Systems and methods for multi-label cancer classification

Abstract

Systems and methods are provided for identifying a diagnosis of a cancer condition for a somatic tumor specimen of a subject. The method receives sequencing information comprising analysis of a plurality of nucleic acids derived from the somatic tumor specimen. The method identifies a plurality of features from the sequencing information, including two or more of RNA, DNA, RNA splicing, viral, and copy number features. The method provides a first subset of features and a second subset of features from the identified plurality of features as inputs to a first classifier and a second classifier, respectively. The method generates, from two or more classifiers, two or more predictions of cancer condition based at least in part on the identified plurality of features. The method combines, at a final classifier, the two or more predictions to identify the diagnosis of the cancer condition for the somatic tumor specimen of the subject.

1 - 104 . (canceled) 
     
     
         105 . A method for identifying a diagnosis of a cancer condition for a subject from among a plurality of cancer conditions, the method comprising:
 sequencing a plurality of DNA molecules from a sample of a tumor from the subject, thereby obtaining a first plurality of sequence reads of DNA from a tumor of the subject;   aligning each respective sequence read in the first plurality of sequence reads to a reference construct for a human genome, thereby generating a corresponding first plurality of aligned sequence reads;   sequencing a plurality of mRNA molecules from the sample of the tumor from the subject, thereby obtaining a second plurality of sequence reads of RNA from the tumor of the subject;   aligning each respective sequence read in the second plurality of sequence reads to a reference construct for a human transcriptome, thereby generating a corresponding second plurality of aligned sequence reads;   identifying a plurality of features from the first plurality of aligned sequence reads and second plurality of aligned sequence reads, collectively, wherein the plurality of features comprises a first subset of features comprising RNA expression features and a second subset of features comprising DNA features, wherein:
 each RNA expression feature is associated with an expression level of a respective target region of the reference human transcriptome and represents a corresponding abundance of sequence reads, in the second plurality of aligned sequence reads, that map to the respective target region of the reference construct for the human transcriptome, and 
 each DNA feature is associated with a respective allele status in a respective target region of the reference human genome and represents a corresponding abundance of sequence reads with a corresponding reference or variant allele, in the first plurality of aligned sequence reads, that map to the respective target region of the reference construct for the human genome; and 
   evaluating the plurality of features using an ensemble classifier comprising (i) a set of intermediate classifiers that includes a first classifier and a second classifier, and (ii) a final classifier, wherein the ensemble classifier uses the plurality of features to form:
 (a) for each respective classifier in the set of intermediate classifiers, a corresponding intermediate prediction by:
 obtaining a first intermediate prediction from among a first plurality of predictions for the cancer condition associated with the first classifier, by providing the first subset of features from the identified plurality of features as inputs to the first classifier, wherein the first classifier evaluates the first subset of features against each cancer condition in the plurality of cancer conditions to provide the first intermediate prediction; and 
 obtaining a second intermediate prediction from among a second plurality of predictions for the cancer condition associated with the second classifier, by providing the second subset of features from the identified plurality of features as inputs to the second classifier, wherein the second classifier evaluates the second subset of features against each cancer condition in the plurality of cancer conditions to provide the second intermediate prediction, thereby forming a plurality of intermediate predictions; and 
 
 (b) a determination of the cancer condition of the subject by combining, at the final classifier, the plurality of intermediate predictions that includes the first and second intermediate predictions to identify the cancer condition for the subject from among the plurality of cancer conditions, 
   wherein the determination of the cancer condition of the subject formed by the ensemble classifier comprises differentiating between general sarcomas, ependymoma, ewing sarcoma, gliosarcoma, leiomyosarcoma, meningioma, mesothelioma, and Rosai-Dorfman.   
     
     
         106 . The method of  claim 105 , wherein combining, at the final classifier, the plurality of intermediate predictions further comprises:
 scaling each intermediate prediction of the plurality of intermediate predictions based at least in part on a respective confidence level in each respective prediction to form a corresponding scaled prediction in a corresponding plurality of scaled predictions; and   generating a combined prediction based at least in part on each scaled prediction by inputting each respective scaled prediction in the corresponding plurality of scaled predictions into the final classifier.   
     
     
         107 . The method of  claim 105 , wherein:
 the plurality of features further comprises a third subset of features comprising RNA splicing features, wherein each RNA splicing feature is associated with a respective splicing event at a respective target region of the first reference genome and represents a corresponding abundance of sequence reads, encompassed by the sequencing information, that map to the respective target region;   the set of intermediate classifiers further comprises a third classifier;   the plurality of intermediate predictions further comprises a third intermediate prediction; and   the ensemble classifier further uses the plurality of features to form the third intermediate prediction from among a third plurality of predictions for the cancer condition associated with the third classifier, by providing the third subset of features from the identified plurality of features as inputs to the third classifier, wherein the third classifier evaluates the third subset of features against each cancer condition in the plurality of cancer conditions to provide the third intermediate prediction.   
     
     
         108 . The method of  claim 105 , wherein:
 the plurality of features further comprises a fourth subset of features comprising viral features, wherein each viral feature is associated with a respective target region of a viral reference genome and represents a corresponding abundance of sequence reads, encompassed by the sequencing information, that map to the respective target region in the viral reference genome;   the set of intermediate classifiers further comprises a fourth classifier;   the plurality of intermediate predictions further comprises a fourth intermediate prediction; and   the ensemble classifier further uses the plurality of features to form the fourth intermediate prediction from among a fourth plurality of predictions for the cancer condition associated with the fourth classifier, by providing the fourth subset of features from the identified plurality of features as inputs to the fourth classifier, wherein the fourth classifier evaluates the fourth subset of features against each cancer condition in the plurality of cancer conditions to provide the fourth intermediate prediction.   
     
     
         109 . The method of  claim 107 , wherein:
 the plurality of features further comprises a fourth subset of features comprising viral features, wherein each viral feature is associated with a respective target region of a viral reference genome and represents a corresponding abundance of sequence reads, encompassed by the sequencing information, that map to the respective target region in the viral reference genome;   the set of intermediate classifiers further comprises a fourth classifier;   the plurality of intermediate predictions further comprises a fourth intermediate prediction; and   the ensemble classifier further uses the plurality of features to form the fourth intermediate prediction from among a fourth plurality of predictions for the cancer condition associated with the fourth classifier, by providing the fourth subset of features from the identified plurality of features as inputs to the fourth classifier, wherein the fourth classifier evaluates the fourth subset of features against each cancer condition in the plurality of cancer conditions to provide the fourth intermediate prediction.   
     
     
         110 . The method of  claim 107 , wherein:
 the plurality of features further comprises a fifth subset of features comprising copy number features, wherein each copy number feature is associated with a target region of the reference human genome and represents a corresponding abundance of sequence reads, in the first plurality of aligned sequence reads, that map to the respective target region of the reference human genome;   the set of intermediate classifiers further comprises a fifth classifier;   the plurality of intermediate predictions further comprises a fifth intermediate prediction; and   the ensemble classifier further uses the plurality of features to form the fifth intermediate prediction from among a fifth plurality of predictions for the cancer condition associated with the fifth classifier, by providing the fifth subset of features from the identified plurality of features as inputs to the fifth classifier, wherein the fifth classifier evaluates the fifth subset of features against each cancer condition in the plurality of cancer conditions to provide the fifth intermediate prediction.   
     
     
         111 . The method of  claim 108 , wherein:
 the plurality of features further comprises a fifth subset of features comprising copy number features, wherein each copy number feature is associated with a target region of the reference human genome and represents a corresponding abundance of sequence reads, in the first plurality of aligned sequence reads, that map to the respective target region of the reference human genome;   the set of intermediate classifiers further comprises a fifth classifier;   the plurality of intermediate predictions further comprises a fifth intermediate prediction; and   the ensemble classifier further uses the plurality of features to form the fifth intermediate prediction from among a fifth plurality of predictions for the cancer condition associated with the fifth classifier, by providing the fifth subset of features from the identified plurality of features as inputs to the fifth classifier, wherein the fifth classifier evaluates the fifth subset of features against each cancer condition in the plurality of cancer conditions to provide the fifth intermediate prediction.   
     
     
         112 . The method of  claim 109 , wherein:
 the plurality of features further comprises a fifth subset of features comprising copy number features, wherein each copy number feature is associated with a target region of the reference human genome and represents a corresponding abundance of sequence reads, in the first plurality of aligned sequence reads, that map to the respective target region of the reference human genome;   the set of intermediate classifiers further comprises a fifth classifier;   the plurality of intermediate predictions further comprises a fifth intermediate prediction; and   the ensemble classifier further uses the plurality of features to form the fifth intermediate prediction from among a fifth plurality of predictions for the cancer condition associated with the fifth classifier, by providing the fifth subset of features from the identified plurality of features as inputs to the fifth classifier, wherein the fifth classifier evaluates the fifth subset of features against each cancer condition in the plurality of cancer conditions to provide the fifth intermediate prediction.   
     
     
         113 . The method of  claim 105 , wherein:
 the target regions of the reference human transcriptome associated with each RNA expression feature collectively represent a plurality of genes, and   the plurality of genes comprises ten or more genes selected from the group consisting of GPM6A, CDX1, SOX2, NAPSA, CDX2, MUC12, SLAMF7, HNF4A, ANXA10, TRPS1, GATA3, SLC34A2, NKX2-1, SLC22A31, ATP10B, STEAP2, CLDN3, SPATA6, NRCAM, USH1C, SOX17, TMPRSS2, MECOM, WT1, CDHR1, HOXA13, SOX10, SALL1, CPE, NPR1, CLRN3, THSD4, ARL14, SFTPB, COL17A1, KLHL14, EPS8L3, NXPE4, FOXA2, SYT11, SPDEF, GRHL2, GBP6, PAX8, ANO1, KRT7, HOXA9, TYR, DCT, LYPD1, MSLN, TP63, CDH1, ESR1, HNF1B, HOXA10, TJP3, NRG3, TMC5, PRLR, GATA2, DCDC2, INS, NDUFA4L2, TBX5, ABCC3, FOLH1, HIST1H3G, S100A1, PTHLH, ACER2, RBBP8NL, TACSTD2, C19orf77, PTPRZ1, BHLHE41, FAM155A, MYCN, DDX3Y, FMN1, HIST1H3F, UPK3B, TRIM29, TXNDC5, BCAM, FAM83A, TCF21, MIA, RNF220, AFAP1, KRT5, SOX21, KANK2, GPM6B, Clorfl 16, FOXF1, MEIS1, EFHD1, and XKRX.   
     
     
         114 . The method of  claim 105 , wherein the first plurality of sequence reads was generated by low pass, whole genome sequencing. 
     
     
         115 . The method of  claim 105 , wherein the second plurality of sequence reads was generated from sequencing of cDNA. 
     
     
         116 . The method of  claim 105 , wherein the ensemble classifier is trained by a method comprising:
 obtaining, for each respective training subject in a plurality of training subjects, (i) the plurality of features, (ii) a respective training label for each respective classifier in the set of intermediate classifiers, and (iii) a respective label for the cancer condition of the respective training subject;   training, for each respective classifier in the set of intermediate classifiers, a respective initial model for the respective classifier that provides a respective initial intermediate prediction for each respective training subject based on at least, for each respective training subject in the plurality of training subjects, (i) a respective subset of features in the three or more subsets of features, and (ii) the respective training label for the respective classifier;   training a respective initial model for the final classifier that provides a corresponding initial diagnosis for the cancer condition based on at least, for each respective training subject in the plurality of training subjects, (i) for each respective classifier in the set of intermediate classifiers, a respective initial classification output from the respective initial model for the respective classifier for the respective training subject, and (ii) the respective label for the cancer condition of the respective training subject;   calculating, for each respective training subject in the plurality of training subjects, a respective entropy score for the respective training subject based at least in part on a respective initial diagnosis output from the initial model for the final classifier;   identifying an entropy threshold based at least in part on the accuracy of the initial model for the final classifier across the plurality of training subjects; and   re-training the ensemble classifier based on respective training subjects in the plurality of training subjects whose respective entropy score satisfies the entropy threshold.   
     
     
         117 . The method of  claim 116 , wherein identifying the entropy threshold comprises identifying a percentile of the accuracy of the initial model for the final classifier across the plurality of training subjects. 
     
     
         118 . The method of  claim 105 , wherein the determination of the cancer condition of the subject formed by the ensemble classifier further comprises differentiating between lung adenocarcinoma, lung squamous, oral adenocarcinoma, and oral adenocarcinoma. 
     
     
         119 . The method of  claim 105 , wherein the determination of the cancer condition of the subject formed by the ensemble classifier further comprises differentiating between a liver metastasis of pancreatic origin, a liver metastasis of upper gastrointestinal origin, and a liver metastasis of cholangio origin. 
     
     
         120 . The method of  claim 105 , wherein the determination of the cancer condition of the subject formed by the ensemble classifier further comprises differentiating between a brain metastasis of glioblastoma origin, a brain metastasis of oligodendroglioma origin, a brain metastasis of astrocytoma origin, and a brain metastasis of medulloblastoma origin. 
     
     
         121 . The method of  claim 105 , wherein the determination of the cancer condition of the subject formed by the ensemble classifier further comprises differentiating between non-small cell lung cancer squamous and adenocarcinoma. 
     
     
         122 . The method of  claim 105 , wherein the determination of the cancer condition of the subject formed by the ensemble classifier further comprises differentiating between one or more sarcomas with carcinoma morphological features or protein expressions, and one or more carcinomas with sarcoma morphologic features or protein expressions. 
     
     
         123 . The method of  claim 105 , wherein the determination of the cancer condition of the subject formed by the ensemble classifier further comprises differentiating between one or more neuroendocrines, one or more carcinomas, and one or more sarcomas. 
     
     
         124 . The method of  claim 105 , further comprising:
 receiving subject information comprising one or more clinical events; and   differentiating the cancer condition between a new tumor and a recurrence of a previous tumor based at least in part on the one or more clinical events.