as generally described above and more specifically below.

To test these models, simulated data were created. A random number generator R was created based upon a normal distribution with a mean MN and standard deviation SD. Then a set of 10 mutations were defined. The naming of these were M1, M2 . . . M10 (this naming scheme is arbitrary). This step simulates the creation of diversity

Each mutation represented an amino acid change at a given position within a protein sequence, and each position is independent of the other positions. Each mutation above had a random activity value A assigned based upon R (MN=0, SD=0.2). Six mutations above were chosen and paired together into three pairs P. These pairs represented epistatic interactions between mutations.

An activity value AP was assigned to each pair P based upon R (MN=0, SD=0.2). A library L of 50 variants was constructed in which each variant contained a random number of mutations M defined aboveâ€”the random number of mutations was defined by the rounded absolute value of R (MN=4, S=0.25). This step simulates library construction and sequencing.

The activity of each variant in L was calculated by first adding to 1.0 (a defined activity of the wild-type, no mutation sequence) the value of the activity from each pair-wise mutation PA (if both mutations were present) followed by adding the values of the remaining single mutations (A). Assay noise was simulated by adding to the final value for each variant a random value from R (MN=0, SD=0.005). This step simulates screening of variants.

A linear model LM was constructed based upon the data from the last step. This model contained ten independent variables/coefficients, each representing one mutation of M. The linear model was then fit using ordinary least squares regression and data obtained above.

A stepwise addition method was used to select a model MM based upon the data obtained above, with the base model being LM, using AIC as the selection criterion, and selecting models which only contain coefficients representing single mutations and pair-wise interactions. See Model Selection description below for further details of the model selection method. The best model selected by AIC was fit using ordinary least squares regression.

To assess the predictive capability of the linear model and nonlinear model, the procedures described above were repeated 20 times. The prediction of the models were plotted against the simulated data, wherein FIG. 3G shows the linear model and FIG. 3H shows the step-wise non-linear model. The models were used to predict the values of single mutations described above. This prediction was performed by using the models to predict a variant containing only one mutation of interest and subtracting 1.0 (wild-type). As apparent from FIGS. 3G and 3H, the non-linear model more accurately predicts values, having a more linear trend and smaller residuals.

iii) Model Selection

In some embodiments, stepwise addition or subtraction methods are used to prepare models with non-linear interaction terms. By implementing operation shown in block 107 of FIG. 1, a final model with high predictive power including interaction terms is provided by stepwise addition or subtraction of interaction terms from a base model. FIG. 4A provides a flow chart of an implementation of operation of block 107 of FIG. 1 by adding interaction terms to a base model and evaluating the new models to create a final best model.

In this example, the base sequence model includes no interaction terms. The method first sets a current sequence model and a best sequence model to the base sequence model, block 409. The method defines a pool of interaction terms for sequence variants. These interaction terms may include any number of pairwise or higher order interactions of two or more amino acid residues. See block 411. Although block 409 is illustrated as occurring before block 411, the order of the two steps is not important. In some embodiments, the pool of interaction terms includes factorial combinations of all amino acid residues of interest. In some additional embodiments, at least all pairwise interaction terms are included. In some further embodiments, pairwise and three-way interaction terms are included.

After creating a base model, the method selects an interaction term that has not yet been tested from the pool. The method then creates a new sequence model by adding the selected interaction term to the current sequence model. See block 413. The method then evaluates the predictive power of the new sequence model using a model selection method having a bias against including additional interaction terms. See block 415. The method determines whether or not the predictive power of the new sequence model is greater than that of the best sequence model. See decision block 417. As an example, the method may use a technique employing â€œlikelihoodâ€ determination (e.g., AIC) as a model selection criterion. In such cases, only a model having an AIC value smaller than the previously tested model is considered to have higher predictive power.

In some embodiments, the selection method biases against models with more parameters. Examples of such selection methods include, but are not limited to Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), and variations thereof. For instance, AIC can be calculated as:

AIC=âˆ’2 log_eL+2k

where L is the likelihood of a model given a data set, and k is the number of free parameters in a model.

In some embodiments, the likelihood of a model given a dataset may be calculated by various methods, including but not limited to the maximum likelihood method. For instance, for a binary dependent variable where an activity either is present or is absent for one observation, the likelihood of the model can be calculated as:

$L (model î¡œ data) = {âˆ}_{i = 1}^{n} î¢ž î¢ž \frac{(a_{i} + b_{i})!}{a_{i}! î¢ž b_{i}!} î¢ž {p_{i}^{a_{i}} î¢ (1 - p_{i})}^{b_{i}}$

where n is the total number of data points in a data set; a_iand b_iare the number of observed trials comprising the i^thcondition; p is the probability of a dependent variable being observed as predicted by the model.

In some embodiments involving a series of nested models, as in regression models with progressively more interaction terms (and associated coefficients) than a base model, more complex models provide equally good or better fits than simpler ones even if the extra coefficients are spurious, because the more complex model enjoys extra degrees of freedom. In some embodiments, AIC penalizes the more complex model to the extent that the gain in goodness of fit is more than offset by the cost of spurious parameters. In model selection, a smaller value of AIC indicates a better model.

In the example shown in FIG. 4A, if the predictive power of the new sequence model is greater than that of the best sequence model, then the method sets the new model as the best model. See block 419. Then, the method checks whether any additional interaction terms are left in the pool that have not been tested. See decision block 421. If so, the process reverts back to block 413, thereby forming an inner loop to test all of the available interaction terms available in the interaction pool. Through iterations of the inner loop, a single best interaction term can be found and added to the model.

After all interaction terms have been tested, and the inner loop ends, a best model having one additional interaction term is identified, given that a model having greater predictive power than the previous best model does exist. See decision block 423. In such embodiments, the method sets the current model to the best model, and excludes the interaction terms of the best model from the available pool of interaction terms. See block 425. Then, the method loops back to block 413. This outer loop searches for the next best interaction term that can improve the model's predictive power. If such an interaction term is found, the search for the next best interaction term continues in the outer loop, until no new model having a predictive power greater than that of the previous best sequence model is identified.

When no more interaction terms can be found to improve the model, the method sets the best model as the final model. See block 427. The search for a best model given the sequence and activity data is finished. The model is then used to predict the activities of new sequences. Such predictions can guide the selection of sequences for further variation and testing.

In certain embodiments, each of the available interaction terms in the pool of interaction terms is treated as having potentially equal impact on the quality or predictive power of the model. In other words, in implementation, each of the available interaction terms in the pool is equally likely to be selected for consideration during a particular iteration. In some embodiments, the available interaction terms are selected randomly or in some arbitrary orders.

In some other embodiments, the interaction terms are biased or weighted in such a manner that some terms are more likely to be selected for consideration than others during a given iteration. The bias or weighting can, in certain embodiments, be applied on the basis of physical or theoretical information about the interactions. For example, it may be known that mutations in two particular areas of a protein are likely to be physically proximate to one another and thereby interact. Interaction terms pertaining to residues in these two general areas could be biased for selection during the iterative process of refining the model.

Pseudo code illustrating processes similar to that for FIG. 4A follows:

SET Coeff = Interaction Terms to TestBest = Baseline Modelcount = 1WHILE count > 0â€ƒcount = 0â€ƒBestFromRound = Bestâ€ƒBestCoefficient = NULLâ€ƒFOR each Interaction Term in Coeffâ€ƒâ€ƒTestModel = (best + Interaction Term)¹â€ƒâ€ƒIF TestModel BETTER THAN BestFromRound THEN²â€ƒâ€ƒâ€ƒBestFromRound = TestModelâ€ƒâ€ƒâ€ƒCount++â€ƒâ€ƒâ€ƒBestCoefficient = Interaction Termâ€ƒâ€ƒENDIFâ€ƒENDFORâ€ƒIF count > 0 THENâ€ƒâ€ƒBest = BestFromRoundâ€ƒâ€ƒRemove BestCoefficient FROM Coeff³â€ƒENDIFENDWHILE

Item 1 adds the test interaction term to the regression model
Item 2 represents Model Comparison, one or more of Akaike Information Criteria (AIC), Bayesian Information Criteria (BIC), Crossvalidation (mean error), ANOVA, or coefficient contribution.
Item 3 is provided to avoid duplicate Interaction Term tests

FIG. 4B provides a flow chart showing an embodiment of operation shown in block 107 of FIG. 1. In this process, interaction terms are subtracted from a base model that includes all possible interaction terms from a pool of such terms to create a final best model.

In this embodiment, the base sequence model includes all interaction terms within a defined pool. The method first sets a current sequence model and a best sequence model to be equal to the base sequence model at the beginning of the process, block 439. This embodiment is similar to the last model described above in that the whole pool of interaction terms may include any number of pairwise or higher order interactions of two or more amino acid residues. In some embodiments, the pool of interaction terms includes factorial combinations of all amino acid residues that are of interest.

After creating a base model, the method selects an interaction term that has not yet been tested from the pool of terms already included in the base model. The method then creates a new sequence model by subtracting the selected interaction term from the current sequence model. See block 441. The method then evaluates the predictive power of the new sequence model using a model selection method having a bias against additional interaction terms. See block 443. The method evaluates whether or not the predictive power of the new sequence model is greater than that of the best sequence model. See decision operation shown in block 445. In some embodiments, AIC is used as the model selection criterion, such that a model having an AIC value smaller than the previously tested model is considered to have higher predictive power.

In this illustrative example, if the predictive power of the new sequence model is greater than that of the best sequence model, then the method sets the new model as the best model. See block 447. Next, the method checks whether any additional interaction terms are left in the pool that have not been tested (i.e., subtracted from the current sequence model). See decision block 449. If there are any untested terms, the method reverts to block 441, thereby forming an inner loop to test all of the available interaction terms available in the interaction pool. Through iterations of the inner loop, a single interaction term is identified. Dropping it from the model improves the model to the greatest extent (and lowering AIC by the largest margin, if AIC is used to measure the predictive power of the model).

After all interaction terms have been tested, and the inner loop ends, a best model having one fewer interaction term is identified, given that a model having greater predictive power than the previous best model does exist. See decision block 451. In this case, the method sets the current model to the best model. See block 453. Then, the method loops back to block 441. This outer loop searches for the next interaction term that can improve the model's predictive power by the largest margin. If such an interaction term is found, the search for the next interaction term to be subtracted continues in the outer loop, until no more new models having predictive powers greater than that of the previous best sequence model is identified.

When an inner loop is completed and no more interaction terms can be found to be subtracted to improve the model (i.e., decision operation shown in block 451 is answered in the negative), the method sets the last best model as the final model. See block 455. The search for a best model given the sequence and activity data is finished.

iv) Alternative Modeling Options

Multiple additional variations on the above approach are within the scope of the disclosure. Indeed, it is not intended that the present invention be limited to any particular model, as any suitable model finds use in the present invention. As one illustrative example, the x_ijvariables are representations of the physical or chemical properties of amino acidsâ€”rather than the exact identities of the amino acids themselves (leucine versus valine versus proline . . . ). Examples of such properties include lipophilicity, bulk, and electronic properties (e.g., formal charge, van der Waals surface area associated a partial charge, etc.). To implement this approach, the x_ijvalues representing amino acid residues can be presented in terms of their properties or principal components constructed from these properties. It is not intended that the present invention be limited to any particular property of amino acids, peptides, and/or polypeptides, as any suitable property finds use in the methods of the present invention.

In some additional embodiments, the x_ijvariables represent nucleotides, rather than amino acid residues. In these embodiments, the goal is to identify nucleic acid sequences that encode proteins for a protein variant library. By using nucleotides rather than amino acids, parameters other than activity (e.g., specific activity) can be optimized, as desired. For example, protein expression in a particular host or vector may be a function of nucleotide sequence. Two different nucleotide sequences may encode a protein having the same amino acid sequence, but one of the nucleotide sequences may lead to production of greater quantities of protein and/or the protein is more active. By using nucleotide sequences rather than amino acid sequences, the methods described herein can be used to optimize strains of microorganisms that exhibit improved gene expression properties and/or improved properties (e.g., specific activity, stability, etc.).

In some embodiments, the nucleotide sequence is represented as a sequence of codons. In some embodiments, models utilize codons as the atomic unit of a nucleotide sequence such that the predicted activities are a function of the various codons present in the nucleotide sequence. Each codon, together with its position in the overall nucleotide sequence serves as an independent variable for generating sequence-activity models. It is noted that in some cases, different codons for a given amino acid are expressed differently in a given organism. In some embodiments, each organism has a preferred codon, or distribution of codon frequencies, for a given amino acid. By using codons as the independent variables, the embodiment accounts for these preferences. Thus, the embodiment can be used to generate a library of expression variants (e.g., where â€œactivityâ€ includes gene expression level of a particular host organism).

In some embodiments, the methods include the following operations: (a) receiving data characterizing a training set of a protein variant library; (b) developing a non-linear sequence-activity model that predicts activity as a function of nucleotide types and corresponding positions in the nucleotide sequence, based on the data obtained in (a); (c) using the sequence-activity model to rank the positions in a nucleotide sequence and/or nucleotide types at specific positions in the nucleotide sequence in order of impact on the desired activity; and (d) using the ranking to identify one or more nucleotides, in the nucleotide sequence, that are to be varied or fixed, in order to improve the desired activity. As indicated, in some embodiments, the nucleotides to be varied encode specific amino acids.

In some other embodiments, the methods involve use of different techniques for ranking or otherwise characterizing residues in terms of their importance related to a certain property. As described above for linear models, the magnitudes of regression coefficients were used to rank residues. Residues having coefficients with large magnitudes (e.g., 166 Ile) were viewed as high-ranking residues. This characterization was used to decide whether or not to vary a particular residue in the generation of a new, optimized library of protein variants. For non-linear models, the sensitivity analysis was more complex, as described herein.

PLS and other techniques provide additional information, beyond regression coefficient magnitude, that can be used to rank specific residues or residue positions. Techniques such as PLS and Principal Component Analysis (PCA) or PCR provide information in the form of principal components or latent vectors. These represent directions or vectors of maximum variation through multi-dimensional data sets such as the protein sequence-activity space employed with the embodiments of the present invention disclosed herein. These latent vectors are functions of the various sequence dimensions; i.e., the individual residues or residue positions that comprise the protein sequences comprising the variant library used to construct the training set. Latent vectors therefore comprise a sum of contributions from each of the residue positions in the training set. Some positions contribute more strongly to the direction of the vector. These are manifested by relatively large â€œloads,â€ i.e., the coefficients used to describe the vector. As a simple illustrative example, a training set may be comprised of tripeptides. In this example, the first latent vector comprises contributions from all three residues.

Vector 1=a1(residue position 1)+a2(residue position 2)+a3(residue position 3)

The coefficients, a1, a2, and a3, are the loads. Because these reflect the importance of the corresponding residue positions to variation in the dataset, they can be used to rank the importance of individual residue positions for purposes of â€œtogglingâ€ decisions, as described above. Loads, like regression coefficients, may be used to rank residues at each toggled position. Various parameters describe the importance of these loads. Some embodiments utilize methods such as Variable Importance in Projection (VIP) to make use of a load matrix. This load matrix is comprised of the loads for multiple latent vectors taken from a training set. In Variable Importance for PLS Projection methods, the importance of a variable (e.g., residue position) is computed by calculating VIP. For a given PLS dimension, a, (VIN)_ak²is equal to the squared PLS weight (w_ak)²of a variable multiplied by the percent explained variability in y (dependent variable, e.g., certain function) by that PLS dimension. (VIN)_ak²is summed over all PLS dimensions (components). VIP is then calculated by dividing the sum by the total percent variability in y explained by the PLS model and multiplying by the number of variables in the model. Variables with a VIP larger than 1 are the most relevant for correlating with a certain function (y) and hence, are highest ranked for purposes of making toggling decisions.

In many embodiments, the present invention utilizes general linear regression methods to identify the effects of mutations in a combinatorial library on a sequence-activity of interest. Alternative modeling options and techniques, e.g., Bayesian regression, ensemble regression, bootstrapping, can be used in combination with or instead of the methods noted above. Indeed, it is not intended that the present invention be limited to any specific modeling options and/or techniques, as any suitable method(s) find use in the present invention.

Bayesian Linear Regression

In some embodiments of the present invention, Bayesian linear regression finds use. This method is an approach to linear regression in which the statistical analysis is undertaken within the context of Bayesian inference. When the regression model has errors that have a normal distribution, and if a particular form of prior distribution is assumed, the posterior probability distributions of the model's parameters can be determined using Bayesian inference techniques.

An ordinary least squares solution of a linear regression model estimates the coefficient vector and model error based on the likelihood function of the data using an analytical calculation method such as the Moore-Penrose pseudo inverse. This is a frequentist approach that assumes that there are sufficient observations of the data to represent the sequence-activity relation for all sequences. However, actual observations of a sample are almost never sufficient to represent all of the members of a population. This is especially problematic when the sample (or training set) size is limited. In the Bayesian approach, the sample data are supplemented with additional information in the form of a prior probability distribution. The prior belief about the parameters is combined with the data's likelihood function according to Bayes theorem to yield the posterior belief about the parameters. The prior belief can take different functional forms depending on the domain and the information that is available a priori.

For instance in some embodiments, Bayesian regression can use prior information to weight coefficients before model fit. In some embodiments, sequence/activity data taken from a previous round of directed evolution, e.g., a round performed using the parental or reference backbone and at least some of the mutations used in the previous rounds, can be used to weight linear coefficients. Further, predictions of the epistatic relationship between two or more mutations can be used to weight non-linear, interaction coefficients. One of the main advantages to this approach is the inclusion of prior information to direct model predictions.

One illustrative example of a source of prior information is a model with independent and interaction terms for each of multiple mutations to a reference backbone. In some embodiments, the data are obtained from a collection of variants that contains one mutation per variant.

Additional examples of prior information that find use in the present invention include, but are not limited to intuitive or physical information about the role of certain mutations or types of mutations. Regardless of the source, the prior information serves as a preconceived notion of the relationship between sequence and activity.

In some embodiments for estimating the parameters of a model, Bayesian linear regression uses Monte Carlo simulations such as Gibbs Sampling or Metropolis algorithms to fit the model given the data. Gibbs Sampling is a Markov chain Monte Carlo algorithm for obtaining a sequence of observations which are approximately from a specified multivariate probability distribution (i.e. from the joint probability distribution of two or more random variables), when direct sampling is difficult.

FIG. 5 is a flow chart illustrating the use of Bayesian regression in guided evolution of variant libraries. Each round of sequence evolution includes mutations based on the sequences from a previous round, which may be guided by knowledge such as a sequence-activity model. At round n of the evolution as in block 501, for instance, there is one mutation per variant. The next or n+1 round of evolution is the current round, as shown in block 503. There is at least one new mutation for each variant, amounting to two or more mutations per variant. Bayesian regression is implemented at this round in this illustrative example.

The sequence variants of round n+1 provides a training set of data for new models. The new models can comprise a base model that includes only linear terms for individual residues, or a complete model that contains all possible interaction terms/coefficients, as indicated in block 507. The new models may also comprise a model selected by various techniques, including the stepwise addition or subtraction techniques explained above, see block 505. The model may alternatively be selected using a genetic algorithm or bootstrap techniques as discussed below. These models are all based on the current/new data from the training set data of round n+1. Bayesian inference technique can be applied to these models, such that a model is based on both the probability function of current data and distribution of prior information. The prior information can come from data of the previous round of sequence variants, as in round n indicated by block 501. Information can also come from sequence-activity data from any previous round of evolution, or other prior intuition on knowledge, as indicated by block 513. The Bayesian regression model indicated by block 509 predicts activity based on information provided by current data and prior information, see block 511. Although FIG. 5 only illustrates application of the Bayesian regression technique to round n+1, it can be applied at various stages. It is also not intended that the present invention be limited to the specific steps provided in FIG. 5, as any suitable methods find use in the present invention.

Ensemble Regression

In some embodiments, the present invention utilizes an ensemble regression technique to prepare the sequence-activity model. An ensemble regression model is based on several regression models. The prediction of each model is weighted based upon a particular information criterion (IC), and the prediction of the ensemble is a weighted sum of the prediction of all the models it contains. In some embodiments, model development starts with a base model containing all of the linear terms. Subsequent models are constructed by adding interaction coefficients in some or all possible combination. In some embodiments, the interaction coefficients are added in a step-wise process. Each model is fit to the data, and an IC is generated. Weight for each model is based upon the IC, which can be the IC itself, or a transformed version, e.g., log value, negated value, etc. Predictions can be made for an observation by generating the prediction of each model in the ensemble, and determining the ensemble prediction by taking the weighted mean of the prediction from each model. A complete ensemble contains all possible models, but can be trimmed to remove poor performing models by setting a threshold on either the number of models it contains or on IC.

The constituent models of the ensemble can be produced using various techniques. For instance, in some embodiments, genetic algorithm is used to create the constituent models. Sequence/activity data is used to produce a plurality of regression models, each of which has its own set of coefficients. The best models are selected according to fitness criterion (e.g., AIC or BIC). These models are â€œmatedâ€ to produce new hybrid models that are then evaluated for fitness and selected accordingly. In some embodiments, this process is repeated for multiple rounds of â€œcomputational evolutionâ€ to produce an ensemble of the best models. Alternatively, in some embodiments, the ensemble constituents are created by stepwise regression as described above, and the best n models are selected to form an ensemble.

FIG. 6 provides a flow chart for a process that implements ensemble regression in directed evolution of sequence variants in accordance with an embodiment of the present invention. In this embodiment, the ensemble regression technique may be applied at any stage of multiple rounds of sequence evolution. For instance, at round n, sequence variants shown in block 601 provide a training set of data for various models to form a model pool as indicated by block 603. The models in the model pool may be models generated by a genetic algorithm and/or stepwise selection. In other embodiments, the model pool comprises n-fold cross validation models and/or bootstrapping models. In some embodiments, only models with superior predictive power are selected to enter the pool based on various model selection criteria, such as AIC or BIC.

Alternatively or additionally in some embodiments, models that have not been screened by model selection also enter the model pool. In one embodiment, all models with all linear and nonlinear terms are entered into the model pool. For a large number of residues and a much larger number of factorial interactions among residues, this embodiment can be very computationally intensive. In some alternative embodiments, only models containing linear terms and pairwise interaction terms are entered into the model pool. Regardless the inclusion method of the model pool, an ensemble model includes all the terms of its constituents. The model pool can contain any number of models, including, but not limited to Bayesian models, in which case, prior information can be incorporated into the ensemble.

In some embodiments, the ensemble predicts sequence activity based on the weighted mean of the coefficients of each model in the pool, wherein the weights are determined by the predictive power of the corresponding models, as indicated by block 605.

In some embodiments, an ensemble regression uses the following work flow: (1) provide an empty ensemble; (2) select a group size n of 1 or greater; (3) categorize data points into groups of size n, where data points are grouped without replacement; and (4) prepare an ensemble model to predict individual and interaction coefficients. In some embodiments, step (4) to prepare an ensemble model further comprises: a) removing data points of each group, wherein remaining data forms a training set and left out data forms a validation set; b) preparing a model by fitting the training set using stepwise regression; c) testing the model using the validation set, which provides an indication of the predictive ability of the model; d) add model to a pool of models that are used to generate an ensemble model as described above.

Bootstrap Approach

Other techniques for characterizing the predictive power of a model under consideration in a given iteration find use in the present invention. In some embodiments, these techniques involve cross validation or bootstrap techniques. In some embodiments, cross validation employs a set of observations used to generate the model but leaves some of the observations out to assess the strength of the model. In some embodiments, the bootstrap technique involves using a set of samples that are tested with replacement. In some embodiments, models generated by cross validation or bootstrapping can be combined into an ensemble model as described above.

In some additional embodiments, the methods rank residues not simply by the magnitudes of their predicted contributions to activity, but by the confidence in those predicted contributions as well. In some cases, the researcher is concerned with the generalizability of the model from one set of data to another set. In other words, the researcher wants to know whether or not the values of the coefficients or principal components are spurious. Cross validation and bootstrapping techniques provide measures to indicate the confidence level that the models are generalizable to various data.

In some embodiments, a more statistically rigorous approach is utilized in which the ranking is based on a combination of magnitude and distribution. In some of these embodiments, coefficients with both high magnitudes and tight distributions give the highest ranking. In some cases, one coefficient with a lower magnitude than another may be given a higher ranking by virtue of having less variation. Thus, some embodiments rank amino acid residues or nucleotides based on both magnitude and standard deviation or variance. Various techniques can be used to accomplish this. Indeed, it is not intended that the present invention be limited to any specific technique for ranking. One embodiment using a bootstrap p-value approach is described below.

An illustrative example of a method that employs a bootstrap method is depicted in FIG. 7. As shown in FIG. 7, the method 725 begins at block 727, where an original data set S is provided. In some embodiments, this is a training set as described above. For example in some embodiments, it is generated by systematically varying the individual residues of a starting sequence in any manner (e.g., as described above). In the case illustrated by method 725, the data set S has M different data points (activity and sequence information collected from amino acid or nucleotide sequences) for use in the analysis.

From data set S, various bootstrap sets B are created. Each of these sets is obtained by sampling, with replacement, from set S to create a new set of M membersâ€”all taken from original set S. See block 729. The â€œwith replacementâ€ condition produces variations on the original set S. The new bootstrap set, B, will sometimes contain replicate samples from S. In some cases, the bootstrap set B also lacks certain samples originally contained in S.

As an illustrative example, a set S of 100 sequences is provided. A bootstrap set B is created by randomly selecting 100 member sequences from the 100 sequences in the original set S. Each bootstrap set B used in the method contains 100 sequences. Thus, it is possible that some sequences will be selected more than once and others will not be selected at all. Using the bootstrap set B produced from the set S of 100 sequences, the method next builds a model. See block 731. The model may be built as described above, using PLS, PCR, a SVM, stepwise regression, etc. Indeed, it is intended that any suitable method will find use in building the model. This model provides coefficients or other indicia of ranking for the residues or nucleotides found in the various samples from set B. As shown at a block 733, these coefficients or other indicia are recorded for subsequent use.

Next, at a decision block 735, the method determines whether another bootstrap set should be created. If yes, the method returns to block 729 where a new bootstrap set B is created as described above. If no, the method proceeds to a block 737 discussed below. The decision at block 735 turns on how many different sets of coefficient values are to be used in assessing the distributions of those values. The number of sets B should be sufficient to generate accurate statistics. In some embodiments, 100 to 1000 bootstrap sets are prepared and analyzed. This is represented by about 100 to 1000 passes through blocks 729, 731, and 733 of method 725. However, it is not intended that the present invention be limited to any particular number of bootstrap sets, as any number suitable for the desired analysis finds use.

After a sufficient number bootstrap sets B have been prepared and analyzed, decision 735 is answered in the negative. As indicated, the method then proceeds to block 737. There, a mean and standard deviation of a coefficient (or other indicator generated by the model) are calculated for each residue or nucleotide (including codons) using the coefficient values (e.g., 100 to 1000 values, one from each bootstrap set). From this information, the method can calculate the t-statistic and determine the confidence interval that the measured value is different from zero. From the t-statistic, it calculates the p-value for the confidence interval. In this illustrative case, the smaller the p-value, the more confidence that the measured regression coefficient is different from zero.

It is noted that the p-value is but one of many different types of characterizations that can account for the statistical variation in a coefficient or other indicator of residue importance. Examples include, but are not limited to calculating 95 percent confidence intervals for regression coefficients and excluding any regression coefficient for consideration for which 95 percent confidence interval crosses line zero. Basically, in some embodiments, any characterization that accounts for standard deviation, variance, or other statistically relevant measure of data distribution finds use. In some embodiments, this characterization step also accounts for the magnitude of the coefficients.

In some embodiments, a large standard deviation results. This large standard deviation may be due to various causes, including but not limited to poor measurements in the data set, and/or limited representation of a particular residue or nucleotide in the original data set. In this latter case, some bootstrap sets will contain no occurrences of a particular residue or nucleotide. In such cases, the value of the coefficient for that residue will be zero. Other bootstrap sets will contain at least some occurrences of the residue or nucleotide and give a non-zero value of the corresponding coefficient. But the sets giving a zero value will cause the standard deviation of the coefficient to become relatively large. This reduces the confidence in the coefficient value and results in a lower rank. But this is to be expected, given that there is relatively little data on the residue or nucleotide involved.

Next, at a block 739, the method ranks the regression coefficients (or other indicators) from lowest (best) p-value to highest (worst) p-value. This ranking correlates highly with the absolute value of the regression coefficients themselves, owing to the fact that the larger are the absolute value, the more standard deviations are removed from zero. Thus, for a given standard deviation, the p-value becomes smaller as the regression coefficient becomes larger. However, the absolute ranking will not always be the same with both p-value and pure magnitude methods, especially when relatively few data points are available to begin with in set S.

Finally, as shown at a block 741, the method fixes and toggles certain residues, based on the rankings observed in the operation of block 739. This is essentially the same use of rankings described above for other embodiments. In one approach, the method fixes the best residues (now those with the lowest p-values) and toggles the others (those with highest p-values).

This method 725 has been shown to perform well in silico. Moreover, in some embodiments, the p-value ranking approach naturally deals with single or few instance residues: the p-values will generally be higher (worse) because in the bootstrap process, those residues that did not appear often in the original data set will be less likely to get picked up at random. Even if their coefficients are large, their variability (measured in standard deviations) will be quite high as well. In some embodiments, this is the desired result, as those residues that are not well represented (i.e., either have not seen with sufficient frequency or have lower regression coefficients) may be good candidates for toggling in the next round of library design.

E. Generating an Optimized Protein Variant Library by Modifying Model-Predicted Sequences

One of the goals of the invention is to generate an optimized protein variant library through directed evolution. Some embodiments of the invention provide methods to guide directed evolution of protein variants using the generated sequence-activity models. The various sequence-activities models prepared and refined according to the methods described above are suitable to guide directed evolution of proteins or biological molecules. As part of the process, the methods may identify sequences that are to be used for generating a new protein variant library. Such sequences include variations on the defined residues identified above, or are precursors used to subsequently introduce such variations. The sequences may be modified by performing mutagenesis or a recombination-based diversity generation mechanism to generate the new library of protein variants. The new library may also be used in developing a new sequence-activity model.

In some embodiments, preparation of oligonucleotides or nucleic acid sequences is achieved by synthesizing the oligonucleotides or nucleic acid sequences using a nucleic acid synthesizer. Some embodiments of the invention include performing a round of directed evolution using the prepared oligonucleotides or protein sequence as building blocks for directed evolution. Various embodiments of the invention can apply recombination and/or mutagenesis to these building blocks to generate diversity.

As one specific example, some embodiments apply recombination techniques to oligonucleotides. In these embodiments, the methods involve selecting one or more mutations for a round of directed evolution by evaluating the coefficients of the terms of the sequence-activity model. Mutations are selected from combinations of defined amino acids or nucleotides of specific types at specific positions based on their contributions to the activity of proteins as predicted by the models. In some embodiments, selection of mutations involves identifying one or more coefficients that are determined to be larger than others of the coefficients, and selecting the defined amino acid or nucleotide at a defined position represented by the one or more coefficients so identified. In some embodiments, after selecting mutations according to the sequence-activity models, the methods involve preparing a plurality of oligonucleotides containing or encoding the one or more mutations, and performing a round of directed evolution using the oligonucleotides prepared. In some embodiments, the directed evolution techniques involve combining and/or recombining the oligonucleotides.

Other embodiments of the invention apply recombination techniques to protein sequences. In some embodiments, the methods involve identifying a new protein or a new nucleic acid sequence, and preparing and assaying the new protein or a protein encoded by the new nucleic acid sequence. In some embodiments, the methods further involve using the new protein or protein encoded by the new nucleic acid sequence as a starting point for further directed evolution. In some embodiments, the directed evolution process involves fragmenting and recombining the protein sequence that is predicted by the model to have a desired level of activity.

In some embodiments, the methods identify and/or prepare a new protein or a new nucleic acid sequence based on individual mutations that are predicted to be important by the model. These methods involve: selecting one or more mutations by evaluating the coefficients of the terms of the sequence-activity model to identify one or more of the defined amino acids or nucleotides at the defined positions that contribute to the activity; identifying a new protein or a new nucleic acid sequence comprising the one or more mutations selected above, and preparing and assaying the new protein or a protein encoded by the new nucleic acid sequence.

In other embodiments, the methods identify and/or prepare a new protein or a new nucleic acid sequence based on the predicted activity of a whole sequence instead of individual mutations. In some of these embodiments, the methods involve applying multiple protein sequences or multiple amino acid sequences to the sequence-activity model and determining activity values predicted by the sequence-activity model for each of the multiple protein sequences or nucleic acid sequences. The methods further involve selecting a new protein sequence or a new nucleic acid sequence from among the multiple protein sequences or multiple amino acid sequences applied above by evaluating the activity values predicted by the sequence-activity model for the multiple sequences. The methods also involve preparing and assaying a protein having the new protein sequence or a protein encoded by the new nucleic acid sequence.

In some embodiments, rather than simply synthesizing the single best-predicted protein, a combinatorial library of proteins is generated based on a sensitivity analysis of the best changes in the residue choices at each location in the protein. In this embodiment, the more sensitive a given residue choice is for the predicted protein, the greater the predicted fitness change will be. In some embodiments these sensitivities are from highest to lowest and the sensitivity scores are used to create combinatorial protein libraries in subsequent rounds (i.e., by incorporating those residues based on sensitivity). In some embodiment, in which a linear model is used, the sensitivity is identified by simply considering the size of the coefficient associated with a given residue term in the model. However, this is not possible for non-linear models. Instead, in embodiments utilizing non-linear models, the residue sensitivity is determined by using the model to calculate changes in activity when a single residue is varied in the â€œbestâ€ predicted sequence.

Some embodiments of the invention include selecting one or more positions in the protein sequence or nucleic acid sequence and conducting saturation mutagenesis at the one or more positions so identified. In some embodiments, the positions are selected by evaluating the coefficients of the terms of the sequence-activity model to identify one or more of the defined amino acids or nucleotides at the defined positions that contribute to the activity. Accordingly, in some embodiments, a round of directed evolution includes performing saturation mutagenesis on a protein sequence at positions selected using the sequence-activity models. In some embodiments involving models comprising one or more interaction terms, the methods involve applying mutagenesis simultaneously at the two or more interacting residues.

In some embodiments, residues are taken into consideration in the order in which they are ranked. In some embodiments, for each residue under consideration, the process determines whether to â€œtoggleâ€ that residue. The term â€œtogglingâ€ refers to the introduction of multiple amino acid residue types into a specific position in the sequences of protein variants in the optimized library. For example, serine may appear in position 166 in one protein variant, whereas phenylalanine may appear in position 166 in another protein variant in the same library. Amino acid residues that do not vary between protein variant sequences in the training set typically remain fixed in the optimized library. However, this is not always the case, as there can be variation in the optimized libraries.

In some embodiments, an optimized protein variant library is designed such that all of the identified â€œhighâ€ ranking regression coefficient residues are fixed, and the remaining lower ranking regression coefficient residues are toggled. The rationale for this embodiment is that the local space surrounding the â€˜bestâ€™ predicted protein should be searched. It is noted that the starting point â€œbackboneâ€ in which the toggles are introduced may be the best protein predicted by a model and/or an already validated â€˜bestâ€™ protein from a screened library. Indeed, it is not intended that the starting point backbone be limited to any particular protein.

In an alternative embodiment, at least one or more, but not all of the identified high-ranking regression coefficient residues are fixed in the optimized library, and the others toggled. This approach is recommended in some embodiments, if there is a desire to not drastically change the context of the other amino acid residues by incorporating too many changes at one time. Again, the starting point for toggling may be the best set of residues as predicted by the model, a best validated protein from an existing library, or an â€œaverageâ€ clone that models well. In the latter case, it may be desirable to toggle the residues predicted to be of higher importance, as a larger space should be explored in the search for activity hills previously omitted from the sampling. This type of library is typically more relevant in early rounds of library production, as it generates a more refined picture for subsequent rounds. It is also not intended that the starting point backbone be limited to any particular protein.

Some alternatives of the above embodiments involve different procedures for using residue importance (rankings) in determining which residues to toggle. In one such alternative embodiment, higher ranked residue positions are more aggressively favored for toggling. The information needed in this approach includes the sequence of a best protein from the training set, a PLS or PCR predicted best sequence, and a ranking of residues from the PLS or PCR model. The â€œbestâ€ protein is a wet-lab validated â€œbestâ€ clone in the dataset (i.e., the clone with the highest measured function that still models well in that it falls relatively close to the predicted value in cross validation). The method compares each residue from this protein with the corresponding residue from a â€œbest predictedâ€ sequence having the highest value of the desired activity. If the residue with the highest load or regression coefficient is not present in the â€˜bestâ€™ clone, the method introduces that position as a toggle position for the subsequent library. If the residue is present in the best clone, the method does not treat the position as a toggle position, and it will move to the next position in succession. The process is repeated for various residues, moving through successively lower load values, until a library of sufficient size is generated.

In some embodiments, the number of regression coefficient residues to retain, and number of regression coefficient residues to toggle, are varied. The determination of which residues to toggle and which to retain is based on various factors including but not limited to the desired library size, the magnitude of difference between regression coefficients, and the degree to which nonlinearity is thought to exist. Retention of residues with small (neutral) coefficients may uncover important nonlinearities in subsequent rounds of evolution. In some embodiments, optimized protein variant libraries contain about 2^Nprotein variants, where N represents the number of positions that are toggled between two residues. Stated another way, the diversity added by each additional toggle doubles the size of the library such that 10 toggle positions produces Ëœ1,000 clones (1,024), 13 positions Ëœ10,000 clones (8,192) and 20 positions Ëœ1,000,000 clones (1,048,576). The appropriate library size depends on factors such as cost of screen, ruggedness of landscape, preferred percentage sampling of space, etc. In some cases, it has been found that a relatively large number of changed residues produce a library in which an inordinately large percentage of the clones are non-functional. Therefore, in some embodiments, the number of residues for toggling ranges from about 2 to about 30; i.e., the library size ranges from between about 4 and 2³⁰Ëœ10⁹clones.

In addition, it is contemplated that various subsequent round library strategies be utilized simultaneously, with some strategies being more aggressive (fixing more â€œbeneficialâ€ residues) and other strategies being more conservative (fixing fewer â€œbeneficialâ€ residues with the goal of exploring the space more thoroughly).

In some embodiments, groups or residues or â€œmotifsâ€ that occur in most naturally occurring or otherwise successful peptides are identified and/or preserved, as they may be important in the functionality of the protein (e.g., activity, stability, etc.). For example, it may be found that Ile at variable position 3 is always coupled with Val at variable position 11 in naturally occurring peptides. Hence, in one embodiment, preservation of such groups is required in any toggling strategy. In other words, the only accepted toggles are those that preserve a particular grouping in the base protein or those that generate a different grouping that is also found in active proteins. In the latter case it is necessary to toggle two or more residues.

In some additional embodiments, a wet-lab validated â€˜bestâ€™ (or one of the few best) protein in the current optimized library (i.e., a protein with the highest, or one of the few highest, measured function that still models well, i.e., falls relatively close to the predicted value in cross validation) serves as a backbone in which various changes are incorporated. In another approach, a wet-lab validated â€˜bestâ€™ (or one of the few best) protein in the current library that may not model well serves as a backbone where various changes are incorporated. In some other approaches, a sequence predicted by the sequence-activity model to have the highest value (or one of the highest values) of the desired activity serves as the backbone. In these approaches, the dataset for the â€œnext generationâ€ library (and possibly a corresponding model) is obtained by changing residues in one or a few of the best proteins. In one embodiment, these changes comprise a systematic variation of the residues in the backbone. In some cases, the changes comprise various mutagenesis, recombination and/or subsequence selection techniques. Each of these may be performed in vitro, in vivo, and/or in silico. Indeed, it is not intended that the present invention be limited to any particular format, as any suitable format finds use.

In some embodiments, while the optimal sequence predicted by a linear model can be identified by inspection as described above, the same is not true for non-linear models. Certain residues appear in both linear and cross product terms and their overall effect on activity in the context of many possible combinations of other residues can be problematic. Thus, as with selection of cross product terms for a non-linear model, the optimal sequence predicted by a non-linear model can be identified by testing all possible sequences with the model (assuming sufficient computational resources) or by utilizing a searching algorithm such as a stepwise algorithm.

In some embodiments, the information contained in the computer-evolved proteins identified as described above is used to synthesize novel proteins and test them on physical assays. An accurate in silico representation of the actual wet lab-determined fitness function, allows researchers to reduce the number of cycles of evolution and/or the number variants needed to be screened in the lab. In some embodiments, optimized protein variant libraries are generated using the recombination methods described herein, or alternatively, by gene synthesis methods, followed by in vivo or in vitro expression. In some embodiments, after the optimized protein variant libraries are screened for desired activity, they are sequenced. As indicated above in the discussion of FIGS. 1 and 2, the activity and sequence information from the optimized protein variant library can be employed to generate another sequence-activity model from which a further optimized library can be designed, using the methods described herein. In one embodiment, all of the proteins from this new library are used as part of the dataset.

III. Digital Apparatus and Systems

As should be apparent, embodiments described herein employ processes acting under control of instructions and/or data stored in or transferred through one or more computer systems. Embodiments disclosed herein also relate to apparatus for performing these operations. In some embodiments, the apparatus is specially designed and/or constructed for the required purposes, or it may be a general-purpose computer selectively activated or reconfigured by a computer program and/or data structure stored in the computer. The processes provided by the present invention are not inherently related to any particular computer or other specific apparatus. In particular, various general-purpose machines find use with programs written in accordance with the teachings herein. However, in some embodiments, a specialized apparatus is constructed to perform the required method operations. One embodiment of a particular structure for a variety of these machines is described below.

In addition, certain embodiments of the present invention relate to computer readable media or computer program products that include program instructions and/or data (including data structures) for performing various computer-implemented operations. Examples of computer-readable media include, but are not limited to, magnetic media such as hard disks, floppy disks, magnetic tape; optical media such as CD-ROM devices and holographic devices; magneto-optical media; semiconductor memory devices; and hardware devices that are specially configured to store and perform program instructions, such as read-only memory devices (ROM) and random access memory (RAM), application-specific integrated circuits (ASICs), and programmable logic devices (PLDs). The data and program instructions may also be embodied on a carrier wave or other transport medium (e.g., optical lines, electrical lines, and/or airwaves). Indeed, it is not intended that the present invention be limited to any particular computer-readable media or any other computer program products that include instructions and/or data for performing computer-implemented operations.

Examples of program instructions include, but are not limited to low-level code such as produced by a compiler, and files containing higher level code that may be executed by the computer using an interpreter. Further, the program instructions include, but are not limited to machine code, source code and any other code that directly or indirectly controls operation of a computing machine in accordance with the present invention. The code may specify input, output, calculations, conditionals, branches, iterative loops, etc.

In one illustrative example, code embodying methods disclosed herein are embodied in a fixed media or transmissible program component containing logic instructions and/or data that when loaded into an appropriately configured computing device causes the device to perform a simulated genetic operation (GO) on one or more character string(s). FIG. 8 shows an example digital device 800 that is a logical apparatus that can read instructions from media 817, network port 819, user input keyboard 809, user input 811, or other inputting means. Apparatus 800 can thereafter use those instructions to direct statistical operations in data space, e.g., to construct one or more data set (e.g., to determine a plurality of representative members of the data space). One type of logical apparatus that can embody disclosed embodiments is a computer system as in computer system 800 comprising CPU 807, optional user input devices keyboard 809, and GUI pointing device 811, as well as peripheral components such as disk drives 815 and monitor 805 (which displays GO modified character strings and provides for simplified selection of subsets of such character strings by a user. Fixed media 817 is optionally used to program the overall system and can include, e.g., a disk-type optical or magnetic media or other electronic memory storage element. Communication port 819 can be used to program the system and can represent any type of communication connection.

In some embodiments, the disclosure provides a computer system including one or more processors; system memory; and one or more computer-readable storage media having stored thereon computer-executable instructions that, when executed by the one or more processors, cause the computer system to implement a method for conducting directed evolution of biological molecules. In some embodiments, the method includes: (a) receiving sequence and activity data for a plurality of biological molecules; (b) preparing a base model from the sequence and activity data, wherein the base model predicts activity as a function of the presence or absence of sub-units of the sequence; (c) preparing at least one new model by adding or subtracting at least one new interaction term to or from the base model, wherein the new interaction term represents the interaction between two or more interacting sub-units; (d) determining the ability of the at least one new model to predict activity as a function of the presence or absence of the sub-units; and (e) determining whether to add or subtract the new interaction term to or from the base model based on the ability of the at least one new model to predict activity as determined in (d) and with a bias against including additional interaction terms.

Certain embodiments can also be embodied within the circuitry of an application specific integrated circuit (ASIC) or programmable logic device (PLD). In such a case, the embodiments are implemented in a computer readable descriptor language that can be used to create an ASIC or PLD. Some embodiments of the present invention are implemented within the circuitry or logic processors of a variety of other digital apparatus, such as PDAs, laptop computer systems, displays, image editing equipment, etc.

In some embodiments, the present invention relates to a computer program product comprising one or more computer-readable storage media having stored thereon computer-executable instructions that, when executed by one or more processors of a computer system, cause the computer system to implement a method for identifying biological molecules to affect a desired activity. Such method be any method described herein such as those encompassed by the Figures and pseudocode. In some embodiments, the method receives sequence and activity data for a plurality of biological molecules, and prepares a base model and an improved model from the sequence and activity data. In some embodiments, the model predicts activity as a function of the presence or absence of sub-units of the sequence.

In some embodiments of the present invention, the method implemented by the computer program product prepares at least one new model by adding or subtracting at least one new interaction term to or from the base model, wherein the new interaction term represents the interaction between two or more interacting sub-units. In some embodiments, the method determines the ability of at least one new model to predict activity as a function of the presence or absence of the sub-units. The method also determines whether to add or subtract the new interaction term to or from the base model based on the ability of at least one new model to predict activity as determined above and with a bias against including additional interaction terms.

While the foregoing has been described in some detail for purposes of clarity and understanding, it will be clear to one skilled in the art from a reading of this disclosure that various changes in form and detail can be made without departing from the true scope of the disclosure. For example, all the techniques and apparatus described above may be used in various combinations. All publications, patents, patent applications, or other documents cited in this application are incorporated by reference in their entirety for all purposes to the same extent as if each individual publication, patent, patent application, or other document were individually indicated to be incorporated by reference for all purposes.