CROSS-REFERENCE TO RELATED APPLICATIONS

The present divisional application claims priority to and the benefit of U.S. patent application Ser. No. 16/870,611, filed May 8, 2020, and titled “ARTIFICIAL INTELLIGENCE ENGINE FOR GENERATING CANDIDATE DRUGS,” which claims priority to and the benefit of U.S. Provisional Application No. 62/975,470, filed Feb. 12, 2020, and titled “ARTIFICIAL INTELLIGENCE ENGINE FOR GENERATING CANDIDATE DRUGS.” The content of these applications are incorporated herein by reference in their entireties for all purposes.

TECHNICAL FIELD

This disclosure relates generally to drug discovery. More specifically, this disclosure relates to an artificial intelligence engine architecture for generating candidate drugs.

BACKGROUND

Therapeutics may refer to a branch of medicine concerned with the treatment of disease and the action of remedial agents (e.g., drugs). Therapeutics includes, but is not limited to, the field of ethical pharmaceuticals. Entities in the therapeutics industry may discover, develop, produce, and market drugs for use as medications to be administered or self-administered to patients. Goals of administering or self-administering the drugs may include curing the patient of a disease, causing an active disease to enter a state of remission, vaccinating the patient by stimulating the immune system to better protect against the disease, and/or alleviating, mitigating or ameliorating a symptom. Existing drug discoveries may be based on any combination of human design, high-throughput screening, synthetic products and natural substances.

SUMMARY

In general, the present disclosure provides an artificial intelligence engine for generating candidate drugs.

In one aspect, a method is disclosed which may include an artificial intelligence engine architecture for generating candidate drugs. In one embodiment, a method includes generating, via a creator module, a candidate drug compound including a sequence of candidate drug compounds, including the candidate drug compound as a node in a knowledge graph; generating, via a descriptor module, a description of the candidate drug compound at the node in the knowledge graph, wherein the description comprises drug compound structural information, drug compound activity information, and drug compound semantic information; based on the description, performing, via a scientist module, a benchmark analysis of a parameter of the creator module; and modifying, based on the benchmark analysis, the creator module to change the parameter in a desired way during a subsequent benchmark analysis.

In another aspect, a system may include a memory device storing instructions and a processing device communicatively coupled to the memory device. The processing device may execute the instructions to perform one or more operations of any method disclosed herein.

In another aspect, a tangible, non-transitory computer-readable medium may store instructions and a processing device may execute the instructions to perform one or more operations of any method disclosed herein.

Other technical features may be readily apparent to one skilled in the art from the following figures, descriptions, and claims.

Before undertaking the DETAILED DESCRIPTION below, it may be advantageous to set forth definitions of certain words and phrases used throughout this patent document. The term “couple” and its derivatives refer to any direct or indirect communication between two or more elements, independent of whether those elements are in physical contact with one another. The terms “transmit,” “receive,” and “communicate,” as well as derivatives thereof, encompass both direct and indirect communication. The terms “transmit,” “receive,” and “communicate,” as well as derivatives thereof, encompass both communication with remote systems and communication within a system, including reading and writing to different portions of a memory device. The terms “include” and “comprise,” as well as derivatives thereof, mean inclusion without limitation. The term “or” is inclusive, meaning and/or. The phrase “associated with,” as well as derivatives thereof, means to include, be included within, interconnect with, contain, be contained within, connect to or with, couple to or with, be communicable with, cooperate with, interleave, juxtapose, be proximate to, be bound to or with, have, have a property of, have a relationship to or with, or the like. The term “translate” may refer to any operation performed wherein data is input in one format, representation, language (computer, purpose-specific, such as drug design or integrated circuit design), structure, appearance or other written, oral or representable instantiation and data is output in a different format, representation, language (computer, purpose-specific, such as drug design or integrated circuit design), structure, appearance or other written, oral or representable instantiation, wherein the data output has a similar or identical meaning, semantically or otherwise, to the data input. Translation as a process includes but is not limited to substitution (including macro substitution), encryption, hashing, encoding, decoding or other mathematical or other operations performed on the input data. The same means of translation performed on the same input data will consistently yield the same output data, while a different means of translation performed on the same input data may yield different output data which nevertheless preserves all or part of the meaning or function of the input data, for a given purpose. Notwithstanding the foregoing, in a mathematically degenerate case, a translation can output data identical to the input data. The term “controller” means any device, system or part thereof that controls at least one operation. Such a controller may be implemented in hardware or a combination of hardware and software and/or firmware. The functionality associated with any particular controller may be centralized or distributed, whether locally or remotely. The phrase “at least one of,” when used with a list of items, means that different combinations of one or more of the listed items may be used, and only one item in the list may be needed. For example, “at least one of: A, B, and C” includes any of the following combinations: A, B, C, A and B, A and C, B and C, and A and B and C.

Moreover, various functions described below can be implemented or supported by one or more computer programs, each of which is formed from computer readable program code and embodied in a computer readable storage medium. The terms “application” and “program” refer to one or more computer programs, software components, sets of instructions, procedures, functions, objects, classes, instances, related data, or a portion thereof adapted for implementation in a suitable computer readable program code. The phrase “computer readable program code” includes any type of computer code, including source code, object code, and executable code. The phrase “computer readable storage medium” includes any type of medium capable of being accessed by a computer, such as read only memory (ROM), random access memory (RAM), a hard disk drive, a compact disc (CD), a digital video disc (DVD), solid state drive (SSD), or any other type of memory. A “non-transitory” computer readable storage medium excludes wired, wireless, optical, or other communication links that transport transitory electrical or other signals. A non-transitory computer readable storage medium includes media where data can be permanently stored and media where data can be stored and later overwritten, such as a rewritable optical disc or an erasable memory device.

The terms “candidate drugs” and “candidate drug compounds” may be used interchangeably herein.

Definitions for other certain words and phrases are provided throughout this patent document. Those of ordinary skill in the art should understand that in many if not most instances, such definitions apply to prior as well as future uses of such defined words and phrases.

BRIEF DESCRIPTION OF THE DRAWINGS

For a more complete understanding of this disclosure and its advantages, reference is now made to the following description, taken in conjunction with the accompanying drawings, in which:

FIG. 1A illustrates a high-level component diagram of an illustrative system architecture according to certain embodiments of this disclosure;

FIG. 1B illustrates an architecture of the artificial intelligence engine according to certain embodiments of this disclosure;

FIG. 1C illustrates first components of an architecture of the creator module according to certain embodiments of this disclosure;

FIG. 1D illustrates second components of the architecture of the creator module according to certain embodiments of this disclosure;

FIG. 1E illustrates an architecture of a variational autoencoder according to certain embodiments of this disclosure;

FIG. 1F illustrates an architecture of a generative adversarial network used to generate candidate drugs according to certain embodiments of this disclosure;

FIG. 1G illustrates types of encodings to represent certain types of drug information according to certain embodiments of this disclosure;

FIG. 1H illustrates an example of concatenating numerous encodings into a candidate drug according to certain embodiments of this disclosure;

FIG. 1I illustrates an example of using a variational autoencoder to generate a latent representation of a candidate drug according to certain embodiments of this disclosure;

FIG. 2 illustrates a data structure storing a biological context representation according to certain embodiments of this disclosure;

FIGS. 3A-3B illustrate a high-level flow diagram according to certain embodiments of this disclosure;

FIG. 4 illustrates example operations of a method for generating and classifying a candidate drug compound according to certain embodiments of this disclosure;

FIGS. 5A-5D provide illustrations of generating a first data structure including a biological context representation of a plurality of drug compounds according to certain embodiments of this disclosure;

FIG. 6 illustrates example operations of a method for translating the first data structure of FIGS. 5A-5D into a second data structure having a second format according to certain embodiments of this disclosure;

FIG. 7 provide illustrations of translating the first data structure of FIGS. 5A-5D into the second data structure having the second format according to certain embodiments of this disclosure;

FIGS. 8A-8C provide illustrations of views of a selected candidate drug compound according to certain embodiments of this disclosure;

FIG. 9 illustrates example operations of a method for presenting a view including a selected candidate drug compound according to certain embodiments of this disclosure;

FIG. 10A illustrates example operations of a method for using causal inference during the generation of candidate drug compounds according to certain embodiments of this disclosure;

FIG. 10B illustrates another example of operations of a method for using causal inference during the generation of candidate drug compounds according to certain embodiments of this disclosure;

FIG. 11 illustrates example operations of a method for using several machine learning models in an artificial intelligence engine architecture to generate peptides according to certain embodiments of this disclosure;

FIG. 12 illustrates example operations of a method for performing a benchmark analysis according to certain embodiments of this disclosure;

FIG. 13 illustrates example operations of a method for slicing a latent representation based on a shape of the latent representation according to certain embodiments of this disclosure; and

FIG. 14 illustrates an example computer system according to certain embodiments of this disclosure.

DETAILED DESCRIPTION

Conventional drug discoveries based on human design, high-throughput screening, and/or natural substances may be inefficient, riven with noise, limited in application, not efficacious, dangerous or poisonous, and/or not defensible. Further, in some instances, there are instances of certain diseases (e.g., instances of prosthetic joint infections) that do not have a corresponding existing therapeutic to treat the certain diseases or which provide temporary results against which the disease is refractory. One reason for the lack of an existing therapeutic may be the conventional drug discovery techniques are incapable of discovering the therapeutic needed to treat the certain diseases. By “treat,” we mean that the disease at hand is cured inter alia, that it is not refractory to treatment. The amount of knowledge, data, assumptions, and queries used to discover a therapeutic to treat the certain disease may be unattainable, overwhelming, and/or inefficiently determined, such that conventional drug discovery techniques cannot overcome these obstacles. Improvement is desired in the field of therapeutics.

Further, conventional techniques for searching for candidate drugs use limited design spaces. For example, some conventional techniques focus on a fact about drugs, where such facts constrain the design space that is searched. The design space may refer to parameterization of limits and constraints in a drug space where candidate drug compounds may be designed. A design space may also refer to a multidimensional combination and interaction of input variables (e.g., material attributes) and process parameters that have been demonstrated to provide assurance of quality. An example of such a fact may include a certain biomedical activity known to be linked to an alpha-helix physical structure of a peptide, where conventional techniques may search for other activities that may result from a peptide having the alpha-helix physical structure. Such a limited design space may limit the results obtained. Thus, it is desirable to enlarge the design space to account for other information such as drug sequence information, drug activity information, drug semantic information, drug chemical information, drug physical information, and so forth. However, enlarging the design space may increase the complexity of searching the design space.

Accordingly, aspects of the present disclosure generally relate to an artificial intelligence engine for generating candidate drugs. By using various encoding types that enable performing searches in the design space in an efficient manner, the artificial intelligence engine (AI) may enlarge the design space to include the combination of drug information (e.g., structural, physical, semantic, activity, sequence, chemical, etc.). The architecture of the AI engine may include various computational techniques that reduce the computational complexity of using a large design space, thereby saving computing resources (e.g., reducing computing time, reducing processing resources, reducing memory resources, etc.). At the same time, the disclosed architecture may generate superior candidate drugs that include desirable features (e.g., structure, semantics, activity, sequence, clinical outcomes, etc.) found in the larger design space as compared to conventional techniques using the smaller design space.

The artificial intelligence (AI) engine may use a combination of rational algorithmic discovery and machine learning models (e.g., generative deep learning methods) to produce enhanced therapeutics that may treat any suitable target disease and/or medical condition. The AI engine may discover, translate, design, generate, create, develop, formulate, classify, and/or test candidate drug compounds that exhibit desired activity (e.g., antimicrobial, immunomodulatory, cytotoxic, neuromodulatory, etc.) in design spaces for target diseases and/or medical conditions. Such candidate drug compounds that exhibit desired activity in a design space may effectively treat the disease and/or medical condition associated with that design space. In some embodiments, a selected candidate drug compound that effectively treats the disease and/or medical condition may be formulated into an actual drug for administration and may be tested in a lab and/or at a clinical stage.

In general, the disclosed embodiments may enable rationally discovery of drug compounds for a larger design space at a larger scale, higher accuracy, and/or higher efficiency than conventional techniques. The AI engine may use various machine learning models to discover, translate, design, generate, create, develop, formulate, classify, and/or test candidate drug compounds. Each of the various machine learning models may perform certain specific operations. The types of machine learning models may include various neural networks that perform deep learning, computational biology, and/or algorithmic discovery. Examples of such neural networks may include generative adversarial networks, recurrent neural networks, convolutional neural networks, fully connected neural networks, etc., as described further below; and such networks may also additionally employ methods of or incorporating causal inference, including counterfactuals, in the process of discovery.

In some embodiments, a biological context representation of a set of drug compounds may be generated. The biological context representation may be a continuous representation of a biological setting that is updated as knowledge is acquired and/or data is updated. The biological context representation may be stored in a first data structure having a format (e.g., a knowledge graph) that includes both various nodes pertaining to health artifacts and various relationships connecting the nodes. The nodes and relationships may form logical structures having subjects and predicates. For example, one logical structure between two nodes having a relation may be “Genes are associated with Diseases” where “Genes” and “Diseases” are the subjects of the logical structure and “are associated with” is the relation. In such a way, the knowledge graph may encompass actual knowledge, rather than simply statistical inferences, pertaining to a biological setting.

The information in the knowledge graph may be continuously or periodically updated and the information may be received from various sources curated by the AI engine. The knowledge in the biological context representation goes well beyond “dumb” data that just includes quantities of a value because the knowledge represents the relationships between or among numerous different types of data, as well as any or all of direct, indirect, causal, counterfactual or inferred relationships. In some embodiments, the biological context representation may not be stored, and instead, based on the stream of knowledge included in the biological context representation, may be streamed from data sources into the AI engine that generates the machine learning models.

The biological context representation may be used to generate candidate drug compounds by translating the first data format to a second data structure having a second format (e.g., a vector). The second format may be more computationally efficient and/or suitable for generating candidate drug compounds that include sequences of ingredients that provide desired activity in a design space. “Ingredients” as used herein may refer, without limitation, to substances, compounds, elements, activities (such as the application or removal of electrical charge or a magnetic field for a specific maximum, minimum or discrete amount of time), and mixtures. Further, the second format may enable generating views of the levels of activity provided by the sequence of ingredients in a certain design space, as described further below.

At a high level, the AI engine may include at least one machine learning model that is trained to use causal inference to generate candidate drug compounds. One of the challenges with discovering new therapeutics may include determining whether certain ingredients are causal agents with respect to certain activity in a design space. The sheer number of possible sequences of ingredients may be extraordinarily large due to mathematical combinatorics, such that identifying a cause and effect relationship between ingredients and activity may be impossible or, at best, extremely unlikely, to identify without the disclosed embodiments. (For example, in public-key encryption, it is theoretically possible to discover and unlock a private key, but doing this would presently require all the computing power in the world to work longer than the age of the universe: this is an example of what is mathematically possible, but impossible within human time frames and computing power. Identifying a cause-and-effect relationship between ingredients and activity, while a different problem, may be similarly mathematically possible, but impossible within human time frames and computer power.) Based on advances in computing hardware (e.g., graphic processing unit processing cores) and the AI techniques using causal inference described herein, the disclosed embodiments may enable the efficient solving of the task of generating candidate drug compounds at scale.

Causal inference may refer to a process, based on conditions of an occurrence of an effect, of drawing a conclusion about a causal connection. Causal inference may analyze a response of an effect variable when a cause is changed. Causation may be defined thusly: a variable X is a cause of Y if Y “listens” to X and determines its response based on what it “hears.” The process of causal inference in the field of AI may be particularly beneficial for generating and testing candidate drug compounds for certain diseases and/or medical conditions because of the use of what are termed counterfactuals. A counterfactual posits and examines conditions contrary to what has actually occurred in reality. For example, if someone takes aspirin for a headache, the headache may go away. The counterfactual asks what would have happened if the person had not taken aspirin, i.e., would the headache still have gone away or would it have remained or even gotten worse? Accordingly, counterfactuals may refer to calculating alternative scenarios based on past actions, occurrences, results, regressions, regression analyses, correlations, or some combination thereof. A counterfactual may enable determining whether a response should stay the same or instead change if something in a sequence does not occur. For example, one counterfactual may include asking: “Would a certain level of activity be the same if a certain ingredient is not included in a sequence of a candidate drug compound?”

By simulating numerous alternative scenarios to further optimize and hone the accuracy of a sequence of ingredients in the candidate drug compounds, such techniques may enable reducing the number of viable candidate drug compounds. As a result, the embodiments may provide technical benefits, such as reducing resources consumed (e.g., processing, memory, network bandwidth) by reducing a number of candidate drug compounds that may be considered for classification as a selected candidate drug compound by another machine learning model.

In some embodiments, one application for the AI engine to design, discover, develop, formulate, create, and/or test candidate drug compounds may pertain to peptide therapeutics. A peptide may refer to a compound consisting of two or more amino acids linked in a chain. Example peptides may include dipeptides, tripeptides, tetrapeptides, etc. Aa polypeptide may refer to a long, continuous, and unbranched peptide chain. Peptides may be simple to manufacture at discovery scale, include drug-like characteristics of small molecules, include safety and high specificity of biologics, and/or provide greater administration flexibility than some other biologics.

The disclosed techniques provide numerous benefits over conventional techniques for designing, developing, and/or testing candidate drug compounds. For example, the AI engine may efficiently use a biological context representation of a set of drug compounds and one or more machine learning models to generate a set of candidate drug compounds and classify one of the set of candidate drug compounds as a selected candidate drug compound. Some embodiments may use causal inference to remove one or more potential candidate drug compounds from classification, thereby reducing the computational complexity and processing burden of classifying a selected candidate drug compound.

In addition, benchmark analysis may be performed for each type of machine learning model that generates candidate drugs. The benchmark analysis may score various parameters of the machine learning models that generate the candidate drugs. The various parameters may refer to candidate drug novelty, candidate drug uniqueness, candidate drug similarity, candidate drug validity, etc. The scores may be used to recursively tune the machine learning models over time to cause one or more of the parameters to increase for the machine learning models. In some embodiments, some of the machine learning models may vary in their effectiveness as it pertains to some of the parameters. In addition, to generate subsequent candidate drug candidates, the benchmark analysis may score the candidate drug candidates generated by the machine learning models, rank the machine learning models that generate the highest scoring candidate drug candidates, and/or select the machine learning models producing the highest scoring candidate drug candidates.

Also, certain markets (e.g., anti-infective, animal, industrial, etc.) may prefer, based on a type of data those markets generate, to use certain machine learning models that generate high scores for a subset of parameters. Accordingly, in some embodiments, the subset of machine learning models that generate the high scores for the subset of parameters may be combined into a package and transmitted to a third party. That is, some embodiments enable custom tailoring of machine learning model packages for particular needs of third parties based on their data.

Further, additional benefits of the embodiments disclosed herein may include using the AI engine to produce algorithmically designed drug compounds that have been validated in vivo and in vitro and that provide (i) a broad-spectrum activity against greater than, e.g., 900 multi-drug resistant bacteria, (ii) at least, e.g., a 2-to-10 times improvement in exposure time required to generate a drug resistance profile, (iii) effectiveness across, e.g., four key animal infection models (both Gram-positive and Gram-negative bacteria), and/or (iv) effectiveness against, e.g., biofilms.

It should be noted that the embodiments disclosed herein may not only apply to the anti-infective market (e.g., for prosthetic joint infections, urinary tract infections, intra-abdominal or peritoneal infections, otitis media, cardiac infections, respiratory infections including but not limited to sequelae from diseases such as cystic fibrosis, neurological infections (e.g., meningitis), dental infections (including periodontal), other organ infections, digestive and intestinal infections (e.g., C. difficile), other physiological system infections, wound and soft tissue infections (e.g., cellulitis), etc.), but to numerous other suitable markets and/or industries. For example, the embodiments may be used in the animal health/veterinary industry, for example, to treat certain animal diseases (e.g., bovine mastitis). Also, the embodiments may be used for industrial applications, such as anti-biofouling, and/or generating optimized control action sequences for machinery. The embodiments may also benefit a market for new therapeutic indications, such as those for eczema, inflammatory bowel disease, Crohn's Disease, rheumatoid arthritis, asthma, auto-immune diseases and disease processes in general, inflammatory disease progressions or processes, and/or oncology treatments and palliatives. The video game industry may also benefit from the disclosed techniques to improve the AI used for generating sequences of decisions that non-player controlled (NPC) characters make during gameplay. The integrated circuit/chip industry may also benefit from the disclosed techniques to improve the mask works generation and routing processes used for generating the most efficient, highest performance, lowest power, lowest heat generating systems on a chip or solid state devices. Accordingly, it should be understood that the disclosed embodiments may benefit any market and/or industry associated with a sequence (e.g., items, objects, decisions, actions, ingredients, etc.) that can be optimized.

FIGS. 1A through 14, discussed below, and the various embodiments used to describe the principles of this disclosure are by way of illustration only and should not be construed in any way to limit the scope of the disclosure.

FIG. 1A illustrates a high-level component diagram of an illustrative system architecture 100 according to certain embodiments of this disclosure. In some embodiments, the system architecture 100 may include a computing device 102 communicatively coupled to a cloud-based computing system 116. Each of the computing device 102 and components included in the cloud-based computing system 116 may include one or more processing devices, memory devices, and/or network interface cards. The network interface cards may enable communication via a wireless protocol for transmitting data over short distances, such as Bluetooth, ZigBee, NFC, etc. Additionally, the network interface cards may enable communicating data over long distances, and in one example, the computing device 102 and the cloud-based computing system 116 may communicate with a network 112. Network 112 may be a public network (e.g., connected to the Internet via wired (Ethernet) or wireless (WiFi)), a private network (e.g., a local area network (LAN) or wide area network (WAN)), or a combination thereof. Network 112 may also comprise a node or nodes on the Internet of Things (IoT).

The computing device 102 may be any suitable computing device, such as a laptop, tablet, smartphone, or computer. The computing device 102 may include a display capable of presenting a user interface of an application 118. The application 118 may be implemented in computer instructions stored on the one or more memory devices of the computing device 102 and executable by the one or more processing devices of the computing device 102. The application 118 may present various screens to a user that present various views (e.g., topographical heatmaps) including measures, gradients, or levels of certain types of activity and optimized sequences of selected candidate drug compounds, information pertaining to the selected candidate drug compounds and/or other candidate drug compounds, options to modify the sequence of ingredients in the selected candidate drug compound, and so forth, as described in more detail below. The computing device 102 may also include instructions stored on the one or more memory devices that, when executed by the one or more processing devices of the computing device 102, perform operations of any of the methods described herein.

In some embodiments, the cloud-based computing system 116 may include one or more servers 128 that form a distributed computing architecture. The servers 128 may be a rackmount server, a router computer, a personal computer, a portable digital assistant, a mobile phone, a laptop computer, a tablet computer, a camera, a video camera, a netbook, a desktop computer, a media center, any other device capable of functioning as a server, or any combination of the above. Each of the servers 128 may include one or more processing devices, memory devices, data storage, and/or network interface cards. The servers 128 may be in communication with one another via any suitable communication protocol. The servers 128 may execute an artificial intelligence (AI) engine 140 that uses one or more machine learning models 132 to perform at least one of the embodiments disclosed herein. The cloud-based computing system 128 may also include a database 150 that stores data, knowledge, and data structures used to perform various embodiments. For example, the database 150 may store a knowledge graph containing the biological context representation described further below. Further, the database 150 may store generated candidate drug compounds, selected candidate drug compounds, information pertaining to the selected candidate drug compounds (e.g., activity for certain types of ingredients, sequences of ingredients, test results, correlations, semantic information, structural information, physical information, chemical information, etc.). Although depicted separately from the server 128, in some embodiments, the database 150 may be hosted on one or more of the servers 128.

In some embodiments the cloud-based computing system 116 may include a training engine 130 capable of generating the one or more machine learning models 132. The machine learning models 132 may be trained to discover, translate, design, generate, create, develop, classify, and/or test candidate drug compounds, among other things. The one or more machine learning models 132 may be generated by the training engine 130 and may be implemented in computer instructions executable by one or more processing devices of the training engine 130 and/or the servers 128. To generate the one or more machine learning models 132, the training engine 130 may train the one or more machine learning models 132. The one or more machine learning models 132 may be used by any of the modules in the AI engine 140 architecture depicted in FIG. 2.

The training engine 130 may be a rackmount server, a router computer, a personal computer, a portable digital assistant, a smartphone, a laptop computer, a tablet computer, a netbook, a desktop computer, an Internet of Things (IoT) device, any other desired computing device, or any combination of the above. The training engine 130 may be cloud-based, be a real-time software platform, include privacy software or protocols, and/or include security software or protocols.

To generate the one or more machine learning models 132, the training engine 130 may train the one or more machine learning models 132. The training engine 130 may use a base data set of biological context representation (e.g., physical properties data, peptide activity data, microbe data, antimicrobial data, anti-neurodegenerative compound data, pro-neuroplasticity compound data, clinical outcome data, etc.) for a set of drug compounds. For example, the biological context representation may include sequences of ingredients for the drug compounds. The results may include information indicating levels of certain types of activity associated with certain design spaces. In one embodiment, the results may include causal inference information pertaining to whether certain ingredients in the drug compounds are correlated with or determined by certain effects (e.g., activity levels) in the design space.

The one or more machine learning models 132 may refer to model artifacts created by the training engine 130 using training data that includes training inputs and corresponding target outputs. The training engine 130 may find patterns in the training data wherein such patterns map the training input to the target output and generate the machine learning models 132 that capture these patterns. Although depicted separately from the server 128, in some embodiments, the training engine 130 may reside on server 128. Further, in some embodiments, the artificial intelligence engine 140, the database 150, and/or the training engine 130 may reside on the computing device 102.

As described in more detail below, the one or more machine learning models 132 may comprise, e.g., a single level of linear or non-linear operations (e.g., a support vector machine [SVM]) or the machine learning models 132 may be a deep network, i.e., a machine learning model comprising multiple levels of non-linear operations. Examples of deep networks are neural networks, including generative adversarial networks, convolutional neural networks, recurrent neural networks with one or more hidden layers, and fully connected neural networks (e.g., each neuron may transmit its output signal to the input of the remaining neurons, as well as to itself). For example, the machine learning model may include numerous layers and/or hidden layers that perform calculations (e.g., dot products) using various neurons. In some embodiments, one or more of the machine learning models 132 may be trained to use causal inference and counterfactuals.

For example, the machine learning model 132 trained to use causal inference may accept one or more inputs, such as (i) assumptions, (ii) queries, and (iii) data. The machine learning model 132 may be trained to output one or more outputs, such as (i) a decision as to whether a query may be answered, (ii) an objective function (also referred to as an estimand) that provides an answer to the query for any received data, and (iii) an estimated answer to the query and an estimated uncertainty of the answer, where the estimated answer is based on the data and the objective function, and the estimated uncertainty reflects the quality of data (i.e., a measure which takes into account the degree and/or salience of incorrect data and/or missing data). The assumptions may also be referred to as constraints and may be simplified into statements used in the machine learning model 132. The queries may refer to scientific questions for which the answers are desired.

The answers estimated using causal inference by the machine learning model may include optimized sequences of ingredients in selected candidate drug compounds. As the machine learning model estimates answers (e.g., candidate drug compounds), certain causal diagrams may be generated, as well as logical statements, and patterns may be detected. For example, one pattern may indicate that “there is no path connecting ingredient D and activity P,” which may translate to a statistical statement “D and P are independent.” If alternative calculations using counterfactuals contradict or do not support that statistical statement, then the machine learning model 132 and/or the biological context representation may be updated. For example, another machine learning model 132 may be used to compute a degree of fitness which represents a degree to which the data is compatible with the assumptions used by the machine learning model that uses causal inference. There are certain techniques that may be employed by the other machine learning model 132 to reduce the uncertainty and increase the degree of compatibility. The techniques may include those for maximum likelihood, propensity scores, confidence indicators, and/or significance tests, among others.

Using causal inference, a generative adversarial network (GAN) may be used to generate a set of candidate drug compounds. A GAN refers to a class of deep learning algorithms including two neural networks, a generator and a discriminator, that both compete with one another to achieve a goal. For example, regarding candidate drug compound generation, the generator goal may include generating candidate drug compounds, including compatible/incompatible sequences of ingredients, and effective/ineffective sequences of ingredients, etc. that the discriminator classifies as feasible candidate drug compounds, including compatible and effective sequences of ingredients that may produce desired activity levels for a design space. In one embodiment, the generator may use causal inference, including counterfactuals, to calculate numerous alternative scenarios that indicate whether a certain result (e.g., activity level) still follows when any element or aspect of a sequence changes. For example, the generator may be a neural network based on Markov models (e.g., Deep Markov Models), which may perform causal inference. In some embodiments, one or more of the counterfactuals used during the causal inference may be determined and provided by the scientist module. The discriminator goal may include distinguishing candidate drug compounds which include undesirable sequences of ingredients from candidate drug compounds which include desirable sequences of ingredients.

In some embodiments, the generator initially generates candidate drug compounds and continues to generate better candidate drug compounds after each iteration until the generator eventually begins to generate candidate drug compounds that are valid drug compounds which produce certain levels of activity within a design space. A candidate drug compound may be “valid” when it produces a certain level of effectiveness (e.g., above a threshold activity level as determined by a standard (e.g., regulatory entity)) in a design space. In order to classify the candidate drug compounds as a valid drug compound or invalid candidate drug compound, the discriminator may receive real drug compound information from a dataset and the candidate drug compounds generated by the generator. “Real drug compound,” as used in this disclosure, may refer to a drug compound that has been approved by any regulatory (governmental) body or agency. The generator obtains the results from the discriminator and applies the results in order to generate better (e.g., valid) candidate drug compounds.

General details regarding the GAN are now discussed. The two neural networks, the generator and the discriminator, may be trained simultaneously. The discriminator may receive an input and then output a scalar indicating whether a candidate drug compound is an actual and/or viable drug compound. In some embodiments, the discriminator may resemble an energy function that outputs a low value (e.g., close to 0) when input is a valid drug compound and a positive value when the input is not a valid drug compound (e.g., if it includes an incorrect sequence of ingredients for certain activity levels pertaining to a design space).

There are two functions that may be used, the generator function (G(V)), and the discriminator function (D(Y)). The generator function may be denoted as G(V), where V is generally a vector randomly sampled in a standard distribution (e.g., Gaussian). The vector may be any suitable dimension and may be referred to as an embedding herein. The role of the generator is to produce candidate drug candidates to train the discriminator function (D(Y)) to output the values indicating the candidate drug candidate is valid (e.g., a low value).

During training, the discriminator is presented with a valid drug compound and adjusts its parameters (e.g., weights and biases) to output a value indicative of the validity of the candidate drug compounds that produce real activity levels in certain design spaces. Next, the discriminator may receive a modified candidate drug compound (e.g., modified using counterfactuals) generated by the generator and adjust its parameters to output a value indicative of whether the modified candidate drug compound provides the same or a different activity level in the design space.

The discriminator may use a gradient of an objective function to increase the value of the output. The discriminator may be trained as an unsupervised “density estimator,” i.e., a contrast function produces a low value for desired data (e.g., candidate drug compounds that include sequences producing desired levels of certain types of activity in a design space) and higher output for undesired data (e.g., candidate drug compounds that include sequences producing undesirable levels of certain types of activity in a design space). The generator may receive the gradient of the discriminator with respect to each modified candidate drug compound it produces. The generator uses the gradient to train itself to produce modified candidate drug compounds that the discriminator determines include sequences producing desired levels of certain types of activity in a design space.

Recurrent neural networks include the functionality, in the context of a hidden layer, to process information sequences and store information about previous computations. As such, recurrent neural networks may have or exhibit a “memory.” Recurrent neural networks may include connections between nodes that form a directed graph along a temporal sequence. Keeping and analyzing information about previous states enables recurrent neural networks to process sequences of inputs to recognize patterns (e.g., such as sequences of ingredients and correlations with certain types of activity level). Recurrent neural networks may be similar to Markov chains. For example, Markov chains may refer to stochastic models describing sequences of possible events in which the probability of any given event depends only on the state information contained in the previous event. Thus, Markov chains also use an internal memory to store at least the state of the previous event. These models may be useful in determining causal inference, such as whether an event at a current node changes as a result of the state of a previous node changing.

The set of candidate drug compounds generated may be input into another machine learning model 132 trained to classify of the set of candidate drug compounds as a selected candidate drug compound. The classifier may be trained to rank the set of candidate drug compounds using any suitable ranking (i.e., for example, non-parametric) technique. For example, in some embodiments, one or more clustering techniques may be used to cluster the set of candidate drug compounds. To classify the selected candidate drug compound, the machine learning model 132 may also perform objective optimization techniques while clustering. To classify the selected candidate drug compound having desired levels of certain types of activity, the objective optimization may include using a minimization and/or maximization function for each candidate drug compound in the clusters.

A cluster may refer to a group of data objects similar to one another within the same cluster, but dissimilar to the objects in the other clusters. Cluster analysis may be used to classify the data into relative groups (clusters). One example of clustering may include K-means clustering where “K” defines the number of clusters. Performing K-means clustering may comprise specifying the number of clusters, specifying the cluster seeds, assigning each point to a centroid, and adjusting the centroid.

Additional clustering techniques may include hierarchical clustering and density based spatial clustering. Hierarchy clustering may be used to identify the groups in the set of candidate drug compounds where there is no set number of clusters to be generated. As a result, a tree-based representation of the objects in the various groups may be generated. Density-based spatial clustering may be used to identify clusters of any shape in a dataset having noise and outliers. This form of clustering also does not require specifying the number of clusters to be generated.

FIG. 1B illustrates an architecture of the artificial intelligence engine according to certain embodiments of this disclosure. The architecture may include a biological context representation 200, a creator module 151, a descriptor module 152, a scientist module 153, a reinforcer module 154, and a conductor module 155. The architecture may provide a platform that improves its machine learning models over time by using benchmark analysis to produce enhanced candidate drug compounds for target design spaces. The platform may also continuously or continually learn new information from literature, clinical trials, studies, research, and/or any suitable data source about drug compounds. The newly learned information may be used to continuously or continually train the machine learning models to evolve with evolving information.

The biological context representation 200 may be implemented in a general manner such that it can be applied to solve different types of problems across different markets. The underlying structure of the biological context representation 200 may include nodes and relationships between the nodes. There may be semantic information, activity information, structural information, chemical information, pathway information, and so forth represented in the biological context representation 200. The biological context representation 200 may include any number of layers (e.g., five) layers of information. The first layer may pertain to molecular structure and physical property information, the second layer may pertain to molecule-to-molecule interactions, the third layer may pertain to molecule pathway interactions, the fourth layer may pertain to molecule cell profile associations, and the fifth layer may pertain to therapeutics (including those using biologics) and indications relevant for molecules. The biological context representation 200 is discussed further below with reference to FIGS. 2 and 5.

Further, to increase computing processing using various encodings, those various encodings may be selected to preferentially represent certain types of data. For example, to effectively capture common backbone structures of molecules, Morgan fingerprints may be used to describe physical properties of the candidate drug compounds. The encodings are discussed further below with reference to FIG. 1G.

Although just one creator module 151 is depicted, there may any suitable number of creator modules 151. Each of the creator modules 151 may include one or more generative machine learning models trained to generate new candidate drug compounds. The new candidate drug compounds are then added to the biological context representation 200. To that end, the term “creator module” and “generative model” may be used interchangeably herein. Each node in the biological context representation 200 may be a candidate drug compound (e.g., a peptide candidate).

The generative machine learning modules included in the creator module 151 may be of different types and perform different functions. The different types and different functions may include a variational autoencoder, structured transformer, Mini Batch Discriminator, dilation, self-attention, upsampling, loss, and the like. Each of these generative machine learning model types and functions is briefly explained below.

Regarding the variational autoencoder, it may simultaneously train two machine learning models, an inference model q_φ(z|x) and a generative model p_θ(x|z)p_θ(z) for data x and a latent variable z. In some embodiments, both the inference model and the generative model may be conditioned on a chosen attribute of the sequences. Both models may be jointly optimized using a tractable variational Bayesian approach which maximizes the evidence lower bound (ELBO) according to the following relationship:

E_{q_{θ}(z|x, a)}[log p_{θ}(x|z, a)]−KL(q_{θ}(x|z, a)∥p_θ(z))

This technique equates to minimizing reconstruction loss on x and a Kullback-Leibler (KL) divergence between the inference model and a prior p(z) usually characterized by an exponential family distribution (e.g., Gaussian).

Regarding the structured transformer, it may perform autoregressive decomposition to decompose the joint probability distribution of the sequence given the structure p=(s|x) autoregressively as:

p(s|x)=Π_ip(s_i|x_<i)

The conditional probability p(s_i|x_<i) of amino acid s_iat position i is conditioned on both the input structure x and the preceding amino acid s_iand the preceding amino acid s_<1={s₁, . . . , s_i−1}. These conditionals may be parameterized in terms of two sub-networks: an encoder that computes embeddings from structure-based features and edge features, and a decoder that autoregressively predicts amino acid letter s_igiven the preceding sequence and structural embeddings from the encoder.

Mode collapse occurs in generative adversarial networks when the generator generates a limited diversity of samples, or even the same sample, regardless of the input. To overcome mode collapse, some embodiments implement a Mini Batch Discriminator (MBD) approach. MBDs each work as an extra layer in the network that computes the standard deviation across the batch of examples (the batch contains only real drug compounds or only candidate drug compounds). If the batch contains a small variety of examples, the standard deviation will be low and the discriminator will be able to use this information to lower the score for each example in the batch. To further reduce mode collapse occurrence, some embodiments balance the sampling frequency of the training dataset clusters.

Regarding dilation, convolution filters may be capable of detecting local features, but they have limitations when it comes to relationships separated by long distances. Accordingly, some embodiments implement convolution filters with dilation. By introducing gaps into convolution kernels, such techniques increase the receptive field without increasing the number of parameters. Dilation rate may be applied to one convolution filter in each residual block of a generator and/or a discriminator. In this way, by the last layer of the generative adversarial network, filters may include a large enough receptive field to learn relationships separated by long-distances. Residual blocks are discussed further below with reference to FIG. 1F.

Regarding self-attention, different areas of a protein have different associations and effects on overall protein behavior. Accordingly, the architecture of the generative adversarial network disclosed herein implements a self-attention mechanism. The self-attention mechanism may include a number of layers that highlight different areas of importance across the entire sequence and allow the discriminator to determine whether parts in distant portions of the protein are consistent with each other.

Regarding upsampling, some embodiments implement techniques best suited for protein generation. For example, nearest-neighbor interpolation, transposed convolution, and sub-pixel shuffle may be used. Any combination of these techniques may be used in the upsampling layers. In some embodiments, transposed convolution by itself may be used for all upsampling layers.

Regarding the loss function, it is a component that aids in the successful performance of a neural network. Various losses, such as non-saturating, non-saturating with R1 regularization, hinge, hinge with relativistic average, and Wassertein and Wassertein with gradient penalty losses, may be used. In some embodiments, due to performance increases, the non-saturating loss with R1 regularization may be used for the generative adversarial network.

Details pertaining to the architecture of the creator module 151 are described below with reference to FIGS. 1C-1I.

The descriptor module 152 may include one or more machine learning models trained to generate descriptions for each of the candidate drug compounds generated by the creator module 151. The descriptor module 152 may be trained to use different encodings to represent the different types of information included in the candidate drug compound. The descriptor module 152 may populate the information in the candidate drug compound with ordinal values, cardinal values, categorical values, etc. depending on the type of information. For example, the descriptor module 152 may include a classifier that analyzes the candidate drug compound and determines whether it is a cancer peptide, an antimicrobial peptide, or a different peptide. The descriptor module 152 describes the structure and the physiochemical properties of the candidate drug compound.

The reinforcer module 154 may include one or more machine learning models trained to analyze, based on the descriptions, the structure and the physiochemical properties of the candidate drug compounds in the biological context representation 200. Based on the analysis, the reinforcer module 154 may identify a set of experiments to perform on the candidate drug compounds to elicit certain desired data (e.g., activity effectiveness, biomedical features, etc.). The identification may be performed by matching a pattern of the structure and physiochemical properties of the candidate drug compounds with the structure and physiochemical properties of other drug compounds and determining which experiments were performed on the other drug compounds to elicit desired data. The experiments may include in vitro or in vivo experiments. Further, the reinforcer module 154 may identify experiments that should not be performed for the candidate drug compounds if a determination is made that those experiments yield useless data for drug compounds.

The conductor module 155 may include one or more machine learning models trained to perform inference queries on the data stored in the biological context representation 200. The inference queries may pertain to performing queries to improve the quality of the data in the biological context representation 200. For example, there may be a gap in data in one of the nodes (e.g., candidate drug compounds) stored in the biological context representation 200. An inference query refers to the process of identifying a first node and a second node similar to the first node, and to obtaining data from the second node to fill a data gap in the first node. An inference query may be executed to search for another node having similarities to the node with the gap and may fill the gap with the data from the another node.

The scientist module 153 may include one or more machine learning models trained to perform benchmark analysis to evaluate various parameters of the creator module 151. In some embodiments, the scientist module 153 may generate scores for the candidate compound drugs generated by the creator module 151. The benchmark analysis may be used to electronically and recursively optimize the creator module 151 to generate candidate drug compounds having improved scores in subsequent generation rounds. There may be several types of benchmarks (e.g., distribution learning benchmarks, goal-directed benchmarks, etc.) used by the scientist module 153 to evaluate generative machine learning models used by the creator module 151. As described herein, one or more parameters (e.g., validity, uniqueness, novelty, Frechet ChemNet Distance (FCD), internal diversity, Kullback-Leiblert (KL) divergence, similarity, rediscovery, isomer capability, median compounds, etc.) of the creator module 151 may be scored during benchmark analysis. The benchmark analysis may also be used to electronically and recursively optimize the creator module 151 to improve scores of the parameters in subsequent generation rounds. Any combination of the benchmarks described below may be used to evaluate the creator module 151.

One type of benchmark used by the scientist module 153 may include a distribution learning benchmark. The distribution learning benchmark evaluates, when given a set of molecules, how well the creator module 151 generates new molecules which follow the same chemical distribution. For example, when provided with therapeutic peptides, the distribution learning benchmark evaluates how well the creator module 151 generates other therapeutic peptides having similar chemical distributions.

The distribution learning benchmark may include generating a score for an ability of the creator module 151 to generate valid candidate drug compounds, a score for an ability of the creator module 151 to generate unique candidate drug compounds, a score for an ability of the creator module 151 to generate novel candidate drug compounds, a Frechet ChemNet Distance (FCD) score for the creator module 151, an internal diversity score for the creator module 151, a KL divergence score for the creator module 151, and so forth. Each of the distribution learning benchmarks is now discussed.

The validity score may be determined as a ratio of valid candidate drug compounds to non-valid candidate drug compounds of generated candidate drug compounds. In some embodiments, the ratio may be determined from a certain number (e.g., 10,000) candidate drug compounds. In some embodiments, candidate drug compounds may be considered valid if their representation (e.g., simplified molecular-input line-entry system (SMILES)) can be successfully parsed using any suitable parser.

The uniqueness score may be determined by sampling candidate drug compounds generated by the creator module 151 until a certain number (e.g., 10,000) of valid molecules are identified by identical representations (e.g., canonical SMILES strings). The uniqueness score may be determined as the number of different representations divided by the certain number (e.g., 10,000).

The novelty score may be determined by generating candidate drug compounds until a certain number (e.g., 10,000) of different representations (e.g., canonical SMILES strings) are obtained and computing the ratio of candidate drug compounds (including real drug compounds) not present in the training dataset.

The Frechet ChemNet Distance (FCD) score may be determined by selecting a random subset of a certain number (e.g., 10,000) of drug compounds from the training dataset, and generating candidate drug compounds using the creator module 151 until a certain number (10,000) of valid candidate drug compounds are obtained. The FCD between the subset of the drug compounds and the candidate drug compounds may be determined. The FCD may consider chemically and biologically relevant information about drug compounds, and also measure the diversity of the set via the distribution of generated candidate drug compounds. The FCD may detect if generated candidate drug compounds are diverse, and the FCD may detect if generated candidate drug compounds have similar chemical and biological properties as real drug compounds. The FCD score (“S”) is determined using the following relationship: S=exp(−0.2*FCD).

The internal diversity score may assess the chemical diversity within a set of generated candidate drug compounds (“GROUP”). The internal diversity score may be determined using the following relationship:

${IntDiv}_{p} (G) = 1 - \sqrt[p]{\frac{1}{{\langle G \rangle}^{2}} \sum_{{m_{1}, m_{2} inG}} {T (m_{1}, m_{2})}^{p}}$

Where T(m₁, m₂) is the Tanimoto Similarity (SNN) between molecule 1, m₁, and molecule 2, m₂. While SNN measures the dissimilarity to external diversity, the internal diversity score may consider dissimilarity between generated candidate drug compounds. The internal diversity score may be used to detect mode collapse in certain generative models. For example, mode collapse may occur when the generative model produces a limited variety of candidate drug compounds while ignoring some areas of a design space. A higher score for the internal diversity corresponds to higher diversity in the set of candidate drug compounds generated.

The KL divergence score may be determined by calculating physiochemical descriptors for both the candidate drug compounds and the real drug compounds. Further, a determination may be made of the distribution of maximum nearest neighbor similarities on fingerprints (e.g., extended connectivity fingerprint up to four bonds (ECFP4)) for both the candidate drug compounds and the real drug compounds. The distribution of these descriptors may be determined via kernel density estimation for continuous descriptors, or as a histogram for discrete descriptors. The KL divergence $D {KL,i}$ may be determined for each descriptor $i$, and is aggregated to determine the KL divergence score $S$ via:

$$S=frac{1}{k}sum_i{circumflex over ( )}k exp(−D_{KL,i})$$

Where $k$ is the number of descriptors (e.g., $k=9$).

The isomer capability score may be determined by whether molecules may be generated that correspond to a target molecular formula (for example C7H8N2O2). The isomers for a given molecular formula can in principle be enumerated, but except for small molecules this number will in general be very large. The isomer capability score represents fully-determined tasks that assess the flexibility of the creator module to generate molecules following a simple pattern (which is a priori unknown).

A second type of benchmark may include a goal-directed benchmark. The goal-direct benchmark may evaluate whether the creator module 151 generates a best possible candidate drug compound to satisfy a pre-defined goal (e.g., activity level in a design space). A resulting benchmark score may be calculated as a weighted average of the candidate drug compound scores. In some embodiments, the candidate drug compounds with the best benchmark scores may be assigned a larger weight. As such, generative models of the creator module 151 may be tuned to deliver a few candidate drug compounds with top scores, while also generating candidate drug compounds with satisfactory scores. For each of the goal-directed benchmarks, one or several average scores may be determined for the given number of top candidate drug compounds and then the resulting benchmark score may be calculated as the mean of these average scores. For example, the resulting benchmark score may be a combination of the top-1, top-10, and top-100 scores, in which the resulting benchmark score is determined by the following relationship:

${IntDiv}_{p} (G) = 1 - \sqrt[p]{\frac{1}{{\langle G \rangle}^{2}} \sum_{?} {T (m_{1}, m_{2})}^{p}}$

$? indicates text missing or illegible when filed$

Where s is an n-dimensional (e.g., 100-dimensional) vector of candidate drug compound scores s_v1≤i≤100 sorted in decreasing order (e.g., s_i≥s_jfor i<j).

The goal-directed benchmark may include generating a score for an ability of the creator module 151 to generate candidate drug compounds similar to a real drug compound, a score for an ability of the creator module 151 to rediscover the potential viability of previously-known drug compounds (e.g., using a drug which is prescribed for certain conditions for a new condition or disease), and the like.

The similarity score may be determined using nearest neighbor scoring, fragment similarity scoring, scaffold similarity scoring, SMARTS scoring, and the like. Nearest neighbor scoring (e.g., nss(G, R)) may refer to a scoring function that determines the similarity of the candidate drug compound to a target real drug compound $g$. The score corresponds to the Tanimoto similarity when considering the fingerprint $r$ and may be determined by the following relationship:

$$NNS(G,R)=frac{1}{|G|}sum_{m_Gin G} max; T(m_G, m_R)$$

Where $m_R$ and $m_G$ are representations of the real drug compounds (R) and the candidate drug compounds (G) as bit strings (e.g., digital fingerprints, e.g., outputs of hash functions, etc.). The resulting score reflects how similar candidate drug compounds are to real drug compounds in terms of chemical structures encoded in these fingerprints. In some embodiments, Morgan fingerprints may be used with a radius of a configurable value (e.g., 2) and an encoding with a configurable number of bits (e.g., 1024). The radius and encoding bits may be configured to produce desirable results in a biochemical space.

The similarity score may be determined using fragment similarity scoring, which itself may be defined as the cosine distance between vectors of fragment frequencies. For a set of candidate drug compounds ($G$), its fragment frequency vector $f_G$ has a size equal to the size of all chemical fragments in the dataset, and elements of $f_G$ represent frequencies with which the corresponding fragments appear in $G$. The distance is determined by the following relationship:

$$Frag(G,R)=1−cos(f_G, f_R)$$

Candidate drug compounds and real drug compounds may be fragmented using any suitable decomposition algorithm. The fragment similarity scoring score represents the similarity of the set of candidate drug compounds and the set of real drug compounds at the level of chemical fragments.

The similarity score may be determined using scaffold similarity scoring, which may be determined in a similar way to the fragment similarity scoring. For example, the scaffold similarity scoring may be determined as a cosine similarity between the vectors $s_G$ and $s_R$ that represent frequencies of scaffolds in a set of candidate drug compounds ($G$) and a set of real drug compound ($R$). The scaffold similarity scoring score may be determined by the following relationship:

$$Frag(G,R)=1−cos(s_G, s_R)$$.

The similarity score may be determined using SMARTS scoring. SMARTS scoring may be implemented according to the relationship: SMART (a, b). The SMARTS scoring may evaluate whether the SMARTS pattern $s$ is present in a candidate drug compound. $b$ is a Boolean value indicating whether the SMARTS pattern should be present (true) or absent (false). When the pattern is desired, a score of 1, for true, is returned if the SMARTS pattern is found. If the pattern is not found, then a score of 0, for false, is returned.

In some embodiments, a goal-directed benchmark may include determining a rediscovery score for the creator module 151. In some embodiments, certain real drug compounds may be removed from the training dataset and the creator module 151 may be retrained using the modified training set lacking the removed real drug compounds. If the creator module 151 is able to generate (“rediscover”) a candidate drug compound that is identical or substantially similar to the removed real drug compounds, then a high rediscovery score may be assigned. Such a technique may be used to validate the creator module 151 is effectively trained and/or tuned.

Various modifiers may be used to modify the scores for the various benchmarks discussed above. For example, a Gaussian modifier may be implemented to target a specific value of some property, while giving high scores when the underlying value is close to the target. It may be adjustable as desired. A minimum Gaussian modifier may correspond to the right half of a Gaussian function and values smaller than a threshold may be given a full score, while values larger than the threshold decrease continuously to zero. A maximum Gaussian modifier may correspond to a left half of the Gaussian function and values larger than the threshold are given a full score, while values smaller than the threshold decrease continuously to zero. A threshold modifier may attribute a full score to values above a given threshold, while values smaller than the threshold decrease linearly to zero.

There are a variety of competing generative models that may be used to evaluate the performance of the creator module 151. For example, the competing generative models may include a random sampling, best of dataset method, SMILES genetic algorithm (GA), graph GA, graph Monte-Carlo tree search (MCTS), SMILES long short-term memory (LSTM), character-level recurrent neural networks (CharRNN), variational autoencoder, adversarial autoencoder, Latent generative adversarial network (LatentGAN), junction tree variational autoencoder (JT-VAE), and objective-reinforced generative adversarial network (ORGAN). Each of these competing generative models will now be discussed briefly.

Regarding random sampling, this baseline samples at random the requested number of molecules (candidate drug compounds) for the dataset. Random sampling may provide a lower bound for the goal-directed benchmarks, because no optimization is performed to obtain the returned molecules. Random sampling may provide an upper bound for the distribution learning benchmarks, because the molecules returned may be taken directly for the original distribution.

Regarding best of dataset method (or “best of dataset” herein), one goal of de novo molecular design is to explore unknown parts of the biochemical space, generating new candidate drug compounds with better properties than the drug compounds already known. The best of dataset scores the entire generated dataset including the candidate drug compounds with a provided scoring function and returns the highest scoring molecules. This effectively provides a lower bound for the goal-directed benchmarks that enables the creator module 151 to create better candidate drug compounds than the real and/or candidate drug compounds provided.

Regarding SMILES GA, this technique may evolve string molecular representations using mutations exploiting the SMILES context-free grammar. For each goal-directed benchmark, a certain number (e.g., 300) of highest scoring molecules in the dataset may be selected as an initial population. In this example, each molecule is represented by 300 genes. During each epoch an offspring of a certain number (e.g., 600) of new molecules may be generated by randomly mutating the population molecules. After deduplication and scoring, these new molecules may be merged with the current population and a new generation is chosen by selecting the top scoring molecules overall. This process may be repeated a certain number of times (e.g., 1000) or until progress has stopped for a certain number (e.g., 5) of consecutive epochs. Distribution-learning benchmarks do not apply to this baseline.

Regarding graph GA, this GA involves molecule evolution at the graph level. For each goal-directed benchmark a certain number (e.g., 100) of highest scoring molecules in the dataset are selected as the initial population. During each epoch, a mating pool of a certain number (e.g., 200) of molecules is sampled with replacement from the population, using scores as weights. This pool may contain many repeated molecules if their score is high. A new population of a certain number (e.g., 100) is then generated by iteratively choosing two molecules at random from the mating pool and applying a crossover operation. With probability of, e.g., 0.5 (i.e., 100/200), a mutation is also applied to the offspring molecule. This process is repeated a certain number (e.g., 1000) of times or until progress has stopped for a certain number (e.g., 5) of consecutive epochs. Distribution-learning benchmarks do not apply to this baseline.

Regarding graph MCTS, the statistics used during sampling may be computed on the training dataset. For this baseline, no initial population is selected for the goal-directed benchmarks. Each new molecule may be generated by running a certain number (e.g., 40) of simulations, starting from a base molecule. At each step, a certain number (e.g., 25) of children are considered and the sampling stops when reaching a certain number (e.g., 60) of atoms. The best-scoring molecule found during the sampling may be returned. A population of a certain number (e.g., 100) of molecules is generated at each epoch. This process may be repeated a certain number (e.g., 1000) of times or until progress has stopped for a certain number (e.g., 5) of consecutive epochs. For the distribution learning benchmark. the generation starts from a base molecule and a new molecule is generated with the same parameters. As for the goal-directed benchmarks, the only difference is that no scoring function is provided, so the first molecule to reach terminal state is returned instead of the highest scoring molecule.

Regarding SMILES LSTM, the technique is a baseline model, consisting of a LSTM neural network which predicts the next character of partial SMILES strings. In some embodiments, a SMILES LSTM may be used with 3 layers of hidden size of 1024. For the goal-directed benchmarks, a certain number (e.g., 20) of iterations of hill-climbing may be performed; at each step the model generated a certain number (e.g., 8192) of molecules and a certain number (e.g., 1024) of the top scoring molecules may be used to fine-tune the model parameters. For the distribution-learning benchmark, the model may generate the requested number of molecules.

Regarding character-level recurrent neural networks (CharRNN), the technique treats the task of generating SMILES as a language model attempting to learn the statistical structure of SMILES syntax by training it on a large corpus of SMILES. The CharRNN parameters may be optimized using maximum likelihood estimation (MLE). CharRNN may be implemented using LSTM RNN cells stacked into three layers with hidden dimension 600 each. To prevent overfitting, a dropout layer may be added between intermediate layers with dropout probability of 0.2. Training may be performed with a batch size of a certain number (e.g., 64) using an optimizer.

Regarding a variational autoencoder (VAE), it is a framework for training two neural networks, an encoder and a decoder, to learn a mapping from a higher-dimensional data representation (e.g., vector) into a lower-dimensional data representation and from the lower-dimensional data representation back to the higher-dimensional data representation. The lower-dimensional space is called the latent space, which is often a continuous vector space with normally distributed latent representation. The latent representation of our data may contain all the important information needed to represent an original data point. The latent representation represents the features of the original data point. In other words, one or more machine learning models may learn the data features of the original data point and simplifies its representation to make it more efficient to analyze. VAE parameters may be optimized to encode and decode data by minimizing the reconstruction loss while also minimizing a KL-divergence term arising from the variational approximation, such that the KL-divergence term may loosely be interpreted as a regularization term. Since molecules are discrete objects, properly trained VAE defines an invertible continuous representation of a molecule.

In some embodiments, aspects from both implementations may be combined. The encoder may implement a bidirectional Gated Recurrent Unit (GRU) with a linear output layer. The decoder may be a 3-layer GRU RNN of 512 hidden dimensions with intermediate dropout layers, the layers having a dropout probability of 0.2. Training may be performed with a batch size of a certain number (e.g., 128), utilizing a gradient clipping of 50 and a KL-term weight of 1, and further optimized with a learning rate of 0.0003 across 50 epochs. Other training parameters may be used to perform the embodiments disclosed herein.

Regarding adversarial autoencoders (AAE), they combine the idea of VAE with that of adversarial training as found in a GAN. In AAE, the KL divergence term is avoided by training a discriminator network to predict whether a given sample came from the latent space of the AE or from a prior distribution of the autoencoder (AE). Parameters may be optimized to minimize the reconstruction loss and to minimize the discriminator loss. The AAE model may consist of an encoder with a 1-layer bidirectional LSTM with 380 hidden dimensions, a decoder with a 2-layer LSTM with 640 hidden dimensions and a shared embedding of size 32. The latent space is of 640 dimensions, and the discriminator networks is a 2-layer fully connected neural network with 640 and 256 nodes respectively, utilizing the ELU activation function. Training may be performed with a batch size of 128, with an optimizer using a learning rate of 0.001 across 25 epochs. Other training parameters may be used to perform the embodiments disclosed herein.

Regarding LatentGAN, the technique encodes SMILES strings into latent vector representations of size 512. A Wasserstein Generative Adversarial network with Gradient Penalty may be trained to generate latent vectors resembling that of the training set, which are then decoded using a heteroencoder.

Regarding a junction tree variational autoencoder (JT-VAE), the model generates molecular graphs in two phases. The model first generates a tree-structured scaffold over chemical substructures, and then combines them into a molecule with a graph message passing network. This approach enables incrementally expanding molecules while maintaining chemical validity at every step.

Regarding an objective-reinforced generative adversarial network (ORGAN), the model is a sequence-generation model based on adversarial training that aims at generating discrete sequences that emulate a data distribution while using reinforcement learning to bias the generation process towards some desired objective rewards. ORGAN incorporates at least 2 networks: a generator network and a discriminator network. The goal of the generator network is to create candidate drug compounds indistinguishable from the empirical data distribution of real drug compounds. The discriminator exists to learn to distinguish a candidate drug compound from real data samples. Both models are trained in alternation.

To properly train a GAN, the gradient must be back-propagated between the generator and discriminator networks. Reinforcement uses an N-depth Monte Carlo tree search, and the reward is a weighted sum of probabilities from the discriminator and objective reward. Both the generator and discriminator may be pre-trained for 250 and 50 epochs, respectively, and then jointly trained for 100 epochs utilizing an optimizer with a learning rate of 0.0001. The learning rate may refer to a hyperparameter of a neural network, and the learning rate may be a number that determines an amount of change (e.g., weights, hidden layers, etc.) to make to a machine learning model in response to an estimated error. Bayesian optimization may be used to determine the optimal learning rate during training of a particular neural network. In some embodiments, validity and uniqueness of candidate drug compounds may be used as rewards.

The scientist module 153 may also include one or more machine learning models trained to perform causal inference using counterfactuals. The causal inference, as described herein, may be used to determine whether the creator module 151 actually generated a candidate drug candidate, including a desired activity in such candidate, or if it was determined because of noisy data (e.g., scarce, incorrect, etc. data).

FIG. 1C illustrates first components of an architecture of the creator module 151 according to certain embodiments of this disclosure. A candidate design space 156 and data 157 may be included in the biological context representation 200, such space 156 and data 157 to include the various sequences of the candidate drug compounds and/or real drug compounds. In some embodiments, the creator module 151 may populate the candidate design space 156. The candidate design space 156 may include a vast amount of information retrieved from numerous sources and/or generated by the AI engine 140. The candidate design space 156 may include information pertaining to antimicrobial peptides, anticancer peptides, peptidomimetics, uProteins and aCRFs, non-ribosomal peptides, and general peptides that are retrieved via genomic screening, literature research, and/or computationally designed using the AI engine 140. The candidate design space 156 may be updated each time the creator module 151 generates a new candidate drug compound. The candidate design space 156 may also be updated continuously or continually as new literature is published and/or genomic screenings are performed.

The creator module 151 may also use data 157 to generate the candidate drug compounds. In some embodiments, the data 157 may be generated and/or provided by the descriptor module 152. In some embodiments, the data may be received from any suitable source. The data may include molecular information pertaining to chemistry/biochemistry, targets, networks, cells, clinical trials, market (e.g., analysis, results, etc.) that result from performing simulations and/or experiments.

The creator module 151 may encode the candidate design space 156 and the data 157 into various encodings. In some embodiments, an attention message-passing neural network may be used to encode molecular graphs. An initial set of states may be constructed, one for each node in a molecular graph. Then, each node may be allowed to exchange information, to “message” with its neighboring nodes. Each message may be a vector describing an atom of a molecule from the atom's perspective in the molecule. After one such step, each node state will contain an awareness of its immediate neighborhood. Repeating the step makes each node aware of its second-order neighborhood, and so forth. During the message-passing stage and based on the total number of occurrences of a message, an attention layer may be used to identify interesting features of a molecule. A certain weight (e.g., heavy, light) may be assigned to a message that occurs more or fewer than a threshold number of times, thereby causing that message to stand out more when the messages are aggregated. For example, a message that occurs a very few amount of times (e.g., less than a threshold) may be more likely to include a desirable feature as opposed to a message that occurs a large number of times. In another example, a message that occurs more than a threshold number of times may be weighted more heavily than a message that occurs fewer than the threshold number of times. Any suitable weighting may be configured to cause a message to stand out more.

Using a summation function to reduce the size of the messages and increase computational efficiency, the attention mechanism may aggregate the messages with their weights. In such a way, the techniques are able to scale to remain computationally efficient as the number of messages increases. Such a technique may be beneficial because it reduces resource (e.g., processing, memory) consumption when performing computations with a large design space, including information in that design space pertaining to structure, semantic, sequence, physiochemical properties, etc.

After a chosen number of “messaging rounds”, all the context-aware node states are collected and converted to a summary representing the whole graph. All the transformations in the steps above may be carried out with machine learning models (e.g., neural networks), yielding a machine learning model that can be trained with known techniques to optimize the summary representation for the current task. The following relationships may be used by the attention message-passing neural network:

$1. Message Passing$

$m_{v}^{(t)} = A_{t} (h_{v}^{(t)}, S_{v}^{(t)}), where S_{v}^{(t)} = {(h_{w}^{(t)}, e_{vw}) | w \in N (v)}$

$A_{t} (h_{v}^{(t)}, {(h_{w}^{(t)}, e_{vw})}) = \sum_{w \in N (v)} f_{NN}^{(e_{vw})} (h_{w}^{(t)}) ⊙ \frac{\exp (? (?))}{\sum \exp (? (?))}$

$2. Node Update$

$h_{v}^{(t + 1)} = u_{t} (h_{v}^{(t)}, m_{v}^{(t)})$

$3. Readout$

$\hat{y} = R ({h_{v}^{(K)} | e \in G})$

$? indicates text missing or illegible when filed$

m^(t)_vis the message function, A_tis the attention function, U_tis the node update function, N(v) is the set of neighbors of node v in graph G, h^(t)_vis the hidden state of node v at time t, and m^(t)_vis a corresponding message vector. For each atom v, messages will be passed from its neighbors and aggregated as the message vector m^(t)from its surrounding environment. Then the hidden state h^(t)_vis updated by the message vector.

y{circumflex over ( )} is a resulting fixed-length feature vector generated for the graph, and R is a readout function invariant to node ordering, a feature allowing the MPNN framework to be invariant to graph isomorphism. The graph feature vector y{circumflex over ( )} then is passed to a fully connected layer to give prediction. All functions M_t, U_t, and R are neural networks and their weights are learned during training.

As depicted, a “Candidates Only Data” encoding 158 may encode just the information from the candidate design space, a “Candidates and Simulated Data” encoding 159 may encode information from the candidate design space 156 and the simulated data from the data 157, and a “Candidates with All Data” encoding 160 may encode information from the candidate design space 156 and both the simulated and experimental data from the data 157. Further, a “Heterologous Networks” encoding 161 may be generated using the “Candidates with All Data” encoding 160. The encodings 158, 159, 160, and 161 may include information pertaining to molecular structure, physiochemical properties, semantics, and so forth.

Each of the encodings 158, 159, 160, and 161 may be input into a separate machine learning model trained to generate an embedding. ML Model A, ML Model B, ML Model C, and ML Model D may be included in a “Single Candidate Embedding” Layer.

“Candidates Only Data” encoding 158 may be input into ML Model A, which outputs a “Candidate Embedding” 162. “Candidates and Simulated Data” encoding 159 may be input into ML Model B, which outputs a “Candidate and Simulated Data Embedding” 163. “Candidates with All Data” encoding 160 may be input into ML Model C, which outputs “Candidate with All Data Embedding” 164. “Heterologous Networks” encoding 161 may be input into ML Model D, which outputs “Graph and Network Embedding” 165. The embeddings 162, 163, 164, and 165 may represent information pertaining to a single candidate drug compound.

FIG. 1D illustrates second components of the architecture of the creator module 151 according to certain embodiments of this disclosure. As depicted, the encodings 158, 159, 160, and 161 are input into ML Model F, which is trained to output a candidate drug compound based on the encodings 158, 159, 160, and 161.

The embeddings 162, 163, 164, and 165 are input into ML Model G, which is trained to output a candidate drug compound based on the embeddings 162, 163, 164, and 165. In some embodiments, the “Heterologous Networks” 161 may be input into ML Model I, which is trained to output a candidate drug compound based on the “Heterologous Networks” 161. The embeddings 162, 163, 164, and 165 are also input into ML Model E in a “Knowledge Landscape Embedding” layer 167. The ML Model E is trained to output a “Latent Representation” based on the embeddings 162, 163, 164, and 165.

The “Latent Representation” 168 may include an “Activity Landscape” 169 and a “Continuous Representation” 170. The “Continuous Representation” 170 may include information (e.g., structural, semantic, etc.) pertaining to all of the molecules (e.g., real drug compounds and candidate drug compounds), and the “Activity Landscape” 169 may include activity information for all of the molecules. In some embodiments, the ML Model E may be a variational autoencoder that receives the embeddings 162, 163, 164, and 165 and outputs lower-dimensional embeddings that are machine-readable and less computationally expensive for processing. The lower-dimensional embeddings may be used to generate the “Latent Representation” 168. An architecture of the variational autoencoder is described further below with reference to FIG. 1E.

The “Latent Representation” 168 is input into the ML Model H. ML Model H may be any suitable type of machine learning model described herein. ML Model H may be trained to analyze the “Latent Representation” 168 and generate a candidate drug compound. The “Latent Representation” 168 may include multiple dimensions (e.g., tens, hundreds, thousands) and may have a particular shape. The shape may be rectangular, cube, cuboid, spherical, an amorphous blob, conical, or any suitable shape having any number of dimensions. The ML Model H may be a generative adversarial network, as described herein. The ML Model H may determine a shape of the “Latent Representation” 168, and may determine an area of the shape from which to obtain a slice based on “interesting” aspects of that area. An interesting aspect may be a peak, valley, a flat portion, or any combination thereof. The ML Model H may use an attention mechanism to determine what is “interesting” and what is not. The interesting aspect may be indicative of a desirable feature, such as a desirable activity for a particular disease or medical condition. The slice may include a combination of a portion of any of the information included in the “Latent Representation” 168, such as the structural information, physiochemical properties, semantic information, and so forth. The information included in the slice may be represented as an eigenvector that includes any number of dimensions from the “Latent Representation” 168. The term “slice” and “candidate drug compound” may be used interchangeably. The slice may be visually presented on a display screen, as shown in FIG. 8A.

A decoder may be used to transform the slice from the lower-dimensional vector to a higher-dimensional vector, which may be analyzed to determine what information is included in that slice. For example, the decoder may obtain a set of coordinates from the higher-dimensional vector which may be back-calculated to determine what information (e.g., structural, physiochemical, semantic, etc.) they represent.

Each of the candidate drug compounds generated by the ML Model F, ML Model G, ML Model H, and ML Model I may be ranked and one of the candidate drug compounds may be classified as a selected candidate drug compound, as described herein. Further, the candidate drug compounds may be input into one or more machine learning models trained to perform benchmark analysis, as described herein. Based on the benchmark analysis, any of the machine learning models in the creator module 151 may be optimized (e.g., tuning weights, adding or removing hidden layers, changing an activation function, etc.) to modify a parameter (e.g., uniqueness, validity, novelty, etc.) score for the machine learning models when generating subsequent candidate drug compounds.

FIG. 1E illustrates an architecture of a variational autoencoder machine learning model according to certain embodiments of this disclosure. In some embodiments, the variational autoencoder may include an input layer, an encoder layer, a latent layer, a decoder layer, and an output layer. The input layer may receive fingerprints of drug compounds and/or candidate drug compounds represented as higher-dimensional vectors, as well as associated drug concentration(s). The encoder layer may include one or more hidden layers, activation functions, and the like. The encoder layer may receive the fingerprint and drug concentration from the input layer and may perform operations to translate the higher-dimensional vectors into lower-dimensional vectors, as described herein. The latent layer may receive the lower-dimensional vectors and represent them in the “Latent Representation” 168. The latent layer may input the “Latent Representation” 168 into the ML Model H, which is a generative adversarial network including a generator and a discriminator, as described herein. The architecture of the generator and the discriminator is discussed further below with reference to FIG. 1F. The generator generates candidate drug compounds and the discriminator analyzes the candidate drug compounds to determine whether they are valid or not.

The candidate drug compounds output by the latent layer may be input into the decoder layer where the lower-dimensional vectors are translated back into the higher-dimensional vectors. The decoder layer may include one or more hidden layers, activation functions, and the like. The decoder layer may output the fingerprints and the drug concentration. The output fingerprint and drug concentration may be analyzed to determine how closely they match the input fingerprint and drug concentration. If the output and input substantially match, the variational autoencoder may be properly trained. If the output and the input do not substantially match, one or more layers of the variational autoencoder may be tuned (e.g., modify weights, add or remove hidden layers).

FIG. 1F illustrates an architecture of a generative adversarial network used to generate candidate drugs according to certain embodiments of this disclosure. As depicted, there is an architecture for the discriminator, discriminator residual block, generator, and generator residual block.

The discriminator architecture may receive a sequence (e.g., candidate drug compound) as an input. The discriminator architecture may include an arrangement of blocks in a particular order that improves computational efficiency when processing the sequence to determine whether the sequence is valid or not. For example, the particular order of blocks includes a first residual block, a self-attention block, a second residual block, a third residual block, a fourth residual block, a fifth residual block, and a sixth residual block. The discriminator may output a score (e.g., 0 or 1) for whether the received sequence is valid or not.

The discriminator residual block architecture may receive an input filtered into two processing pathways. A first processing pathway performs a conversion operation on the input. The second processing pathway performs several operations, including a conversion, a batch normalization operation, a leaky ReLu operation, a conversion operation, and another batch normalization operation. The output from the first and second processing pathways is summed and then output.

The generator architecture may receive a noise (e.g., biological context representation 200) as an input. The generator architecture may include an arrangement of blocks in a particular order that improves computational efficiency when processing the noise to generate a sequence (e.g., candidate drug compound). For example, the particular order of blocks includes a first residual block, a second residual block, a third residual block, a fourth residual block, a fifth residual block, a self-attention block, and a sixth residual block. The generator may output a score (e.g., 0 or 1) for whether the received sequence is valid or not.

The generator residual block architecture may receive an input filtered into two processing pathways. A first processing pathway performs a de-conversion operation on the input. The second processing pathway performs several operations, including a conversion, a batch normalization operation, a leaky ReLu operation, a de-conversion operation, and another batch normalization operation. The output from the first and second processing pathways is summed and then output.

FIG. 1G illustrates types of encodings to represent certain types of drug information according to certain embodiments of this disclosure. A table 180 includes three columns labeled “Encoding”, “Compressed?”, and “Information”. The “Encoding” column includes rows storing a type of encoding used to represent a certain type of information; the “Compressed?” column includes rows storing an indication of whether the encoding in that row is compressed; and the “Information” column includes rows storing a type of information represented by the encoding in each respective row. The descriptor module 152 may include a machine learning module trained to analyze a candidate drug compound and identify various structural properties, physiochemical properties, and the like. The descriptor module 152 may be trained to represent the type of structural and physiochemical properties using an encoding that increases computational efficiency and to store a description including the encodings at a node representing the candidate drug compound. During processing, the encodings may be aggregated for each candidate drug compound.

For example, using an alphanumeric string, SMILES encoding spells out molecular structure from a beginning portion to an ending portion. Morgan Fingerprints are useful for temporal molecular structures and the descriptor module 152 may include a machine learning module trained to output a compressed vector. Morgan Fingerprints may include the isomer for a particular molecule, and common backbone structures for molecules.

As depicted, SMILES, Morgan Fingerprints, InChl, One-Hot, N-gram, Graph-based Graphic Processing Unit Nearest Neighbor Search (GGNN), Gene regulatory network (GRN), M-P Neural Network (MPNN), and Knowledge Graph (Structural/Semantic) encodings represent structural information of molecules (drug compounds). The Morgan Fingerprints, GGNN, GRN, and MPNN are also compressed to improve computations, while the SMILES, InChl, One-Hot, N-gram, and the Knowledge Graph are not compressed.

Quantitative structure-activity relationship (QSAR), Z-descriptors, and the Knowledge Graph encodings may represent physiochemical properties of molecules. These encodings may not be compressed. The QSAR encoding may include the type of activity (e.g., and without limitation to a particular physiological or anatomical organ, organ, state or states, or to a particular disease-process, antiviral, antimicrobial, antifungal, antiemetic, antineoplastic, anti-inflammatory, leukotriene inhibitory, neurotransmitter inhibitory, etc.) the molecule provides. The encodings selected for each type of information may optimize the computations when considering such a large design space with information pertaining to structure, physiochemical properties, and semantic information. The large design space referred to may include not only a string of amino acid sequences, and physiochemical properties, but also the semantic information, such as system biology and ontological information, including relationships between nodes, molecular pathways, molecular interactions, molecular family, and the like.

FIG. 1H illustrates an example of concatenating (merging) numerous encodings into a candidate drug compound according to certain embodiments of this disclosure. A concatenated vector 191 may represent an embedding for a candidate drug compound. In some embodiments, an ensemble learning approach may be implemented by using different types of techniques to generate unique encodings and merge those unique encodings to improve generated candidate drug compounds. As depicted, various encoding techniques may be used to represent different types of information. The different types of information (e.g., structural, semantic, etc.) may be represented by unique encodings. For example, molecular graphs and Morgan Fingerprints may represent structural and physical molecular information. Activity data (e.g., QSAR) may represent molecular structural knowledge and/or molecular physiochemical knowledge, and a knowledge graph may represent molecular semantic knowledge. Attention message passing neural network (AMPNN) and/or long short-term memory may receive the molecular graph and Morgan Fingerprints as input and output the structural/physical information represented by 1's and 0's. One-hot may receive the activity data as input and output the structural knowledge represented by 1's and 0's. AMPNN may receive a knowledge graph as input and output semantic knowledge represented by 1's and 0's. The resulting concatenated vector 191 is a combination of each type of information for a single candidate drug compound. Accordingly, the single candidate drug compound may include better properties and more robust information than conventional techniques.

FIG. 1I illustrates an example of using a variational autoencoder (VAE) to generate a Latent Representation 168 of a candidate drug compound according to certain embodiments of this disclosure. The concatenated vector 191 (e.g., embedding) may be higher-dimensional prior to being input to the VAE. The VAE may be trained to translate the higher-dimensional concatenated vector 191 to a lower-dimensional concatenated vector that represents the Latent Representation 168.

FIG. 2 illustrates a data structure storing a biological context representation 200 according to certain embodiments of this disclosure. Biology is context-dependent and dynamic. For example, the same molecule can manifest multiple, potentially competing, phenotypes. Further, data on an existing drug labeled as antimicrobial can suggest a null behavior in applications against different microbes or even against the same microbes but in different contexts, e.g., temperature, pressure, environmental, contextual, comorbid. To accurately predict candidate drug compounds that provide desirable activity levels in design spaces, the machine learning models 132 are trained to handle evolving knowledge maps of biology and drug compounds. Further, conventional techniques for discovery and generating drug compounds may be ineffective for biological data because such data is non-Euclidian. For example, machine learning models used for computer vision, image classification, and language models compute on Euclidian data and cannot therefore be applied to make useful inferences about non-Euclidian data in biology.

In some embodiments, the biological context representation 200 generated by the disclosed techniques may be used to graphically model the continually or continuously modifying biological and drug compound knowledge. That is, the biology may be represented as graphs within a comprehensive knowledge graph (e.g., biological context representation 200), where the graphs have complex relationships and interdependencies between nodes.

The biological context representation 200 may be stored in a first data structure having a first format. The first format may be a graph, an array, a linked list, or any suitable data format capable of storing the biological context representation. In particular, FIG. 2 illustrates various types of data received from various sources, including physical properties data 202, peptide activity data 204, microbe data 206, antimicrobial compound data 208, clinical outcome data 210, evidence-based guidelines 212, disease association data 214, pathway data 216, compound data 218, gene interaction data 220, anti-neurodegenerative compound data 222, and/or pro-neuroplasticity compound data 224.

These example data may be curated by the AI engine 140 and/or a person having a certain degree (e.g., a degree in data science, molecular biology, microbiology, etc.), certification, license (e.g., a licensed medical doctor (e.g., M.D. or D.O.), and/or credential. Further, the data in the biological context representation 200 may be retrieved from any suitable data source (e.g., digital libraries, websites, databases, files, or the like). These examples are not meant to be limiting. Thus, the example types of data are also not meant to be limiting and other types of data may be stored within the biological context representation without departing from the scope of this disclosure. Further, the various data included in the biological context representation 200 may be linked based on one or more relationships between or among the data, in order to represent knowledge pertaining to the biological context and/or drug compound.

The physical properties data 202 includes physical properties exhibited by the drug compound. The physical properties may refer to characteristics that provide a physical description of the drug such as color, particle size, crystalline structure, melting point, solubility. In some instances, the physical properties data 202 may also include chemical property data, such as the structure, form, and reactivity of a substance. In some embodiments, biological data may also be included (e.g., anti-neurodegenerative compound data, pro-neuroplasticity compound data, anti-cancer data) in the biological context representation 200.

The peptide activity data 204 may include various types of activity exhibited by the drug. For example, the activity may be hormonal, antimicrobial, immunomodulatory, cytotoxic, neurological, and the like. A peptide may refer to a short chain of amino acids linked by peptide bonds.

The microbe data 206 may include information pertaining to cellular structure (e.g., unicellular, multicellular, etc.) of a microscopic organism. The microbes may refer to bacteria, parasites, fungi, viruses, prions, or any combination of these, etc.

The antimicrobial compound data 208 may include information pertaining to agents that kill microbes or stop their growth. This data may include classifications based on the microorganisms against which the antimicrobial compound acts (e.g., antibiotics act against bacteria but not against viruses; anti-virals act against viruses but not against bacteria). The antimicrobial compound may also be classified according to function (e.g., microbicidal, meaning “that which kills, vitiates, inactivates or otherwise impairs the activity of certain microbes”).

The clinical outcome data 210 may include information pertaining to the administration of a drug compound to a subject in a clinical setting. For example, upon or subsequent to administration of the drug compound, the outcome may be a prevented disease, cured disease, treated symptom, etc.

The evidence-based guidelines 212 may include information pertaining to guidelines based upon clinical studies for acceptable treatment and/or therapeutics for certain diseases and/or medical conditions. Evidence-based guidelines data 212 may include data specific to various specialties within healthcare such as, for example, obstetrics, anesthesiology, hepatology, gastroenterology, neurology, pulmonology, orthopaedics, pediatrics, trauma care (including but not limited to burns and post-burn infections), histology, oncology, ophthalmology, endocrinology, rheumatology, internal medicine, surgery (including reconstructive (plastic) and cosmetic), vascular medicine, radiology, psychiatry, cardiology, urology, gynecology, genetics, and dermatology. In the example described herein, the evidence-based guidelines 212 include systematically developed statements to assist practitioner and patient decisions about appropriate health care (e.g., types of drugs to prescribe for treatment) for specific clinical circumstances.

The disease association data 214 may include information about which disease and/or medical condition the drug compounds are associated with. For example, the drug compound Metformin may be associated with the disease type 2 diabetes.

The pathway data 216 may include information pertaining in a design space to the relationships or paths between ingredients (e.g., chemicals) and activity levels.

The compound data 218 may include information pertaining to the compound such as the sequence of ingredients (e.g., type, amount, etc.) in the compound. In the therapeutics industry, for example, the compound data 218 can include data specific to the various types of drug compounds that are designed, defined, developed, and/or distributed.

The gene interaction data 220 may include information pertaining to which gene the drug compound and/or a disease may interact with.

The anti-neurodegenerative compound data 222 may include information pertaining to characteristics of anti-neurodegenerative compounds, such as their physical and chemical properties and activities on portions of tissue. For example, the activity may include anti-inflammatory and/or neuro-protective actions.

The pro-neuroplasticity compound data 224 may include information pertaining to characteristics of pro-neuroplasticity compound, such as their physical and chemical properties and activities on portions of tissue. For example, the activity may enhance the capacity of motor systems by upregulation of neurotrophins.

FIGS. 3A-3B illustrate a high-level flow diagram according to certain embodiments of this disclosure. Regarding FIG. 3A, a flow diagram 300 begins with obtaining heterogeneous datasets, such as the biological context representation 200. Heterogeneous datasets may refer to populations or samples of data that are different (e.g., as opposed to homogenous datasets where the data is the same). The heterogeneous datasets may include compound data (e.g., peptide sequence data), clinical outcome data, and/or activity data (in vitro and in vivo activity), as well as any other suitable data depicted in FIG. 2.

The data structure storing the heterogeneous datasets may be translated to a second data structure having a second format (e.g., a 2-dimensional vector) that the AI engine 140 may use to generate the candidate drug compounds. The next step in the flow diagram 300 includes training the one or more machine learning models 132 using the heterogeneous datasets. The one or more machine learning models 132 (e.g., generative models) may generate a set of candidate drug compounds based on the heterogeneous datasets. As described herein, a machine learning model may use causal inference and counterfactuals when generating the set of candidate drug compounds. Further, a GAN may be used in conjunction with causal inference to generate the set of candidate drug compounds. In some embodiments, a certain number (e.g., over 100,000 candidate drug compounds) of novel candidate drug compounds may be generated in a set. That is, each candidate drug compound in the set of candidate drug compounds is intended to be unique.

The next step in the flow diagram 300 includes inputting the set of candidate drug compounds into one or more machine learning models 132 trained to classify the set of candidate drug compounds. The machine learning models 132 may perform supervised and/or unsupervised filtering. In some embodiments, the machine learning models 132 may perform clustering to rank the various candidate drug compounds to classify one candidate drug compound as a selected candidate drug compound. In some embodiments, the machine learning models 132 may output a subset (e.g., 1,000 to 10,000, or more, or fewer) of candidate drug compounds.

The next step in the flow diagram 300 may include performing experimental validation by validating whether each candidate drug compound in the subset of candidate drug compounds provides the desired level of certain types of activity in a design space. The results of the experimental validation may be fed back into the heterogeneous dataset to reinforce and expand the experimental dataset.

The next step in the flow diagram 300 may include performing peptide drug optimization. The optimizations may include performing gradient descent and/or ascent using the sequence of ingredients in the candidate drug compounds to attempt to increase and/or decrease certain activity levels in a design space. The results of the peptide drug optimization may be fed back into the heterogeneous datasets to reinforce and expand the experimental dataset.

FIG. 3B illustrates another high-level flow diagram 310 according to some embodiments. As depicted, a heterogeneous network of biology may be included in a knowledge graph of a biological context representation 200. Various paths or meta-paths may be expressed between nodes in the biological context representation 200. For example, the meta-paths may include indications for compound upregulates, pathway participates, disease associations, gene interactions, and compound data.

The biological context representation 200 may be translated from a first format (e.g., knowledge graph) to a format (e.g., vector) that may be processed by the AI engine 140. The AI engine 140 may use one or more machine learning models to traverse the knowledge graph by performing random walks until a corpus of random walks is generated, wherein such random walks include the indications associated with the meta-paths representing sequences of ingredients. The corpus of random walks may be referred to as a set of candidate drug compounds. A generative adversarial network using causal inference may be used to generate the set of candidate drug compounds. The set of candidate drug compounds may be stored in a higher-dimensional vector.

The AI engine 140 may compress the higher-dimensional vector of the set of candidate drug compounds into a lower-dimensional vector of the set of candidate drug compounds, depicted as biological embeddings in FIG. 3B. In some embodiments, the lower-dimensional vector may include fewer dimensions (e.g., 2, 3, N) than the higher-dimensional vector (e.g., greater than N). As depicted, the nodes may be organized by the meta-path indicators and by dimension.

To output a subset of candidate drug compounds, the lower-dimensional vector of the set of candidate drug compounds may be input to one or more machine learning models 132 trained to perform classification. The classification techniques may include using clustering to filter out candidate drug compounds that produce undesirable levels of types of activity. In some embodiments, to enable the AI engine 140 to perform the classification, views presenting the levels of types of activity of each candidate drug compound in a design space may be generated using the lower-dimensional vectors. These views may also be presented to a user via the computing device 102. The machine learning models 132 may output a candidate drug candidate classified as a selected candidate drug candidate based on the clustering. For example, the selected candidate drug candidate may include an optimized sequence of ingredients that provides the most desirable levels of a certain type of activity in a design space.

FIG. 4 illustrates example operations of a method 400 for generating and classifying a candidate drug candidate compound according to certain embodiments of this disclosure. The method 400 is performed by processing logic that may include hardware (circuitry, dedicated logic, etc.), software (such as is run on a general purpose computer system or a dedicated machine), or a combination of both. The method 400 and/or each of their individual functions, routines, subroutines, or operations may be performed by one or more processors of a computing device (e.g., any component of FIG. 1, such as server 128 executing the artificial intelligence engine 140). In certain implementations, the method 400 may be performed by a single processing thread. Alternatively, the method 400 may be performed by two or more processing threads, each thread implementing one or more individual functions, routines, subroutines, or operations of the methods. One or more operations of the method 400 may be performed by the training engine 130 of FIG. 1.

For simplicity of explanation, the method 400 is depicted and described as a series of operations. However, operations in accordance with this disclosure can occur in various orders and/or concurrently, and with other operations not presented and described herein. For example, the operations depicted in the method 400 may occur in combination with any other operation of any other method disclosed herein. Furthermore, not all illustrated operations may be required to implement the method 400 in accordance with the disclosed subject matter. In addition, those skilled in the art will understand and appreciate that the method 400 could alternatively be represented as a series of interrelated states via a state diagram or events.

At 402, the processing device may generate a biological context representation 200 of a set of drug compounds. The biological context representation 200 may include a first data structure having a first format (e.g., a knowledge graph). The biological context representation 200 may include, for each drug compound of the set of drug compounds, one or more relationships between or among, without limitation, (i) physical properties data 202, (ii) peptide activity data 204, (iii) microbe data 206, (iv) antimicrobial compound data 208, (v) clinical outcome data 210, (vi) evidence-based guidelines 212, (vii) disease association data 214, (viii) pathway data 216, (ix), compound data 218, (x) gene interaction data 220, (xi) antimicrobial compound data, (xii) pro-neuroplasticity data 224, or some combination thereof.

At 404, the processing device may translate, by the artificial intelligence engine 140, the first data structure having the first format to a second data structure having a second format. The translating may include converting the first data structure having the first format (e.g., knowledge graph) to the second data structure having the second format (e.g., vector) according to a specific set of rules executed by the artificial intelligence engine 140. In some embodiments, the translating may be performed by one or more of the machine learning models 132. For example, a recurrent neural network may perform at least a portion of the translating.

The translating may include obtaining a higher-dimensional vector and compressing the higher-dimensional vector into a lower-dimensional vector (e.g., two-dimensional, three-dimensional, four-dimensional), referred to as an embedding herein. In some embodiments, one or more embeddings may be created from the first data structure having the first format. There may be any suitable number of dimensions of the embeddings. When used for classifying candidate drug compounds, the number of dimensions may be selected based on a desired performance to process the embeddings. The lower-dimensional vector may have at least one fewer dimension than the higher-dimensional vector.

At 406, the processing device may generate, based on the second data structure having the second format, a set of candidate drug compounds. In some embodiments, the generating may be performed by one or more of the machine learning models 132. For example, a generative adversarial network may perform the generating of the set of candidate drug compounds. In some embodiments, the set of candidate drug compounds may be associated with design spaces pertaining to antimicrobial, anti-cancer, anti-biofilm, or the like. A biofilm may include any syntrophic consortium of microorganisms in which cells stick to each other and often also to a surface. These adherent cells may become embedded within an extracellular matrix that is composed of extracellular polymeric substances (EPS). Cancer may refer to a disease caused or correlated with an uncontrolled division of abnormal cells in a part of the body.

At 408, the processing device may classify a candidate drug compound from the set of candidate drug compounds as a selected candidate drug compound. In some embodiments, the classifying may be performed by one or more of the machine learning models 132. For example, a classifier trained using supervised or unsupervised learning may perform the classifying. In some embodiments, the classifier may use clustering techniques to rank and classify the selected candidate drug compound.

In some embodiments, the processing device may generate a set of views including a representation of a design space. The design space may be antimicrobial. The processing device may cause the set of views to be presented on a computing device (e.g., computing device 102). The representation of the design space may pertain to, without limitation, (i) antimicrobial activity, (ii) immunomodulatory activity, (iii) neuromodulatory activity, (iv) cytotoxic activity, or some combination thereof. Each view of the set of views may present an optimized sequence representing the selected candidate drug compound.

The optimized sequence in each view may be generated using any suitable optimization technique. The optimization technique may include maximizing or minimizing an objective function by systematically selecting input values from a domain of values and computing the value using the objective function. The domain of values may include a subset of values from a Euclidean space. The subset of values may satisfy one or more constraints, equalities, and/or inequalities. A value that minimizes or maximizes the objective function may be referred to as an optimal solution. Certain values in the subset may result in a gradient of the objective function being zero. Those certain values may be at stationary points, where a first derivative at those points with respect to time (dt) is zero. The gradient may refer to a scalar-valued differentiable function (e.g., objective function) of several variables, where a point p is a vector whose components are the partial derivatives of the objective function. If the gradient is not a zero vector at a certain point p, then a direction of the gradient is the direction of fastest increase of the objective function at the certain point p.

Gradients may be used in gradient descent, which refers to a first-order iterative optimization algorithm for finding the local minimum of an objective function. To find the local minimum, gradient descent may proceed by performing operations proportional to the negative of the gradient of the objective function at a current point. In some embodiments, the optimized sequence may be found for a candidate drug compound performing gradient descent in the design space. Additionally, gradient ascent, which is the algorithm opposite to gradient descent, may determine a local maximum of the objective function at various points in the design space.

The views generated may include a topographical heatmap, itself including indicators for the least activity at points in the design space and the most activity at points in the design space. The indicator associated with the most activity may represent a local maximum obtained using gradient ascent. The indicator associated with the least activity may represent a local minimum obtained using gradient descent. The optimal sequence may be generated by navigating points between the local minima and local maxima. The optimized sequence may be overlaid on the indicators ranging from at least one least active property to an at least one most active property.

In some embodiments, the processing device may cause the selected candidate drug compound to be formulated. In some embodiments, the processing device may cause the selected candidate drug compound to be created, manufactured, developed, synthesized, or the like. In some embodiments, the processing device may cause the selected candidate drug compound to be presented on a computing device (e.g., computing device 102). The selected candidate drug compound may include one or more active ingredients (e.g., chemicals) at a specified amount.

FIGS. 5A-5D provide illustrations of generating a first data structure including a biological context representation 200 of a plurality of drug compound devices according to certain embodiments of this disclosure. The first data format may include a knowledge graph. The biological context representation 200 may capture an entire biological context by integrating every known association or relationship for each drug compound into a comprehensive knowledge graph.

FIG. 5A presents the biological context representation 200 including biomedical and domain knowledge on peptide activity, microbes, antimicrobial compounds, clinical outcomes, and any relevant information depicted in FIG. 2. A table 500 may include rows representing various categories (A, B, C, D, and E) pertaining to a biological context for each drug compound and columns representing sub-categories (1, 2, 3, 4, and 5). For example, the table includes subcategories for category A: A1 2D fingerprints, A2 3D fingerprints, A3 Scaffolds, A4 Struct. Keys, A5 Physicochem./B: B1 Mech. Of act, B2 Metab. Genes, B3 Crystals, B4 Binding, B5 HTS bioassays/C: C1 S. mol. Roles, C2 S. mol. Path., C3 Signal. Path., C4 Biol. Proc., C5 Interactome/D: D1 Transcript, D2 Can. Cell lines, D3 Ch. Genetics, D4 Morphology, D5 Cell bioassays/E: E1 Therap. Areas, E2 Indications, E3 Side effects, E4 Dis. & Toxicol., E5 Drug-drug inter.

Charts 502, 504, and 506 represent characteristics for each subcategory. The characteristics for chart 502 include the size of molecules, for chart 504 the complexity of variables, and for 506 the correlation with mechanism of action. Another chart 508 may represent the various characteristics of the subcategories using an indicator (such as a range of colors from 0 to 1) to express the values of the characteristics in relation to each other.

FIG. 5B illustrates a different representation 520 of characteristics for several subcategories (e.g., A1, B1, C5, D1, and E3) across different subject matter areas (e.g., neurology and psychiatry, infectious disease, gastroenterology, cardiology, ophthalmology, oncology, endocrinology, pulmonary, rheumatology, and malignant hematology.). Accordingly, the representation 520 provides an even more granular representation of the biological context representation 200 than does the chart 508. Flowchart 530 represents the process for generating candidate drugs as described further herein.

FIG. 5C illustrates a knowledge graph 540 representing the biological context representation 200. The knowledge graph 540 may refer to a cognitive map. In particular, the knowledge graph 540 represents a graph traversed by the AI engine 140, when generating candidate drug compounds having desired levels of certain types of activity in a design space. Individual nodes in the knowledge graph 540 represent a health artifact (health-related information) or relationship (predicate) gleaned and curated from numerous data sources. Further, the knowledge represented in the knowledge graph 540 may be improved over time as the machine learning models discover new associations, correlations, and/or relationships. The nodes and relationships may form logical structures that represent knowledge (e.g., Genes Participates Pathways). FIG. 5D illustrates another representation of the knowledge graph 540 that more clearly identifies all the various relationships among the nodes.

FIG. 6 illustrates example operations of a method 600 for translating the first data structure of FIGS. 5A-5B a second data structure according to certain embodiments of this disclosure. Method 600 includes operations performed by processors of a computing device (e.g., any component of FIG. 1, such as server 128 executing the artificial intelligence engine 140). In some embodiments, one or more operations of the method 600 are implemented in computer instructions that are stored on a memory device and executed by a processing device. The method 600 may be performed in the same or a similar manner as described above in regards to method 400. The operations of the method 600 may be performed in some combination with any of the operations of any of the methods described herein.

The method 600 may include operation 404 from the previously-described method 400 depicted in FIG. 4. For example, at 404 in the method 600, the processing device may translate, by the artificial intelligence engine 140, the first data structure having the first format (e.g., knowledge graph) to the second data structure having the second format (e.g., vector). The method 600 in FIG. 6 includes operations 602 and 604.

At 602, the processing device may obtain a higher dimensional vector from the biological context representation 200. This process is further illustrated in FIG. 7.

At 604, the processing device may compress the higher-dimensional vector to a lower dimensional-vector. The compressing may be performed by a first machine learning model 132 trained to perform deep auto-encoding via a recurrent neural network configured to output the lower-dimensional vector.

At 606, the processing device may train the first machine learning model 132 by using a second machine learning model 132 to recreate the first data structure having the first format. The second machine learning model 132 is trained to perform a decoding operation to recreate the first data structure having the first format. The decoding operation may be performed on the second data structure having the second data format (e.g., two-dimensional vector).

FIG. 7 provides illustrations of translating the first data structure of FIGS. 5A-5B to the second data structure according to certain embodiments of this disclosure. Aggregated biological data may be difficult to model and format correctly for an AI engine to process. Aspects of the present disclosure overcome the hurdle of modeling and formatting the aggregated biological data to enable the AI engine 140 to generate candidate drug compounds accurately and efficiently.

As depicted, a higher-dimensional vector 700 may be obtained from the biological context representation 200. Using a recurrent neural network performing autoencoding, the higher-dimensional vector is compressed to a lower-dimensional vector 702. The recurrent neural network performing autoencoding is trained using another machine learning model 132 that recreates the higher-dimensional vector 704. If the other machine learning model 132 is unable to recreate higher-dimensional vector 704 from the lower-dimensional vector 702, then the other machine learning model 132 provides feedback to the recurrent neural network performing autoencoding in order to update its weights, biases, or any suitable parameters.

FIGS. 8A-8C provide illustrations of views of a selected candidate drug compound according to certain embodiments of this disclosure. As depicted, FIG. 8A illustrates a view 800 including antimicrobial activity, FIG. 8B illustrates a view 802 including immunomodulatory activity, and FIG. 8C illustrates a view 804 including cytotoxic activity. Each view presents a topographical heatmap where one axis is for sequence parameter y and the other axis is for sequence parameter x. Each view includes an indicator ranging from a least active property to a most active property. Further each view includes an optimized sequence 806 for a selected candidate drug compound classified by the classifier (machine learning model 132). These views may be presented to the user on a computing device 102. Further, the selected candidate drug compound 806 may be formulated, generated, created, manufactured, developed, and/or tested.

FIG. 9 illustrates example operations of a method 900 for presenting a view including a selected candidate drug compound according to certain embodiments of this disclosure. Method 900 includes operations performed by processors of a computing device (e.g., any component of FIG. 1, such as computing device 102). In some embodiments, one or more operations of the method 1000 are implemented in computer instructions that are stored on a memory device and executed by a processing device. The method 1000 may be performed in the same or a similar manner as described above in regards to method 400. The operations of the method 1000 may be performed in some combination with any of the operations of any of the methods described herein.

At 902, the processing device may receive, from the artificial intelligence engine 140, a candidate drug compound generated by the artificial intelligence engine 140.

At 904, the processing device may generate a view including the candidate drug compound overlaid on a representation of a design space. The view may present a topographical heatmap of the representation of the design space. The topographical heatmap may include the candidate drug compound overlaid on indicators ranging from an at least one least active property to an at least one most active property.

At 906, the processing device may present the view on a display screen of a computing device (e.g., computing device 102).

FIG. 10A illustrates example operations of a method 1000 for using causal inference during the generation of candidate drug compounds according to certain embodiments of this disclosure. Method 1000 includes operations performed by processors of a computing device (e.g., any component of FIG. 1, such as server 128 executing the artificial intelligence engine 140). In some embodiments, one or more operations of the method 1000 are implemented in computer instructions that are stored on a memory device and executed by a processing device. The method 1000 may be performed in the same or a similar manner as described above in regards to method 400. The operations of the method 1000 may be performed in some combination with any of the operations of any of the methods described herein.

At 1002, the processing device may perform one or more modifications pertaining to the biological context representation 200, the second data structure having the second format, or some combination thereof.

At 1004, the processing device may use causal inference to determine whether the one or more modifications provide one or more desired performance results. In some embodiments, using causal inference may further include using 1006 counterfactuals to calculate alternative scenarios based on past actions, occurrences, results, regressions, regression analyses, correlations, or some combination thereof. The term “calculate” may be used interchangeably with any of the following terms: simulate, emulate, determine, generate, formulate, execute, and/or obtain. A counterfactual may refer to determining whether the desired performance still results if something does not occur during the calculation. For example, in a scenario, a person may improve their health after taking a medication. The counterfactual may be used in causal inference to calculate an alternative scenario to see whether the person's health improved without taking the medication. If the person's health still improved without taking the medication it may be inferred that the medication did not cause the health of the person to improve. However, if the person's health did not improve without taking the medication, it may be inferred that the medication is correlated with causing the health of the person to improve. There may, however, be other factors involved in conjunction with taking the medication that actually cause the health of the person to improve.

FIG. 10B illustrates another example of operations of method 1050 for using causal inference during the generation of candidate drug compounds according to certain embodiments of this disclosure. Method 1050 includes operations performed by processors of a computing device (e.g., any component of FIG. 1, such as server 128 executing the artificial intelligence engine 140). In some embodiments, one or more operations of the method 1050 are implemented in computer instructions that are stored on a memory device and executed by a processing device. The method 1050 may be performed in the same or a similar manner as described above in regards to method 400. The operations of the method 1050 may be performed in some combination with any of the operations of any of the methods described herein.

At 1052, the processing device may generate a set of candidate drug compounds by performing a modification using causal inference based on a counterfactual. For example, the counterfactual may include removing an ingredient from a sequence of ingredients to determine whether a candidate drug compound provides the same level and/or type of activity it previously provided when the ingredient was included in the sequence. If the same level and/or type of activity is still provided after application of the counterfactual (e.g., removal of the ingredient), then the processing device may use causal inference to determine that the ingredient is not correlated with the level and/or type of activity. If the same level and/or type of activity is not present after application of the counterfactual (e.g., removal of the ingredient), then the processing device may use causal inference to determine that the ingredient is correlated with the level and/or type of activity.

At 1054, the processing device may classify a candidate dug compound from the set of candidate drug compounds as a selected candidate drug compound, as previously described herein.

FIG. 11 illustrates example operations of a method 1100 for using several machine learning models in an artificial intelligence engine architecture to generate peptides according to certain embodiments of this disclosure. Method 1100 includes operations performed by processors of a computing device (e.g., any component of FIG. 1, such as server 128 executing the artificial intelligence engine 140). In some embodiments, one or more operations of the method 1100 are implemented in computer instructions stored on a memory device and executed by a processing device. The method 1100 may be performed in the same or a similar manner as described above in regards to method 400. The operations of the method 1100 may be performed in some combination with any of the operations of any of the methods described herein.

At block 1102, the processing device may generate, via a creator module 151, a candidate drug compound including a sequence for candidate drug compound. The sequence for the candidate drug compound includes a concatenated vector that may include drug compound sequence information, drug compound activity information, drug compound structure information, and drug compound semantic information.

In some embodiments, the candidate drug compound may be generated using a GAN. In some embodiments, the processing device may use an attention message passing neural network including an attention mechanism that identifies and assigns a weight to a desired feature in a portion of the knowledge graph. The desired feature may be include in the candidate drug compound as drug compound semantic information, drug compound structural information, drug compound activity information, or some combination thereof.

In some embodiments, the creator module 151 may generate the candidate drug compound by performing ensemble learning by concatenating a set of encodings. The encodings may each respective sequences represented in a vector. A first encoding of the set of encodings may pertain to drug compound sequence information. A second encoding of the set of encodings may pertain to drug compound structural information. A third encoding of the set of encodings may pertain to peptide activity information. A fourth encoding of the set of encodings may pertain to drug compound semantic information.

In some embodiments, the creator module 151 may generate the candidate drug compound using an autoencoder machine learning model trained to receive a higher-dimensional vector encoding representing the candidate drug compound and output a lower-dimensional vector embedding representing the candidate drug compound. The creator module 151 may generate a latent representation using the lower-dimensional vector embedding representing the candidate drug compound.

At block 1104, the processing device may include, via the creator module 151, the candidate for the candidate drug compound as a node in a knowledge graph (e.g., biological context representation 200). In some embodiments, the knowledge graph may include a first layer including structure and physical properties of molecules, a second layer including molecule-to-molecule interactions, a third layer including molecular pathway interactions, a fourth layer including molecular cell profile associations, and a fifth layer including molecular therapeutics and indications. Indications may refer to drug indications, or the disease which gives a valid reason for clinicians to administer a specific drug.

At block 1106, the processing device may generate, via a descriptor module 152, a description of the candidate drug compound at the node in the knowledge graph. The description may include drug compound sequence information, drug compound structural information, drug compound activity information, and drug compound semantic information.

At block 1108, based on the description, the processing device may perform, via a scientist module 153, a benchmark analysis of a parameter of the creator module 151. In some embodiments, the scientist module 153 may perform causal inference using the candidate drug compound in a design space pertaining to biomedical activity (e.g., antimicrobial, anti-cancer, etc.) to determine if the candidate drug compound still provides a desired effect regarding the type of biomedical activity if the candidate drug compound, or the design space, is changed.

At block 1110, the processing device may modify, based on the benchmark analysis, the creator module 151 to change the parameter in a desired way during a subsequent benchmark analysis. Changing the parameter in a desired way may refer to changing a value of the parameter in a desired way. Changing the value of the parameter in the desired way may refer to increasing or decreasing the value of the parameter. Accordingly, a self-improving AI engine 140 is disclosed that increasingly generates better candidate drug components over time by recursively updating the creator module 151 based on baselines. In some embodiments, “change the parameter” means change a value of the parameter as desired (e.g., either increase or decrease).

In some embodiments, the processing device may generate, via a reinforcer module 154 based on the candidate drug compound and the description, experiments that produce desired data for the candidate drug compound. The experiments may be generated in response to the candidate drug compound and the description being similar to a real drug compound and another description of the real drug compound. For example, the reinforce module 154 may determine that certain experiments for the real drug compound elicited desired data and may select those experiments to perform for the candidate drug compound. The processing device may perform the experiments (e.g., by running simulations) to collect data pertaining to the candidate drug compound. The processing device may determine, based on the data, an effectiveness of the candidate drug compound.

FIG. 12 illustrates example operations of a method 1200 for performing a benchmark analysis according to certain embodiments of this disclosure. Method 1200 includes operations performed by processors of a computing device (e.g., any component of FIG. 1, such as server 128 executing the artificial intelligence engine 140). In some embodiments, one or more operations of the method 1200 are implemented in computer instructions that are stored on a memory device and executed by a processing device. The method 1200 may be performed in the same or a similar manner as described above in regards to method 400. The operations of the method 1200 may be performed in some combination with any of the operations of any of the methods described herein.

The method 1200 includes additional operations included in block 1108 of FIG. 11. At block 1202, the processing device generates, via the scientist module 143, a score for a parameter of the creator module 151 that generated the candidate drug compound. The parameter may include a validity of the candidate drug compound, uniqueness of the candidate drug compound, novelty of the candidate drug compound, similarity of the candidate drug compound to another candidate drug compound, or some combination thereof.

At block 1204, the processing device may rank a set of creator modules 151 based on the score, where the set of creator modules comprises the creator module. For example, other creator modules in the set of creator modules may be scored based on the candidate drug compounds they generated. The set of creator modules may be ranked for each respective category from highest scoring to lowest scoring or vice versa.

At block 1206, the processing device may determine which creator module 151 of the set of creator modules performs better for each respective parameter. The scores of the parameters for each of the set of creator modules 151 may be presented on a display screen of a computing device. The best performing creator modules for each parameter may also be presented on the display screen.

At block 1208, the processing device may tune the set of creator modules 151 to cause the set of creator modules 151 to receive higher scores for certain parameters during subsequent benchmark analysis. The tuning may optimize certain weights, activation functions, hidden layer number, loss, and the like of one or more generative modules included in the creator modules.

At block 1210, the processing device may select, based on the parameters, a subset of the set of creator modules 151 to use to generate subsequent candidate drug compounds having desired parameter scores. For example, it may be desired to generate drug candidate compounds that result in a high uniqueness score. The creator module(s) 151 associated with high uniqueness scores may be selected in the subset of creator modules 151.

At block 1212, the processing device may transmit the subset of the set of creator modules as a package to a third-party to be used with data of the third-party. The subset of the set of creator modules may be trained to process a type of the data of the third-party. Other modules, such as the reinforce module, the descriptor module, the scientist module, and the conductor module may be included in the package delivered to the third-party. Also, a knowledge graph including data pertaining to the third-party may be included in the package. In such a way, the disclosed techniques may provide custom tailored packages that may be used by the third-party to perform the embodiments disclosed herein.

FIG. 13 illustrates example operations of a method 1300 for slicing a latent representation based on a shape of the latent representation according to certain embodiments of this disclosure. Method 1300 includes operations performed by processors of a computing device (e.g., any component of FIG. 1, such as server 128 executing the artificial intelligence engine 140). In some embodiments, one or more operations of the method 1300 are implemented in computer instructions stored on a memory device and executed by a processing device. The method 1300 may be performed in the same or a similar manner as described above in regards to method 400. The operations of the method 1300 may be performed in some combination with any of the operations of any of the methods described herein.

At block 1302, the processing device may determine a shape of the multi-dimensional, continuous representation of the set of candidates. At block 1304, the processing device may determine, based on the shape, a slice to obtain from the multi-dimensional, multi-dimensional, continuous representation of the set of candidates. At block 1306, the processing device may determine, using a decoder, which dimensions are included in the slice. The dimensions may pertain to peptide sequence information, peptide structural information, peptide activity information, peptide semantic information, or some combination thereof. At block 1308, the processing device may determine, based on the dimensions, an effectiveness of a biomedical feature of the slice.

FIG. 14 illustrates example computer system 1400 which can perform any one or more of the methods described herein, in accordance with one or more aspects of the present disclosure. In one example, computer system 1400 may correspond to the computing device 102 (e.g., user computing device), one or more servers 128 of the cloud-based computing system 116, the training engine 130, or any suitable component of FIG. 1. The computer system 1400 may be capable of executing application 118 and/or the one or more machine learning models 132 of FIG. 1. The computer system may be connected (e.g., networked) to other computer systems in a LAN, an intranet, an extranet, or the Internet. The computer system may operate in the capacity of a server in a client-server network environment. The computer system may be a personal computer (PC), a tablet computer, a wearable (e.g., wristband), a set-top box (STB), a personal Digital Assistant (PDA), a mobile phone, a camera, a video camera, or any device capable of executing a set of instructions (sequential or otherwise) that specify actions to be taken by that device. Further, while only a single computer system is illustrated, the term “computer” shall also be taken to include any collection of computers that individually or jointly execute a set (or multiple sets) of instructions to perform any one or more of the methods discussed herein.

The computer system 1400 includes a processing device 1402, a main memory 1404 (e.g., read-only memory (ROM), flash memory, solid state drives (SSDs), dynamic random access memory (DRAM) such as synchronous DRAM (SDRAM)), a static memory 1406 (e.g., flash memory, solid state drives (SSDs), static random access memory (SRAM)), and a data storage device 1108, which communicate with each other via a bus 1410.

Processing device 1402 represents one or more general-purpose processing devices such as a microprocessor, central processing unit, or the like. More particularly, the processing device 1402 may be a complex instruction set computing (CISC) microprocessor, reduced instruction set computing (RISC) microprocessor, very long instruction word (VLIW) microprocessor, or a processor implementing other instruction sets or processors implementing a combination of instruction sets. The processing device 1402 may also be one or more special-purpose processing devices such as an application specific integrated circuit (ASIC), a system on a chip, a field programmable gate array (FPGA), a digital signal processor (DSP), network processor, or the like. The processing device 1402 is configured to execute instructions for performing any of the operations and steps discussed herein.

The computer system 1400 may further include a network interface device 1412. The computer system 1400 also may include a video display 1414 (e.g., a liquid crystal display (LCD), a light-emitting diode (LED), an organic light-emitting diode (OLED), a quantum LED, a cathode ray tube (CRT), a shadow mask CRT, an aperture grille CRT, a monochrome CRT), one or more input devices 1416 (e.g., a keyboard and/or a mouse), and one or more speakers 1418 (e.g., a speaker). In one illustrative example, the video display 1414 and the input device(s) 1416 may be combined into a single component or device (e.g., an LCD touch screen).

The data storage device 1416 may include a computer-readable medium 1420 on which the instructions 1422 embodying any one or more of the methods, operations, or functions described herein is stored. The instructions 1422 may also reside, completely or at least partially, within the main memory 1404 and/or within the processing device 1402 during execution thereof by the computer system 1400. As such, the main memory 1404 and the processing device 1402 also constitute computer-readable media. The instructions 1422 may further be transmitted or received over a network via the network interface device 1412.

While the computer-readable storage medium 1420 is shown in the illustrative examples to be a single medium, the term “computer-readable storage medium” should be taken to include a single medium or multiple media (e.g., a centralized or distributed database, and/or associated caches and servers) that store the one or more sets of instructions. The term “computer-readable storage medium” shall also be taken to include any medium that is capable of storing, encoding or carrying a set of instructions for execution by the machine and that cause the machine to perform any one or more of the methodologies of the present disclosure. The term “computer-readable storage medium” shall accordingly be taken to include, but not be limited to, solid-state memories, optical media, and magnetic media.

None of the description in this application should be read as implying that any particular element, step, or function is an essential element that must be included in the claim scope. The scope of patented subject matter is defined only by the claims. Moreover, none of the claims is intended to invoke 35 U.S.C. § 112(f) unless the exact words “means for” are followed by a participle.

Consistent with the above disclosure, the examples of systems and method enumerated in the following clauses are specifically contemplated and are intended as a non-limiting set of examples.

Clause 1: A method comprising:

generating, via a creator module, a candidate drug compound comprising a sequence;

including the candidate drug compound as a node in a knowledge graph;

generating, via a descriptor module, a description of the candidate drug compound at the node in the knowledge graph, wherein the description comprises drug compound structural information, drug compound activity information, and drug compound semantic information;

based on the description, performing, via a scientist module, a benchmark analysis of a parameter of the creator module; and

modifying, based on the benchmark analysis, the creator module to change the parameter in a desired way during a subsequent benchmark analysis.

Clause 2. The method of clause 1, further comprising:

performing, via the scientist module, causal inference using the candidate drug compound in a design space pertaining to biomedical activity to determine if the candidate drug compound still provides a desired effect regarding the biomedical activity if the candidate drug compound, or the design space, is changed.

Clause 3. The method of clause 1, wherein the knowledge graph comprises a multi-dimensional, continuous representation of a plurality of candidate drug compounds, and the method further comprising:

determining a shape of the multi-dimensional, continuous representation of the plurality of candidate drug compounds;

determining, based on the shape, a slice to obtain from the multi-dimensional, continuous representation of the plurality of candidate drug compounds;

determining, using a decoder, which dimensions are included in the slice, wherein the dimensions pertain to drug compound sequence information, drug compound structural information, drug compound activity information, drug compound semantic information, or some combination thereof; and

determining, based on the dimensions, an effectiveness of a biomedical feature of the slice.

Clause 4. The method of clause 1, wherein performing, via the scientist module, based on the candidate drug compound, the benchmark analysis of the parameter of the creator module further comprises:

generating a score for the parameter, wherein the parameter comprises a validity of the candidate drug compound, uniqueness of the candidate drug compound, novelty of the candidate drug compound, similarity of the candidate drug compound to another candidate drug compound, or some combination thereof; and

ranking a plurality of creator modules based on the score, wherein the plurality of creator modules comprises the creator module.

Clause 5. The method of clause 1, further comprising:

generating, via a reinforcer module based on the candidate drug compound and the description, experiments that produce desired data for the candidate drug compound, wherein the experiments are generated in response to the candidate drug compound and the description being determined to be similar to a real drug compound and another description of the real drug compound;

performing the experiments to collect data pertaining to the candidate drug compound; and

determining, based on the data, an effectiveness of the candidate drug compound.

Clause 6. The method of clause 1, wherein the knowledge graph comprises:

a first layer including structural and physical properties of molecules;

a second layer including molecule-to-molecule interactions;

a third layer including molecular pathway interactions;

a fourth layer including molecular cell profile associations; and

a fifth layer including biologic drug therapeutics and indications.

Clause 7. The method of clause 1, wherein generating, via the creator module, the candidate drug compound further comprises:

using a generative adversarial network to generate the candidate drug compound.

Clause 8. The method of clause 1, wherein generating, via the creator module, the candidate drug compound further comprises:

using an attention message passing neural network including an attention mechanism that identifies and assigns a weight to a desired feature in a portion of the knowledge graph, configured to include in the candidate drug compound drug compound semantic information, drug compound structural information, drug compound activity information, or some combination thereof.

Clause 9. The method of clause 1, wherein generating, via the creator module, the candidate drug compound further comprises:

concatenating a plurality of encodings, wherein:

the encodings are each respective sequences represented in a vector,

a first encoding of the plurality of encodings pertains to drug compound structural information,

a second encoding of the plurality of encodings pertains to peptide activity information, and

a third encoding of the plurality of encodings pertains to drug compound semantic information.

Clause 10. The method of clause 9, wherein a fourth encoding of the plurality of encodings pertains to drug compound sequence information.

Clause 11. The method of clause 1, wherein generating, via the creator module, the candidate drug compound further comprises:

using an autoencoder machine learning model trained to receive a higher-dimensional vector encoding representing the candidate drug compound and output a lower-dimensional vector embedding such embedding representing the candidate drug compound; and

generating a latent representation using the lower-dimensional vector embedding, such embedding representing the candidate drug compound.

Clause 12. The method of clause 1, wherein the sequence for the candidate drug compound comprises a second vector including drug compound sequence information, drug compound activity information, drug compound structure information, and drug compound semantic information.

Clause 13. The method of clause 1, wherein the definition further comprises drug compound sequence information.

Clause 14. A tangible, non-transitory computer-readable medium storing instructions that, when executed, cause a processing device to:

generate, via a creator module, a candidate drug compound comprising a sequence;

include the candidate drug compound as a node in a knowledge graph;

generate, via a descriptor module, a description of the candidate drug compound at the node in the knowledge graph, wherein the description comprises drug compound structural information, drug compound activity information, and drug compound semantic information;

based on the description, perform, via a scientist module, a benchmark analysis of a parameter of the creator module; and

modify, based on the benchmark analysis, the creator module to change the parameter in a desired way during a subsequent benchmark analysis.

Clause 15. The computer-readable medium of clause 14, wherein the processing device is further to:

perform, via the scientist module, causal inference using the candidate drug compound in a design space pertaining to biomedical activity to determine if the candidate drug compound still provides a desired effect regarding the biomedical activity if the candidate drug compound or the design space is changed.

Clause 16. The computer-readable medium of clause 14, wherein the knowledge graph comprises a multi-dimensional, continuous representation of a plurality of candidate drug compounds, and the processing device is further to:

determine a shape of the multi-dimensional, continuous representation of the plurality of candidate drug compounds;

determine, based on the shape, a slice to obtain from the multi-dimensional, continuous representation of the plurality of candidate drug compounds;

determine, using a decoder, which dimensions are included in the slice, wherein the dimensions pertain to drug compound sequence information, drug compound structural information, drug compound activity information, drug compound semantic information, or some combination thereof; and

determine, based on the dimensions, an effectiveness of a biomedical feature of the slice.

Clause 17. The computer-readable medium of clause 14, wherein performing, via the scientist module, based on the candidate drug compound, the benchmark analysis of the parameter of the creator module further comprises:

ranking a plurality of creator modules based on the score, wherein the plurality of creator modules comprises the creator module.

Clause 18. The computer-readable medium of clause 14, wherein the processor is further to:

generate, via a reinforcer module based on the candidate drug compound and the description, experiments that produce desired data for the candidate drug compound, wherein the experiments are generated in response to the candidate drug compound and the description being determined to be similar to a real drug compound and another description of the real drug compound;

perform the experiments to collect data pertaining to the candidate drug compound; and

determine, based on the data, an effectiveness of the candidate drug compound.

Clause 19. The computer-readable medium of clause 14, wherein the knowledge graph comprises:

a first layer including structural and physical properties of molecules;

a second layer including molecule-to-molecule interactions;

a third layer including molecular pathway interactions;

a fourth layer including molecular cell profile associations; and

a fifth layer including biologic drugs and indications.

Clause 20. A system comprising:

a memory device storing instructions; and

a processing device communicatively coupled to the memory device, the processing device executes the instructions to: