In some embodiments, aspects of the disclosure relate to a computer-implemented method for processing genomic information, by obtaining data representing expression levels in a biological sample of at least two informative-genes (e.g., at least two informative-mRNAs from Table 11), wherein the biological sample was obtained of a subject, and using the expression levels to assist in determining the likelihood that the subject has lung cancer. A computer-implemented method can include inputting data via a user interface, computing (e.g., calculating, comparing, or otherwise analyzing) using a processor, and/or outputting results via a display or other user interface.

In some embodiments, the step of determining comprises calculating a risk-score indicative of the likelihood that the subject has lung cancer. In some embodiments, computing the risk-score involves determining the combination of weighted expression levels (e.g., expression levels of one or more informative-genes alone or together with one of more genomic correlate genes), in which the expression levels are weighted by their relative contribution to predicting increased likelihood of having lung cancer. In some embodiments, genomic correlate genes are genes related to or correlated with specific clinical variables (e.g., self-reportable variables). In some embodiments, such clinical variables are correlated with cancer, e.g., lung cancer. In some embodiments, rather than using expression levels of genes, groups of related genes that vary collinearly (e.g., are correlated with one another) within a population of subjects may be combined or collapsed into a single value (e.g., the mean value of a group of related genes). In some embodiments, a computer-implemented method comprises generating a report that indicates the risk-score. In some embodiments, the report is transmitted to a health care provider of the subject.

In some embodiments, a computer-implemented method comprises obtaining data representting expression levels in a biological sample of at least 2, 3, 4, 5, 6, 7, 8, 9, or 10 genes selected from the set of genes identified in cluster 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, or 11 in table 11. In some embodiments, the genes comprise MYOT.

It should be appreciated that in any embodiment or aspect described herein, a biological sample can be obtained from the respiratory epithelium of the subject. The respiratory epithelium can be of the mouth, nose, pharynx, trachea, bronchi, bronchioles, or alveoli. However, other sources of respiratory epithelium also can be used. The biological sample can comprise histologically normal tissue. The biological sample can be obtained using bronchial brushings, such as cytobrush or histobrush; broncho-alveolar lavage; bronchial biopsy; oral washings; touch preps; fine needle aspirate; or sputum collection. The subject can exhibit one or more symptoms of lung cancer and/or have a lesion that is observable by computer-aided tomography or chest X-ray. In some cases, the subject has not been diagnosed with primary lung cancer prior to being evaluating by methods disclosed herein.

In any of the embodiments or aspects described herein, the expression levels can be determined using a quantitative reverse transcription polymerase chain reaction, a bead-based nucleic acid detection assay or an oligonucleotide array assay (e.g., a microarray assay) or other technique.

In any of the embodiments or aspects described herein, the lung cancer can be a adenocarcinoma, squamous cell carcinoma, small cell cancer or non-small cell cancer.

In some embodiments, aspects of the disclosure relate to a composition consisting essentially of at least one nucleic acid probe, wherein each of the at least one nucleic acid probes specifically hybridizes with an informative-gene (e.g., at least one informative-mRNA selected from Table 11).

In some embodiments, aspects of the disclosure relate to a composition comprising up to 5, up to 10, up to 25, up to 50, up to 100, or up to 200 nucleic acid probes, wherein each of the nucleic acid probes specifically hybridizes with an informative-gene (e.g., at least one informative-mRNA selected from Table 1 or 11).

In some embodiments, a composition comprises at least 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10 nucleic acid probes. In some embodiments, at least 2, 3, 4, 5, 6, 7, 8, 9, or 10 of the nucleic acid probes hybridize with an mRNA expressed from a different gene selected from clusters 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, or 11 of Table 11.

In some embodiments, nucleic acid probes are conjugated directly or indirectly to a bead. In some embodiments, the bead is a magnetic bead. In some embodiments, the nucleic acid probes are immobilized to a solid support. In some embodiments, the solid support is a glass, plastic or silicon chip.

In some embodiments, aspects of the disclosure relate to a kit comprising at least one container or package housing any nucleic acid probe composition described herein.

In some embodiments, expression levels are determined using a quantitative reverse transcription polymerase chain reaction.

In some embodiments, aspects of the disclosure relate to genes for which expression levels can be used to determine the likelihood that a subject (e.g., a human subject) has lung cancer. In some embodiments, the expression levels (e.g., mRNA levels) of one or more genes described herein can be determined in airway samples (e.g., epithelial cells or other samples obtained during a bronchoscopy or from an appropriate bronchial lavage samples). In some embodiments, the patterns of increased and/or decreased mRNA expression levels for one or more subsets of informative-genes (e.g., 1-5, 5-10, 10-15, 15-20, 20-25, 25-50, 50-80, or more genes) described herein can be determined and used for diagnostic, prognostic, and/or therapeutic purposes. It should be appreciated that one or more expression patterns described herein can be used alone, or can be helpful along with one or more additional patient-specific indicia or symptoms, to provide personalized diagnostic, prognostic, and/or therapeutic predictions or recommendations for a patient. In some embodiments, sets of informative-genes that distinguish smokers (current or former) with and without lung cancer are provided that are useful for predicting the risk of lung cancer with high accuracy. In some embodiments, the informative-genes are selected from Table 1 or 11.

In some embodiments, methods provided herein for determining the likelihood that a subject has lung cancer involve subjecting a biological sample obtained from a subject to a gene expression analysis that comprises determining mRNA expression levels in the biological sample of one or more informative-genes that relate to lung cancer status (e.g., an informative gene selected from Table 1 or 11). In some embodiments, the methods comprise determining mRNA expression levels in the biological sample of one or more genomic correlate genes that relate to one or more self-reportable characteristics of the subject. In some embodiments, the methods further comprise transforming the expression levels determined above into a lung cancer risk-score that is indicative of the likelihood that the subject has lung cancer. In some embodiments, the one or more self-reportable characteristics of the subject are selected from: smoking pack years, smoking status, age and gender. In some embodiments, the lung cancer risk-score is determined according to the follow equation:

Score
=

ex

Score
1

⁢

or
⁢

Score
⁢

2

(

1
+

ex

Score
1

⁢

or
⁢

Score
⁢

2

)

wherein
:

X

score
⁢

1

=

W
0

+

W
1

×
G
⁢
G

+

W
2

×
G
⁢
S

+

W
3

×
G
⁢
P
⁢
Y

+

W
4

×
G
⁢
A

+

W
5

×
 
C
⁢
1
⁢
A

+

W
6

×
C
⁢
1
⁢
B

+

W
7

×
C
⁢
2

+

W
8

×
C
⁢
3

+

W
9

×
C
⁢
4
⁢
A

+

W

1
⁢
0

×
C
⁢
4
⁢
B

and

x

score
⁢

2

=

W
0

+

W
1

×
G
⁢
G

+

W
2

×
G
⁢
S

+

W
3

×
G
⁢
P
⁢
Y

+

W
4

×
Reported
⁢

Age

+
 

W
5

×
C
⁢
1
⁢
A

+

W
6

×
C
⁢
1
⁢
B

+

W
7

×
C
⁢
2

+

W
8

×
C
⁢
3

+

W
9

×
C
⁢
4
⁢
A

+

W

1
⁢
0

×
C
⁢
4
⁢
B

,

- and in which GG, GS, GPY, GA, C1A, C1B, C2, C3, C4A and C4B are determined according to the equations disclosed herein.

In some embodiments, informative-genes are selected from Table 1 or 11. In some embodiments, groups of related genes that vary collinearly (e.g., are correlated with one another) within a population of subjects may be combined or collapsed into a single value (e.g., the mean value of a group of related genes). In some embodiments, groups of related genes are correlated because they are associated with the same cellular and/or molecular pathways. In some embodiments, at least 2, at least 3, at least 4, at least 5 or more related genes (e.g., correlated genes, genes within a common cluster) are combined together in a single value. In some embodiments, groups of related genes are identified by performing a cluster analysis of expression levels obtained from multiple subjects (e.g., 2 to 100, 2 to 500, 2 to 1000 or more subjects). Any appropriate cluster analysis may be used to identify such related genes including, for example, centroid based clustering (e.g., k-means clustering), connectivity based clustering (e.g., hierarchical clustering) and other suitable approaches. Non-limiting examples of such clusters are identified in Table 11 with the values in column 2 specifying the cluster within which each gene resides such that related genes (e.g., correlated genes) are within the same cluster. In some embodiments, a value reflecting the expression status of a set of related genes is the mean expression level of the set of related genes. For example, one or more of the following values may be used: C1A, C1B, C2, C3, C4A, and C4B in a model for predicting the likelihood that a subject has cancer, in which

- C1A=mean of (BST1, CD177.1, CD177.2),
- C1B=mean of (ATP12A, TSPAN2),
- C2=mean of (GABBRI, MCAM, NOVA1, SDC2),
- C3=mean of (CDR1, CGREF1, CLND22, NKX3-1),
- C4A=mean of (EPHX3, LYPD2), and
- C4B=mean of (MIA, RNF150).

In some embodiments genes within a cluster can be substituted for each other. Thus, in some embodiments, all genes within a cluster need to be evaluated or used in a prediction model. In some embodiments, only 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10 genes within a cluster are independently selected for analysis as described herein. In some embodiments, at least 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10 genes within a cluster of table 11 are identified.

In some embodiments, one or more informative-genes are selected from the set of genes identified as cluster 1 in Table 11. In some embodiments, one or more informative-genes are selected from the set of genes identified as cluster 2 in Table 11. In some embodiments, one or more informative-genes are selected from the set of genes identified as cluster 3 in Table 11. In some embodiments, one or more informative-genes are selected from the set of genes identified as cluster 4 in Table 11. In some embodiments, one or more informative-genes are selected from the set of genes identified as cluster 5 in Table 11. In some embodiments, one or more informative-genes are selected from the set of genes identified as cluster 6 in Table 11. In some embodiments, one or more informative-genes are selected from the set of genes identified as cluster 7 in Table 11. In some embodiments, one or more informative-genes are the set of genes identified as cluster 8 in Table 11. In some embodiments, one or more informative-genes are selected from the set of genes identified as cluster 9 in Table 11. In some embodiments, one or more informative-genes are selected from the set of genes identified as cluster 10 in Table 11. In some embodiments, one or more informative-genes are selected from the set of genes identified as cluster 11 in Table 11. In some embodiments, the informative-genes comrise MYOT. In some embodiments, genes selected from a cluster are reduced to a single value, such as, for example, the mean, median, mode or other summary statistic of the expression levels of the selected genes.

In some embodiments, provided herein are methods for establishing appropriate diagnostic intervention plans and/or treatment plans for subjects and for aiding healthcare providers in establishing appropriate diagnostic intervention plans and/or treatment plans. In some embodiments, methods are provided that involve making a risk assessment based on expression levels of informative-genes in a biological sample obtained from a subject during a routine cell or tissue sampling procedure. In some embodiments, methods are provided that involve establishing lung cancer risk scores based on expression levels of informative genes. In some embodiments, appropriate diagnostic intervention plans are established based at least in part on the lung cancer risk scores. In some embodiments, methods provided herein assist health care providers with making early and accurate diagnoses. In some embodiments, methods provided herein assist health care providers with establishing appropriate therapeutic interventions early on in patients' clinical evaluations. In some embodiments, methods provided herein involve evaluating biological samples obtained during bronchoscopies procedure. In some embodiments, the methods are beneficial because they enable health care providers to make informative decisions regarding patient diagnosis and/or treatment from otherwise uninformative bronchoscopies. In some embodiments, the risk assessment leads to appropriate surveillance for monitoring low risk lesions. In some embodiments, the risk assessment leads to faster diagnosis, and thus, faster therapy for certain cancers.

Provided herein are methods for determining the likelihood that a subject has lung cancer, such as adenocarcinoma, squamous cell carcinoma, small cell cancer or non-small cell cancer. The methods alone or in combination with other methods provide useful information for health care providers to assist them in making diagnostic and therapeutic decisions for a patient. The methods disclosed herein are often employed in instances where other methods have failed to provide useful information regarding the lung cancer status of a patient. For example, approximately 50% of bronchoscopy procedures result in indeterminate or non-diagnostic information. There are multiple sources of indeterminate results, and may depend on the training and procedures available at different medical centers. However, in certain embodiments, molecular methods in combination with bronchoscopy are expected to improve cancer detection accuracy.

In some embodiments, provided herein are methods of determining the likelihood that a subject has lung cancer. In some embodiments, methods are provided that involve subjecting a biological sample obtained from a subject to a gene expression analysis, wherein the gene expression analysis comprises measuring cDNA levels of one or more informative-genes that relate to lung cancer status, and measuring cDNA levels of ore or more genomic correlate genes that relate to one or more self-reportable characteristics of the subject; and determining a lung cancer risk-score based on the cDNA levels determined in (a) and (b), that is indicative of the likelihood that the subject has lung cancer; wherein the cDNA is prepared from mRNA from the biological sample.

In some embodiments, the methods of the present disclosure include the conversion of mRNA into cDNA. In further embodiments, cDNA is amplified.

These and other aspects are described in more detail herein and are illustrated by the non-limiting figures and examples.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a non-limiting example of a plot of correlation coefficients from pairwise correlation of all gene expression data of all qualified AEGIS I samples; samples with a correlation coefficient <0.955 were identified as outliers and excluded from further analysis; a total of 597 samples were retained;

FIG. 2 is a non-limiting example of ROC curves for prediction models based on a set of training samples; and

FIG. 3 is a non-limiting example of ROC curves for prediction models based on a set of bronchoscopy negative training samples.

FIG. 4 depicts the following color-coding: patients that met inclusion criteria of the study (blue); patients who were excluded (yellow); patients who were included in the final analysis (green).

FIG. 5 depicts ROC curves for total patients (light gray) and the subset of patients with a non-diagnostic bronchoscopy (black) in the AEGIS 1 (left) and AEGIS 2 (right) cohorts is shown. In AEGIS 1 the AUC=0.78 (95% CI, 0.73 to 0.83) and AUC=0.76 (95% CI, 0.68 to 0.83) for the two groups, respectively (p=0.31). In AEGIS 2 the AUC=0.74 (95% CI, 0.68 to 0.80) and AUC=0.75 (95% CI, 0.68 to 0.82) in the two groups (p=0.85). The AUC was also not significantly different for patients with a non-diagnostic bronchoscopy comparing AEGIS 1 and AEGIS 2 (p=0.61).

FIG. 6A and FIG. 6B. FIG. 6A depicts a post-test POM related to pre-test POM based on a negative classifier call (solid line; adjusted using the negative likelihood ratio) and a positive classifier call (dotted line; adjusted using the positive likelihood ratio) calculated for the classifier in combination with bronchoscopy. The negative classifier call curve shows that for patients with a pre-test POM of <66%, the post-test POM is <10% when bronchoscopy is negative and the classifier is negative. For patients with a negative bronchoscopy and a positive classifier score, the post-test likelihood of cancer is >10% when the pre-test likelihood is greater than 5%. FIG. 6B depicts post test probability of cancer based on the pretest probability and the negative likelihood ratio of the classifier and bronchoscopy. The posttest probability of lung cancer is shown in relation to the pretest probability based on a nondiagnostic bronchoscopic examination and a negative classifier score (adjusted with the use of the negative likelihood ratio). The curve shows that for patients with a pretest probability of cancer of less than 66% (short vertical line), the posttest probability is less than 10% (broken line) when bronchoscopic findings are negative and the classifier score is negative.

FIG. 7 depicts a pairwise correlation of genes with cancer-associated gene expression. The correlation between all possible pairs of genes with cancer-associated gene expression (n=232) were assessed to identify groups of genes that share a similar pattern of gene expression. Unsupervised hierarchical clustering was used to group correlated genes into 11 clusters, with the dendrogram threshold level to establish clusters indicated on the y-axis (green line). Genes were selected from the clusters in a parsimonious manner to predict lung cancer status using linear regression. The classifier genes came from specific clusters (outlined in blue), using 2-4 genes from each cluster. Clusters 4 and 7 contain genes which were up-regulated in lung cancer, and clusters 1, 2, 9, and 10 were down-regulated in lung cancer.

FIG. 8 depicts an ROC curve of patients with a non-diagnostic bronchoscopy in the test set. The AUC=0.81 for the 123 patients whose bronchoscopy did not result in a diagnosis of lung cancer (in which the prevalence of lung cancer=31%).

FIG. 9 depicts gene expression data corresponding to all patients in the training set (black line), and the subset of patients with a non-diagnostic bronchoscopy (grey line) were analyzed using the locked classifier. The AUC was calculated as 0.78 (95% CI, 0.73-0.82) and 0.78 (95% CI, 0.71-0.85), for the two groups respectively.

FIG. 10A and FIG. 10B depicts nucleic acid probes used in hybridizing to nucleic acid sequences represented by gene classifier CD177. FIG. 10A discloses the 19 nucleic acid probes in CD177.1 (SEQ ID NOs: 24-42 in order from top to bottom) and FIG. 10B discloses the 4 nucleic acid probes in CD177.2 (SEQ ID NOs: 43-46 from top to bottom).

DETAILED DESCRIPTION OF THE INVENTION

Methods disclosed herein provide alternative or complementary approaches for evaluating cell or tissue samples obtained by bronchoscopy procedures (or other procedures for evaluating respiratory tissue), and increase the likelihood that the procedures will result in useful information for managing the patient's care. The methods disclosed herein are highly sensitive, and produce information regarding the likelihood that a subject has lung cancer from cell or tissue samples (e.g., bronchial brushings of airway epithelial cells), which are often obtained from regions in the airway that are remote from malignant lung tissue. In general, the methods disclosed herein involve subjecting a biological sample obtained from a subject to a gene expression analysis to evaluate gene expression levels. However, in some embodiments, the likelihood that the subject has lung cancer is determined in further part based on the results of a histological examination of the biological sample or by considering other diagnostic indicia such as protein levels, mRNA levels, imaging results, chest X-ray exam results etc.

The term “subject,” as used herein, generally refers to a mammal. Typically the subject is a human. However, the term embraces other species, e.g., pigs, mice, rats, dogs, cats, or other primates. In certain embodiments, the subject is an experimental subject such as a mouse or rat.

The subject may be a male or female. The subject may be an infant, a toddler, a child, a young adult, an adult or a geriatric. The subject may be a smoker, a former smoker or a non-smoker. The subject may have a personal or family history of cancer. The subject may have a cancer-free personal or family history. The subject may exhibit one or more symptoms of lung cancer or other lung disorder (e.g., emphysema, COPD). For example, the subject may have a new or persistent cough, worsening of an existing chronic cough, blood in the sputum, persistent bronchitis or repeated respiratory infections, chest pain, unexplained weight loss and/or fatigue, or breathing difficulties such as shortness of breath or wheezing. The subject may have a lesion, which may be observable by computer-aided tomography or chest X-ray. The subject may be an individual who has undergone a bronchoscopy or who has been identified as a candidate for bronchoscopy (e.g., because of the presence of a detectable lesion or suspicious imaging result). In some embodiments, a subject has or has been diagnosed with chronic obstructive pulmonary disease (COPD). In some embodiments, a subject does not have or has not been diagnosed with COPD. A subject under the care of a physician or other health care provider may be referred to as a “patient.”

The term “about”, as used herein, refers to plus or minus ten percent of the object that “about” modifies. Thus, the phrase “about 10, 20, or 30” encompasses 8-11, 18-22, 27-33,respectively.

Informative-Genes

The expression levels of genes of the present disclosure have been identified as providing useful information regarding the lung cancer status of a subject. These genes are referred to herein as “informative-genes.” Informative-genes include protein coding genes and non-protein coding genes. It will be appreciated by the skilled artisan that the expression levels of informative-genes may be determined by evaluating the levels of appropriate gene products (e.g., mRNAs, miRNAs, proteins etc.) Accordingly, the expression levels of certain mRNAs have been identified as providing useful information regarding the lung cancer status of a subject. These mRNAs are referred to herein as “informative-mRNAs.”

Table 11 provides a listing of informative-genes that are differentially expressed in cancer. In some embodiments, informative-genes that are differentially expressed in lung cancer are selected from: BST1, CD177.1, CD177.2, ATP12A, TSPAN2, GABBR1, MCAM, NOVA1, SDC2, CDR1, CGREF1, CLND22, NKX3-1, EPHX3, LYPD2, MIA, RNF150. In some embodiments, informative-genes that are differentially expressed in lung cancer are selected from: TMEM51, CR1L, PDZKIIP1, MICAL2, VWASA, ACAD8, SAA4, GLYATL2, ETV6, CD177, CEACAM7, QPCT, CASP10, PI3, BSTI, MTNR1A, STARD4, CFB, SLC26A8, VNN2, HDAC9, SLC26A4, and LCN2. In some embodiments, informative-genes that are differentially expressed in lung cancer are selected from: CCDC18, FAM72D, NUF2, FBXO28, GPR137B, STIL, DEPDC1, TSPAN2, ASPM, KIF14, KIF20B, RAD51AP1, GAS2L3, SPIC, SMAGP, ATP12A, BRCA2, BORA, SKA3, DLGAP5, CASC5, LRRC28, PYCARD, TXNL4B, EFCAB5, SPAG5, ABCA12, AURKA, SGOL1, BANK1, CENPE, CASP6, MAD2L1, CCNA2, CCNB1, KIF20A, CENPK, ERAP1, FAM54A, PHTF2, CLDN12, BPGM, PCMTD1, MELK, and MST4. In some embodiments, informative-genes that are differentially expressed in lung cancer are selected from: CR1, GOS2, CSF3R, S100A12, SELL, NCF2, LIPN, ZNF438, NAMPT, CBL, CASP5, CARD16, CARD17, CLEC4A, LRRK2, HMGN2P46, AQP9, BCL2A1, ITGAX, GPR97, CCL4, PSTPIP2, IFI30, FFAR2, EMR3, FPR1, LILRA5, PLEK, MXD1, TNFAIP6, CXCR2, IL1B, CXCR1, SIRPB1, NCF4, IRAK2, PROK2, TLR2, TREM1, SOD2, CREB5, TNFRSF10C, CSGALNACT1, and ASAP 1. In some embodiments, informative-genes that are differentially expressed in lung cancer are selected from: PLA2G2A, NFYC, RASSF10, GLB1L3, TRIM3, MCAM, MSRB3, SLITRK5, GAS6, NOVA1, GABRG3, ABCA3, LPO, FSCN2, RASD1, HILS1, SDK2, NTN5, KCNA7, ATOH8, KCNIP3, INHBB, VSTM2L, ZNRF3, PLEKHG4B, GNMT, GABBR1, ARHGEF10, SDC2, CRB2, GAS1, PNPLA7, and RAI2.

Certain methods disclosed herein involve determining expression levels in the biological sample of at least one informative-gene. However, in some embodiments, the expression analysis involves determining the expression levels in the biological sample of at least 2, at least 3, at least 4, at least 5, at least 6, at least 7, at least 8, at least 9, at least 10, at least 20, at least 30, at least 40, at least 50, at least 60, at least 70, or least 80 informative-genes. In some embodiments, the expression analysis involves determining expression levels in the biological sample of 1 to 5, 1 to 10, 5 to 10, 5 to 15, 10 to 15, 10 to 20, 15 to 20, 15 to 25, 20 to 30, 25 to 50, 25 to 75, 50 to 100, 50 to 200 or more informative-genes, such as those in table 11. In some embodiments, the expression analysis involves determining expression levels in the biological sample to at least 1 to 5, 1 to 10, 2 to 10, 5 to 10, 5 to 15, 10 to 15, 10 to 20, 15 to 20, 15 to 25, 20 to 30, 25 to 50, 25 to 75, 50 to 100, 50 to 200 or more informative-genes, such as those in table 11.

In some embodiments, the number of informative-genes for an expression analysis are sufficient to provide a level of confidence in a prediction outcome that is clinically useful. This level of confidence (e.g., strength of a prediction model) may be assessed by a variety of performance parameters including, but not limited to, the accuracy, sensitivity specificity, and area under the curve (AUC) of the receiver operator characteristic (ROC) curve. These parameters may be assessed with varying numbers of features (e.g., number of genes, mRNAs) to determine an optimum number and set of informative-genes. An accuracy, sensitivity or specificity of at least 60%, 70%, 80%, 90%, may be useful when used alone or in combination with other information.

Any appropriate system or method may be used for determining expression levels of informative-genes. Gene expression levels may be determined through the use of a hybridization-based assay. As used herein, the term, “hybridization-based assay” refers to any assay that involves nucleic acid hybridization. A hybridization-based assay may or may not involve amplification of nucleic acids. Hybridization-based assays are well known in the art and include, but are not limited to, array-based assays (e.g., oligonucleotide arrays, microarrays), oligonucleotide conjugated bead assays (e.g., Multiplex Bead-based Luminex® Assays), molecular inversion probe assays, and quantitative RT-PCR assays. Multiplex systems, such as oligonucleotide arrays or bead-based nucleic acid assay systems are particularly useful for evaluating levels of a plurality of genes simultaneously. Other appropriate methods for determining levels of nucleic acids will be apparent to the skilled artisan.

As used herein, a “level” refers to a value indicative of the amount or occurrence of a substance, e.g., an mRNA. A level may be an absolute value, e.g., a quantity of mRNA in a sample, or a relative value, e.g., a quantity of mRNA in a sample relative to the quantity of the mRNA in a reference sample (control sample). The level may also be a binary value indicating the presence or absence of a substance. For example, a substance may be identified as being present in a sample when a measurement of the quantity of the substance in the sample, e.g., a fluorescence measurement from a PCR reaction or microarray, exceeds a background value. Similarly, a substance may be identified as being absent from a sample (or undetectable in the sample) when a measurement of the quantity of the molecule in the sample is at or below background value. It should be appreciated that the level of a substance may be determined directly or indirectly.

Further non-limiting examples of informative mRNAs are disclosed in, for example, the following patent applications, the contents of which are incorporated herein by reference in their entirety for all purposes: U.S. Patent Publication No. US2007/148650, filed on May 12, 2006, entitled ISOLATION OF NUCLEIC ACID FROM MOUTH EPITHELIAL CELLS; U.S. Patent Publication No. US2009/311692, filed Jan. 9, 2009, entitled ISOLATION OF NUCLEIC ACID FROM MOUTH EPITHELIAL CELLS; U.S. application Ser. No. 12/884,714, filed Sep. 17, 2010, entitled ISOLATION OF NUCLEIC ACID FROM MOUTH EPITHELIAL CELLS; U.S. Patent Publication No. US2006/154278, filed Dec. 6, 2005, entitled DETECTION METHODS FOR DISORDER OF THE LUNG; U.S. Patent Publication No. US2010/035244, filed Feb. 8, 2008, entitled, DIAGNOSTIC FOR LUNG DISORDERS USING CLASS PREDICTION; U.S. application Ser. No. 12/869,525, filed Aug. 26, 2010, entitled, DIAGNOSTIC FOR LUNG DISORDERS USING CLASS PREDICTION; U.S. application Ser. No. 12/234,368, filed Sep. 19, 2008, entitled, BIOMARKERS FOR SMOKE EXPOSURE; U.S. application Ser. No. 12/905,897, filed Oct. 154, 2010, entitled BIOMARKERS FOR SMOKE EXPOSURE; U.S. Patent Application No. US2009/186951, filed Sep. 19, 2008, entitled IDENTIFICATION OF NOVEL PATHWAYS FOR DRUG DEVELOPMENT FOR LUNG DISEASE; U.S. Publication No. US2009/061454, filed Sep. 9, 2008, entitled, DIAGNOSTIC AND PROGNOSTIC METHODS FOR LUNG DISORDERS USING GENE EXPRESSION PROFILES; U.S. application Ser. No. 12/940,840, filed Nov. 5, 2010, entitled, DIAGNOSTIC AND PROGNOSTIC METHODS FOR LUNG DISORDERS USING GENE EXPRESSION PROFILES; U.S. Publication No. US2010/055689, filed Mar. 30, 2009, entitled, MULTIFACTORIAL METHODS FOR DETECTING LUNG DISORDERS; and International Patent Application No. PCT/US13/38449, filed Apr. 26, 2013, entitled METHODS FOR EVALUATING LUNG CANCER STATUS.

CDNA

cDNA molecules are non-naturally occurring polynucleotide sequences that are synthesized from mRNA molecules by one possessing ordinary skill in the art. In some embodiments, cDNA molecules of the present invention are obtained or acquired. The conversion of RNA to cDNA utilizing a reverse transcriptase enzyme creates cDNA, a non-naturally occurring molecule that lacks introns. Methods that rely on cDNA are necessarily relying on an artificial molecule that does not naturally occur in nature, e.g. protein expression of cDNA molecules or hybridization of cDNA molecules.

In certain aspects, mRNA in a biological sample is used to produce cDNA from a sample by reverse transcription of mRNA using at least one primer; amplifying the cDNA using polynucleotides as sense and antisense primers to amplify cDNAs therein; and detecting the presence of the amplified cDNA. In further aspects, the sequence of the amplified cDNA can be determined by any suitable method.

In one embodiment, once the mRNA is obtained from a sample, it is converted to complementary DNA (cDNA). cDNA does not exist in vivo and therefore is a non-natural molecule. In a further embodiment, the cDNA is then amplified, for example, by the polymerase chain reaction (PCR) or other amplification method known to those of ordinary skill in the art. The product of this amplification reaction, i.e., amplified cDNA is necessarily a non-natural product. As mentioned above, cDNA is a non-natural molecule. Second, in the case of PCR, the amplification process serves to create hundreds of millions of cDNA copies for every individual cDNA molecule of starting material. The number of copies generated are far removed from the number of copies of mRNA that are present in vivo.

In one embodiment, cDNA is amplified with primers that introduce an additional DNA sequence (adapter sequence) onto the fragments (with the use of adapter-specific primers). Amplification therefore serves to create non-natural double stranded molecules from the non-natural single stranded cDNA, by introducing barcode, adapter and/or reporter sequences onto the already non-natural cDNA. In one embodiment, during amplification with the adapter-specific primers, a detectable label, e.g., a fluorophore, is added to single strand cDNA molecules. Amplification therefore also serves to create DNA complexes that do not occur in nature, at least because (i) cDNA does not exist in vivo, (i) adapter sequences are added to the ends of cDNA molecules to make DNA sequences that do not exist in vivo, (ii) the error rate associated with amplification further creates DNA sequences that do not exist in vivo, (iii) the disparate structure of the cDNA molecules as compared to what exists in nature and (iv) the chemical addition of a detectable label to the cDNA molecules.

In one embodiment, the synthesized cDNA (for example, amplified cDNA) is immobilized on a solid surface via hybridization with a probe, e.g., via a microarray. In another embodiment, cDNA products are detected via real-time polymerase chain reaction (PCR) via the introduction of fluorescent probes that hybridize with the cDNA products. For example, in one embodiment, biomarker detection is assessed by quantitative fluorogenic RT-PCR (e.g., with TaqMan® probes). For PCR analysis, well known methods are available in the art for the determination of primer sequences for use in the analysis.

In one embodiment, to synthesize and amplify cDNAs, the 5′Ampli FINDER RACE kit (Manufactured by Clontech) and the 5′-RACE method using PCR (Frohman, M. A. et al., Proc. Natl. Acad. Sci. USA (1988) 85: 8998-9002; Belyaysky, A. et al., Nucleic Acids Research. (1989) 17: 2919-2932) can be used. In the process of such cDNA synthesis, restriction enzyme sites can be introduced into both ends of the cDNA.

Genomic Correlates

As disclosed herein, the expression levels of certain genes have been identified as being related to (correlated with) certain self-reportable characteristics of a subject. Such genes are referred to herein as “genomic correlate genes” or “genomic correlates” and are useful because they provide a surrogate marker for characteristics of a subject that could otherwise be incorrectly and/or inaccurately reported. For example, in some embodiments, a subject may incorrectly estimate information such as pack years, smoking status or age (e.g., by providing an underestimate of such information). In such embodiments, the use of a prediction model based on genomic correlate genes can reduce or eliminate variability associated with incorrect reporting because it is based on the expression of the genomic correlate genes rather than a subject's decision making about what information to report and/or a subject's recollection of circumstances. It will be appreciated by the skilled artisan that the expression levels of such genomic correlate genes may be determined by evaluating the levels of appropriate gene products (e.g., mRNAs, miRNAs, proteins etc.) Expression levels of genomic correlate genes may be determined in parallel with informative-genes of lung cancer status (e.g., an informative gene selected from Table 11) or independently of such genes.

In some embodiments, genomic correlates reflect a response of an individual to an environmental hazard (e.g., cigarette smoke). In some embodiments, genomic correlates reflect exposure to a hazard.

In some embodiment gender of a subject is determined based on one or more genomic correlate genes. In some embodiments, a genomic correlate gene related to gender is RPS4Y1. In some embodiments, if the expression of RPS4Y1 is below a threshold then the subject is identified as being a male and if the expression of RPS4Y1 is above the threshold the subject is identified as being a female. In some embodiments, a threshold is a relative expression level that accurately differentiates males and females for the gene(s) of interest.

In some embodiment smoking status (e.g., current or former) of a subject is determined based on one or more genomic correlate genes. In some embodiments, a genomic correlate gene related to smoking status is SLC7A11, CLND10 or TKT. In some embodiments, the smoking status of a subject is determined according to the following model: smoking status (also, referred to as Genomic Smoking (GS))=exp(x)/(1+exp(x)), in which

x
=

β
⁢

G
⁢
S

0

+

β
⁢

G
⁢
S

1

*
SLC
⁢
7
⁢
A
⁢
11

+

β
⁢

G
⁢
S

2

*
CLND
⁢
10

+

β
⁢

G
⁢
S

3

*
TKT

,

- in which

β
⁢

G
⁢
S

n

- are regression weights for the regression model and gene symbols represent the relative expression intensity of each respective gene. In some embodiments, a smoker is a subject who has smoked at least 100 cigarettes in a lifetime. In some embodiments, a former smoker is a subject who quit or who has not smoked a cigarette within 1 month prior to bronchoscopy.

In some embodiment, smoking history of a subject is determined based on one or more genomic correlate genes. In some embodiments, a genomic correlate gene related to smoking history is AKR1C2 or RUNX1T1. In some embodiments, the smoking history of a subject is determined according to the following model: smoking history (also, referred to as Genomic Pack Years (GPY))=exp(x)/(1+exp(x)), in which

x
=

β
⁢

G
⁢
P
⁢
Y

0

+

β
⁢

G
⁢
P
⁢
Y

1

*
RUNX
⁢
1
⁢
T
⁢
1

+

β
⁢

G
⁢
P
⁢
Y

2

*
AKR
⁢
1
⁢
C
⁢
2

,

- in which

β
⁢

G
⁢
P
⁢
Y

n

,

- are regression weights for the model and gene symbols represent the relative expression intensity of each respective gene.

In some embodiment, age of a subject is determined based on one or more genomic correlate genes. In some embodiments, a genomic correlate gene related to age is CD52, SYT8, TNNT3, ALX1, KLRK1, RASA3, CERS3, ASPA, GRP, APOC1, EPHX3, REEP1, FAM198B, PCDHB4, PCDHB16, FOXD1, SPARC, NKAPL, or GPR110. In some embodiments, the age of a subject is determined according to the following model: age (also referred to as genomic age

(

G
⁢
A

)

)

=

β
⁢

G
⁢
A

0

+

β
⁢

G
⁢
A

1

*
CD
⁢
52

+

β
⁢

G
⁢
A

2

*
SYT
⁢
8

+

β
⁢

G
⁢
A

3

*
TNNT
⁢
3

+
 

β
⁢

G
⁢
A

4

*
ALX
⁢
1

+

β
⁢

G
⁢
A

5

*
KLRK
⁢
1

+

β
⁢

G
⁢
A

6

*
RASA
⁢
3

+
 

β
⁢

G
⁢
A

7

*
CERS
⁢
3

+

β
⁢

G
⁢
A

8

*
ASPA

+

β
⁢

G
⁢
A

9

*
G
⁢
R
⁢
P

+

β
⁢

G
⁢
A

1
⁢
0

*
APOC
⁢
1

+
 

β
⁢

G
⁢
A

1
⁢
1

*
EPHX
⁢
3

+

β
⁢

G
⁢
A

1
⁢
2

*
REEP
⁢
1

+

β
⁢

G
⁢
A

1
⁢
3

*
FAM
⁢
198
⁢
B

+
 

β
⁢

G
⁢
A

1
⁢
4

*
PCDHB
⁢
4

+

β
⁢

G
⁢
A

1
⁢
5

*
PCDHB
⁢
16

+

β
⁢

G
⁢
A

1
⁢
6

*
FOXD
⁢
1

+
 

β
⁢

G
⁢
A

1
⁢
7

*
SPARC
⁢

β
⁢

G
⁢
A

1
⁢
8

*
NKAPL

+

β
⁢

G
⁢
A

1
⁢
9

*
GPR
⁢
110

,

- in which

β
⁢

G
⁢
A

n

- are regression weights for the model and gene symbols represent the relative expression intensity of each respective gene.

Biological Samples

The methods generally involve obtaining a biological sample from a subject. As used herein, the phrase “obtaining a biological sample” refers to any process for directly or indirectly acquiring a biological sample from a subject. For example, a biological sample may be obtained (e.g., at a point-of-care facility, a physician's office, a hospital) by procuring a tissue or fluid sample from a subject. Alternatively, a biological sample may be obtained by receiving the sample (e.g., at a laboratory facility) from one or more persons who procured the sample directly from the subject.

The term “biological sample” refers to a sample derived from a subject, e.g., a patient. A biological sample typically comprises a tissue, cells and/or biomolecules. In some embodiments, a biological sample is obtained on the basis that it is histologically normal, e.g., as determined by endoscopy, e.g., bronchoscopy. In some embodiments, biological samples are obtained from a region, e.g., the bronchus or other area or region, that is not suspected of containing cancerous cells. In some embodiments, a histological or cytological examination is performed. However, it should be appreciated that a histological or cytological examination may be optional. In some embodiments, the biological sample is a sample of respiratory epithelium. The respiratory epithelium may be of the mouth, nose, pharynx, trachea, bronchi, bronchioles, or alveoli of the subject. The biological sample may comprise epithelium of the bronchi. In some embodiments, the biological sample is free of detectable cancer cells, e.g., as determined by standard histological or cytological methods. In some embodiments, histologically normal samples are obtained for evaluation. Often biological samples are obtained by scrapings or brushings, e.g., bronchial brushings. However, it should be appreciated that other procedures may be used, including, for example, brushings, scrapings, broncho-alveolar lavage, a bronchial biopsy or a transbronchial needle aspiration.

It is to be understood that a biological sample may be processed in any appropriate manner to facilitate determining expression levels. For example, biochemical, mechanical and/or thermal processing methods may be appropriately used to isolate a biomolecule of interest, e.g., RNA, from a biological sample. Accordingly, a RNA or other molecules may be isolated from a biological sample by processing the sample using methods well known in the art.

Lung Cancer Assessment

Methods disclosed herein may involve comparing expression levels of informative-genes with one or more appropriate references. An “appropriate reference” is an expression level (or range of expression levels) of a particular informative-gene that is indicative of a known lung cancer status. An appropriate reference can be determined experimentally by a practitioner of the methods or can be a pre-existing value or range of values. An appropriate reference represents an expression level (or range of expression levels) indicative of lung cancer. For example, an appropriate reference may be representative of the expression level of an informative-gene in a reference (control) biological sample obtained from a subject who is known to have lung cancer. When an appropriate reference is indicative of lung cancer, a lack of a detectable difference (e.g., lack of a statistically significant difference) between an expression level determined from a subject in need of characterization or diagnosis of lung cancer and the appropriate reference may be indicative of lung cancer in the subject. When an appropriate reference is indicative of lung cancer, a difference between an expression level determined from a subject in need of characterization or diagnosis of lung cancer and the appropriate reference may be indicative of the subject being free of lung cancer.

Alternatively, an appropriate reference may be an expression level (or range of expression levels) of a gene that is indicative of a subject being free of lung cancer. For example, an appropriate reference may be representative of the expression level of a particular informative-gene in a reference (control) biological sample obtained from a subject who is known to be free of lung cancer. When an appropriate reference is indicative of a subject being free of lung cancer, a difference between an expression level determined from a subject in need of diagnosis of lung cancer and the appropriate reference may be indicative of lung cancer in the subject. Alternatively, when an appropriate reference is indicative of the subject being free of lung cancer, a lack of a detectable difference (e.g., lack of a statistically significant difference) between an expression level determined from a subject in need of diagnosis of lung cancer and the appropriate reference level may be indicative of the subject being free of lung cancer.

In some embodiments, the reference standard provides a threshold level of change, such that if the expression level of a gene in a sample is within a threshold level of change (increase or decrease depending on the particular marker) then the subject is identified as free of lung cancer, but if the levels are above the threshold then the subject is identified as being at risk of having lung cancer.

In some embodiments, the methods involve comparing the expression level of an informative-gene to a reference standard that represents the expression level of the informative-gene in a control subject who is identified as not having lung cancer. This reference standard may be, for example, the average expression level of the informative-gene in a population of control subjects who are identified as not having lung cancer.

The magnitude of difference between a expression level and an appropriate reference that is statistically significant may vary. For example, a significant difference that indicates lung cancer may be detected when the expression level of an informative-gene in a biological sample is at least 1%, at least 5%, at least 10%, at least 25%, at least 50%, at least 100%, at least 250%, at least 500%, or at least 1000% higher, or lower, than an appropriate reference of that gene. Similarly, a significant difference may be detected when the expression level of informative-gene in a biological sample is at least 1.1-fold, 1.2-fold, 1.5-fold, 2-fold, at least 3-fold, at least 4-fold, at least 5-fold, at least 6-fold, at least 7-fold, at least 8-fold, at least 9-fold, at least 10-fold, at least 20-fold, at least 30-fold, at least 40-fold, at least 50-fold, at least 100-fold, or more higher, or lower, than the appropriate reference of that gene. In some embodiments, at least a 20% to 50% difference in expression between an informative-gene and appropriate reference is significant. Significant differences may be identified by using an appropriate statistical test. Tests for statistical significance are well known in the art and are exemplified in Applied Statistics for Engineers and Scientists by Petruccelli, Chen and Nandram 1999 Reprint Ed.

It is to be understood that a plurality of expression levels may be compared with plurality of appropriate reference levels, e.g., on a gene-by-gene basis, in order to assess the lung cancer status of the subject. The comparison may be made as a vector difference. In such cases, Multivariate Tests, e.g., Hotelling's T2 test, may be used to evaluate the significance of observed differences. Such multivariate tests are well known in the art and are exemplified in Applied Multivariate Statistical Analysis by Richard Arnold Johnson and Dean W. Wichern Prentice Hall; 6th edition (Apr. 2, 2007).

Classification Methods

The methods may also involve comparing a set of expression levels (referred to as an expression pattern or profile) of informative-genes in a biological sample obtained from a subject with a plurality of sets of reference levels (referred to as reference patterns), each reference pattern being associated with a known lung cancer status, identifying the reference pattern that most closely resembles the expression pattern, and associating the known lung cancer status of the reference pattern with the expression pattern, thereby classifying (characterizing) the lung cancer status of the subject.

The methods may also involve building or constructing a prediction model, which may also be referred to as a classifier or predictor, that can be used to classify the disease status of a subject. As used herein, a “lung cancer-classifier” is a prediction model that characterizes the lung cancer status of a subject based on expression levels determined in a biological sample obtained from the subject. Typically the model is built using samples for which the classification (lung cancer status) has already been ascertained. Once the model (classifier) is built, it may then be applied to expression levels obtained from a biological sample of a subject whose lung cancer status is unknown in order to predict the lung cancer status of the subject. Thus, the methods may involve applying a lung cancer-classifier to the expression levels, such that the lung cancer-classifier characterizes the lung cancer status of a subject based on the expression levels. The subject may be further treated or evaluated, e.g., by a health care provider, based on the predicted lung cancer status.

The classification methods may involve transforming the expression levels into a lung cancer risk-score that is indicative of the likelihood that the subject has lung cancer. In some embodiments, such as, for example, when a linear discriminant classifier is used, the lung cancer risk-score may be obtained as the combination (e.g., sum, product, or other combination) of weighted expression levels, in which the expression levels are weighted by their relative contribution to predicting increased likelihood of having lung cancer.

It should be appreciated that a variety of prediction models known in the art may be used as a lung cancer-classifier. For example, a lung cancer-classifier may comprises an algorithm selected from logistic regression, partial least squares, linear discriminant analysis, quadratic discriminant analysis, neural network, naïve Bayes, C4.5 decision tree, k-nearest neighbor, random forest, support vector machine, or other appropriate method.

The lung cancer-classifier may be trained on a data set comprising expression levels of the plurality of informative-genes in biological samples obtained from a plurality of subjects identified as having lung cancer. For example, the lung cancer-classifier may be trained on a data set comprising expression levels of a plurality of informative-genes in biological samples obtained from a plurality of subjects identified as having lung cancer based histological findings. The training set will typically also comprise control subjects identified as not having lung cancer. As will be appreciated by the skilled artisan, the population of subjects of the training data set may have a variety of characteristics by design, e.g., the characteristics of the population may depend on the characteristics of the subjects for whom diagnostic methods that use the classifier may be useful. For example, the population may consist of all males, all females or may consist of both males and females. The population may consist of subjects with history of cancer, subjects without a history of cancer, or subjects from both categories. The population may include subjects who are smokers, former smokers, and/or non-smokers.

A class prediction strength can also be measured to determine the degree of confidence with which the model classifies a biological sample. This degree of confidence may serve as an estimate of the likelihood that the subject is of a particular class predicted by the model.

Accordingly, the prediction strength conveys the degree of confidence of the classification of the sample and evaluates when a sample cannot be classified. There may be instances in which a sample is tested, but does not belong, or cannot be reliably assigned to, a particular class. This may be accomplished, for example, by utilizing a threshold, or range, wherein a sample which scores above or below the determined threshold, or within the particular range, is not a sample that can be classified (e.g., a “no call”).

Once a model is built, the validity of the model can be tested using methods known in the art. One way to test the validity of the model is by cross-validation of the dataset. To perform cross-validation, one, or a subset, of the samples is eliminated and the model is built, as described above, without the eliminated sample, forming a “cross-validation model.” The eliminated sample is then classified according to the model, as described herein. This process is done with all the samples, or subsets, of the initial dataset and an error rate is determined. The accuracy the model is then assessed. This model classifies samples to be tested with high accuracy for classes that are known, or classes have been previously ascertained. Another way to validate the model is to apply the model to an independent data set, such as a new biological sample having an unknown lung cancer status.

As will be appreciated by the skilled artisan, the strength of the model may be assessed by a variety of parameters including, but not limited to, the accuracy, sensitivity and specificity. Methods for computing accuracy, sensitivity and specificity are known in the art and described herein (See, e.g., the Examples). The lung cancer-classifier may have an accuracy of at least 60%, at least 65%, at least 70%, at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 99%, or more. The lung cancer-classifier may have an accuracy in a range of about 60% to 70%, 70% to 80%, 80% to 90%, or 90% to 100%. The lung cancer-classifier may have a sensitivity of at least 60%, at least 65%, at least 70%, at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 99%, or more. The lung cancer-classifier may have a sensitivity in a range of about 60% to 70%, 70% to 80%, 80% to 90%, or 90% to 100%. The lung cancer-classifier may have a specificity of at least 60%, at least 65%, at least 70%, at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 99%, or more. The lung cancer-classifier may have a specificity in a range of about 60% to 70%, 70% to 80%, 80% to 90%, or 90% to 100%.

The Negative Predictive Value (NPV) may be greater than 40%, 41%, 42%, 43%, 44%, 45%, 46%, 47%, 48%, 49%, 50%, 51%, 52%, 53%, 54%, 55%, 56%, 57%, 58%, 59%, 60%, 61%, 62%, 63%, 64%, 65%, 66%, 67%, 68%, 69%, 70%, 71%, 72%, 73%, 74%, 75%, 76%, 77%, 78%, 79%, 80%, 81%, 82%, 83%, 84%, 85%, 86%, 87%, 88%, 89%, 90%, 91%, 92%, 93%, 94%, 95%, 96%, 97%, 98%, or 99% for ruling out lung cancer in an intended use population.

The intended use population may have a prevalence of cancer at or about 40%, 41%, 42%, 43%, 44%, 45%, 46%, 47%, 48%, 49%, 50%, 51%, 52%, 53%, 54%, 55%, 56%, 57%, 58%, 59%, 60%, 61%, 62%, 63%, 64%, 65%, 66%, 67%, 68%, 69%, 70%, 71%, 72%, 73%, 74%, 75%, 76%, 77%, 78%, 79%, 80%, 81%, 82%, 83%, 84%, 85%, 86%, 87%, 88%, 89%, 90%, 91%, 92%, 93%, 94%, 95%, 96%, 97%, 98%, or 99%.

Clinical Treatment/Management

In certain aspects, methods are provided for determining a treatment course for a subject. The methods typically involve determining the expression levels in a biological sample obtained from the subject of one or more informative-genes, and determining a treatment course for the subject based on the expression levels. Often the treatment course is determined based on a lung cancer risk-score derived from the expression levels. The subject may be identified as a candidate for a lung cancer therapy based on a lung cancer risk-score that indicates the subject has a relatively high likelihood of having lung cancer. The subject may be identified as a candidate for an invasive lung procedure (e.g., transthoracic needle aspiration, mediastinoscopy, or thoracotomy) based on a lung cancer risk-score that indicates the subject has a relatively high likelihood of having lung cancer (e.g., greater than 60%, greater than 70%, greater than 80%, greater than 90%). The subject may be identified as not being a candidate for a lung cancer therapy or an invasive lung procedure based on a lung cancer risk-score that indicates the subject has a relatively low likelihood (e.g., less than 50%, less than 40%, less than 30%, less than 20%) of having lung cancer. In some cases, an intermediate risk-score is obtained and the subject is not indicated as being in the high risk or the low risk categories. In some embodiments, a health care provider may engage in “watchful waiting” and repeat the analysis on biological samples taken at one or more later points in time, or undertake further diagnostics procedures to rule out lung cancer, or make a determination that cancer is present, soon after the risk determination was made. In a particular example, a subject is identified as intermediate risk due to a non-diagnostic bronchoscopy and is reassigned to non-invasive monitoring (such as CT surveillance) following a determination, using the methods herein, that the patient is at low-risk of cancer. In another particular example, the samples assayed as described herein may be used The methods may also involve creating a report that summarizes the results of the gene expression analysis. Typically the report would also include an indication of the lung cancer risk-score.

Computer Implemented Methods

Methods disclosed herein may be implemented in any of numerous ways. For example, certain embodiments may be implemented using hardware, software or a combination thereof. When implemented in software, the software code can be executed on any suitable processor or collection of processors, whether provided in a single computer or distributed among multiple computers. Such processors may be implemented as integrated circuits, with one or more processors in an integrated circuit component. Though, a processor may be implemented using circuitry in any suitable format.

Further, it should be appreciated that a computer may be embodied in any of a number of forms, such as a rack-mounted computer, a desktop computer, a laptop computer, or a tablet computer. Additionally, a computer may be embedded in a device not generally regarded as a computer but with suitable processing capabilities, including a Personal Digital Assistant (PDA), a smart phone or any other suitable portable or fixed electronic device.

Also, a computer may have one or more input and output devices. These devices can be used, among other things, to present a user interface. Examples of output devices that can be used to provide a user interface include printers or display screens for visual presentation of output and speakers or other sound generating devices for audible presentation of output. Examples of input devices that can be used for a user interface include keyboards, and pointing devices, such as mice, touch pads, and digitizing tablets. As another example, a computer may receive input information through speech recognition or in other audible format.

Such computers may be interconnected by one or more networks in any suitable form, including as a local area network or a wide area network, such as an enterprise network or the Internet. Such networks may be based on any suitable technology and may operate according to any suitable protocol and may include wireless networks, wired networks or fiber optic networks.

Also, the various methods or processes outlined herein may be coded as software that is executable on one or more processors that employ any one of a variety of operating systems or platforms. Additionally, such software may be written using any of a number of suitable programming languages and/or programming or scripting tools, and also may be compiled as executable machine language code or intermediate code that is executed on a framework or virtual machine.

In this respect, aspects of the disclosure may be embodied as a computer readable medium (or multiple computer readable media) (e.g., a computer memory, one or more floppy discs, compact discs (CD), optical discs, digital video disks (DVD), magnetic tapes, flash memories, circuit configurations in Field Programmable Gate Arrays or other semiconductor devices, or other non-transitory, tangible computer storage medium) encoded with one or more programs that, when executed on one or more computers or other processors, perform methods that implement the various embodiments of the disclosure discussed above. The computer readable medium or media can be transportable, such that the program or programs stored thereon can be loaded onto one or more different computers or other processors to implement various aspects of the present disclosure as discussed above. As used herein, the term “non-transitory computer-readable storage medium” encompasses only a computer-readable medium that can be considered to be a manufacture (i.e., article of manufacture) or a machine.

The terms “program” or “software” are used herein in a generic sense to refer to any type of computer code or set of computer-executable instructions that can be employed to program a computer or other processor to implement various aspects of the present disclosure as discussed above. Additionally, it should be appreciated that according to one aspect of this embodiment, one or more computer programs that when executed perform methods of the present disclosure need not reside on a single computer or processor, but may be distributed in a modular fashion amongst a number of different computers or processors to implement various aspects of the present disclosure.

As used herein, the term “database” generally refers to a collection of data arranged for ease and speed of search and retrieval. Further, a database typically comprises logical and physical data structures. Those skilled in the art will recognize the methods described herein may be used with any type of database including a relational database, an object-relational database and an XML-based database, where XML stands for “extensible-Markup-Language”. For example, the gene expression information may be stored in and retrieved from a database. The gene expression information may be stored in or indexed in a manner that relates the gene expression information with a variety of other relevant information (e.g., information relevant for creating a report or document that aids a physician in establishing treatment protocols and/or making diagnostic determinations, or information that aids in tracking patient samples). Such relevant information may include, for example, patient identification information, ordering physician identification information, information regarding an ordering physician's office (e.g., address, telephone number), information regarding the origin of a biological sample (e.g., tissue type, date of sampling), biological sample processing information, sample quality control information, biological sample storage information, gene annotation information, lung-cancer risk classifier information, lung cancer risk factor information, payment information, order date information, etc.

Computer-executable instructions may be in many forms, such as program modules, executed by one or more computers or other devices. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Typically the functionality of the program modules may be combined or distributed as desired in various embodiments.

In some aspects of the disclosure, computer implemented methods for processing genomic information are provided. The methods generally involve obtaining data representing expression levels in a biological sample of one or more informative-genes and determining the likelihood that the subject has lung cancer based at least in part on the expression levels. Any of the statistical or classification methods disclosed herein may be incorporated into the computer implemented methods. In some embodiments, the methods involve calculating a risk-score indicative of the likelihood that the subject has lung cancer. Computing the risk-score may involve a determination of the combination (e.g., sum, product or other combination) of weighted expression levels, in which the expression levels are weighted by their relative contribution to predicting increased likelihood of having lung cancer. The computer implemented methods may also involve generating a report that summarizes the results of the gene expression analysis, such as by specifying the risk-score. Such methods may also involve transmitting the report to a health care provider of the subject.

Affymetrix Array

In some aspects, the Affymetrix Human Gene 1.0 ST array (Affymetrix Cat. #901087) is used to identify the mRNA or cDNA in a biological sample. The Affymetrix Human Gene 1.0 ST array utilizes numerous probes that are disclosed at the world wide web address: affymetrix.com/site/include/byproduct.affx?product=hugene-1_0-st-v1. Multiple probes are present that correspond to segments of specific genes, which is to say that that is not a 1:1 ratio of one probe per gene, rather multiple probes are present that correspond to multiple segments of a single gene. In one example, the LYPD2 gene is represented by three probe sets in the Human Gene 1.0 ST array (Release 32), probeset IDs 8153343, 8153344, and 8153345 as disclosed in the Affymetrix Human Gene 1.0 ST array (HuGene-1_0-st-v1 Probeset Annotations. Exemplary suitable builds of the array include release 32 (Sep. 30, 2011), release 33 (Mar. 27, 2013), release 34 Apr. 7, 2014), release 35 on (Apr. 15, 2015), and release 36. Additional releases, including future releases may also be used. Information correlating probe-sets and nucleic acid sequences can be found at Affymetrix.com, including at the world wide web address: affymetrix.com/site/include/byproduct.affx?product-hugene-1_0-st-v1. Furthermore, data sets are available on the NCBI Gene Expression Omnibus website under Platform GPL6244 detailing the probes and genes that may be utilized in practicing the methods of the present disclosure at the world wide web address: ncbi.nlm.nih.gov/geo/query/acc.cgi?acc-GPL6244. These documents, including those correlating the probesets and gene symbols are incorporated herein by reference.

Compositions and Kits

In some aspects, compositions and related methods are provided that are useful for determining expression levels of informative-genes. For example, compositions are provided that consist essentially of nucleic acid probes that specifically hybridize with informative-genes or with nucleic acids having sequences complementary to informative-genes. These compositions may also include probes that specifically hybridize with control genes or nucleic acids complementary thereto. These compositions may also include appropriate buffers, salts or detection reagents. The nucleic acid probes may be fixed directly or indirectly to a solid support (e.g., a glass, plastic or silicon chip) or a bead (e.g., a magnetic bead). The nucleic acid probes may be customized for used in a bead-based nucleic acid detection assay.

In some embodiments, compositions are provided that comprise up to 5, up to 10, up to 25, up to 50, up to 100, or up to 200 nucleic acid probes. In some cases, each of the nucleic acid probes specifically hybridizes with an mRNA selected from Table 11 or with a nucleic acid having a sequence complementary to the mRNA. In some embodiments, probes that detect informative-mRNAs are also included. In some cases, each of at least 2, at least 3, at least 4, at least 5, at least 6, at least 7, at least 8, at least 9, at least 10, or at least 20 of the nucleic acid probes specifically hybridizes with an mRNA selected from Table 11 or with a nucleic acid having a sequence complementary to the mRNA. In some embodiments, the compositions are prepared for detecting different genes in biochemically separate reactions, or for detecting multiple genes in the same biochemical reactions. In some embodiments, the compositions are prepared for performing a multiplex reaction.

Also provided herein are oligonucleotide (nucleic acid) arrays that are useful in the methods for determining levels of multiple informative-genes simultaneously. Such arrays may be obtained or produced from commercial sources. Methods for producing nucleic acid arrays are also well known in the art. For example, nucleic acid arrays may be constructed by immobilizing to a solid support large numbers of oligonucleotides, polynucleotides, or cDNAs capable of hybridizing to nucleic acids corresponding to genes, or portions thereof. The skilled artisan is referred to Chapter 22 “Nucleic Acid Arrays” of Current Protocols In Molecular Biology (Eds. Ausubel et al. John Wiley and #38; Sons NY, 2000) or Liu C G, et al., An oligonucleotide microchip for genome-wide microRNA profiling in human and mouse tissues. Proc Nall Acad Sci USA. 2004 Jun. 29;101 (26): 9740-4, which provide non-limiting examples of methods relating to nucleic acid array construction and use in detection of nucleic acids of interest. In some embodiments, the arrays comprise, or consist essentially of, binding probes for at least 2, at least 5, at least 10, at least 20, at least 50, at least 60, at least 70 or more informative-genes. In some embodiments, the arrays comprise, or consist essentially of, binding probes for up to 2, up to 5, up to 10, up to 20, up to 50, up to 60, up to 70 or more informative-genes. In some embodiments, an array comprises or consists of 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10 of the mRNAs selected from Table 11. In some embodiments, an array comprises or consists of 4, 5, or 6 of the mRNAs selected from Table 11. Kits comprising the oligonucleotide arrays are also provided. Kits may include nucleic acid labeling reagents and instructions for determining expression levels using the arrays.

The compositions described herein can be provided as a kit for determining and evaluating expression levels of informative-genes. The compositions may be assembled into diagnostic or research kits to facilitate their use in diagnostic or research applications. A kit may include one or more containers housing the components of the disclosure and instructions for use. Specifically, such kits may include one or more compositions described herein, along with instructions describing the intended application and the proper use of these compositions. Kits may contain the components in appropriate concentrations or quantities for running various experiments.

The kit may be designed to facilitate use of the methods described herein by researchers, health care providers, diagnostic laboratories, or other entities and can take many forms. Each of the compositions of the kit, where applicable, may be provided in liquid form (e.g., in solution), or in solid form, (e.g., a dry powder). In certain cases, some of the compositions may be constitutable or otherwise processable, for example, by the addition of a suitable solvent or other substance, which may or may not be provided with the kit. As used herein, “instructions” can define a component of instruction and/or promotion, and typically involve written instructions on or associated with packaging of the disclosure. Instructions also can include any oral or electronic instructions provided in any manner such that a user will clearly recognize that the instructions are to be associated with the kit, for example, audiovisual (e.g., videotape, DVD, etc.), Internet, and/or web-based communications, etc. The written instructions may be in a form prescribed by a governmental agency regulating the manufacture, use or sale of diagnostic or biological products, which instructions can also reflect approval by the agency.

A kit may contain any one or more of the components described herein in one or more containers. As an example, in one embodiment, the kit may include instructions for mixing one or more components of the kit and/or isolating and mixing a sample and applying to a subject. The kit may include a container housing agents described herein. The components may be in the form of a liquid, gel or solid (e.g., powder). The components may be prepared sterilely and shipped refrigerated. Alternatively they may be housed in a vial or other container for storage. A second container may have other components prepared sterilely.

As used herein, the terms “approximately” or “about” in reference to a number are generally taken to include numbers that fall within a range of 1%, 5%, 10%, 15%, or 20% in either direction (greater than or less than) of the number unless otherwise stated or otherwise evident from the context (except where such number would be less than 0% or exceed 100% of a possible value).

All references described herein are incorporated by reference for the purposes described herein.

Exemplary embodiments of the disclosure will be described in more detail by the following examples. These embodiments are exemplary of the disclosure, which one skilled in the art will recognize is not limited to the exemplary embodiments.

EXAMPLES
Example 1: Development of a Microarray Based Prediction Model Introduction

This example describes a method for developing a prediction algorithm. A final optimized model is described, including the combination of genes used in the model. The method uses Clinical Factor Genomic Correlates (CFGC) to aid in the selection of a cancer-specific signature.

The objectives were to develop and characterize a new cancer prediction model that effectively predicts lung cancer status after accounting gene expression signal attributed more specifically to clinical factors by deriving genomic correlates. Genomic correlates are defined herein as gene expression algorithms to predict the specific clinical characteristics, such as subject gender, smoking status, and smoking history.

An objective was to develop a prediction algorithm that meets the following specific performance criteria:

- Negative Predictive Value (NPV) of greater than 90% for ruling out lung cancer in an intended use population, and
- NPV of greater than 85% for subjects diagnosed with COPD.

Materials and Methods
Sample Processing and Analysis

Clinical specimens were collected from patients scheduled to undergo bronchoscopy for the suspicion of lung cancer. Bronchial epithelial cells (BECs) were collected from the mainstem bronchus of subjects using standard bronchial brushes during a scheduled bronchoscopy procedure.

Samples were analyzed on gene-expression microarrays using Gene 1.0 ST microarrays (Affymetrix). A pairwise correlation analysis of the array data was conducted to identify outliers (described herein). A total of 597 samples were retained as the final data set. The data set were then split into equivalently sized Training and Validation sets in a randomized manner.

Microarray CEL files were used for the development of a prediction algorithm. Subjects were first designated to independent Training and Validation sets, and gene selection and optimization of the prediction model was conducted within the Training set. The model was optimized and locked prior to predicting the cancer status of validation set samples.

Normalization/Batch Adjustment

RMA was used to compute gene expression values from Gene 1.0 ST (Affymetrix) CEL files. ComBat batch adjustment was used to correct for batch effects. All samples were analyzed within 5 separate microarray experiments (i.e., batches). Training and test samples were combined in RMA and ComBat pre-processing. Subsequent development was restricted to the Training set as previously described. CEL files corresponding to samples with a RIN score less than 4 were excluded, as were samples with average pairwise correlation less than. 955 (see FIG. 1).

Genomic Correlates

Genomic correlates were established as a gene expression signature that accurately predict the corresponding clinical characteristic. The genomic correlates, in combination with separate genes to predict cancer status, are combined in a prediction algorithm. Correlates for the following clinical characteristics were developed and evaluated:

- Gender
- Smoking status (current versus former)
- Smoking history (pack-year; PY)
- Age

The genomic correlates were developed by selecting top-ranked genes differentiating the clinical characteristic of interest and fitting those genes to a model using logistic regression. Scoring of clinical characteristics was based on the gene expression of those selected genes.

Two models were developed in parallel to be tested simultaneously in a Validation set. A first model (Score 1) was based on the methods described herein and factoring the reported Age into the prediction algorithm as well as the genomic signal. A second model (Score 2) was developed using a genomic correlate for age, which was then incorporated into the prediction algorithm.

Initial Gene Selection

A clinical factor model was developed, using logistic regression of cancer status (0/1) on age, gender, smoking status, and pack years. The residuals from the clinical factor model were used to select genes using an empirical Bayes linear model to test association of each gene with the residuals. The top 232 genes were selected based on the p-value and fold change from this model. The top 232 genes are listed in Table 11.

Clustering/Final Gene Selection

Gene selection and model fitting were conducted in an automated cross-validation approach in order to minimize bias during the selection process and to provide a robust final selection. The gene selection consisted of the following steps. Hierarchical clustering was used to divide the genes into 11 clusters. The cluster membership of each gene is identified in column 2 of the Table 11. For each of the clusters, cluster means were computed using all of the genes within each cluster. A combination of LASSO and backwards selection were used in repeated random subsets of the data to identify six clusters that were consistently selected to have independent predictive association with cancer status. Cross validation was then used to determine the approximate number of genes within each cluster that would retain the predictive strength of the cluster means.

A gene titration analysis was done to determine the sensitivity of the models to increased numbers of genes from the selected clusters within the final model. This was included as part of the optimization to determine if complementary genes could add additional clinical sensitivity to the model. Software

R (Version 3.01) was used for the analysis, including the packages rms, limma, verification, and sva.

Results
Derivation of Genomic Correlates

Genes were selected to represent clinical characteristics using the whole Training set. These “genomic correlate genes” were based on a minimal set of genes to represent the clinical factor as accurately as possible. Genomic Gender (GG) was defined with 100% accuracy using a single gene. The values are set forth as follows: GG=1 (male) if RPS4Y1<threshold, and GG=0 (female) otherwise.

For genomic smoking status, genes were screened based on an empirical Bayes t-test. The top genes by p-value were included in a logistic regression model where smoking status was the dependent variable. The resulting predicted genomic smoking (GS) value was derived from this model (using gene symbols to represent the relative expression intensity), where,

x
=

β
⁢

G
⁢
S

0

+

β
⁢

G
⁢
S

1

*
SLC
⁢
7
⁢
A
⁢
11

+

β
⁢

G
⁢
S

2

*
CLND
⁢
10

+

β
⁢

G
⁢
S

3

*
TKT

,

- in which

β
⁢

G
⁢
S

n

- are regressopm weights for the genomic regression model, and GS=exp(x)/(1+exp(x)).

For genomic pack years, genes were screened based on an empirical Bayes t-test. The top genes by p-value were included in a logistic regression model where pack years <10 was the dependent variable. The resulting predicted genomic pack years (GPY) value was derived from this model. where

x
=

β
⁢

G
⁢
P
⁢
Y

0

+

β
⁢

G
⁢
P
⁢
Y

1

*
RUNX
⁢
1
⁢
T
⁢
1

+

β
⁢

G
⁢
P
⁢
Y

2

*
AKR
⁢
1
⁢
C
⁢
2

,

- in which

β
⁢

G
⁢
P
⁢
Y

n

- are regression weights for the genomic pack years regression model, and GPY=exp(x)/(1+exp(x)).

For genomic age, genes were screened based on an empirical Bayes linear model. The top genes by p-value were included in a penalized linear regression model (LASSO) where age (in years) was the dependent variable. The resulting predicted genomic age (GA) value was derived from this model, where,

G
⁢
A

=

β
⁢

G
⁢
A

0

+

β
⁢

G
⁢
A

1

*
CD
⁢
52

+

β
⁢

G
⁢
A

2

*
SYT
⁢
8

+

β
⁢

G
⁢
A

3

*
TNNT
⁢
3

+
 

β
⁢

G
⁢
A

4

*
ALX
⁢
1

+

β
⁢

G
⁢
A

5

*
KLRK
⁢
1

+

β
⁢

G
⁢
A

6

*
RASA
⁢
3

+
 

β
⁢

G
⁢
A

7

*
CERS
⁢
3

+

β
⁢

G
⁢
A

8

*
ASPA

+

β
⁢

G
⁢
A

9

*
G
⁢
R
⁢
P

+

β
⁢

G
⁢
A

1
⁢
0

*
APOC
⁢
1

+
 

β
⁢

G
⁢
A

1
⁢
1

*
EPHX
⁢
3

+

β
⁢

G
⁢
A

1
⁢
2

*
REEP
⁢
1

+

β
⁢

G
⁢
A

1
⁢
3

*
FAM
⁢
198
⁢
B

+
 

β
⁢

G
⁢
A

1
⁢
4

*
PCDHB
⁢
4

+

β
⁢

G
⁢
A

1
⁢
5

*
PCDHB
⁢
16

+

β
⁢

G
⁢
A

1
⁢
6

*
FOXD
⁢
1

+
 

β
⁢

G
⁢
A

1
⁢
7

*
SPARC
⁢

β
⁢

G
⁢
A

1
⁢
8

*
NKAPL

+

β
⁢

G
⁢
A

1
⁢
9

*
GPR
⁢
110

,

- in which

β
⁢

G
⁢
A

n

- are regression weights for the genomic age regression model.

Derivation of the Cancer Genes

The top 232 gene were selected initially, followed by a down-selection of genes using the clustering analysis described in the methods section.

Cross validation was then used to determine the approximate number of genes within each cluster that would retain the predictive strength of the cluster means. This number was found to be between 2 and 4 genes per cluster. Final gene selection within each cluster was based on p-value, fold change, and strength of evidence for cancer association from the literature. Cluster means were recomputed using the reduced gene sets within each cluster, as follows:

- C1A=mean of (BST1, CD177.1, CD177.2)
- C1B=mean of (ATP 12A, TSPAN2)
- C2=mean of (GABBRI, MCAM, NOVA1, SDC2)
- C3=mean of (CDR1, CGREF1, CLND22, NKX3-1) C4A=mean of (EPHX3, LYPD2)
- C4B=mean of (MIA, RNF150).

Description of the Finalized Model

To estimate the final model coefficients, a penalized logistic regression model was used, with Age (in years), genomic gender (GG), genomic smoking status (GS), genomic pack years (GPY), and the six reduced gene cluster means (labeled C1A, C1B, C2, C3, C4A, C4B) as the independent predictors and cancer status (0/1) as the dependent variable. The penalization factor (lambda) was 0 for the clinical/genomic correlates and 10 for each of the gene expression clusters. The second model was built using the same approach, but replacing Age with genomic age (GA) as defined above. The model coefficients were then re-estimated.

The final classification algorithm was of the form,

x

score
⁢

1

=

W
0

+

W
1

×
G
⁢
G

+

W
2

×
G
⁢
S

+

W
3

×
G
⁢
P
⁢
Y

+

W
4

×
Reported
⁢

Age

+
 

W
5

×
C
⁢
1
⁢
A

+

W
6

×
C
⁢
1
⁢
B

+

W
7

×
C
⁢
2

+

W
8

×
C
⁢
3

+

W
9

×
C
⁢
4
⁢
A

+

W

1
⁢
0

×
C
⁢
4
⁢
B

X

score
⁢

2

=

W
0

+

W
1

×
G
⁢
G

+

W
2

×
G
⁢
S

+

W
3

×
G
⁢
P
⁢
Y

+

W
4

×
G
⁢
A

+
 

W
5

×
C
⁢
1
⁢
A

+

W
6

×
C
⁢
1
⁢
B

+

W
7

×
C
⁢
2

+

W
8

×
C
⁢
3

+

W
9

×
C
⁢
4
⁢
A

+

W

1
⁢
0

×
C
⁢
4
⁢

B
.

The logistic regression score is then converted to a prediction score, ranging from 0 to 1, using the equation

Prediction
⁢

Score

=

ex

Score
1

⁢

or
⁢

Score
⁢

2

(

1
+

ex

Score
1

⁢

or
⁢

Score
⁢

2

)

Evaluation of Prediction Models

A cross-validation approach was used in which the Training set samples were randomly split (90:10) into training and testing groups, respectively. The clinical accuracy was recorded and the process was repeated 100-fold. Scores for the training set were reported as the average of the total iterations for each sample. Results were reported as ROC curves, AUC, and sensitivity and specificity after defining the score threshold to yield 50% specificity and to maximize the clinical sensitivity.

The scores for all training set samples were generated and compared to recorded clinical status. ROC curves for Score 1 and Score 2 are shown in for all samples in FIG. 2, and for bronchoscopy-negative samples only in FIG. 3. The AUC's were calculated as, Score 1=0.803, and Score 2=0.785, for all samples, and Score 1=0.808, and Score 2=0.778 for the bronch-negative set. Table 1 provides a list of informative genes used in the prediction score models. Table 1.1 provides a non-limiting list of probe sequences for detecting expression of such genes.

TABLE 1

Gene list and function

Gene Symbol
Function
Gene Symbol
Function

RPS4Y1
GG
TSPAN2
CA

SLC7A11
GS
MCAM
CA

CLDN10
GS
ATP12A
CA

TKT
GS
NOVA1
CA

RUNX1T1
GPY
MIA
CA

AKR1C2
GPY
CD177.1
CA

CD52
GA
EPHX3
CA

SYT8
GA
CD177.2
CA

TNNT3
GA
CGREF1
CA

ALX1
GA
BST1
CA

KLRK1
GA
RNF150
CA

RASA3
GA
CLND22
CA

CERS3
GA
GABBR1
CA

ASPA
GA
SDC2
CA

GRP
GA
NKX3-1
CA

APOC1
GA
LYPD2
CA

EPHX3
GA, CA
CDR1
CA

REEP1
GA

FAM198B
GA

PCDHB4
GA

PCDHB16
GA

FOXD1
GA

SPARC
GA

NKAPL
GA

GPR110
GA

Table 1.1: Non-Limiting Examples of Probe Sequences for Detecting Informative-Genes

TABLE 1.1

Non-limiting examples of probe sequences for detecting

informative-genes.

Gene:

Cancer
Probe Sequence
SEQ ID NO:

TSPAN2
TCAACATTAAGAAGTCTTAATTCAG
1

MCAM
GCTTTAATCCCCATGAAGGACAGTG
2

ATP12A
GCGGTGGAGATACCGGCGGCGGCGC
3

NOVA1
GACGAAATTCAGACATGGAGCATCA
4

MIA
ATGACACCAAGGCACACCAGGGACC
5

CD177.1
GACCCGTCTGTGGCTGGTAATCTCT
6

EPHX3
GCTCCACTGGAAGAGAGGTATACCC
7

CD177.2
TTCCTGTGTCCCATTGAGCAGGTTG
8

CGREF1
TAGGGTACAGCACTTAACGCAATCT
9

BST1
TGTTTGCCGTTTCCCGTTCCAGACA
10

RNF150
TGGTTAATCCAAGCCGCAGCCTGGT
11

CLDN22
GGTGTTTCTCGTCTCCAGTTCTTGA
12

GABBR1
TACGGAGCCATTACCTGGGCAGTGC
13

SDC2 NKX3-1
TGAGCCTGCTTCTCCGGGCTCCCCT
14

LYPD2 CDR1
TTTGTGCTGGCTAGTACTCCGGTCG
15

TTCTTCAAGGCATTCGGGGCTGGGC
16

AACAACTCCGGGTCTTCCAGCGACT
17

Gene:

Gender
Probe sequence
SEQ ID NO:

RP S4Y1
TAAACCGCAGGAAGTCAGATGAGTG
18

Gene:

Smoking
Probe sequence
SEQ ID NO:

SLC7A11
GGTTGAAGCAACTAGAAGCGTGACA
19

CLDN10 TKT
GACAGCGTTTCATGCTCGGATGGCC
20

GGTTTATTCTCTCCAGACGGTCAGG
21

Gene:

PY
Probe sequence
SEQ ID NO:

RUNX1T1
TAACAGGGAGGAGGTCAAATCTATC
22

AKR1 C2
TAGCTGTAGCTTACTGAAGTCGCCA
23

Table 2 below provides a summary of performance characteristics for the two models for bronchoscopy-negative subjects in the training set. The number of bronch-negative subjects (N) corresponds to CA+subjects for sensitivity and CA-subjects for specificity.

TABLE 2

N
Score 1
Score 2

Sensitivity
68
91.2%
82.4%

Specificity
76
56.6%
48.7%

AUC

80.8%
77.8%

NPV*

95.5%
90.0%

*NPV is calculated assuming a 50% prevalence of cancer combined with the observed sensitivity and specificity of bronchoscopy and prediction model.

Model Performance

Calculation of the sensitivity and specificity of the prediction models (based on scores 1 and 2) was done using a specific score threshold, to differentiate prediction of CA+ and CA− samples. The same threshold that was selected for both models in the Training Set (Model Score=0.65) was used in the Validation Set. Prediction accuracy was first determined in the bronchoscopy-negative samples (the intended-use cases) for both models and is summarized in

The sensitivity of the prediction model was also calculated for several subgroup categories within the Validation Set. Results are shown in Tables 3-9 for both models. Sub-categories contain different numbers of samples which affect confidence intervals.

TABLE 3

Sensitivity of prediction model as a function of

the mass size of the observed lesion- Score 1

Mass Size
N
Sens Sens.
Sens,
PM + BR Sens.

<1
13
100.00%
83.30%
100.00%

1 to 2
35
87.50%
50.00%
100.00%

2 to 3
41
90.60%
53.10%
96.90%

>3
155
85.10%
79.40%
95.00%

Infiltrate
32
100.00%
100.00%
100.00%

Unknown
21
100.00%
80.00%
100.00%

PM = Prediction Model;

BR = Bronchoscopy

TABLE 4

Sensitivity of prediction model as a function of

the mass size of the observed lesion- Score 2

N

BR Sens.

Mass Size
PM + BR S
PMs.
Senens.

<1
13
66.70%
83.30%
100.00%

1 to 2
35
87.50%
50.00%
93.80%

2 to 3
41
81.30%
53.10%
93.80%

>3
155
83.00%
79.40%
95.70%

Infiltrate
32
90.00%
100.00%
100.00%

Unknown
21
100.00%
80.00%
100.00%

TABLE 5

Sensitivity of prediction model as a function of the

location of the observed lesion in the lung (Score 1)

Mass Size
N
PM Sens
BR Sens
PM + BR Sens

Central
97
84.40%
79.20%
93.50%

Peripheral
86
90.70%
61.10%
96.30%

Both
90
90.50%
78.40%
100.00%

Unknown
25
86.70%
80.00%
93.30%

TABLE 6

Sensitivity of prediction model as a function of the

location of the observed lesion in the lung (Score 2)

Mass Size
N
PM Sens.
BR Sens.
PM + BR Sens.

Central
97
79.20%
79.20%
93.50%

Peripheral
86
81.50%
61.10%
96.30%

Both
90
89.20%
78.40%
98.60%

Unknown
25
93.30%
80.00%
93.30%

TABLE 7

Sensitivity of prediction model as a function of the

cancer sub-type as determined in pathology (Score 1)

Cancer Type
N
P M Sens.
BR Sens.
PM + BR Sens.

AC
67
83.60%
70.10%
94.00%

SCC
72
94.40%
75.00%
95.80%

LC
7
71.40%
85.70%
100.00%

NSCLC
26
92.30%
73.10%
76.90%

SCLC
37
89.20%
86.50%
97.30%

NSCLC/SCLC
2
50.00%
100.00%
100.00%

Unknown
86
75.00%
37.50%
100.00%

TABLE 8

Sensitivity of prediction model as a function of the

cancer sub-type as determined in pathology (Score 2)

Cancer Type
N
PM Sens.
BR Sens.
P M + BR Sens.

AC
67
80.60%
70.10%
94.00%

SCC
72
87.50%
75.00%
97.20%

LC
7
85.70%
85.70%
100.00%

NSCLC
26
69.20%
73.10%
80.80%

SCLC
37
89.20%
86.50%
97.30%

NSCLC/SCLC
2
0.00%
100.00%
100.00%

Unknown
86
62.50%
37.50%
87.50%

PM = Prediction Model;

BR = Bronchoscopy

TABLE 9

Sensitivity of prediction model as a function of cancer stage

Score 1

PM
BR

Stage
N
Sens.
Sens.
PM + BR

I
13
92.3%
38.5%
100.0%

Ila
2
100.0%
0.0%
100.0%

IIb
13
92.3%
76.9%
100.0%

Early
28
92.9%
53.6%
100.0%

Ina
27
92.6%
74.1%
96.3%

Illb
19
89.5%
89.5%
100.0%

IV
47
87.2%
89.4%
100.0%

Extensive
15
73.3%
73.3%
93.3%

Score 2

PM
BR
PM + BR

Stage
N
Sens,
Sens.
Sens.

I
13
84.6%
38.5%
92.3%

IIa
2
100.0%
0.0%
100.0%

Jib
13
92.3%
76.9%
100.0%

Early
28
89.3%
53.6%
96.4%

Ina
27
92.6%
74.1%
100.0%

Illb
19
78.9%
89.5%
94.7%

IV
47
80.9%
89.4%
100.0%

Extensive
15
80.0%
73.3%
93.3%

Comparison of Performance in Training and Validation Sets

The overall prediction accuracy of the prediction models in the Training and Validation Sets was compared in order to evaluate the reproducibility of the models in independent cohorts. Results are provided in Table 10.

TABLE 10

Comparison of sensitivity, specificity and AUC

for two models in Training and Validation sets

Score 1
Score 2

Training
Validation
Training
Validation

All
BR-neg
All
BR-neg
All
BR-neg
All
BR-neg

samples
samples
samples
samples
samples
samples
samples
samples

Sensitivity
90.6%
91.2%
88.2%
85.7%
85.7%
82.4%
84.1%
83.9%

Specificity
56.6%
56.6%
47.4%
47.4%
48.7%
48.7%
47.4%
47.4%

AUC
0.803
0.808
0.793
0.764
0.785
0.778
0.751
0.739

Significant difference (p>0.05) in sensitivity or specificity were not observed between the Training and Validation Sets for either model. Likewise the small differences in AUC for each model between the two cohorts was not statistically significant (based on p>0.05). The sensitivity and specificity is also similar for the bronch-negative samples compared to all samples (bronch-neg and bronch-pos combined). The was a relatively small drop in overall performance between Score 1 and Score 2, with the latter showing a 2-4% point drop in AUC compared to score 1.

TABLE 11

232 differentially expressed genes and cluster membership

Symbol
Cluster
Tstat
p.value

TMEM51
1
−3.73334
0.000227

CR1L
1
−2.84036
0.00482

PDZKlIP 1
1
−3.65419
0.000305

MICAL2
1
−3.08175
0.002252

VWA5A
1
−3.01006
0.002838

ACAD8
1
−3.25793
0.001253

SAA4
1
−2.84375
0.00477

GLYATL2
1
−3.02638
0.002693

ETV6
1
−4.46414
1.15E−05

CD177
1
−3.68455
0.000273

CEACAM7
1
−5.40343
1.35E−07

CD177
1
−3.85704
0.000141

QPCT
1
−3.16572
0.001709

CASP10
1
−3.07674
0.002289

PI3
1
−5.27363
2.59E−07

BST1
1
−4.29031
2.42E−05

MTNRIA
1
−4.47628
1.09E−05

STARD4
1
−2.94146
0.003525

CFB
1
−2.87447
0.004341

SLC26A8
1
−3.09768
0.002138

VNN2
1
−2.88831
0.00416

HDAC9
1
−3.27926
0.001165

SLC26A4
1
−3.93362
0.000104

LCN2
1
−3.64257
0.000319

CFB
1
−2.86774
0.004432

CCDC18
2
−2.89401
0.004087

FAM72D
2
−3.41712
0.000722

NUF2
2
−3.35293
0.000904

FAM72D
2
−3.65638
0.000303

FBXO28
2
−2.93189
0.003633

GPR137B
2
−3.29637
0.001099

STIL
2
−3.05607
0.002448

DEPDCI
2
−3.10412
0.002094

TSPAN2
2
−3.92967
0.000106

FAM72D
2
−3.18098
0.001624

ASPM
2
−3.21677
0.001441

KIF14
2
−3.09106
0.002185

KIF20B
2
−2.95904
0.003336

RAD51AP1
2
−3.38028
0.000822

GAS2L3
2
−3.22465
0.001403

SPIC
2
−3.00214
0.00291

SMAGP
2
−3.38429
0.00081

ATP12A
2
−3.49107
0.000555

BRCA2
2
−2.9535
0.003395

BORA
2
−2.81443
0.005215

SKA3
2
−2.94422
0.003495

DLGAP5
2
−2.8962
0.00406

CASC5
2
−2.98473
0.003076

LRRC28
2
−3.78219
0.000188

PYCARD
2
−3.0296
0.002666

TXNL4B
2
−3.75986
0.000205

EFCAB5
2
−4.15228
4.31E−05

SPAG5
2
−3.28148
0.001157

FAM72D
2
−3.66614
0.000292

ABCAl2
2
−3.25495
0.001266

AURKA
2
−3.06413
0.002385

SGOL1
2
−2.87447
0.004341

BANK1
2
−3.22839
0.001385

CENPE
2
−2.90302
0.003975

CASP6
2
−2.96202
0.003305

MAD2L1
2
−3.2685
0.001209

CCNA2
2
−3.12554
0.001952

CCNB1
2
−3.43884
0.000668

KIF20A
2
−3.16017
0.001741

CENPK
2
−3.38809
0.000799

ERAP1
2
−3.0003
0.002927

FAM54A
2
−3.5307
0.000481

PHTF2
2
−2.82226
0.005093

CLDN12
2
−3.10361
0.002097

BPGM
2
−2.89166
0.004117

PCMTD1
2
−2.92723
0.003686

MELK
2
−2.90007
0.004011

MST4
2
−3.46215
0.000615

CR1
3
−3.1375
0.001876

GOS2
3
−3.08943
0.002197

CSF3R
3
−3.326
0.000992

S100Al2
3
−2.93619
0.003584

SELL
3
−2.81834
0.005154

NCF2
3
−2.8535
0.00463

LIPN
3
−3.16551
0.00171

ZNF438
3
−3.31249
0.00104

NAMPT
3
−2.85014
0.004678

CBL
3
−3.77272
0.000195

CASP5
3
−2.96813
0.003242

CARD16
3
−3.32829
0.000985

CARD17
3
−2.96301
0.003295

CLEC4A
3
−2.9666
0.003257

LRRK2
3
−3.38419
0.00081

HMGN2P46
3
−3.56798
0.00042

AQP9
3
−3.18953
0.001578

BCL2A1
3
−2.88499
0.004203

ITGAX
3
−2.92718
0.003687

GPR97
3
−3.39828
0.000771

CCL4
3
−2.86648
0.004449

PSTPIP2
3
−2.88175
0.004245

IFI30
3
−2.81741
0.005168

FFAR2
3
−3.20036
0.001522

EMR3
3
−3.09579
0.002151

FPR1
3
−3.03905
0.002586

LILRA5
3
−2.87193
0.004375

PLEK
3
−3.08569
0.002223

MXD1
3
−2.9859
0.003064

TNFAIP6
3
−3.05049
0.002492

CXCR2
3
−3.51316
0.000512

IL1B
3
−3.27702
0.001174

CXCR1
3
−3.38462
0.000809

SIRPB1
3
−3.65291
0.000307

NCF4
3
−3.14344
0.00184

IRAK2
3
−3.27512
0.001182

PROK2
3
−3.43542
0.000677

TLR2
3
−2.82581
0.005038

TREM1
3
−2.96731
0.00325

SOD2
3
−3.16918
0.001689

CREB 5
3
−3.32746
0.000987

NAMPT
3
−2.80465
0.005372

TNFRSF10C
3
−2.80794
0.005318

CSGALNACT1
3
−3.54557
0.000455

ASAP1
3
−2.80888
0.005303

PLA2G2A
4
3.087887
0.002208

NFYC
4
3.20083
0.00152

RASSF10
4
3.105541
0.002084

GLB1L3
4
2.800517
0.005439

TRIM3
4
3.207674
0.001485

MCAM
4
2.70666
0.007192

MSRB3
4
3.432075
0.000685

SLITRK5
4
3.963535
9.27E−05

GAS6
4
2.820157
0.005125

NOVA1
4
2.729315
0.006728

GABRG3
4
2.904108
0.003961

ABCA3
4
3.321624
0.001008

LPO
4
3.513723
0.000511

FSCN2
4
2.781525
0.005759

RASD1
4
3.198556
0.001531

HILS1
4
3.002738
0.002905

SDK2
4
3.45176
0.000638

NTN5
4
3.159291
0.001746

KCNA7
4
3.631462
0.000332

ATOH8
4
3.062376
0.002398

KCNIP3
4
2.994468
0.002982

INHBB
4
3.056753
0.002442

VSTM2L
4
3.520433
0.000499

ZNRF3
4
3.592073
0.000384

PLEKHG4B
4
2.830968
0.00496

GNMT
4
3.274623
0.001184

GABBR1
4
2.881256
0.004252

ARHGEF10
4
3.270419
0.001201

SDC2
4
2.847433
0.004717

CRB2
4
3.452184
0.000637

GAS1
4
3.470173
0.000598

PNPLA7
4
2.715581
0.007006

RAI2
4
3.25911
0.001248

PLA2G2A
4
2.899771
0.004015

ID3
5
3.213565
0.001456

PGLYRP4
5
−3.64634
0.000314

SFTPA1
5
3.189676
0.001578

SFTPA1
5
3.189676
0.001578

LIPK
5
−2.99802
0.002949

SFTPA2
5
3.858853
0.00014

SFTPA2
5
3.858853
0.00014

ASCL3
5
2.764568
0.006059

RPPH1
5
2.832278
0.00494

CD209
5
3.331793
0.000973

GPR32
5
−2.98245
0.003098

UGT2A3
5
−2.84993
0.004681

CD58
6
−3.35673
0.000892

LBR
6
−3.62457
0.000341

ARHGDIB
6
−3.22018
0.001424

SLC12A6
6
−4.12377
4.85E−05

LPCAT2
6
−3.5142
0.00051

PLEKHB2
6
−3.02046
0.002745

KYNU
6
−4.13269
4.68E−05

ANKRD36B
6
−3.84654
0.000147

ANKRD36B
6
−3.49272
0.000551

ANKRD36B
6
−3.55319
0.000443

SRD5A3
6
−4.25485
2.81E−05

NEIL3
6
−3.23716
0.001345

TLR1
6
−2.90118
0.003997

BCAP29
6
−3.44601
0.000652

MGAM
6
−3.11565
0.002016

TPK1
6
−3.21156
0.001466

ATP6V1B2
6
−3.57198
0.000414

LYN
6
−3.1938
0.001556

SDCBP
6
−2.80868
0.005307

GK
6
−2.91501
0.003829

GLA
6
−3.30241
0.001077

ADRA2A
7
3.550905
0.000447

PRKCDBP
7
2.741146
0.006496

PRR4
7
3.898998
0.00012

PRB4
7
2.988547
0.003039

PRB3
7
3.690625
0.000266

PRB1
7
2.976402
0.003158

PRB2
7
3.201779
0.001515

BMP4
7
2.886357
0.004185

PRKCA
7
3.80558
0.000172

CYP1B1-AS1
7
2.720927
0.006897

CGREF1
7
3.275505
0.00118

RPRM
7
2.944897
0.003488

SDPR
7
2.84669
0.004728

BPIFB2
7
3.949534
9.80E−05

BPIFB 6
7
2.98498
0.003073

SNCA
7
2.777053
0.005837

CLDN22
7
3.502336
0.000533

COBL
7
3.1512
0.001793

NKX3-1
7
2.92659
0.003693

CDR1
7
4.307308
2.25E−05

CH25H
8
3.168911
0.001691

FXCl
8
3.17821
0.001639

DLG2
8
2.948964
0.003443

NRXN3
8
2.863338
0.004493

CES1P1
8
3.630652
0.000333

CES1
8
3.122943
0.001968

KCNJ16
8
3.53821
0.000468

APCDD1
8
3.103106
0.002101

TMEM178
8
2.7868
0.005668

MYRIP
8
2.958247
0.003344

FLNB
8
2.911823
0.003867

ENPP5
8
2.788207
0.005644

SEMA3E
8
2.940987
0.003531

SLC7A2
8
3.321666
0.001007

ARHGAP6
8
3.220932
0.001421

ANO3
9
−2.98289
0.003094

SLC22A10
9
−3.04806
0.002512

UFM1
9
−3.15354
0.001779

EPHX3
9
−3.73504
0.000225

KLF7
9
−2.84977
0.004683

LGSN
9
−3.5566
0.000438

LYPD2
9
−2.93177
0.003634

CES3
10
−3.3613
0.000878

MIA
10
−3.23844
0.001339

RNF150
10
−4.32839
2.06E−05

SLC9A3
10
−2.88577
0.004193

MYOT
11
−3.40228
0.000761

Example 2: Validation of Bronchial Genomic Classifier for Lung Cancer Patients Undergoaing Diagnostic Chronchoscopy
Introduction

Bronchoscopy is frequently non-diagnostic in patients with pulmonary lesions suspicious for lung cancer. This often results in additional invasive testing, although many lesions are benign. We sought to validate a bronchial gene expression classifier that could improve the diagnostic performance of bronchoscopy.

Lesions suspicious for lung cancer are frequently identified on chest imaging. The decision of whether to pursue surveillance imaging or an invasive evaluation requiring tissue sampling is complex and requires assessment of the likelihood of malignancy, ability to biopsy, surgical risk, and patient preferences [1]. When a biopsy is required, the approach can include bronchoscopy, transthoracic needle biopsy (TTNB), or surgical lung biopsy (SLB). The choice between these modalities is usually determined by considerations such as lesion size and location, presence of adenopathy, risk of procedure and local expertise. Bronchoscopy is a safe procedure, with less than 1% complicated by pneumothorax [2]. There are approximately 500,000 bronchoscopies performed per year in the U.S. [3], of which roughly half are for the diagnostic workup of lung cancer. However, bronchoscopy is limited by its sensitivity, ranging between 34-88% depending on the location and size of the lesion [4]. Even with newer bronchoscopic guidance techniques, the sensitivity is only ˜70% for peripheral lesions [5].

Patients with a non-diagnostic bronchoscopy often undergo further invasive testing to establish a definitive diagnosis. SLB is not the initial preferred approach given the inherent risks, with a complication rate of approximately of 5% and 30-day mortality of ˜1% [6]. Importantly, 20-25% of SLBs are performed in patients ultimately diagnosed with benign lesions [7,8].Furthermore, TTNB is associated with significant morbidity including a 15% pneumothorax rate [9], of which 6% require chest tube drainage [10,11]. Given the pitfalls of invasive procedures, alternative approaches are needed to identify patients with lower likelihood of malignancy who are appropriate for imaging surveillance.

Classification of biological disease states, including cancer, using gene expression measurements of clinical specimens is well established [12]. In the setting of lung cancer, there are distinct cancer-associated gene expression patterns in cytologically-normal epithelium collected from the proximal airways of smokers [13]. Recently, we developed a gene expression classifier in bronchial epithelial cells (BECs) collected from the mainstem bronchus via bronchoscopy that distinguishes patients with and without lung cancer amongst current and former smokers (manuscript submitted). We undertook the present studies to validate this classifier in two prospective multicenter trials of patients undergoing bronchoscopy for suspected lung cancer and to assess how this classifier alters the diagnostic performance of bronchoscopy.

Materials and Methods
Study Design, Population, and Protocol

Current and former smokers undergoing bronchoscopy for suspicion of lung cancer were enrolled in AEGIS 1 and AEGIS 2, two independent, prospective, multicenter, observational studies (NCT01309087 and NCT00746759). Patients were enrolled at 28 sites in the U.S, Canada, and Ireland. Cytology brushes were used to collect normal appearing BECs from the mainstem bronchus and submerged in an RNA preservative. Results of the classifier were not reported to physicians or patients. Patients were screened prior to bronchoscopy to determine if they met the requirements of the study protocol. Exclusion criteria included subjects <21 year old, never smokers (defined as smoking <100 cigarettes in lifetime), and patients with a concurrent cancer or a history of lung cancer. Patients who had been on a mechanical ventilator for >24 hours immediately prior to bronchoscopy, or could not consent or comply with the study, were excluded. Patients were followed until a final diagnosis was established or until 12 months post bronchoscopy. A diagnosis of lung cancer was established at the time of index bronchoscopy or by subsequent biopsy using TTNB, SLB, a second bronchoscopy, or other invasive procedures. The specific bronchoscopic methods used and subsequent surveillance imaging or procedures performed after non-diagnostic bronchoscopy was at the discretion of the treating physician. Patients diagnosed as cancer-free had a specific benign diagnosis or radiographic stability/resolution at 12 months of follow-up. Patients without a definitive diagnosis of cancer, a specific benign diagnosis or stability/resolution at 12 months of follow-up were excluded from further analysis. The treating physician assessed each patient's pre-test probability of malignancy (POM) prior to bronchoscopy using a five level scale (<10%, 10-39%, 40-60%, 61-85%, and >85%). The study protocol was approved by the institutional review board at each participating center, and all patients provided written informed consent before enrollment.

A total of 855 patients in AEGIS 1 and 502 in AEGIS 2 qualified and were enrolled between January 2009 and August 2012 (FIG. 4) at twenty-eight medical centers in the United States, Canada, and Ireland (Table 12). The sites were a mix of tertiary/academic medical centers (n=20), community based hospitals (n=6), and Veteran's Administration hospitals (n=2).

Additional patients were excluded from the study after enrollment based on the following criteria (see FIG. 4). First, a total of 111 patients in AEGIS 1 and 71 patients in AEGIS 2 were ineligible due to the following protocol deviations: In AEGIS 1, 24 patients either did not meet the study enrollment criteria upon review, had specimens collected that did not meet the acceptance criteria in accessioning, or were enrolled but did not have a specimen collected. An additional 87 samples did not meet minimum QC criteria after RNA isolation, either due to a RIN score <4, or RNA yield <1μg. In AEGIS 2, there were 3 ineligible patients due to having specimens collected that did not meet the acceptance criteria in accessioning, and 68 samples that did not meet the minimum QC criteria for RNA. Second, patients diagnosed with a non-primary lung cancer, including 40 patients in AEGIS 1 and 10 patients in AEGIS 2 were excluded. Additionally, upon review at 12 months, patients without a final diagnosis were also excluded from the study, including 91 subjects in AEGIS 1 and 71 patients in AEGIS 2. Approximately two thirds of these 162 patients (n=106) from AEGIS 1 and 2 were lost to follow-up after their initial bronchoscopy, while the remaining third (n=56) did not have a definitive diagnosis at 12 months post-bronchoscopy. Finally, we excluded 16 patients in AEGIS 1 and 9 patients in AEGIS 2 with gene expression data that failed the assay QC criteria (described below under Microarray Processing methods).

The AEGIS 1 cohort had previously been divided into equal training and test sets in a randomized manner. The training set (n=299 patients) was used to derive the gene expression classifier composed of 23 genes plus age and has been described elsewhere (manuscript submitted). In the training set, the classifier was found to have an AUC of 0.78 (95% CI, 0.71-0.85) in patients whose bronchoscopy did not lead to a diagnosis of lung cancer (n=134), with a sensitivity of 93% and specificity of 57%. The current study is focused on validation of the locked classifier in two independent validation sets.

The baseline demographics and clinical characteristics of the final AEGIS 1 and AEGIS 2 validation sets are compared in Table 1. A separate comparison of patients diagnosed with cancer and benign disease within each cohort is provided in Table 12.

Table 12—Characteristics of Patients in the AEGIS 1 and AEGIS 2 Test Sets

TABLE 12

Characteristics of patients in the AEGIS 1 and AEGIS 2 test sets.

AEGIS 1 Test Set
AEGIS 2 Test Set

Ca+
Ca−
P
Ca+
Ca−
P

N
220
78

267
74

Sex

0.506

0.091

Female
95
30

77
29

Male
125
48

190
45

Age (IQR) ^a
64
57
<0.001
65
60
<0.001

(15)
(14)

(13)
(18)

Race ^b

0.878

0.426

White
166
60

206
61

Black
42
13

55
11

Other/unknown
12
5

6
2

Smoking Status

0.065

0.026

Current
115
31

141
28

Former
105
47

126
46

Smoking History
45
30
<0.001
50
20
<0.001

(PY) (IQR) ^a
(30)
(36)

(35)
(30)

^aReported as the median value (and interquartile range; IQR). P-values calculated using the Mann-Whitney test.

^bP-value calculated for white vs. non-white.

Clinical and demographic data was collected for each patient and recorded on a study clinical report form (CRF). Additional pathology and radiology reports were maintained in the medical records of each patient and were available for review. All source documents were monitored and entered into databases maintained by the study sponsor. Size and location of the pulmonary lesions were obtained from the CT scan report. Subjects were followed up to twelve months post bronchoscopy to collect data for a clinical diagnosis. A diagnosis of lung cancer was based on results from pathology and copies of pathology reports were collected from the medical centers. The specimen leading to a clinical diagnosis of cancer was either obtained during bronchoscopy or from a subsequent invasive when bronchoscopy was non-diagnostic. Data on all clinical procedures performed after a non-diagnostic bronchoscopy was collected during the study. The subsequent evaluation of patients with non-diagnostic bronchoscopies was at the discretion of the treating physician at each study center. In some cases, more than one procedure after bronchoscopy was required to successfully render a diagnosis, but in all cases the procedure that led to the diagnosis of lung cancer was collected. Invasive follow-up procedures were defined as either a second bronchoscopy, TTNB, or surgical lung biopsy (SLB), which could include a mediastinoscopy, thoracoscopy, thoracotomy, or video-assisted thorascopic surgery (VATS).

The records of patients who were not diagnosed with cancer underwent an adjudication process in order to declare subjects as cancer-free. The process consisted of a review of the available medical records for each patient by a panel of five pulmonologists. Two pulmonologists from this panel independently reviewed each case and the patient was determined to be cancer-free if he/she met one of the following criteria: patient was diagnosed with an alternative diagnosis that explained the initial suspicious abnormality, the abnormality was determined to be stable, or the abnormality resolved. Patients whom did not meet these criteria at the completion of the 12-month follow-up period had no final diagnosis and were excluded from further analysis in the study.

Histology (Table 13) and cancer stage (Table 14) data was analyzed for patients diagnosed with primary lung cancer. Data was extracted from the study CRFs and were confirmed by review of complete medical records. Of the total cancers, 80% (175/220; 95% CI, 74-84%) in AEGIS 1 and 83% (222/267; 95% CI, 78-87%) in AEGIS 2 were NSCLC. There were 17% (38/220; 95% CI, 13-23%) and 16% (42/267 95% CI, 12-21%) SCLC cancers, respectively. Of the patients with NSCLC with known cancer stage data, early stage (stage I or 2) was present in 40% (53/139 95% CI, 32-48%) in AEGIS 1 and 37% (67/199; 95% CI, 30-44%) in AEGIS 2.

Table 13-Summary demographics and clinical characteristics of study participants: A comparison of AEGIS 1 vs. AEGIS 2 patients was performed across all clinical characteristics. Significant differences are indicated by *p<05, **p<001. The p-value for race was calculated for white versus non-white, and for lung cancer histology using NSCLC versus SCLC.

AEGIS 1
AEGIS 2

N
298
341

Sex*

Female
125
106

Male
173
235

Age*, median(IQR)
62 (16)
64 (15)

Race

White
226
267

Black
55
66

Other
15
4

Unknown
2
4

Smoking Status

Current
146
169

Former
152
172

Tobacco Pack-years, median, (IQR)
40 (36)
45 (38)

Lesion Size**

<2
cm
48
83

2 to 3
cm
41
39

>3
cm
155
188

Infiltrate
32
28

Unknown
22
3

Lesion Location**

Central
98
127

Peripheral
86
108

Central & Peripheral
90
102

Unknown
24
4

Lung Cancer Histology
220
267

Small cell lung cancer
38
42

Nonsmall cell lung cancer
175
222

Adenocarcinoma
69
100

Squamous
72
81

Large cell
8
8

NSCLC, not otherwise specified
26
33

Unknown
7
3

Benign Diagnoses
78
74

Infection
18
14

Sarcoidosis
16
15

Resolution or Stability
27
24

Other
17
21

TABLE 14

Stage data of patients diagnosed with primary lung cancer.

Histology
Stage
AEGIS 1
AEGIS 2

SCLC

38
42

Limited
16
14

Extensive
19
23

Unknown
3
5

NSCLC

175
222

1
37
49

2
16
18

3
44
65

4
42
67

Unknown
36
23

Unknown

7
3

Total

220
267

Data about each of the benign diagnoses was extracted from medical records and summarized in Table 15. In both studies approximately two-thirds of cancer-free subjects had alternative diagnoses, with the specific alternative diagnoses enumerated in Table S4 for each study. Diagnosis of patients with abnormalities that were determined to have resolved or were stable was based on review of medical records, including follow-up imaging, at 12 months post-bronchoscopy, and approximately one-third of cancer-free patients were in this category in both studies.

Table 15—Alternative Diagnoses of Patients With Benign Disease

TABLE 15

Alternative diagnosis of patients with benign disease

Category
AEGIS 1
AEGIS 2

Resolution or Stability*
27
24

Alternative Diagnosis
51
50

Sarcoidosis
16
15

Inflammation
4
4

Benign growth
5
6

Fibrosis
4
5

Other
4
6

Infection
18
14

Fungal
8
3

Mycobacteria
5
4

Bacterial
5
7

Total
78
74

During the clinically indicated bronchoscopy, study participants underwent collection of mucosal brushings of the right or left main-stem bronchus using a standard, disposable cytology brush rubbed against the bronchial wall. Two brushings were obtained from each subject; physicians were trained to obtain brushings of normal appearing epithelial tissue and to avoid sampling tumor tissue or dysplastic cells. All samples were stored in an RNA preservative (RNAprotect; Qiagen) and were shipped to a central CLIA-certified laboratory for accessioning and processing. Centers were requested to store samples at 4° C. for up to 14 days, and to ship at sub-ambient temperatures (4-20° C.) using 2-day shipping and insulated shipping containers (NanoCool) provided by the sponsor.

Laboratory Methods

All BEC specimens were processed to isolate and analyze RNA for quality and yield prior to gene expression analysis; Only specimens with an RNA yield of at least 1 μg and RIN score of >4 were run on Gene-ST 1.0 microarrays. All microarray data has been deposited in GEO as GSE66499.

The samples were lysed and RNA was isolated using a column-based method (miRNeasy Kit; Qiagen) and the manufacturer's recommended protocol. The large RNA fraction was analyzed on a spectrophotometer (Nanodrop; ThermoFisher) to determine the concentration, purity, and yield. Samples were also analyzed for RNA integrity (Bioanalyzer), reported as the RIN score. Samples with a yield of <1 μg and RIN score <4 were excluded from the study. Approximately 10% of specimens in AEGIS 1 and 13% in AEGIS 2 were excluded due to insufficient RNA quality or quantity. The RNA was then stored frozen (−80° C.) until further processing.

200 ng of total RNA was converted to sense strand cDNA using Ambion WT Expression kit (Life technologies Cat. #4440536), and subsequently labeled with Affymetrix GeneChip WT terminal labeling kit (Affymetrix Cat. #900671). For hybridization, a hybridization cocktail was prepared and added to the labeled cDNA target using the Hybridization, Wash and Stain kit (Affymetrix Cat. #900720), applied to Human Gene 1.0 ST arrays (Affymetrix Cat. #901087), and incubated at 45° C. for 16 hours. Following hybridization, arrays were washed and stained using standard Affymetrix procedures before they were scanned on the Affymetrix GeneChip Scanner. Data was extracted using Expression Console software (Affymetrix).

The Affymetrix Human Gene 1.0 ST array (Affymetrix Cat. #901087) utilizes numerous probes that are disclosed at the world wide web address: affymetrix.com/site/include/byproduct.affx?product-hugene-1_0-st-v1. Multiple probes are present that correspond to segments of specific genes, which is to say that that is not a 1:1 ratio of one probe per gene, rather multiple probes are present that correspond to multiple segments of a single gene. In one example, the LYPD2 gene is represented by three probe sets in the Human Gene 1.0 ST array (Release 32), probeset IDs 8153343, 8153344, and 8153345 as disclosed in the Affymetrix Human Gene 1.0 ST array (HuGene-1_0-st-v1 Probeset Annotations, release 32 on Sep. 30, 2011, release 33 on Mar. 27, 2013, and release 34 on Apr. 7, 2014). Information correlating probe-sets and nucleic acid sequences can be found at Affymetrix.com, including at the world wide web address: affymetrix.com/site/include/byproduct.affx?product-hugene-1_0-st-v1. Furthermore, data sets are available on the NCBI Gene Expression Omnibus website under Platform GPL6244 detailing the probes and genes that may be utilized in practicing the methods of the present disclosure at the world wide web address: ncbi.nlm.nih.gov/geo/query/acc.cgi?acc-GPL6244. These documents, including those correlating the probesets and gene symbols are incorporated herein by reference.

Microarray data (CEL files) corresponding to the samples in the final data set of each cohort were normalized using RMA [22]. The AEGIS 1 samples were run in a total of 5 batches and ComBat was used to correct for batch effects. The AEGIS 2 samples were all run in a single microarray batch. After normalization, outliers were identified as having a genome-wide pairwise correlation of less than 0.955 in global gene expression. All microarray data has been deposited in GEO under accession #GSE66499.

Normalization and preprocessing of the microarray data are described in the Supplementary Appendix. Subjects enrolled in AEGIS 1 had previously been randomly assigned into independent training and validation sets (FIG. 4) and the classifier algorithm was derived strictly within the AEGIS 1 training set and locked, as described previously. Scores for each sample in the AEGIS 1 validation set and for all AEGIS 2 samples were generated using this pre-specified classifier that was based on the expression of 23 genes and patient age. These scores were dichotomized as test-positive and test-negative using a pre-specified threshold value.

Performance of the classifier was evaluated using receiver operator characteristic (ROC) curves, calculation of area under the curve (AUC) [14], and estimates of sensitivity, specificity, negative predictive value (NPV), positive predictive value (PPV) and the negative likelihood ratio (NLR) which was defined as (1-sensitivity)/specificity. A Mann-Whitney non-parametric test was used for analysis of continuous variables and Fisher's exact test was used for categorical variables. All confidence intervals are reported as the two-sided binomial 95% confidence intervals (95% CI). Statistical analysis was performed using R software (Version 3.01).

Physicians were asked to assess the pre-test POM for all subjects enrolled based on expert opinion and available medical records prior to bronchoscopy, using one of the following categories: <10%, 10-39%, 40-60%, 61-85% , and >85%. The results were collapsed into categories of <10%, 10-60%, and >60%. We then calculated the actual prevalence of cancer in each of the POM categories in order to compare the actual prevalence of cancer in the stratified POM levels (Table 16). Additionally, we found that the prevalence of lung cancer post-non diagnostic bronchoscopy was 3%, 29% and 80% in the <10%, 10-60% and >60% POM groups respectively.

Table 16—Correlation of Physician Assessed POM and the Prevalence of Cancer

TABLE 16

Correlation of physician assessed

POM and the prevalence of cancer.

Pre-test POM ^a
CA + patients
Total patients
Prevalence ^b

<10%
3
62
5%

10-60%
41
101
41%

>60%
404
425
95%

Unknown
39
51
77%

^aPOM = probability of malignancy; assessed by treating physicians prior to bronchoscopy.

^bPrevalence was calculated as the fraction diagnosed with lung cancer within each of the POM categories.

The prediction accuracy of the bronchial genomic classifier stratified by lesion size is summarized in Table 2, and by pre-test POM in Table 3. The sensitivity of the classifier is also reported for patients diagnosed with lung cancer stratified by stage and histology, in Tables S5 and S6. The classifier in combination with bronchoscopy leads to a sensitivity >90 % for all categories.

Table 17-Sensitivity of the classifier, bronchoscopy, and the combined approaches according to radiological imaging characteristics. Sensitivity of bronchoscopy was determined for lung cancer patients in each category. Sensitivity of the classifier was determined for the patients with lung cancer who were not diagnosed during bronchoscopy. Sensitivity of the classifier combined with bronchoscopy was calculated for all lung cancer patients in each category.

TABLE 17

Sensitivity of the classifier, bronchoscopy, and the combined approaches according

to radiological imaging characteristics. Sensitivity of bronchoscopy was

determined for lung cancer patients in each category. Sensitivity of the

classifier was determined for the patients with lung cancer who were not

diagnosed during bronchoscopy. Sensitivity of the classifier combined with

bronchoscopy was calculated for all lung cancer patients in each category.

Classifier &

bronchoscopy

Bronchoscopy
Classifier
combined

Size (mm)
N_total
N_cancer
(95% CI)
(95% CI)
(95% CI)

All patients
639
487
75% (71-79)
89% (82-94)
97% (95-98)

Size (mm)

<2
cm
131
73
55% (43-66)
91% (76-98)
96% (88-99)

2-3
cm
80
60
58% (46-70)
92% (74-99)
97% (88-100)

>3
cm
343
313
82% (78-86)
85% (74-93)
97% (95-99)

Infiltrate
60
25
84% (65-94)
100% (45-100)
100% (84-100)

Unknown
25
16
80% (54-94)
100% (38-100)
100% (76-100)

Location

Central
225
174
84% (78-89)
81% (63-92)
97% (93-99)

Peripheral
194
133
55% (46-63)
90% (80-96)
95% (90-98)

Both
192
164
82% (75-87)
97% (82-100)
99% (96-100)

Unknown
28
16
81% (56-94)
67% (20-94)
94% (70-100)

Results
Characteristics of the Study Participants

298 patients from AEGIS 1 served as a first validation set and all 341 patients from AEGIS 2 meeting study criteria were used as a second validation set (FIG. 4 & Table 13). The prevalence of lung cancer was 74% and 78% for the AEGIS 1 and AEGIS 2 cohorts, respectively. Patients with lung cancer were older (p<0.001), had higher cumulative tobacco exposure compared to patients without cancer (p<0.001) and were more likely to be current smokers (p=0.07 in AEGIS 1 and p=0.03 in AEGIS 2; Table 12). A summary of cancer stage and categories of benign diagnoses are shown in Tables 15 and 16, respectively.

Performance of Bronchoscopy

A total of 639 patients underwent bronchoscopy for suspected lung cancer. Of those, 272 (43%; 95%CI, 39 to 46) were non-diagnostic, including 120 of 487 patients (25%; 95%CI, 21 to 29) ultimately diagnosed with lung cancer. The sensitivity of bronchoscopy for lung cancer was 74% (95%CI, 68 to 79) and 76% (95%CI, 71 to 81) in AEGIS 1 and AEGIS 2, respectively. Follow-up procedure data was available for 98% (267 of 272) of the patients with a non-diagnostic bronchoscopy. Invasive procedures following non-diagnostic bronchoscopy were performed in 170 of 267 patients (64%; 95%CI, 58 to 69), including 52 of 147 (35%; 95%CI, 24 to 38) with benign lesions and 118 of 120 (98%; 95%CI, 94 to 99) with cancer. SLB was performed in 76 patients, of which 27 (36%; 95%CI, 26 to 47) had benign lesions.

Performance of Gene Expression Classifier

The classifier alone had an AUC-0.78 (95% CI, 0.73 to 0.83) and accurately identified 194 of 220 patients with cancer (88% sensitivity; 95% CI, 83 to 92), and 37 of 78 patients without cancer (47% specificity; 95% CI, 37 to 58) in AEGIS 1 (FIG. 5). In AEGIS 2, the classifier had an AUC-0.74 (95% CI, 0.68 to 0.80) and correctly identified 237 of 267 patients with cancer (89% sensitivity; 95% CI, 84 to 92), and 35 of 74 without cancer (47% specificity; 95%, 36 to 59). The combination of the classifier with bronchoscopy increased the sensitivity to 96% (95%CI, 93 to 98) and 98% (95%CI, 96 to 99) in AEGIS 1 and 2, respectively compared to 74% and 76%, respectively, for bronchoscopy alone (p<0.001).

In patients with a non-diagnostic bronchoscopy, the classifier accurately identified cancer in 49 of 57 patients in AEGIS 1 (86% sensitivity; 95% CI, 74 to 94) and in 62 of 67 patients in AEGIS 2 (92% sensitivity; 95% CI, 82 to 97). As there was no significant difference between patients in the two cohorts with regard to the classifier AUC either across all patients (p=0.32) or for patients with a non-diagnostic bronchoscopy (p=0.61) (FIG. 5), we combined the two cohorts for subsequent analyses of sub-groups. The sensitivity of bronchoscopy alone was lower in lesions that were <3 cm (p<0.001) or peripherally located (p<0.001) (Table 17). In contrast, the sensitivity of the classifier alone and the classifier combined with bronchoscopy were consistently high and not significantly associated with size or location of the lesion (Table 17), cancer stage (Table 18) or histological subtype (Table 19).

Table 18—Sensitivity of the Classifier and Bronchoscopy According to Cancer Stage

TABLE 18

Sensitivity of the classifier and bronchoscopy

according to cancer stage.

Broncho-

Histology
Stage
N
scopy
Classifier
Combined

NSCLC
1
86
41%
88%
93%

2
34
68%
82%
94%

3
109
83%
94%
99%

4
109
86%
93%
99%

Unknown
59
80%
92%
98%

Total NSCLC

397
73%
89%
97%

SCLC
Limited
30
80%
83%
97%

Extensive
42
93%
100%
100%

Unknown
8
88%
n/a
88%

Total SCLC

80
88%
80%
98%

Unknown

10
70%
100%
100%

Total Cancers

487
74%
86%
97%

Table 19—Sensitivity of the Classifier and Bronchoscopy According to Histology

TABLE 19

Sensitivity of the classifier and bronchoscopy

according to histology.

Histology
N
Bronchoscopy
Classifier
Combined

Adenocarcinoma
169
65%
86%
95%

Squamous
153
80%
90%
98%

Large Cell
16
75%
100%
100%

NSCLC-NOS
59
78%
100%
100%

SCLC
80
88%
80%
98%

Unknown
10
70%
100%
100%

Accuracy of the Classifier in Patients With Intermediate Probability of Cancer

We combined the physician-assessed probability of malignancy (POM) into categories of low (<10%), intermediate (10-60%), and high (>60%) POM (Table 3), to align with guideline recommendations for assessing lung cancer risk [1]. Bronchoscopy was non-diagnostic for cancer in 83% of patients with an intermediate pre-test POM (n=101), despite a 41% cancer prevalence rate. In this group of patients, the classifier achieved a NPV of 91% (95%CI, 75 to 98) among those with a non-diagnostic bronchoscopy and a PPV of 40% (95%CI, 28 to 54) (Table 20).

Table 20—Performance of Bronchoscopy and the Classifier Stratified Pre-Test POM

TABLE 20

Performance of bronchoscopy and the classifier stratified pre-test POM.

Pre-test POM Category

N
<10%
10-60%
>60%
Unknown

Patient population

Total Patients ^a
639
62
101
425
51

Lung Cancer
487
3 (5%)
41 (41%)
404 (95%)
39 (76%)

Benign
152
59 (95%)
60 (59%)
21 (5%)
12 (24%)

Bronchoscopy performance

Bronchoscopy sensitivity (95% CI)

33% (6-80)
41% (28-57)
79% (74-82)
79% (64-89)

Patients with non-diagnostic bronchoscopy ^b
272
61 (98%)
84 (83%)
107 (25%)
20 (39%)

Classifier performance

Classifier sensitivity (95% CI) ^c

100% (29-100)
88% (68-96)
90% (82-95)
88% (51-100)

Classifier specificity (95% CI) ^d

56% (43-68)
48% (36-61)
29% (14-50)
33% (14-61)

Classifier NPV (95% CI) ^e

100% (88-100)
91% (75-98)
40% (20-64)
80% (36-98)

Classifier PPV (95% CI) ^f

7% (1-24)
40% (28-54)
84% (75-90)
47% (25-70)

Combined classifier & bronchoscopy sensitivity

100% (38-100)
93% (80-98)
98% (96-99)
97% (86-100)

^aThere were 639 total patients across the POM categories shown.

^bBronchoscopy was non-diagnostic for 272 of 639 patients (43%; 95% CI, 39 to 46), including 120 of 487 patients (25%; 95% CI, 21 to 29) diagnosed with lung cancer.

^cThe classifier accurately predicted 107 of 120 total patients overall (89%; 95% CI, 82 to 94) with cancer. Sensitivity is reported in each POM category for patients with non-diagnostic bronchoscopy procedures.

^dThe classifier accurately predicted 72 of 152 patients overall (47%; 95% CI, 40 to 55) without cancer. Specificity is reported in each POM category for patients with non-diagnostic bronchoscopy procedures.

^eNPV, and ^fPPV is reported for patients with non-diagnostic bronchoscopy procedures.

Although the classifier had a high NPV in patients with a non-diagnostic bronchoscopy, there were 13 patients with a non-diagnostic bronchoscopy who had lung cancer and a negative classifier score (i.e. false negatives). The majority (10 of 13) had a high (>60%) POM with only three patients in the 10-60% pre-test POM group.

The NLR of the classifier in combination with bronchoscopy was calculated to determine the range of pre-test POM in which the post-test probability would be <10%. The NLR of bronchoscopy (0.244; 95% CI, 0.21 to 0.29) improves when combined with the classifier to 0.056 (95% CI, 0.03 to 0.10). As a result, when both bronchoscopy and the classifier are negative, the post-test POM is reduced to <10% for patients with a pre-test POM up to 66% (FIG. 6).

Evaluation of Results

This study describes the validation of a bronchial genomic classifier that identifies patients without lung cancer among those undergoing bronchoscopy in two independent prospective cohorts. We find that the gene-expression classifier has high sensitivity across different sizes, locations, stages, and cell types of lung cancer in the combined cohorts. The combination of the classifier and bronchoscopy has a sensitivity of 96% and 98% in the AEGIS-1 and AEGIS-2 validation cohorts respectively. We also report several additional findings that support the clinical need for this type of classifier. First, our studies confirm the previously reported observations that non-diagnostic bronchoscopy is common (particularly in patients of intermediate pre-test POM) and leads to further invasive testing including SLB, often in patients ultimately found to not have lung cancer. Second, in contrast to the high sensitivity of the classifier, we find that bronchoscopy performed poorly in small, peripheral or early stage cancers. Third, the classifier has a high NPV in patients with intermediate POM and a non-diagnostic bronchoscopy. These findings suggest that this classifier has the potential to assist clinical decision making in patients with intermediate POM in whom the prevalence of lung cancer is 41% but the sensitivity of bronchoscopy is only 41%. Due to the high NPV, a negative classifier score in patients with a non-diagnostic bronchoscopy and intermediate POM warrants a more conservative diagnostic strategy with active surveillance via imaging.

Although the high NPV of the classifier would help avoid unnecessary invasive procedures in patients with an intermediate POM that are classifier negative, there were a small number of patients in this group who have lung cancer; the negative gene-expression classifier result may delay further invasive testing in these patients. However, this group of patients would undergo active surveillance via imaging, which is the standard practice when an immediate invasive strategy is not employed [1,15]. This would allow for identification of lesion growth, triggering additional invasive testing to establish a definitive diagnosis. In contrast to the high NPV observed in patients with an intermediate POM, the classifier has a modest PPV of 40% in this setting. Thus, a positive result with the classifier does not warrant alteration in the diagnostic strategy; further testing would need to be based on traditional factors used to choose between an invasive versus an imaging surveillance strategy.

This gene expression classifier is measured in proximal BECs and not from cells within the pulmonary lesion. The ability of gene-expression alterations in cytologically-normal proximal airway to detect the presence of lung cancer within the lung parenchyma stems directly from the “field of injury” paradigm [13]. Spira, et al. has previously shown that there is a distinct pattern of gene-expression alterations in cytologically-normal bronchial epithelial cells among current and previous smokers with lung cancer [13,16]. Additionally oncogenic signaling pathways are activated in the proximal airway epithelium of smokers with lung cancer and smokers with premalignant airway lesions [17]. More recently, Kadara, et al. [18] demonstrated that genes whose expression is altered in non-small cell lung cancer itself and the adjacent small airway epithelium are enriched among those genes that are altered in the proximal airway epithelium, suggesting that the gene-expression changes within the proximal airway reflect, in part, the altered transcriptome observed in lung tumors.

A bronchoscopy was considered as “diagnostic” only when the procedure yielded a lung cancer diagnosis. There were a relatively small number of bronchoscopies that were potentially diagnostic of a specific benign etiology, but most of these patients received further invasive testing including patients ultimately diagnosed with lung cancer, suggesting that the concern for lung cancer remained elevated despite the initial benign finding on bronchoscopy. Finally, we did not assess the accuracy of a model incorporating the classifier in combination with clinical variables. Although clinical risk prediction models have been developed for solitary pulmonary nodules [1,20,21], there are no validated models for patients selected to undergo diagnostic bronchoscopy, which includes patients with a broad range of findings, including larger lesions (i.e. >3 cm), infiltrates or other features such as lymphadenopathy. Thus, most patients are selected for bronchoscopy based on the physician's qualitative assessment of lung cancer probability. Importantly, we demonstrate that our classifier performs well in patients with intermediate POM by physician assessment, a process that incorporates available clinical risk factors.

The potential impact of this work is bolstered by a number of key strengths in its study design. First, these were two independent prospective validation studies in which the classifier was measured in the setting in which the test would be used clinically (i.e. prior to diagnosis). This is a critical step in moving molecular biomarkers from discovery studies to their ultimate clinical application. Second, the large multicenter design enabled inclusion of patients undergoing bronchoscopy from different practice settings and geographic locations. Third, our data demonstrates the high prevalence of lung cancer among patients with a non-diagnostic bronchoscopy and intermediate pre-test POM, and show that the classifier has an NPV of 91% in this setting where there is the greatest uncertainty about cancer status. In this setting, the opportunity to use a classifier with high sensitivity for lung cancer allows for the confident identification of patients without lung cancer that might be followed by active surveillance, thereby avoiding potentially harmful invasive follow up testing.

Example 3: Derivation of Bronchial Genomic Classifier for Lung Cancer Patients In Patients Undergoing Diagnostic Chronchoscopy
Introduction

Lung cancer remains the leading cause of cancer mortality in the United States, with an estimated 224,000 new diagnoses, and 160,000 deaths in 2014, 90% of which are due to

smoking [24]. Recently, the National Lung Cancer Screening Trial showed that low dose Computed Tomography (CT) screening results in a 20% relative mortality reduction in high

risk individuals [25]. The mortality reduction, however, was accompanied by a high rate (˜96%) of false-positive CT findings, which in turn has generated concern for the overutilization of invasive diagnostic procedures [26].

Patients with suspected lung cancer are often referred for bronchoscopy where the primary aim is to sample a suspicious pulmonary lesion for pathological analysis. It is estimated that 500,000 bronchoscopies are performed per year in the U.S. [27], of which roughly half are for the diagnosis of lung cancer. Bronchoscopy is considered to be safer than other invasive sampling methods, such as transthoracic needle biopsy (TTNB), or surgical techniques. However the diagnostic sensitivity of bronchoscopy is sub-optimal, ranging from 34% (for <2 cm peripheral nodules) to 88% (for larger, centrally located lesions) [28]. Adoption of guidance techniques has expanded the applicability of bronchoscopy to more challenging suspicious lesions (i.e., solitary pulmonary nodules which are often peripheral in the lung), but the overall clinical sensitivity of bronchoscopy for lung cancer has not improved substantially [29,30]. When bronchoscopy is non-diagnostic, physicians are often left with the ambiguity of whether to pursue further invasive diagnostic procedures, with associated complications [31,32], or choose imaging surveillance. In current practice when these invasive procedures are performed, approximately a third of patients are determined to have benign disease [33], suggesting that these procedures are avoidable. Methods that reduce this ambiguity by substantially improving the diagnostic yield of bronchoscopy could improve patient care.

It has previously been demonstrated that cigarette smoke creates a molecular field of injury in airway epithelial cells that line the entire respiratory tract [34]. The reversible and irreversible impact of cigarette smoke on the bronchial airway transcriptome has been characterized and a set of gene-expression alterations in the bronchial epithelium have been identified in current and former smokers with lung cancer [35]. These cancer-associated gene expression profiles have previously been shown to yield a sensitive classifier for detecting lung cancer when bronchoscopy is non-diagnostic. The high sensitivity of this classifier, measured in a biospecimen readily accessible during bronchoscopy, results in a very low probability of lung cancer when the test result is negative, and suggests that physicians might be enabled to confidently pursue active surveillance and reduce risky invasive procedures in subjects without lung cancer.

Materials and Methods
Training Set Patient Population

Patients were enrolled in the AEGIS trials (Airway Epithelium Gene Expression In the DiagnosiS of Lung Cancer), designed as prospective, observational, cohort studies (registered as NCT01309087 and NCT00746759) of current and former cigarette smokers with a suspicion of lung cancer undergoing bronchoscopy as part of their diagnostic workup. A set of patients from one of the cohorts (“AEGIS 1”) was selected for the exclusive purpose of training a gene expression classifier. The study was approved by IRB at each of the participating medical centers, and all patients signed an informed consent prior to enrollment. All enrolled patients were followed post-bronchoscopy until a final diagnosis was made, or for 12 months. Patients were diagnosed as having primary lung cancer based on cytopathology obtained at bronchoscopy or upon subsequent lung biopsy (such as TTNB or surgical lung biopsy (SLB) when bronchoscopy did not lead to a diagnosis of lung cancer). Patients were diagnosed as having benign disease based on a review of medical records and follow-up procedures at 12 months post-bronchoscopy (described in more detail in Additional File 1). Bronchoscopy was considered “diagnostic” when clinical samples collected at the time of the bronchoscopy procedure yielded a confirmed lung cancer diagnosis via cytology or pathology.

Patients enrolled in the NIH-registered AEGIS studies (NCT01309087 and NCT00746759) were patients undergoing clinically indicated bronchoscopy for suspicion of lung cancer who were at least 21 years of age and had smoked at least 100 cigarettes in their lifetime. Study exclusions included patients who had previously been diagnosed with primary lung cancer, who had been on a mechanical ventilator for ≥24 consecutive hours immediately prior to bronchoscopy, or who could not consent or comply with the study. Additional patients were excluded prior to training the classifier to exclude patients with a malignancy other than primary lung cancer. This included the exclusion of patients with a history of any malignancy, confirmed metastatic cancer to the lung, or found to have an active non-lung primary cancer after enrollment. Also, patients without a final definitive diagnosis were excluded. Finally, after specimen processing, those with insufficient yield (<1 μg) or quality (RIN<4) of RNA were excluded from further analysis.

Patients were followed for up to twelve months post bronchoscopy and records were reviewed to confirm or determine a final clinical diagnosis. A diagnosis of cancer was based on cytopathology of cells/tissue collected either during bronchoscopy, or in follow-up procedures when bronchoscopy was non-diagnostic. Follow-up procedures leading to diagnosis consisted of a second bronchoscopy, transthoracic needle aspiration (TTNA), surgical lung biopsy (SLB), or a combination of procedures. Records of patients who were not diagnosed with cancer, and who had been followed for 12 months, underwent an adjudication process by a panel of five pulmonologists. The process consisted of a review of the available medical records and patients were only declared to be cancer-free if the patient met one of the following criteria: diagnosed with an alternative diagnosis that explained the initial suspicious abnormality, the abnormality was determined to be stable, or the abnormality resolved. Patients who did not meet these criteria at the completion of the 12-month follow-up period were labeled as “indeterminate” and were excluded from training, due to lack of diagnostic “truth”.

Sample Collection

Physicians at each of 25 participating medical centers were instructed to collect normal appearing bronchial epithelial cells (BEC) from the right mainstem bronchus (or the left side if any abnormalities were observed on the right) during bronchoscopy using standard bronchoscopic cytology brushes. Following collection, the cytology brushes were cut and placed in an RNA preservative (Qiagen RNAProtect, Cat. 76526) immediately after collection and stored at 4° C. Specimens were then shipped at 4-20° C. to a central laboratory for further processing.

A shipping container was provided to all sites enabling the transport of specimens at 4-20° C. within a 48 hour period. Sites were asked to send specimens using 2-day shipping services. Upon receipt in the central laboratory, specimens were inspected and accessioned into a laboratory information system. Accepted specimens were stored at 4° C. prior to RNA isolation, which was typically conducted within 7 days of receipt. Records of all storage, and shipping times were retained, and the cumulative time between specimen collection and RNA isolation was less than 30 days (consistent with manufacturer's recommendations for the RNA preservative).

RNA Isolation

BECs were separated from cytology brushes using a vortex mixer and were then pelleted and processed using QIAzol lysis reagent (Qiagen). RNA was isolated by phenol/chloroform extractions and purified on a silica membrane spin-column (Qiagen miRNeasy kit, Cat. #217004) according to manufacturer's recommendations. RNA was analyzed on a NanoDrop ND-1000 spectrophotometer (Thermo Scientific) to determine concentration and purity, and RNA integrity (RIN) was measured on a 2100 Bioanalyzer (Agilent Technologies). Each sample was then stored at −80° C. until processing further on microarrays.

cDNA Preparation

Total RNA was converted to sense strand cDNA, amplified using the Ambion WT Expression kit (Life Technologies Cat. #4440536) designed for use with Affymetrix microarrays. Starting with 200ng of total RNA, single stranded cDNA was prepared through reverse transcription using T7 promoter primers protocol. Single-strand cDNA was converted to double stranded cDNA using DNA polymerase.

Microarray Processing

cDNA obtained from the total RNA was labeled with Affymetrix GeneChip WT terminal labeling kit (Affymetrix Cat. #900671). The labeled cDNA was hybridized to Gene 1.0 ST microarrays (Affymetrix Cat. #901085) and analyzed on an Affymetrix GeneChip Scanner. Individual CEL files for each of the patient samples were normalized using the standard Affymetrix Gene 1.0 ST CDF and RMA [38].

Total RNA was converted to sense-strand cDNA using a commercial kit (Ambion WT; Life Technologies, Cat. #4440536) designed for use with Affymetrix microarrays. Starting with 200ng of total RNA, single stranded cDNA was prepared through reverse transcription using T7 promoter primers protocol. Single-strand cDNA was converted to double stranded cDNA using DNA polymerase. Double stranded cDNA acts as a template for in vitro transcription of cRNA which was then purified to remove enzymes, salts, inorganic phosphates and unincorporated nucleotides. The yield of cRNA was measured using UV-adsorption and labeled sense-stranded cDNA was then generated using 10 μg of the purified CRNA by reverse transcription with random primers and a mix of dUTP/dNTPs, fragmented, and labeled using the GeneChip WT Terminal labeling kit (Affymetrix, Cat. #900671). The labeled cDNA was hybridized to Gene 1.0 ST microarrays (Affymetrix Cat. #901085) using the Hybridization, Wash and Stain kit (Affymetrix Cat. #900720), and incubated at 45° C. for 16 hours. Following hybridization, arrays were washed and stained using standard Affymetrix procedures before being scanned on the Affymetrix GeneChip Scanner, and data was extracted using Expression Console software (Affymetrix). Thus, microarrays described herein are not measuring the expression of natural molecules, rather the microarrays measure the expression of non-naturally occuring cDNA molecules.

Classifier Development

A gene expression classifier was derived in a multi-step process. Initial modeling consisted of using the training data to select genes (“gene expression correlates”) which were associated with three clinical covariates (gender, tobacco use, and smoking history) to identify gene expression correlates of these clinical variables. Lung cancer-associated genes were then selected, and finally a classifier for predicting the likelihood of lung cancer based on the combination of the cancer genes, the gene expression correlates, and patient age was determined. All aspects of this classifier development procedure were determined using cross validation and using only data from the training set samples.

Clinical Factor Gene Expression Correlates (CFGC)

Covariates of lung cancer in this study population, including sex (male/female), smoking status (current/former), and pack years (<10/>10), were modeled to identify gene expression correlates for the clinical factors. Empirical Bayes t-tests were used to identify genes whose expression was significantly associated with each of the clinical factors. Next,

genes were selected in a sparse manner that could be used to predict the value of each clinical factor. Finally, the predicted values from the gene expression correlates for gender (GG), smoking status (GS), and pack-years (GPY) were computed for each patient and used in selecting genes with lung-cancer associated gene expression and in the lung cancer classifier described below.

Selection of Lung Cancer Genes

A logistic regression model with lung cancer status (1=cancer-positive and 0=cancer-negative) as the dependent variable was fit using the training data, CFGC's, and patient age as predictors. Next an empirical Bayes linear model was fit using gene expression values as the independent variable and the logistic regression model residuals as the dependent variable. This was used to select genes most directly correlated with disease status and independent of clinical covariates. The top lung cancer-associated genes from this analysis were grouped using hierarchical clustering. Genes were selected in an iterative manner to maximize AUC using cross-validation to estimate prediction accuracy. The aim was to select clusters that cumulatively provide the best classifier performance, and specific genes that best represent each of the clusters. A gene titration analysis was also performed to determine the number of genes per cluster providing optimal performance. For the clusters selected, the top genes were averaged, yielding cluster mean estimates for each patient/cluster combination. Functional

analysis of genes within each of the cancer clusters was performed using DAVID to identify biological terms describing the cancer-associated genes in the classifier.

Lung Cancer Classifier

A lung cancer classifier was developed using lung cancer status as the outcome variable and the cancer gene expression estimates, patient age, and CFGC's for gender (GG), smoking status (GS), pack years (GPY) as predictors. The model was fit using a penalized logistic regression model; the penalization factor (lambda) was 0 for the clinical/gene expression correlates and 10 for each of the gene expression cluster estimates. The resulting score is on a 0 to 1 scale. A score threshold for predicting lung cancer status was established to achieve a sensitivity of approximately 90% for patients with a non-diagnostic bronchoscopy. An evaluation of the benefit of the gene expression classifier to predict lung cancer compared to clinical factors alone was performed by generating a “clinical model” that included age, gender, smoking status, and pack-years (determined clinically) in a logistic regression model to predict lung cancer status. The difference in performance between the complete gene expression classifier and the clinical factors classifier to predict lung cancer status was assessed by comparing the AUC's of each model in the training set.

Analysis of an Independent Test Set

Data from a prior study were used as an independent test set to assess the performance of the locked classifier derived in this study. In that study BECs were collected at bronchoscopy from patients undergoing bronchoscopy for suspicion of lung cancer, and RNA was analyzed on microarrays (Affymetrix HG-U133A). CEL files from that study (n=163) were re-normalized to produce gene-level expression values using Robust Multiarray Average (RMA) [38] in the Bioconductor R package (version 1.28.1). This processing used the Entrez Gene-specific probeset chip definition file (CDF) [39] in place of the standard U133A CDF provided by Affymetrix in order to facilitate cross-platform analyses. Analyses were performed using the R environment for statistical computing (version 2.9.2).

The classifier was applied to patients in the test set with two modifications to account for the difference in microarray platforms. First, the HG-U133A RMA expression values were adjusted by a gene-wise constant which shifted the mean of each gene's expression levels in the test set to the mean observed in the training set. Second, for the classifier genes where a corresponding HG-U133A probeset was not available (LYPD2 and RNF150), the gene's mean expression value in the training set was used for all of the test set samples.

Statistical Methods

Classifier accuracy was assessed using standard measures of prediction accuracy: the area under the curve (AUC), sensitivity, specificity, NPV and PPV. Cross-validation, using a 10% sample hold-out set, was used in the training set to estimate the performance of the

prediction classifiers generated using these approaches [40]. These performance estimates were used to guide the development of the classifier discovery procedure. A final model was set prior to performing a one-time analysis of the test set. Fisher's exact test was used to calculate statistical significance of all categorical variables and a t-test was used for continuous variables.

Results
Study Populations

A set of 299 patients from AEGIS 1 consisting of 223 patients diagnosed with lung cancer and 76 patients diagnosed with benign disease (Table 1) were used to derive our gene expression classifier. Characteristics of the independent test set have been previously described and are summarized here. Although the study design was similar to the one described here, there were some differences in the study populations. The patients were older on average in the training set compared to the test set (p<0.001) (although there was no significant difference in age (p=0.959) for patients diagnosed with lung cancer). The training set also consisted of fewer current smokers (p=0.050); and a lower proportion of patients with <3 cm lesions (p<0.001). In addition, the prevalence of lung cancer was higher in the training set (75% versus 48%; p<0.001).

Table 21—Clinical and Demographic Characteristics of the Patients Used to Train the Classifier

TABLE 21

Clinical and demographic characteristics of

the patients used to train the classifier.

Lung
Benign

Category
Sub-category
cancer
disease
p

N

223
76

Sex
Female
97
26
0.178

Male
126
50

Age (median years)

65
56
<0.0001

Race
Caucasian
168
59
0.757

African-American
47
13

Other
5
3

Unknown
3
1

Smoking Status
Current
101
26
0.107

Former
122
50

Smoking History

43
30
<0.0001

(median PY)

Mass size
<2
cm
46
23

>2 to <3
cm
30
12

≥3
cm
122
19

ill-defined
10
13

infiltrate

Unknown
15
9

Mass Location
Central
86
16

Peripheral
60
30

Central &
60
18

peripheral

Unknown
17
12

Histology
Sub-type

SCLC

40

NSCLC

180

Adenocarcinoma
83

Squamous
73

Large cell
6

Mixed/undefined
18

Unknown

3

Histology
Stage

SCLC
Limited
16

Extensive
18

Unknown
6

NSCLC
1
28

2
16

3
42

4
62

Unknown
32

Benign disease
Sub-category

Alternative Diagnosis

54

Infection
23

Sarcoid
14

Inflammation
7

Fibrosis
4

Other
4

Benign growths
2

Resolution/Stability

22

Derivation of the Classifier and Evaluation of Performance

Gene expression was associated with current smoking status for a large fraction of the genes on the array (6477 genes with p<0.001; top 10 genes reported in Table 22). Three of the top ranked genes (SLC7A11, TKT, and CLND10) were selected to serve as a logistic regression-based smoking status classifier based on cross-validation. This smoking status classifier had an AUC of 0.93 within the training set. An additional CFGC was derived for smoking history, independent of smoking status, and was based on cumulative smoke exposure, measured in pack-years. Smoking history (<10 PY vs >10 PY) was significantly associated with the expression of 531 genes (p<0.001; top 10 genes reported in Table 23). Two of the top genes were selected to serve as a logistic regression-based smoking history classifier (RUNXIT1, AKR1C2) which had an AUC of 0.78 within the training set. Sex was significantly associated with 339 genes (p<0.001; top 10 genes reported in Table 24). The top ranked gene (RPS4Y1) was a perfect classifier (AUC=1) of sex within the training set.

Table 22: Top Differentially Expressed Genes Associated With Smoking Status

TABLE 22

Top differentially expressed genes associated with smoking status

ID
Symbol
logFC
AveExpr
T
P. Value
GS term

8102800
SLC7A11
−2.31513
7.80246
−19.0302
1.44E−53
YES

8088106
TKT
−0.70341
9.137779
−18.3095
7.46E−51
YES

8084630
NA
−1.39188
6.897858
−18.2898
8.86E−51

8136336
AKR1B10
−2.27454
6.839364
−18.2346
1.43E−50

7969640
CLDN10
−0.94118
9.579322
−18.1835
2.23E−50
YES

8171435
PIR
−0.80298
8.924225
−18.0114
9.97E−50

7937465
TALDO1
−0.62614
10.07378
−17.8271
4.95E−49

8051583
CYP1B1
−2.8955
8.179293
−17.7765
7.69E−49

8020653
CABYR
−1.19744
8.009791
−17.6885
1.65E−48

7979658
GPX2
−1.10719
10.62466
−17.524
6.93E−48

Table 23: Top Differentially Expressed Genes Associated With Smoking History

TABLE 23

Top differentially expressed genes associated with smoking history

ID
Symbol
logFC
AveExpr
T
P. Value
GPY term

8151768
RUNX1T1
0.435653
5.905711
8.547091
6.44E−16
Yes

8077989
TPRXL
−0.36913
8.733733
−6.22283
1.63E−09

7994058
SCNN1G
0.685624
8.450023
5.90033
9.75E−09

8069764
NA
−0.32511
7.23309
−5.82162
1.49E−08

8145470
DPYSL2
0.260084
7.62917
5.749689
2.19E−08

7931832
AKR1C2
−0.81612
10.93402
−5.72526
2.50E−08
Yes

8039674
ZNF154
0.372911
7.366637
5.724589
2.51E−08

8150978
CA8
0.415392
6.182963
5.638727
3.95E−08

8129497
EPB41L2
0.4248
6.895569
5.589681
5.10E−08

8039672
NA
0.437196
4.729487
5.515997
7.48E−08

Table 24: Top Differentially Expressed Genes Associated With Gender

TABLE 24

Top differentially expressed genes associated with gender

ID
Symbol
logFC
AveExpr
T
P. Value
GS term

8176375
RPS4Y1
−3.16276
7.851579
−84.6047
6.46E−212
YES

8176624
DDX3Y
−4.05255
7.702569
−84.0882
3.79E−211

8177232
KDM5D
−2.23316
7.443775
−80.4804
1.17E−205

8176578
USP9Y
−3.35997
7.485794
−79.3437
7.01E−204

8177137
UTY
−3.38728
7.655898
−78.9665
2.76E−203

8176698
TXLNG2P
−2.82024
6.904035
−70.7168
1.36E−189

8176709
CYorf15B
−2.63878
7.07696
−68.9368
1.87E−186

8176719
EIF1AY
−3.02926
7.079489
−66.6741
2.34E−182

8176384
ZFY
−1.6795
6.679083
−59.2385
5.47E−168

8176460
PRKY
−1.33643
7.761388
−52.0418
1.48E−152

As described in the methods, we identified genes whose expression is significantly associated with the residuals from the CFGC model for lung cancer. A total of 232 cancer associated genes (Table 25) met the significance criteria (T score>2.7). A pairwise correlation of the 232 genes followed by hierarchical clustering was examined to identify genes with similar expression patterns and partitioned the genes into 11 clusters (FIG. 7). Since genes were correlated within each cluster, we hypothesized that the mean of a small set of genes within each cluster could be used to represent the cluster in a sparse manner. We optimized the classifier, using cross validation to estimate the AUC. We selected genes to represent the gene clusters whose expression was most strongly associated with lung cancer and determined that inclusion of clusters 1, 2, 4, 7, 9 and 10 gave the best AUC. We also determined that beyond 2-4 genes per cluster the performance of the test did not improve. In cross-validation, AUC=0.80 (95% CI 0.75-0.84) for all patients in the training set (n=299); for the subset of patients with non-diagnostic bronchoscopy (n=134) the performance was similar (AUC=0.81; 95% CI 0.74-0.87).

Table 25: Genes Associated With Cancer Which are Included in the Classifier

TABLE 25

Genes associated with cancer which are included in the classifier

Final

ID
Symbol
T
p. value
FC
Cluster
Model

8094228
BST1
−4.29031
2.41E−05
0.89208
1
Yes

8037298
CD177
−3.85704
0.00014
0.715357
1
Yes

8029280
CD177
−3.68455
0.000272
0.840725
1
Yes

7918857
TSPAN2
−3.92967
0.000106
0.845904
2
Yes

7968062
ATP12A
−3.49107
0.000553
0.794623
2
Yes

8124654
GABBR1
2.881256
0.004247
1.071879
4
Yes

8147461
SDC2
2.847433
0.004712
1.089453
4
Yes

7978391
NOVA1
2.729315
0.006721
1.094912
4
Yes

7952205
MCAM
2.70666
0.007186
1.06072
4
Yes

8175531
CDR1
4.307308
2.24E−05
1.468199
7
Yes

8103877
CLDN22
3.502336
0.000531
1.329189
7
Yes

8051001
CGREF1
3.275505
0.001178
1.056672
7
Yes

8149811
NKX3-1
2.92659
0.003689
1.175825
7
Yes

8034974
EPHX3
−3.73504
0.000225
0.898923
9
Yes

8153342
LYPD2
−2.93177
0.00363
0.887468
9
Yes

8102938
RNF150
−4.32839
2.05E−05
0.880745
10
Yes

8028924
MIA
−3.23844
0.001337
0.906368
10
Yes

The final lung cancer classifier was then determined using the finalized classifier discovery procedure on the entire training set. The classifier consisted of a combination of the six cancer gene clusters (represented by 17 genes in total), patient age, and the gene expression correlates (GG, GS, GPY) (Table 26) as predictors. Dichotomous classification was performed using a score threshold of 0.65 (patients with scores >/=0.65 were predicted as cancer-positive and <0.65, cancer-negative). The classifier had a sensitivity of 93% and specificity of 57% in the training set and there was no difference in the AUC of the classifier for the entire training set (0.78; 95% CI, 0.73-0.82), compared with the subset of patients whose bronchoscopy was non-diagnostic for lung cancer (AUC=0.78; 95% 0.71-0.85), (see, FIG. 9).

TABLE 26

Description of the gene expression classifier ^a

Feature ^b,
Coeffi-

(x_i)
cient, (b_i)
Informative Genes within features

Age
0.0623

GG
0.5450
RPS4Y1

GS
0.1661
SLC7A11
CLDN10
TKT

GPY
3.0205
RUNX1T1
AKR1C2

CA (1)
−0.4406
BST1
CD177.1
CD177.2

CA (2)
−0.3402
ATP12A
TSPAN2

CA (4)
0.1725
GABBR1
MCAM
NOVA1
SDC2

CA (7)
0.5670
CDR1
CGREF1
CLDN22
NKX3-1

CA (9)
−0.3160
EPHX3
LYPD2

CA (10)
−0.3791
MIA
RNF150

Inter-
3.3173

cept (b₀)

a) Genomic gender was defined as GG=1 (female) if RPS4Y1<7.5, 0 (male) otherwise. The predicted genomic smoking (GS) value was derived, where x=40.8579-0.4462*SLC7A11-2.1298*CLND10-1.8256*TKT, and genomic smoking GS=e^x/(1+e^x). The predicted genomic pack years (GPY) value was derived, where x=−5.1429+2.1891*RUNXIT1−0.9506*AKRIC2, and genomic pack years GPY=exp(x)/(1+exp(x)). The generalized equation for the prediction classifier was: Score=e^y/(1+e^y), where, y=b₀+Σ(b_i*x_i), where b₀is the intercept, b_iis the coefficient, and x_iis the feature (as shown).

b) Features include patient age (as reported), GG, GS, GPY as described in the methods, and CA (i), the lung cancer gene clusters (shown in FIG. 7).

The gene expression classifier performed significantly better (AUC=0.78; 95% CI, 0.73-0.82) than a model using clinical factors alone (AUC=0.72; 95% CI, 0.67-0.77) in the training set (p<0.001). Functional analysis of the 17 cancer genes is summarized separately (Table 27). Nine of the genes are down-regulated and 8 are up-regulated in association with cancer.

Table 27: Biological Characterization of Classifier Genes

TABLE 27

Biological characterization of classifier genes.

Direction in

Cluster
Cancer
Biomarker genes
Biological themes

1
Down
BST1, CD177.1, CD177.2
Innate immune response

2
Down
ATP12A, TSPAN2
Mitotic cell cycle

4
Up
GABBR1, MCAM, NOVA1, SDC2
Response to retinoic acid, cell cycle

7
Up
CGREF1, CDR1, CLDN22, NKX3-1
Submucosal gland markers

9
Down
EPHX3, LYPD2
Xenobiotic detoxification

10
Down
MIA, RNF150
Cartilaginous markers

Validation in an Independent Test Set

In the patients with non-diagnostic bronchoscopy (n=123) of the independent test set, the AUC of the classifier was 0.81 (95% CI, 0.73-0.88), (FIG. 8) which was similar to the performance in patients with non-diagnostic bronchoscopy in the training set (AUC-0.78; 95% 0.71-0.85; p=0.495). The sensitivity was 92% and with a specificity of 55%, the NPV was 94% (95% CI, 83-99%), (see Table 28). Interestingly we did not observe any effect of cancer histology or stage (Table 29, or lesion size (Table 30) on the classifier's sensitivity for cancer. Moreover, in the test set the classifier had an AUC of 0.79 in current smokers and 0.82 in former smokers, suggesting that smoking status does not have a dramatic effect on classifier performance (p=0.710). When compared with bronchoscopy alone, the combination of the gene expression classifier with bronchoscopy improved the sensitivity from 51% to 95% (p<0.001).

Table 28—Performance of Bronchoscopy, Classifier, and the Combined Procedures in the Test Set

TABLE 28

Performance of bronchoscopy, classifier, and

the combined procedures in the test set.

Classifier &

bronchoscopy

Category ^a
Bronchoscopy
Classifier ^b
combined

N, total
163
123
163

N, Lung cancer
78
38
78

N, Benign disease
85
85
85

Sensitivity (95% CI)
51% (40-62%)
92% (78-98%)
96% (89-99%)

Specificity (95% CI)
100% (95-100%)
53% (42-63%)
53% (42-63%)

NPV (95% CI)
69% (60-77%)
94% (83-98%)
94% (83-98%)

PPV (95% CI)
100% (90-100%)
47% (36-58%)
65% (56-73%)

^aPatients diagnosed with cancer = CA+ and benign disease = CA−

^bThe performance of the classifier was evaluated in patients in which bronchoscopy did not result in a finding of cancer (n = 123).

TABLE 29

Sensitivity of bronchoscopy, the classifier, and the combined

procedures for patients with lung cancer in the test set.

Bronchoscopy
Classifier
Combined

Histology
Sub-type
N
Sensitivity
Sensitivity
Sensitivity

All Cancers

78
51% ^a
92% ^b
96% ^c

SCLC

14
64%
100%
100%

NSCLC

64
48%
91%
95%

Adenocarcinoma
18
33%
83%
89%

Squamous
27
56%
92%
96%

Large Cell
4
25%
100%
100%

Undefined
15
60%
83%
93%

Histology
Stage

SCLC

Limited
9
78%
100%
100%

Extensive
5
40%
100%
100%

NSCLC

1
14
36%
100%
100%

2
2
50%
100%
100%

3
25
52%
92%
96%

4
22
55%
80%
91%

Unknown
1
0%
100%
100%

Of 163 patients who underwent a diagnostic bronchoscopy procedure for suspicion of lung cancer, 78 were diagnosed with cancer. A lung cancer diagnosis was made at bronchoscopy (a) in 40 patients (51%; 95%CI, 40-62%), and in the remaining lung cancer patients where no diagnosis was made at bronchoscopy, the classifier correctly predicted 34 (b) of them (89%; 95%CI, 75-96%). The classifier combined with bronchoscopy yielded a detection of 74 of 78 (95%; 95%CI, 87-98%) patients with lung cancer (c). The sensitivities of bronchoscopy, the classifier, and the combined procedures are also shown for lung cancers according to sub-type and stage.

Table 30—Sensitivity of Bronchoscopy, the Classifier, and the Combined Procedures in the Test Set Stratified by Size of Suspicious Lesions

TABLE 30

Sensitivity of bronchoscopy, the classifier, and the combined procedures

in the test set stratified by size of suspicious lesions.

Bronchoscopy
Classifier
Combined

Mass size ^a
N
Sensitivity
Sensitivity
Sensitivity

<3 cm
99
44%
87%
93%

>3 cm
48
58%
94%
98%

Ill-def Infiltrate
16
38%
100%
100%

Includes patients diagnosed with lung cancer and those with benign disease.

Table 26 recites multiple gene expression classifiers of interest. The following table, Table 31, provides a Gene ID, as available on NCBI, providing descriptive support for the gene expression classifiers. Gene classifier CD177 is depicted in Table 31 with two designations, CD177.1 and CD177.2. The 0.1 and 0.2 designations identify that two different probe sets are used in the arrays which detect differential expression of the genes represented by the gene classifiers. FIG. 10A discloses 19 probes utilized in hybridizing to CD177, accounting for CD177.1. FIG. 10B discloses 4 probes utilized in hybridizing to CD177, accounting for CD177.2.

Table 31—NCBI Gene ID Numbers Corresponding to Gene Expression Classifiers

TABLE 31

NCBI Gene ID Numbers corresponding

to gene expression clssifiers.

Gene

Gene
Gene ID

Classifier
Gene ID Number
Classifier
Number

RPS4Y1
6192
MCAM
4162

SLC7A11
23657
NOVA1
4857

CLDN10
9071
SDC1
6382

TKT
7086
CDR1
1038

AKR1C2
1646
CGREF1
10669

BST1
683
CLDN22
53842

CD177.1
57126 (See FIG. 10A)
NKX3-1
4824

CD177.2
57126 (See FIG. 10B)
EPHX3
79852

ATP12A
479
LYPD2
137797

TSPAN2
10100
MIA
8190

GABBR1
2550
RNF150
57484

RUNX1T1
862

Evaluation of Results

Work has demonstrated that there are persistent gene-expression alterations in normal epithelial cells from the bronchial airway that are associated with exposure to cigarette smoke and the presence of lung cancer in current and former smokers [32,41,42,43]. These cancer-associated differences can be used to derive classifiers capable of accurately detecting lung cancer in these relatively non-invasively collected biospecimens obtained during bronchoscopy [35]. In current practice it is challenging to rule out lung cancer when bronchoscopy does not lead to a finding of malignancy, and the false-negative rate can range from 20-70% [28]. Current guidelines suggest that patients with elevated risk of disease should be pursued with more invasive follow-up diagnostic procedures [28], which carry increased risk of complications [31]. However due to uncertainty these procedures often performed in patients found to have benign disease [33]. Therefore our goal was to derive a gene-expression classifier from the proximal airway during bronchoscopy to increase the overall sensitivity, minimize ambiguity when bronchoscopy is non-diagnostic, and reduce the need for unnecessary invasive procedures.

In this study, we leveraged a cohort of current and former smokers undergoing bronchoscopy for suspected lung cancer from a larger multicenter study to derive a gene-expression classifier for lung cancer. The classifier is a multivariate logistic regression model that has high sensitivity and high NPV. Importantly, we have validated the performance of the classifier in an independent cohort, using data from a previously published study of airway samples collected from smokers undergoing bronchoscopy for suspected lung cancer. The sensitivity is 92% in patients whose bronchoscopy is non-diagnostic in the test set with a specificity of 55%. The NPV is 94% in the test set compared to an NPV of 69% for bronchoscopy alone suggesting that the classifier could help physicians reliably identify patients unlikely to have lung cancer after a non-diagnostic bronchoscopy.

The functions of the differentially expressed genes in the normal appearing airway epithelium in current and former smokers with lung cancer provide insight into the biology underlying the field of injury. Among genes that are suppressed, there are a number involved in the immune response, including CD177 and BST1, suggesting an impaired immune response in the airway of smokers with lung cancer. The gene TSPAN2, whose expression is depressed by p53 knockdown and is associated with poor prognosis in lung adenocarcinomas [44] was also expressed at lower levels in patients with cancer. Also EPHX3, a gene involved in xenobiotic metabolism, processing of carcinogens in tobacco smoke, and carcinogenesis in other epithelial cancers is down-regulated [45]. Among the classifier genes that are up-regulated in lung cancer, NOVA1 and CDR1 are predominantly expressed in neurons, but are also expressed in tumors and are associated with para-neoplastic antibodies in several malignancies, including small-cell lung cancer [46,47,48,49,50]. Furthermore, MCAM which is up-regulated in lung cancer, is expressed in basal bronchial epithelial cells [51]. MCAM is also strongly and transiently up-regulated in tracheal epithelium during repair [52], is required for tracheal epithelial regeneration [53], and is up-regulated in the bronchial epithelium of patients with COPD [54] and asthma [55]. A number of classifier genes that regulate cell growth and proliferation are up-regulated in patients with lung cancer, including SDC2, and NKX3-1 as well as the cell-cycle-arrest mediator CGREF1. Finally the CFGC genes selected to predict smoking status (SLC7A11, CLDN10, TKT) and smoking history (RUNXIT1, AKR1C2) in our classifier have been previously reported as being altered by tobacco smoke exposure, confirming the robust effect of smoking on airway epithelium biology [11,18,33].

Our discovery approach extends earlier work on gene-expression based lung cancer diagnostics primarily in the explicit modeling of clinical covariates as components of the predictive model prior to selection of features with lung cancer-associated expression. It is known that the response to environmental insults and other clinical factors can vary substantially between individuals. Therefore our approach was to use gene expression to capture the patient-level physiological response to an environmental insult (e.g., cumulative smoke exposure), as this response may be more reflective of disease risk than the actual reported values [57]. Another component of our approach was selecting genes whose

expression is associated with cancer after accounting for the modeled clinical factors. We hypothesized that this approach would help ensure that the information about the likelihood of cancer captured by the genes with cancer-associated gene expression is independent from the information about cancer captured by the modeled clinical factors. An additional important aspect of our classifier discovery approach was our methodology to identify patterns of independent cancer-associated gene expression through clustering and then to model cancer as the additive effects of each of the cancer-associated gene expression modules. This is in contrast to selecting only genes that are globally top-ranked according to their association with cancer which could potentially result in selecting an entire panel of genes that reflect a single cancer-associated molecular process. Previous studies to derive a gene expression classifier to predict risk of lung cancer in normal appearing airway epithelial cells have described similar results with high sensitivity and NPV when bronchoscopy is non-diagnostic [35]. While there are no common genes in that classifier compared to the one described here, we believe that our new classifier represents similar mechanisms of action given the strong performance in the independent test set. However, the differences in the specific genes selected this may be due to differences in the feature selection process.

We have derived a gene expression classifier for lung cancer using cells from the proximal airway that can be used in conjunction with bronchoscopy for suspected lung cancer. We have validated the performance of this classifier in an independent test set. The classifier adds substantial sensitivity to the bronchoscopy procedure resulting in high NPV. This classifier can be used to aid in decision-making when bronchoscopy is non-diagnostic by identifying patients who are at low risk of having lung cancer.

Having thus described several aspects of at least one embodiment of this disclosure, it is to be appreciated that various alterations, modifications, and improvements will readily occur to those skilled in the art. Such alterations, modifications, and improvements are intended to be part of this disclosure, and are intended to be within the spirit and scope of the disclosure. Accordingly, the foregoing description and drawings are by way of example only and the disclosure is described in detail by the claims that follow.

Use of ordinal terms such as “first,” “second,” “third,” etc., in the claims to modify a claim element does not by itself connote any priority, precedence, or order of one claim element over another or the temporal order in which acts of a method are performed, but are used merely as labels to distinguish one claim element having a certain name from another element having a same name (but for use of the ordinal term) to distinguish the claim elements.