Machine learning for multi-cancer metastasis risk

Abstract

Comprehensive molecular profiling provides a wealth of data concerning the molecular status of patient samples. Such data can be used to train machine learning models using disease outcomes This approach has been applied to train machine learning models that predict metastatic potential.

Description

MACHINE LEARNING FOR MULTI-CANCER METASTASIS RISKCROSS REFERENCE

[0001] The present application is a PCT application and claims priority from U.S. Provisional Patent Application Serial No. 63 / 468,982, entitled “Multi-Cancer Brain Metastasis Risk Score” filed on May 25, 2023, the entire contents of which are herein incorporated by reference for all purposes.BACKGROUND

[0002] Metastasis refers to the “spread of cancer cells from the place where they first formed to another part of the body.” See cancer.gov / publications / dictionaries / cancer- terms / def / metastasis. Metastatic cancer is also commonly referred to as stage IV cancer. The organ location where a tumor originally forms is the primary tumor origin or site, and may define a tumor’s lineage. For example, a primary breast tumor forms in the breast, a primary colon cancer forms in the colon, etc. Cancer cells continuously break off of a primary tumor and may spread throughout the body via the blood or lymph systems. These cells may then form a tumor in a new location, thereby creating a metastatic tumor or lesion. The site of the metastatic tumors may vary based on the primary origin. For example, in addition to lymph nodes, breast tumors most often metastasize to the bones, lung, liver and / or brain, colorectal cancer most often metastasizes to the lungs and / or liver, lung cancer tends to spread to the brain, bones, liver, and / or adrenal glands, and prostate cancer tends to spread to the bones. The metastatic tumors are of the same type, or lineage, as the primary tumor, not the location of spread. For example, a breast tumor that has metastasized to the brain is a breast cancer, not a brain cancer.

[0003] Treatment of a metastatic cancer depends on a number of factors, including the primary location, the extent and location of the spreading, past treatments, and patient characteristics such as age and general health. Treatment for metastatic cancer may differ from that of the primary origin. Treatments can include therapeutic agents such as chemotherapy, immunotherapy, hormone therapy, targeted therapies, or various combinations thereof. Treatments may also include localized therapy such as surgery and / or radiation.

[0004] Most cancer deaths result from metastasis. Treating stage IV cancers is complicated by many factors. For example, the spread may occur to several locations. In addition, the metastatic lesions may not respond as well to the same treatments as the primary tumor. Without being bound by theory, differential treatment efficacy may result from differing microenvironments between primary and metastatic sites, genetic alterations, or other unknown factors. Cancer patients remain at risk of developing metastatic lesions for years after initial treatment.

[0005] Predicting whether a cancer will metastasize can guide personalized treatment plans tailored to a specific patient to prevent metastasis and to help avoid under- or over-treatment. Factors underlying the prediction may also prove to be therapeutic targets and / or suggest interventions. A physician may propose an altered, e.g., more aggressive, treatment regimen in a patient whose cancer is more likely to metastasize. Taken together, there is a need to better identify those cancers more likely to metastasize to achieve better patient outcomes and to avoid unnecessary adverse events and high costs.

[0006] Machine learning (ML) models can be configured to analyze labeled training data and then draw inferences from the training data. Once the machine learning model has been trained, sets of data that are not labeled may be provided to the machine learning model as an input. These unlabeled sets of data are often referred to as “test” sets. The machine learning model may process the input data, e.g., molecular profiling data, and make predictions about the input based on inferences learned during training.

[0007] Comprehensive molecular profiling provides a wealth of data concerning the molecular status of patient samples. We have performed such profiling on well many thousands of tumor patients from practically all cancer lineages and have tracked patient outcomes and responses to treatments in thousands of these patients. For example, our molecular profiling data can be compared to patient benefit or decreased / lack of benefit to treatments and processed using machine learning algorithms to train ML models that predict to the effectiveness of various treatments. Here, we provide ML models that predict risk of metastasis.BRIEF SUMMARY

[0008] Provided herein are systems, methods and medium that predict risk of metastasis in a cancer patient. Comprehensive molecular profiling provides a wealth of data concerning themolecular status of patient samples. Such data can be compared to train machine learning models using disease outcomes. Described herein are methods for training and employing machine learning models to predict probability of metastasis based on the disease outcomes. In some embodiments, the trained machine-learning models can employ dimensionality reduction techniques within the machine learning platform to predict the risk of metastasis.

[0009] In an aspect, provided herein is a system for predicting risk that a cancer in a subject will metastasize, the system comprising: one or more computers and one or more memory devices storing instructions that, when executed by the one or more computers, cause the one or more computers to perform operations, the operations comprising: obtaining, by the one or more computers, molecular data obtained by sequencing a panel of mRNA transcripts in a biological sample from the subject, wherein the panel of mRNA transcripts comprises transcripts for at least 10, 20, 50, 100, 200, 500, 1000, 2000, 5000, 10000, 20000, 21000, 22000 or 23000 genes; generating, by the one or more computers, input data that includes a set of features extracted from the obtained molecular data; providing, by the one or more computers, the generated input data as input to a predictive model, the predictive model comprising at least one machine learning model, wherein each particular machine learning model of the at least one machine learning models was trained to generate output data that indicates risk that a cancer in a subject will metastasize based on the particular machine learning model processing of the set of features extracted from the molecular data of a cohort of cancer patients comprising at least 1000, 2000, 5000, 10000, 25000, 50000, 100000, 150000 or 200000 patients; processing, by the one or more computers, the generated input data through the at least one machine learning model to generate output data indicating risk that the cancer in the subject will metastasize; and determining, by the one or more computers and based on the generated output data, risk that the cancer in the subject will metastasize.

[0010] These and other embodiments of the disclosure are described in detail below. For example, other embodiments are directed to systems, devices, and computer readable media associated with methods described herein.

[0011] A better understanding of the nature and advantages of embodiments of the present disclosure may be gained with reference to the following detailed description and the accompanying drawings.BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The novel features of the invention are set forth with particularity in the appended claims. A better understanding of the features and advantages of the present invention will be obtained by reference to the following detailed description that sets forth illustrative embodiments, in which the principles of the invention are utilized, and the accompanying drawings of which:

[0013] FIG. 1A is a block diagram of an example system for training a machine learning model.

[0014] FIG. IB is a block diagram of a system that generates training data structures for training a machine learning model to predict metastatic potential of a tumor in a subject having a particular set of biomarkers.

[0015] FIG. 1C is a block diagram of a system for using a machine learning model that has been trained to predict metastatic potential of a tumor in a subject having a particular set of biomarkers.

[0016] FIG. ID is a block diagram of a system for predicting metastatic potential of a tumor in a subject having a particular set of biomarkers by using a voting unit to interpret output generated by multiple machine learning models.

[0017] FIGs. 2A-E illustrate artificial intelligence (Al) feature generation and model training. FIG. 2A is an overview wherein embedding of transcriptomic data (e.g., whole transcriptome sequencing; WTS) is obtained from patient samples using a combination of Transcripts Per Million (TPM), Trimmed Mean of the M-values (TMM), and Gene Set Enrichment Analysis (GSEA) and is delivered to three layers of Al feature generation algorithms to reduce dimensionality and identify underlying clusters where the coordinates can serve as inputs into secondary AI / ML models. Features from the primary Al layers are used as inputs into a multilayer deliberation analytic Al platform which learns to identify patients that are more or less likely to develop a brain metastasis. There were 7 populations identified based on risk probabilities. FIGs. 2B-E illustrate that more data, e.g., derived from a larger patient cohort, results in better refined cluster coordinates for GSEA. FIG. 2B illustrates cluster coordinates for 1000 patients, FIG. 2C illustrates cluster coordinates for 50,000 patients, FIG. 2D illustratescluster coordinates for 100,000 patients, and FIG. 2E illustrates cluster coordinates for 200,000 patients. Enhancement in detail is readily observed as the cohort size is increased. Similar results to FIGs. 2B-E were observed for cluster refinement with TPM and TMM (data not shown).

[0018] FIGs. 3A-C illustrate predictions of the risk of brain metastasis on an independent set of patients. The plots show the relative likelihood of developing a brain met with all cancers (FIG. 3A), breast cancer (FIG. 3B), and non-small cell lung cancer (NSCLC; FIG. 3C). All samples are from the primary tumor site. In FIG. 3A, the median time to brain metastasis development for groups 1-5 was unevaluable while groups 6-7 were 424 and 244 days, respectively.

[0019] FIG. 4 is a flowchart illustrating a method for predicting metastatic potential of a tumor in a subject having a particular set of biomarkers according to embodiments of the present disclosure.

[0020] FIG. 5 illustrates a measurement system 1200 according to an embodiment of the present disclosure.

[0021] FIG. 6 shows a block diagram of an example computer system usable with systems and methods according to embodiments of the present disclosure.TERMS

[0022] A sample as used herein includes any relevant biological sample that can be used for molecular profiling, e.g., sections of tissues such as biopsy or tissue removed during surgical or other procedures, bodily fluids, autopsy samples, and frozen sections taken for histological purposes. Any biopsy technique known in the art can be applied. The biopsy technique applied can depend on the tissue type to be evaluated (e.g., colon, prostate, kidney, bladder, lymph node, liver, bone marrow, blood cell, lung, breast, etc.), the size and type of the tumor (e.g., solid or suspended, blood or ascites), among other factors. Representative biopsy techniques include, but are not limited to, excisional biopsy, incisional biopsy, needle biopsy (e.g., core-needle or fine- needle), surgical biopsy, and bone marrow biopsy. Such samples can also include blood and blood fractions or products (e.g., serum, buffy coat, plasma, platelets, red blood cells, and the like), sputum, malignant effusion, cheek cells tissue, cultured cells (e.g., primary cultures, explants, and transformed cells), stool, urine, other biological or bodily fluids (e.g., prostaticfluid, gastric fluid, intestinal fluid, renal fluid, lung fluid, cerebrospinal fluid, and the like), etc. The sample can comprise biological material that is a fresh frozen & formalin fixed paraffin embedded (FFPE) block, formalin-fixed paraffin embedded, or is within an RNA preservative + formalin fixative. The sample can comprise a fixed tumor sample. Unless otherwise noted, a “sample” as referred to herein for molecular profiling of a patient may comprise more than one physical specimen. As still another non-limiting example, a molecular profile may be generated for a subject using a “sample” comprising a solid tumor specimen and a bodily fluid specimen. In some embodiments, a sample is a unitary sample, i.e., a single physical specimen. More than one sample of more than one type can be used for each patient.

[0023] A molecular profile as used herein can refer to molecular characteristics of a biological sample, including without limitation genetic, epigenetic, proteomic, and / or expression characteristics of molecules in the biological sample. The biological sample may be a biopsy of a tumor. A summary and / or analysis of a patient / disease state based on a molecular profile may be included in a molecular profiling report, as further described herein.

[0024] A “cell-free sample” as used herein can mean any sample where a majority of the nucleic acids are cell-free prior to any artificial fragmentation processes, e.g., greater than 50%, 60%, 70%, 80%, 90%, 95%, or 99% of the nucleic acids can be cell-free. A cell-free sample can be a cell free portion of a bodily fluid such as peripheral blood, sera / serum, plasma, ascites, urine, cerebrospinal fluid (CSF), sputum, saliva, bone marrow, synovial fluid, aqueous humor, amniotic fluid, cerumen, breast milk, broncheoalveolar lavage fluid, semen, prostatic fluid, Cowper’s fluid, pre-ejaculatory fluid, female ejaculate, sweat, fecal matter, tears, cyst fluid, pleural fluid, peritoneal fluid, pericardial fluid, lymph, chyme, chyle, bile, interstitial fluid, menses, pus, sebum, vomit, vaginal secretions, mucosal secretion, stool water, pancreatic juice, lavage fluids from sinus cavities, bronchopulmonary aspirates, or blastocyst cavity fluid. As an example, blood can be fractionated via centrifugation into an upper layer of cell-free clear fluid known as plasma, a lower layer of red fluid comprising granulocytes and red blood cells, and a thin layer in between known as the buffy coat, which comprises mainly white blood cells and platelets.

[0025] A “cellular sample” as used herein can mean any sample where a majority of the nucleic acids or other biological entities of interest are within an intact cell, e g., greater than50%, 60%, 70%, 80%, 90%, 95%, or 99% of the DNA can be cellular. Examples of such samples include the huffy coat of a blood sample, any bodily fluid sample subjected to an enrichment process (e.g., centrifuging or filtration) to enrich for cells, a tissue sample, tumor biopsy, a hair follicle, and the like.

[0026] A biological sample (e.g., blood) can include both a cell-free portion and a cellular portion, which can be separated from each other. After a separation process (e.g., centrifuging), the separated portions can be considered a cell-free sample and a cellular sample, e.g., plasma and buffy coat, respectively, both derived from a same biological sample. In other embodiments, the cell-free sample and the cellular sample can be derived from different biological samples. Regardless of the sample type, in various embodiments, at least 1,000, 5,000, 10,000 or 50,000 or 100,000 or 500,000 or 1,000,000 or 5,000,000 nucleic acid molecules, or more, can be analyzed. At least a same number of sequence reads can be analyzed. Additionally, amounts of sequence reads determined for embodiments of the present disclosure can be at least 1,000, 5,000, 10,000, 50,000, 100,000, 500,000, 1,000,000, or 5,000,000. Example sizes of a sample can include 30, 50, 100, 200, 300, 500, 1,000, 5,000, or 10,000 or more nanograms, or 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10 ml.

[0027] The term “phenotype” as used herein can mean any trait or characteristic that can be identified in part or in whole by using the systems and / or methods provided herein. In some embodiments, the systems can include one or more computer programs on one or more computers in one or more locations, e.g., configured for use in a method described herein. Phenotypes may be determined by analyzing a biological sample obtained from a subject. Phenotypes to be characterized can be any phenotype of interest, including without limitation a tissue, tissue origin, tumor type, anatomical origin of a sample (or sample origin), medical condition, ailment, disease, disorder, or useful combinations thereof. A phenotype can be any observable characteristic or trait of, such as a disease or disorder, a stage of a disease or disorder, susceptibility to a disease or disorder, prognosis of a disease stage or disorder, a physiological state, or response / potential response (or lack thereof) to interventions such as therapeutics, e.g., a stage or grade of a tumor, or a tumor type / origin (e.g., tissue origin). A phenotype can result from a subject’s genetic makeup as well as the influence of environmental factors and the interactions between the two, as well as from epigenetic modifications to nucleic acid sequences.In various embodiments, a phenotype in a subject is characterized by obtaining a biological sample from a subject and analyzing the sample using the systems and / or methods provided herein. For example, characterizing a phenotype for a subject or individual can include detecting a disease or disorder (including pre-symptomatic early-stage detection), determining a prognosis, diagnosis, or theranosis of a disease or disorder, or determining the stage or progression of a disease or disorder. Characterizing a phenotype can include identifying appropriate treatments or treatment efficacy for specific diseases, conditions, disease stages and condition stages, predictions and likelihood analysis of disease progression, particularly disease recurrence, metastatic spread or disease relapse. A phenotype can also be a clinically distinct type or subtype of a condition or disease, such as a cancer or tumor. Phenotype determination can also be a determination of a physiological condition, or an assessment of organ distress or organ rejection, such as post-transplantation. The compositions and methods described herein allow assessment of a subject on an individual basis, which can provide benefits of more efficient and economical decisions in treatment. In some case, a “+” symbol (or the word “positive”) is used to signify that a sample a particular phenotype is present, whereas a symbol (of the word “negative”) is used to indicate that a particular phenotype is absent. As a non-limiting example, HER2+ may indicate overexpression or amplification of the HER2 gene while HER- may indicated no expression or deletion of the HERZ gene.

[0028] A medical condition as used herein can refer to a disease or disorder of a subject, as well as a stage or severity of the medical condition.

[0029] A subject (individual, patient, or the like) can be any animal which may benefit from the methods described herein. A subject or individual can be any animal which may benefit from the methods described herein, including, e.g., humans and non-human mammals, such as primates, rodents, horses, dogs and cats. Subjects include without limitation a eukaryotic organisms, most preferably a mammal such as a primate, e.g., chimpanzee or human, cow; dog; cat; a rodent, e.g., guinea pig, rat, mouse; rabbit; or a bird; reptile; or fish. In preferred embodiments, a subject is human. Subjects specifically intended for treatment using the methods described herein include humans. A subject may also be referred to herein as an individual or a patient. The subject can have a pre-existing disease or disorder, including without limitationcancer. Alternatively, the subject may not have any known pre-existing condition. The subject may also be non-responsive to an existing or past treatment, such as a treatment for cancer.

[0030] Theranostics as used herein can include therapy-related diagnostic testing that provides the ability to affect therapy or treatment of a medical condition such as a disease or disease state. Theranostics testing provides a theranosis in a similar manner that diagnostics or prognostic testing provides a diagnosis or prognosis, respectively. As used herein, theranostics encompasses any desired form of therapy related testing, including predictive medicine, personalized medicine, precision medicine, integrated medicine, pharmacodiagnostics and Dx / Rx partnering. Therapy related tests can be used to predict and assess drug response in individual subjects, thereby providing personalized medical recommendations. Predicting a likelihood of response can be determining whether a subject is a likely responder or a likely non-responder to a candidate therapeutic agent, e.g., before the subject has been exposed or otherwise treated with the treatment. Assessing a therapeutic response can be monitoring a response to a treatment, e.g., monitoring the subject’s improvement or lack thereof over a time course after initiating the treatment. Therapy related tests are useful to select a subject for treatment who is particularly likely to benefit or lack benefit from the treatment or to provide an early and objective indication of treatment efficacy in an individual subject. Characterization using the systems and methods provided herein may indicate that treatment should be altered to select a more promising treatment, thereby avoiding the expense of delaying beneficial treatment and avoiding the financial and morbidity costs of less efficacious or ineffective treatment(s).

[0031] Theranosis can comprise predicting a treatment efficacy or lack thereof, classifying a patient as a responder or non-responder to treatment. A predicted “responder” can refer to a patient likely to receive a benefit from a treatment whereas a predicted “non-responder” can be a patient unlikely to receive a benefit from the treatment. Unless specified otherwise, a benefit can be any clinical benefit of interest, including without limitation cure in whole or in part, remission, or any improvement, reduction or decline in progression of the condition or symptoms. The theranosis can be directed to any appropriate treatment, e.g., the treatment may comprise at least one of chemotherapy, immunotherapy, targeted cancer therapy, a monoclonal antibody, small molecule, surgery, radiation, or any useful combinations thereof.

[0032] A classification of a medical condition can include a diagnosis, prognosis, or theranosis of the subject, including a tumor type, e.g., tumor origin. The term “classification” as used herein refers to any number(s) or other characters(s) that are associated with a particular property of a sample, e.g., a medicinal condition of a subject from whom the sample was obtained. For example, a “+” symbol (or the word “positive”) could signify that a sample is classified as having deletions or amplifications. The classification can be binary (e.g., positive or negative) or have more levels of classification (e.g., a scale from 1 to 10 or 0 to 1), including probabilities. Different techniques for determining a classification can be combined to obtain a final classification from the initial or intermediate classification for each of the different techniques, e.g., by majority vote or a requirement that all initial / intermediate classifications are the same (e.g., positive). In some embodiments, a classification provided herein comprises a risk of metastasis.

[0033] The terms “cutoff” and “threshold” refer to predetermined numbers used in an operation. For example, a cutoff size can refer to a size above which fragments are excluded. As another example, a threshold value may be a value above or below which a particular classification applies. Either of these terms can be used in either of these contexts. A cutoff or threshold may be “a reference value” or derived from a reference value that is representative of a particular classification or discriminates between two or more classifications. A cutoff may be predetermined with or without reference to the characteristics of the sample or the subject. For example, cutoffs may be chosen based on the age or sex of the tested subject. A cutoff may be chosen after and based on output of the test data. For example, certain cutoffs may be used when the sequencing of a sample reaches a certain depth. As another example, reference subjects with known classifications of one or more conditions and measured characteristic values (e.g., a methylation level, a statistical size value, or a count) can be used to determine reference levels to discriminate between the different conditions and / or classifications of a condition (e.g., whether the subject has the condition). A reference value can be selected as representative of one classification (e.g., a mean) or a value that is between two clusters of the metrics (e.g., chosen to obtain a desired sensitivity and specificity). As another example, a reference value can be determined based on statistical simulations of samples. Any of these terms can be used in any of these contexts. Such a reference value can be determined in various ways, as will be appreciated by the skilled person. For example, metrics can be determined for two different cohorts ofsubjects with different known classifications, and a reference value can be selected as representative of one classification (e.g., a mean) or a value that is between two clusters of the metrics (e.g., chosen to obtain a desired sensitivity and specificity). As another example, a reference value can be determined based on statistical simulations of samples. A particular value for a cutoff, threshold, reference, etc. can be determined based on a desired accuracy (e.g., a sensitivity and specificity).

[0034] The terms “genetic variant,” “nucleotide variant,” and “mutation” may be used herein interchangeably to refer to changes or alterations to the reference human gene sequence at a particular locus, including, but not limited to, nucleotide base point mutations, polymorphisms, deletions, insertions, inversions, substitutions, translocations, rearrangements, fusions, breaks, duplications, amplifications, or repeats in the coding and non-coding regions. Certain mutations may be detectable on a given molecule, e.g., using next -generation sequencing techniques (see, e.g., Slatko BE, et al. Overview of Next-Generation Sequencing Technologies. Curr Protoc Mol Biol. 2018 Apr;122(l):e59), or in some cases long read techniques, such as single molecule sequencing such as Single-molecule real-time (SMRT) sequencing and nanopore sequencing. Deletions may be of a single nucleotide base, a portion or a region of the nucleotide sequence of the gene, or of the entire gene sequence. Insertions may be of one or more nucleotide bases. The variants may occur in transcriptional regulatory regions, untranslated regions of mRNA, exons, introns, exon / intron junctions, etc. Variants can potentially result in stop codons, frame shifts, deletions of amino acids, altered gene transcript splice forms or altered amino acid sequence.

[0035] An allele or gene allele comprises generally a naturally occurring gene having a reference sequence or a gene containing a specific nucleotide variant.

[0036] A haplotype refers to a combination of genetic (nucleotide) variants in a region of an mRNA or a genomic DNA on a chromosome found in an individual. Thus, a haplotype includes a number of genetically linked polymorphic variants which are typically inherited together as a unit.

[0037] The term “locus” refers to a specific position or site in a gene sequence or protein. Thus, there may be one or more contiguous nucleotides in a particular gene locus, or one or more amino acids at a particular locus in a polypeptide. Moreover, a locus may refer to a particular position in a gene where one or more nucleotides have been deleted, inserted, or inverted.

[0038] A “biomarker” or “marker” can comprise various biological entities of interest, e.g., a gene and / or gene product depending on the context. “Biomarkers” or “sets of biomarkers” can be used to train and test machine learning models and classify samples. Particular biomarkers may be used, such as particular nucleic acids (e.g., DNA or RNA, such as mRNA and microRNA), proteins, lipids, carbohydrates and metabolites and optionally also include a state of such molecules. Examples of the state of a biomarker include various aspects that can be queried such as presence, level (quantity, concentration, etc., such as an expression level), sequence, location, activity, structure, modifications, covalent or non-covalent binding partners, and the like. As a non-limiting examples, a set of biomarkers may include a gene or gene product (i.e., mRNA or protein) having a specified sequence (e.g., KRAS mutant), and / or a gene or gene product and a level thereof (e.g., amplified ERBB2 gene or overexpressed HER2 protein), In some cases, biomarkers may be circulating biomarkers that are detectable in body fluids, such as blood, plasma, and serum. Various numbers of biomarkers can be used to generate an input data structure (also referred to as an input feature vector), e.g., at least 10, 20, 50, 100, 200, 500, 1,000, 2,000, 5,000, 10,000, 20,000, 50,000, or 100,000 biomarkers can be used.

[0039] In the machine learning context, a “label” can refer to a known classification of a training sample. For example, a label can include a first-level classification (such as a tumor origin) and potentially one or more additional classifications for one or more additional layers when a tumor type has multiple subtypes.

[0040] The terms “primer”, “probe,” and “oligonucleotide” may be used herein interchangeably to refer to a relatively short nucleic acid fragment or sequence. They can comprise DNA, RNA, or a hybrid thereof, or chemically modified analog or derivatives thereof. Typically, they are single-stranded. However, they can also be double-stranded having two complementing strands which can be separated by denaturation. Normally, primers, probes and oligonucleotides have a length of from about 8 nucleotides to about 200 nucleotides, preferably from about 12 nucleotides to about 100 nucleotides, and more preferably about 18 to about 50 nucleotides. They can be labeled with detectable markers or modified using conventional manners for various molecular biological applications.

[0041] The term “isolated” when used in reference to nucleic acids (e.g., genomic DNAs, cDNAs, mRNAs, or fragments thereof) is intended to mean that a nucleic acid molecule ispresent in a form that is substantially separated from other naturally occurring nucleic acids that are normally associated with the molecule. Because a naturally existing chromosome (or a viral equivalent thereof) includes a long nucleic acid sequence, an isolated nucleic acid can be a nucleic acid molecule having only a portion of the nucleic acid sequence in the chromosome but not one or more other portions present on the same chromosome. More specifically, an isolated nucleic acid can include naturally occurring nucleic acid sequences that flank the nucleic acid in the naturally existing chromosome (or a viral equivalent thereof). An isolated nucleic acid can be substantially separated from other naturally occurring nucleic acids that are on a different chromosome of the same organism. An isolated nucleic acid can also be a composition in which the specified nucleic acid molecule is significantly enriched so as to constitute at least 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, 95%, or at least 99% of the total nucleic acids in the composition.

[0042] An isolated nucleic acid can be a hybrid nucleic acid having the specified nucleic acid molecule covalently linked to one or more nucleic acid molecules that are not the nucleic acids naturally flanking the specified nucleic acid. For example, an isolated nucleic acid can be in a vector. In addition, the specified nucleic acid may have a nucleotide sequence that is identical to a naturally occurring nucleic acid or a modified form or mutein thereof having one or more mutations such as nucleotide substitution, deletion / insertion, inversion, and the like.

[0043] An isolated nucleic acid can be extracted from a tissue sample, prepared from a recombinant host cell (in which the nucleic acids have been recombinantly amplified and / or expressed), or can be a chemically synthesized nucleic acid having a naturally occurring nucleotide sequence or an artificially modified form thereof.

[0044] The term “mapping" or “aligning” refers to a process that relates a sequence to a location or coordinate (e.g., a genomic coordinate) in a reference (e.g., a reference genome) having a known reference sequence, where the sequence is similar to the known reference sequence at the location in the reference. The degree of similarity can be measured or reported in terms of a “mapping quality"" In one example of a mapping quality used herein, a mapping quality of X for a sequence with respect to a reported location or coordinate in a reference indicates that the probability of the sequence mapping to a different location is no greater than1 OA(-X / 10). For instance, a mapping quality of 30 indicates a less than 0.1% probability of the sequence mapping to an alternate location.

[0045] A “reference genome" or “reference sequence” may be an entire genome sequence of a reference organism, one or more portions of a reference genome that may or may not be contiguous, a consensus sequence of many reference organisms, a compilation sequence based on different components of different organisms, or any other appropriate reference sequence. As examples, a reference genome / sequence can at least 1,000, 10,000, 50,000, 100,000, 500,000, 1,000,000, 5,000,000, 10,000,000, 50,000,000, 100,000,000, 500,000,000, one billions, or 3 billion nucleotides long, e.g., a full human genome or a repeat masked human genome. A reference may also include information regarding variations of the reference known to be found in a population of organisms.

[0046] For the purpose of comparing two different nucleic acid or polypeptide sequences, one sequence (test sequence) may be described to be a specific percentage identical to another sequence (comparison sequence). The percentage identity can be determined by the algorithm of Karlin and Altschul, Proc. Natl. Acad. Sci. USA, 90:5873-5877 (1993), which is incorporated into various BLAST programs. The percentage identity can be determined by the “BLAST 2 Sequences” tool, which is available at the National Center for Biotechnology Information (NCBI) website. See Tatusova and Madden, FEMS Microbiol. Lett., 174(2):247-250 (1999). For pairwise DNA-DNA comparison, the BLASTN program is used with default parameters (e.g., Match: 1; Mismatch: -2; Open gap: 5 penalties; extension gap: 2 penalties; gap x_dropoff 50; expect: 10; and word size: 11, with filter). For pairwise protein-protein sequence comparison, the BLASTP program can be employed using default parameters (e.g., Matrix: BLOSUM62; gap open: 11; gap extension: 1; x_dropoff: 15; expect: 10.0; and wordsize: 3, with filter). Percent identity of two sequences is calculated by aligning a test sequence with a comparison sequence using BLAST, determining the number of amino acids or nucleotides in the aligned test sequence that are identical to amino acids or nucleotides in the same position of the comparison sequence, and dividing the number of identical amino acids or nucleotides by the number of amino acids or nucleotides in the comparison sequence. When BLAST is used to compare two sequences, it aligns the sequences and yields the percent identity over defined, aligned regions. If the two sequences are aligned across their entire length, the percent identity yielded by the BLAST is thepercent identity of the two sequences. If BLAST does not align the two sequences over their entire length, then the number of identical amino acids or nucleotides in the unaligned regions of the test sequence and comparison sequence is considered to be zero and the percent identity is calculated by adding the number of identical amino acids or nucleotides in the aligned regions and dividing that number by the length of the comparison sequence. Various versions of the BLAST programs can be used to compare sequences, e.g., BLAST 2.1.2 or BLAST+ 2.2.22.

[0047] A ''machine learning model” (ML model) can refer to a software module configured to be run on one or more processors to provide a classification or numerical value of a property of one or more samples. An ML model can include various parameters (e.g., for coefficients, weights, thresholds, functional properties of function, such as activation functions). As examples, an ML model can include at least 10, 100, 1,000, 5,000, 10,000, 50,000, 100,000, or one million parameters. An ML model can be generated using sample data (e.g., training samples) to make predictions on test data. Various number of training samples can be used, e.g., at least 10, 100, 1,000, 5,000, 10,000, 50,000, 100,000, or at least 200,000 training samples. One example is an unsupervised learning model, such as clustering. Another example type of model is supervised learning. Example supervised learning models may include different approaches and algorithms including analytical learning, statistical models, artificial neural network, backpropagation, boosting (meta-algorithm), Bayesian statistics, case-based reasoning, decision tree learning, inductive logic programming, Gaussian process regression, genetic programming, group method of data handling, kernel estimators, learning automata, learning classifier systems, minimum message length (decision trees, decision graphs, etc.), multilinear subspace learning, naive Bayes classifier, maximum entropy classifier, conditional random field, nearest neighbor algorithm, probably approximately correct learning (PAC) learning, ripple down rules, a knowledge acquisition methodology, symbolic machine learning algorithms, subsymbolic machine learning algorithms, minimum complexity machines (MCM), random forests, ensembles of classifiers, ordinal classification, data pre-processing, handling imbalanced datasets, statistical relational learning, or Proaftn, a multicriteria classification algorithm. The model may include linear regression, logistic regression, deep recurrent neural network (e.g., long short term memory, LSTM), hidden Markov model (HMM), linear discriminant analysis (LDA), k-means clustering, density -based spatial clustering of applications with noise (DBSCAN), random forest algorithm, support vector machine (SVM), or any model described herein. Supervised learning models canbe trained in various ways using various cost / loss functions that define the error from the known label (e.g., least squares and absolute difference from known classification) and various optimization techniques, e.g., using backpropagation, steepest descent, conjugate gradient, and Newton and quasi-Newton techniques.

[0048] The term “about” or “approximately” can mean within an acceptable error range for the particular value as determined by one of ordinary skill in the art, which will depend in part on how the value is measured or determined, i.e., the limitations of the measurement system. For example, “about” can mean within 1 or more than 1 standard deviation, per the practice in the art. Alternatively, “about” can mean a range of up to 20%, up to 10%, up to 5%, or up to 1% of a given value. Alternatively, particularly with respect to biological systems or processes, the term “about” or “approximately” can mean within an order of magnitude, within 5-fold, and more preferably within 2-fold, of a value. Where particular values are described in the application and claims, unless otherwise stated the term “about” meaning within an acceptable error range for the particular value should be assumed. The term “about” can have the meaning as commonly understood by one of ordinary skill in the art. The term “about” can refer to ±10%. The term “about” can refer to ±5%.

[0049] Where a range of values is provided, it is understood that each intervening value, to the tenth of the unit of the lower limit unless the context clearly dictates otherwise, between the upper and lower limits of that range is also specifically disclosed. Each smaller range between any stated value or intervening value in a stated range and any other stated or intervening value in that stated range is encompassed within embodiments of the present disclosure. The upper and lower limits of these smaller ranges may independently be included or excluded in the range, and each range where either, neither, or both limits are included in the smaller ranges is also encompassed within the present disclosure, subject to any specifically excluded limit in the stated range. Where the stated range includes one or both of the limits, ranges excluding either or both of those included limits are also included in the present disclosure.DETAILED DESCRIPTION

[0050] Application of artificial intelligence (Al) to large molecular data sets covering transcribed genes has been limited. To date, there have not been large enough transcriptomic data sets to take advantage of the power of Al technology to predict disease progression for patientswith cancer using large amount of transcriptomic data. Provided herein is an application of machine learning (ML) and Al methodologies to a large collection of molecular and clinical data that generates insight into disease progression. The molecular data comprises whole transcriptome sequencing (WTS) data obtained using sequencing technology (e.g., next-generation sequencing). As the transcriptome contains over 22,000 genes, the use of very large (e.g., >>100,000) patient cohorts are used herein to maximize accuracy for applications in precision medicine and to advance the science and medicine of cancer care. Our models directly impact the care of current patients as well as provide insight into future drug develop efforts.

[0051] Comprehensive molecular profding provides a wealth of data concerning the molecular status of patient samples. Such molecular profding data (e.g., using transcriptome data) can be compared with any variety of attributes, such as patient characteristics and clinical data such as treatment outcomes, and processed using machine learning models to identify models that make predictions based on the input attributes, including, without limitation, disease characteristics (e.g., risk of metastasis), disease progression, and the effectiveness of various treatments. Here, this approach has been applied to develop a risk score for metastasis.I. MACHINE LEARNING AND ENSEMBLE MODELS

[0052] "Machine learning" can include an artificial intelligence process in which software applications may be trained to make accurate predictions through learning. The predictions can be generated by applying input data to a predictive model formed from performing statistical analyses on aggregated data. A model can be trained using training data, such that the model may be used to make accurate predictions. Once trained, the trained model can be used to process input data and make predictions based on the processed input data. The prediction can be, for example, a classification of a risk that the cancer in the subject will metastasize based on the analysis of a biological sample of the subject. In some cases, the predication can be a relative indication of a metastatic potential, such as a likelihood or other metric whether a cancer is more likely (higher metastatic potential) or less likely (lower metastatic potential) to become metastatic.

[0053] A “machine learning model” may include an application of artificial intelligence that provides systems with the ability to automatically learn and improve from experience withoutexplicitly being programmed. A machine learning model may include a set of software routines and parameters that can predict an output of a process (e.g., classification of a biological sample or of a subject from which the biological sample was obtained) based on feature vectors or other input data. A structure of the software routines (e.g., number of subroutines and the relation between them) and / or the values of the parameters can be determined in a training process, which can use actual results of the process that is being modeled, e.g., the identification of different classes of input data. An example type of the machine learning model is supervised learning that can be used with embodiments of the present disclosure. Example supervised learning models may include different approaches and algorithms including analytical learning, artificial neural network, backpropagation, boosting (meta-algorithm), Bayesian statistics, casebased reasoning, decision tree learning, inductive logic programming, Gaussian process regression, genetic programming, group method of data handling, kernel estimators, learning automata, learning classifier systems, minimum message length (decision trees, decision graphs, etc.), multilinear subspace learning, naive Bayes classifier, maximum entropy classifier, conditional random field, nearest neighbor algorithm, probably approximately correct learning (PAC) learning, ripple down rules, a knowledge acquisition methodology, symbolic machine learning algorithms, subsymbolic machine learning algorithms, minimum complexity machines (MCM), random forests, ensembles of classifiers, ordinal classification, data pre-processing, handling imbalanced datasets, statistical relational learning, or Proaftn, a multicriteria classification algorithm.

[0054] The model may include linear regression, logistic regression, convolutional neural network (CNN), deep recurrent neural network (e g., fully-connected recurrent neural network (RNN), Gated Recurrent Unit (GRU), long short-term memory, (LSTM)), transformed-based methods (e.g. XLNet, BERT, XLM, RoBERTa), hidden Markov model (HMM), linear discriminant analysis (LDA), k-means clustering, density-based spatial clustering of applications with noise (DBSCAN), random forest algorithm, adaptive boosting (AdaBoost), extreme Gradient Boosting (XGBoost), support vector machine (SVM), or any model described herein. A machine learning model can be trained using “training data” (e.g., to identify patterns in the training data) and then apply this training when it is used for its intended purpose. A machine learning model may be defined by “model parameters,” which can comprise numerical values that define how the machine learning model performs its function. Training a machine learningmodel can comprise an iterative process used to determine a set of model parameters that achieve the best performance for the model. Supervised learning models can be trained in various ways using various cost / loss functions that define the error from the known label (e.g., least squares and absolute difference from known classification) and various optimization techniques, e.g., using b ckpropagation, steepest descent, conjugate gradient, and Newton and quasi-Newton techniques

[0055] In broad terms, the process of training a machine learning model can involve determining the set of parameters that achieve the “best” performance, often based on the output of a loss or error function. A loss function typically relates the expected or ideal performance of a machine learning model to its actual performance. For example, if there is a training data set comprising input data paired with corresponding labels (indicating, e.g., whether that input data corresponds to normal data or anomalous data), then a loss function or error function can relate or compare the classifications or labels produced by the machine learning model to the known labels corresponding to the training data set. If the machine learning model produces labels that are identical to the labels associated with the training data, then the loss function may take on a small value (e.g., zero), while if the machine learning model produces labels that are inconsistent with the labels corresponding to the training data set, then the loss function may take on a larger value (e.g., one, a value greater than one, infinity, etc.).

[0056] In some cases, the training process can involve iteratively updating the model parameters during training rounds, batches, epochs, or other suitable divisions of the training process. In each round, a computer system can evaluate the current performance of the model for a particular set of parameters using the loss or error function. Then, the computer system can use metrics such as the gradient and techniques such as stochastic gradient descent to update the model parameters based on the loss or error function. In summary terms, the computer system can predict which change to the model parameters result in the fastest decrease in the loss function, then change the model parameters based on that prediction. This process can be repeated in subsequent training rounds, epochs, etc. The computer system can perform this iterative process until training is complete. In some cases, the training process involve a set number of training rounds or epochs. In other cases, training can proceed until the model “converges”, i.e., the model shows little to no further improvement in successive training rounds,or that the difference in the output of the error or loss function between successive training rounds approaches zero.

[0057] Once training has been completed, a trained classifier can be used to classify unlabeled input data based on the parameters determined during training, e.g., determining at least one attribute of a biological sample of a subject, such as the risk that the cancer in the subject will metastasize.

[0058] Aspects of the present disclosure are directed towards a system that generates a set of one or more training data structures that can be used to train a machine learning model to provide various classifications, such as characterizing a phenotype of a biological sample. Characterizing a phenotype can include providing a diagnosis, prognosis, theranosis or other relevant classification. For example, the classification can be predicting a disease state, such as a state or other characteristic of a cancer (e.g., tissue-of-origin, metastatic potential, etc), or effectiveness of a treatment for a disease or disorder of a subject having a particular set of biomarkers. Once trained, the trained machine learning model can be used to process input data provided by the system and make predictions based on the processed input data. The input data may include a set of features related to a subject such as data representing one or more subject biomarkers and data representing a disease or disorder or related characteristic. In some embodiments, the input data may include features representing an observable characteristic of a biological sample and make a prediction about the sample, such as the tissue-of-origin of the sample, or, in the case of a cancer sample, the cancer’s metastatic potential, which potential is the tendency of a primary tumor to form secondary metastatic lesions. In some embodiments, the input data may include features representing a proposed treatment type and make a prediction describing the subject’s likely responsive to the treatment. The prediction may include data that is output by the machine learning model based on the machine learning model’s processing of a specific set of features provided as an input to the machine learning model. The data may include data representing one or more subject biomarkers, data representing a disease or disorder, data representing characteristics of a disease or disorder within a particular subject, or data representing a proposed treatment type as desired.

[0059] The system is configured to obtain output data generated by the trained machine learning model based on the machine learning model’s processing of the data. In variousembodiments, the data comprises biological data representing one or more biomarkers, data representing a disease or disorder, data representing characteristics of a disease or disorder within a particular subject, and / or data representing a treatment type. The system may make a prediction for a subject having a particular set of biomarkers, including but not limited to effectiveness of a treatment or metastatic potential. In some implementations, the disease or disorder may include a type of cancer.

[0060] Consider an example wherein the system is making a prediction about efficacy of a treatment. The treatment for the subject may include one or more therapeutic agents, e.g., small molecule drugs, biologies, and various combinations thereof. In this setting, output of the trained machine learning model that is generated based on trained machine learning model processing of the input data that includes the set of biomarkers, the disease or disorder and the treatment type includes data representing the level of responsiveness that the subject will have to the treatment for the disease or disorder.

[0061] Consider another example wherein the system is making a prediction about metastatic potential of a cancer in a subject. In this setting, output of the trained machine learning model that is generated based on trained machine learning model processing of the input data that includes the set of biomarkers, the disease or disorder, and other relevant sample or patient data, includes data representing the metastatic potential of the cancer.

[0062] In some implementations, the output data generated by the trained machine learning model may include a probability of the desired classification. By way of illustration, such probability may be a probability that the subject will favorably respond to the treatment for the disease or disorder. As another illustration, such probability may be a probability that the cancer in the subject will metastasize. In other implementations, the output data may include any output data generated by the trained machine learning model based on the trained machine learning model’s processing of the input data.

[0063] In some implementations, the training data structures generated by the present disclosure may include a plurality of training data structures that each include fields representing feature vector corresponding to a particular training sample. The feature vector includes a set of features derived from, and representative of, a training sample. The training sample may include, for example, one or more biomarkers of a subject, a disease or disorder of the subject, variouscharacteristics of the disease or disorder of the subject (e.g., whether metastasis occurred), and / or a proposed treatment for the disease or disorder. The training data structures are flexible because each respective training data structure may be assigned a weight representing each respective feature of the feature vector. Thus, each training data structure of the plurality of training data structures can be particularly configured to cause certain inferences to be made by a machine learning model during training.

[0064] Consider a non-limiting example wherein the model is trained to make a prediction of likely metastatic spread of a cancer in a subject. As a result, the novel training data structures that are generated in accordance with this specification are designed to improve the performance of a machine learning model because they can be used to train a machine learning model to predict metastasis in a cancer in a subject having a particular set of biomarkers. By way of example, a machine learning model that could not perform predictions regarding the metastatic potential of a cancer in a subject having a particular set of biomarkers prior to being trained using the training data structures, system, and operations described by this disclosure can learn to make predictions regarding metastatic potential being trained using the training data structures, systems and operations described by the present disclosure. Accordingly, this process takes an otherwise general purpose machine learning model and changes the general purpose machine leaning model into a specific computer for performing a specific task of predicting the metastatic potential of a cancer in a subject having a particular set of biomarkers.

[0065] FIG. 1A is a block diagram of an example system for training a machine learning model 110. The machine learning may employ any useful predictive modelling approach. The machine learning model may be, for example, a decision tree, such as a random forest model or gradient boosted tree. In some embodiments, the machine learning model may include a support vector machine, neural network model, a linear regression model, a logistic regression model, a naive Bayes model, a quadratic discriminant analysis model, a K-nearest neighbor model, or the like. The machine learning model training system 100 may be implemented as computer programs on one or more computers in one or more locations, in which the systems, components, and techniques described below can be implemented. The machine learning model training system 100 trains the machine learning model 110 using training data items from a database (or data set) 120 of training data items. The training data items may include a plurality of featurevectors. Each training vector may include a plurality of values that each correspond to a particular feature of a training sample that the training vector represents. The training features may be referred to as independent variables. In addition, the system 100 maintains a respective weight for each feature that is included in the feature vectors.

[0066] The machine learning model 110 is configured to receive an input training data item 122 and to process the input training data item 122 to generate an output 118. The input training data item may include a plurality of features (or independent variables “X”) and a training label (or dependent variable “Y”). The machine learning model may be trained using the training items, and once trained, is capable of predicting X = (Y).

[0067] To enable machine learning model 110 to generate accurate outputs for received data items, the machine learning model training system 100 may train the machine learning model 110 to adjust the values of the parameters of the machine learning model 110, e.g., to determine trained values of the parameters from initial values. These parameters derived from the training steps may include weights that can be used during the prediction stage using the fully trained machine learning model 110.

[0068] In training, the machine learning model 110, the machine learning model training system 100 uses training data items stored in the database (data set) 120 of labeled training data items. The database 120 stores a set of multiple training data items, with each training data item in the set of multiple training items being associated with a respective label. Generally, the label for the training data item identifies a correct classification (or prediction) for the training data item, i.e., the classification that should be identified as the classification of the training data item by the output values generated by the machine learning model 110. With reference to FIG. 1A, a training data item 122 may be associated with a training label 122a.

[0069] The machine learning model training system 100 trains the machine learning model 110 to optimize an objective function. Optimizing an objective function may include, for example, minimizing a loss function 130. Generally, the loss function 130 is a function that depends on the (i) output 118 generated by the machine learning model 110 by processing a given training data item 122 and (ii) the label 122a for the training data item 122, i.e., the target output that the machine learning model 110 should have generated by processing the training data item 122.

[0070] The machine learning model training system 100 can train the machine learning model 110 to minimize the (cumulative) loss function 130 by performing multiple iterations of conventional machine learning model training techniques on training data items from the database 120, e.g., hinge loss, stochastic gradient methods, stochastic gradient descent with backpropagation, or the like, to iteratively adjust the values of the parameters of the machine learning model 110. A fully trained machine learning model 110 may then be deployed as a predicting model that can be used to make predictions based on input data that is not labeled.

[0071] FIG. IB is a block diagram of a system 200 that generates training data structures for training a machine learning model to predict metastatic potential or risk of metastasis of a cancer in a subject having a particular set of biomarkers.

[0072] The system 200 includes two or more distributed computers 210, 310, a network 230, and an application server 240. The application server 240 includes an extraction unit 242, a memory unit 244, a vector generation unit 250, and a machine learning model 270. The machine learning model 270 may include one or more of a support vector machine, a neural network model, a linear regression model, a decision tree (including without limitation gradient boosted tree, random forest model), a logistic regression model, a naive Bayes model, a quadratic discriminant analysis, model, a K-nearest neighbor model, or the like. Each distributed computer 210, 310 may include a smartphone, a tablet computer, laptop computer, or a desktop computer, or the like. Alternatively, the distributed computers 210, 310 may include server computers that receive data input by one or more terminals 205, 305, respectively. The terminal computers 205, 305 may include any user device including a smartphone, a tablet computer, a laptop computer, a desktop computer or the like. The network 230 may include one or more networks 230 such as a LAN, a WAN, a wired Ethernet network, a wireless network, a cellular network, the Internet, or any combination thereof.

[0073] The application server 240 is configured to obtain, or otherwise receive, data records 220, 222, 224, 320 provided by one or more distributed computers such as the first distributed computer 210 and the second distributed computer 310 using the network 230. In some implementations, each respective distributed computer 210, 310 may provide different types of data records 220, 222, 224, 320. For example, the first distributed computer 210 may provide biomarker data records 220, 222, 224 representing biomarkers for a subject and the seconddistributed computer 310 may provide outcome data 320 representing outcome data for a subject obtained from the outcomes database 312. In this context, outcome can refer to whether or a not a cancer becomes metastatic and location of any such secondary (metastatic) tumors.

[0074] The biomarker data records 220, 222, 224 may include any type of biomarker data that describes biometric attributes of a subject. By way of example, the example of FIG. IB shows the biomarker data records as including data records representing DNA biomarkers 220, protein biomarkers 222, and RNA data biomarkers 224. These biomarker data records may each include data structures having fields that structure information 220a, 222a, 224a describing biomarkers of a subject such as a subject’s DNA biomarkers 220a, protein biomarkers 222a, or RNA biomarkers 224a (e.g., a number of transcripts for each of a set of genes or all known genes). For example, the biomarker data records 220, 222, 224 may include next generation sequencing data such as DNA alterations. Such next generation sequencing data may include single variants / mutations, insertions and deletions, substitutions, translocations, fusions, breaks, duplications, amplification, loss, copy number, repeats, tumor mutational burden (TMB, also referred to as total mutational burden or tumor mutation load (TML)), microsatellite instability (MSI), or the like. Next generation sequencing data can be for specific sets of genes or any other desired genomic loci, or can comprise whole exome, whole genome, or whole transcriptome data, or a combination such as whole exome data with a boosted gene set. Alternatively, or in addition, the biomarker data records 220, 222, 224 may also include in situ hybridization data such as DNA copy number or chromosomal loss or rearrangements. Alternatively, or in addition, the biomarker data records 220, 222, 224 may include RNA data such as gene expression or gene fusion, including without limitation RNA-seq, including whole transcriptome sequencing (WTS), with or without boosted sets of transcripts. RNA data may be reported, normalized, or otherwise processed in various ways such as transcripts per million (TPM), trimmed mean of M values (TMM; see Johnson, K.A., Krishnan, A. Robust normalization and transformation techniques for constructing gene coexpression networks from RNA-seq data. Genome Biol 23, 1 (2022)), or Gene Set Enrichment Analysis (GSEA). Alternatively, or in addition, the biomarker data records 220, 222, 224 may include protein expression data such as obtained using immunohistochemistry (IHC). However, the present disclosure need not be so limited and any biomarker data can be employed, such as described herein.

[0075] In some implementations, the set of one or more biomarkers include one or more biomarkers obtained by whole genome sequencing (WGS), whole exome sequencing (WES), whole transcriptome sequencing (WTS), or any useful combination thereof. In some embodiments, the sequencing information for certain biomarkers (e.g., genes or regions) of interest is boosted, e.g., by higher depth of sequencing for such certain biomarkers of interest as compared to others. For example, WES may be supplemented with additional baits that capture certain genes of interest, e.g., cancer genes, to provide higher depth of sequencing for those genes. Similarly, WTS may be supplemented with additional baits that capture certain transcripts of interest, e.g., transcripts for cancer genes, to provide higher depth of sequencing for those transcripts.

[0076] The outcome data records 320 may describe outcomes of a cancer for a subject, such as metastasis. For example, the outcome data records 320 obtained from the outcome database 312 may include one or more data structures having fields that structure data attributes of a subject such as a cancer, primary location of the tumor, secondary tumor location (or none), or a combination thereof. In addition, the outcome data records 320 may also include fields that structure data attributes describing details of the metastasis and secondary tumor locations. An example of a cancer may include, for example, a lineage of cancer, such as breast cancer, prostate cancer, brain cancer, or other type of cancer such as described herein. A location of metastasis may include, for example, a location of a secondary tumor included in the outcome data records 320, such as brain, bones, liver, lungs, and / or adrenal glands. A metastasis result may include data representing the outcome of a subject’s cancer such as whether the cancer did or did not metastasize, and a time frame used to make such determination. Accordingly, although the example of FIG. IB indicates that outcome data may include a cancer, a primary and / or secondary tumor location, and a metastatic result, the outcome data may include other types of information, as described herein. Moreover, there is no requirements that the outcome data be limited to human “patients.” Instead, the outcome data records 220, 222, 224 and biometric data records 320 may be associated with any desired subject including any non-human organism.

[0077] In some implementations, each of the data records 220, 222, 224, 320 may include keyed data that enables the data records from each respective distributed computer to be correlated by application server 240. The keyed data may include, for example, data representinga subject identifier. The subject identifier may include any form of data that identifies a subject and that can associate biomarker for the subject with outcome data for the subject.

[0078] The first distributed computer 210 may provide 208 the biomarker data records 220, 222, 224 to the application server 240. The second distributed computer 310 may provide the outcome data records 320 to the application server 240. The application server 240 can provide the biomarker data records 220 and the outcome data records 220, 222, 224 to the extraction unit 242.

[0079] The extraction unit 242 can process the received biomarker data 220, 222, 224 and outcome data records 320 in order to extract data 220a- 1, 222a- 1, 224a- 1, 320-1, 320-2, 320-3 that can be used to train the machine learning model. For example, the extraction unit 242 can obtain data structured by fields of the data structures of the biometric data records 220, 222, 224, obtain data structured by fields of the data structures of the outcome data records 320, or a combination thereof. The extraction unit 242 may perform one or more information extraction algorithms such as keyed data extraction, pattern matching, natural language processing, or the like to identify and obtain data 220a-l, 222a-l, 224a-l, 320-1, 320-2, 320-3 from the biometric data records 220, 222, 224 and outcome data records 320, respectively. The extraction unit 242 may provide the extracted data to the memory unit 244. The extracted data unit may be stored in the memory unit 244 such as flash memory (as opposed to a hard disk) to improve data access times and reduce latency in accessing the extracted data to improve system performance. In some implementations, the extracted data may be stored in the memory unit 244 as an in-memory data grid.

[0080] In more detail, the extraction unit 242 may be configured to filter a portion of the biomarker data records 220, 222, 224 and the outcome data records 320 that will be used to generate an input data structure 260 for processing by the machine learning model 270 from the portion of the outcome data records 320 that will be used as a label for the generated input data structure 260. Such filtering includes the extraction unit 242 separating the biomarker data and a first portion of the outcome data that includes a cancer, primary location, secondary location (or none if the cancer did not spread), or a combination thereof, from the metastasis result. The application server 240 can then use the biomarker data 220a- 1, 222a- 1, 224a- 1, and the first portion of the outcome data that includes the cancer 320-1, metastasis result 320-3, anyadditional phenotypic details (not shown in FIG. IB), or a combination thereof, to generate the input data structure 260. In addition, the application server 240 can use the second portion of the outcome data describing the metastatic result 320-3 as the label for the generated data structure.

[0081] The application server 240 may process the extracted data stored in the memory unit 244 correlate the biomarker data 220a- 1, 222a- 1, 224a- 1 extracted from biomarker data records 220, 222, 224 with the first portion of the outcome data 320-1, 320-2. The purpose of this correlation is to cluster biomarker data with outcome data so that the outcome data for the subject is clustered with the biomarker data for the subject. In some implementations, the correlation of the biomarker data and the first portion of the outcome data may be based on keyed data associated with each of the biomarker data records 220, 222, 224 and the outcome data records 320. For example, the keyed data may include a subject identifier.

[0082] The application server 240 provides the extracted biomarker data 220a- 1, 222a- 1, 224a- 1 and the extracted first portion of the outcome data 320-1, 320-2 as an input to a vector generation unit 250. The vector generation unit 250 is used to generate a data structure based on the extracted biomarker data 220a- 1, 222a- 1, 224a- 1 and the extracted first portion of the outcome data 320-1, 320-2. The generated data structure is a feature vector 260 that includes a plurality of values that numerical represents the extracted biomarker data 220a- 1, 222a- 1, 224a- 1 and the extracted first portion of the outcome data 320-1, 320-2. The feature vector 260 may include a field for each type of biomarker and each type of outcome data. For example, the feature vector 260 may include one or more fields corresponding to (i) one or more types of next generation sequencing data such as single variants / mutations, insertions and deletions, substitution, translocation, fusion, break, duplication, amplification, loss, copy number, repeat, TMB, MSI status, (ii) one or more types of in situ hybridization data such as DNA copies, gene copies, gene translocations, (iii) one or more types of RNA data such as gene sequence, gene expression, gene fusion, TPM, TMM and / or GSEA, (iv) one or more types of protein data such as level and localization obtained using immunohistochemistry, (v) one or more types of outcomes data such as cancer, primary tumor location, secondary tumor location (or none), or the like.

[0083] The vector generation unit 250 is configured to assign a weight to each field of the feature vector 260 that indicates an extent to which the extracted biomarker data 220a- 1, 222a- 1,224a-l and the extracted first portion of the outcome data 320-1, 320-2 includes the data represented by each field. In one implementation, for example, the vector generation unit 250 may assign a ‘ 1 ’ to each field of the feature vector that corresponds to a feature found in the extracted biomarker data 220a- 1, 222a- 1, 224a- 1 and the extracted first portion of the outcome data 320-1, 320-2. In such implementations, the vector generation unit 250 may, for example, also assign a ‘0’ to each field of the feature vector that corresponds to a feature not found in the extracted biomarker data 220a- 1, 222a- 1, 224a- 1 and the extracted first portion of the outcome data 320-1, 320-2. The output of the vector generation unit 250 may include a data structures such as a feature vector 260 that can be used to train the machine learning model 270.

[0084] The application server 240 can label the training feature vector 260. Specifically, the application server can use the extracted second portion of the outcome data 320-3 to label the generated feature vector 260 with a metastatic result 320-3. The label of the training feature vector 260 generated based on the metastatic result 320-3 can provide an indication of metastatic potential of a primary tumor 320-2 for a cancer 320-1 of a subject defined by the specific set of biomarkers 220a-l, 222a-l, 224a-l, each of which is described by described in the training data structure 260.

[0085] The application server 240 can train the machine learning model 270 by providing the feature vector 260 as an input to the machine learning model 270. The machine learning model 270 may process the generated feature vector 260 and generate an output 272. The application server 240 can use a loss function 280 to determine the amount of error between the output 272 of the machine learning model 270 and the value specified by the training label, which is generated based on the second portion of the extracted patient outcome data describing the metastatic result 320-3. The output 282 of the loss function 280 can be used to adjust the parameters of the machine learning model 270.

[0086] In some implementations, adjusting the parameters of the machine learning model 270 may include manually tuning of the machine learning model parameters model parameters. Alternatively, in some implementations, the parameters of the machine learning model 270 may be automatically tuned by one or more algorithms of executed by the application server 240.

[0087] The application server 240 may perform multiple iterations of the process described above with reference to FIG. IB for each outcome data record 320 stored in the outcomesdatabase that correspond to a set of biomarker data for a subject. This may include hundreds of iterations, thousands of iterations, tens of thousands of iterations, hundreds of thousands of iterations, millions of iterations, or more, until each of the outcomes data records 320 stored in the outcomes database 312 and having a corresponding set of biomarker data for a subject are exhausted, until the machine learning model 270 is trained to within a particular margin of error, or a combination thereof. A machine learning model 270 is trained within a particular margin of error when, for example, the machine learning model 270 is able to predict, based upon a set of unlabeled biomarker data, cancer data, primary tumor data, and / or any other desired attributes, such as a metastatic potential of a cancer in a subject having the biomarker data. The prediction may include, for example, a risk or probability of such metastasis, or a similar metric.

[0088] In some implementations, the machine learning model 270 is trained to predict metastasis for particular primary tumor locations and / or secondary tumor locations. If the model is trained using primary tumor location data, the prediction may be that of metastatic potential of tumors from that primary location. As a non-limiting example, the model may be trained to predict metastasis of a primary breast tumor. If the model is trained using secondary tumor location data, the prediction may be that of metastasis to one or more particular secondary location. As a non-limiting example, the model may be trained to predict metastasis to the brain. In some embodiments, the model is trained using both primary and secondary tumor location data. As a non-limiting example, the model may be trained to predict metastasis of a breast cancer (primary tumor) to the brain (secondary tumor). In some embodiments, the model is trained to more generically predict metastatic potential of any number of different cancer lineages to any number of different secondary locations.

[0089] FIG. 1C is a block diagram of a system for using a machine learning model that has been trained to predict metastatic potential or risk of metastasis of a cancer in a subject having a particular set of biomarkers.

[0090] The machine learning model 370 includes a machine learning model that has been trained using the process described with reference to the system of FIG. IB above. The trained machine learning model 370 is capable of predicting, based on an input feature vector representative of a set of one or more biomarkers, cancer data, primary tumor data, and / or any other desired attributes, a metastatic potential of a cancer in a subject having the biomarker data.

[0091] The application server 240 hosting the machine learning model 370 is configured to receive unlabeled biomarker data records 320, 322, 324. The biomarker data records 320, 322, 324 include one or more data structures that have fields structuring data that represents one or more particular biomarkers such as DNA biomarkers 320a, protein biomarkers 322a, RNA biomarkers 324a, or any combination thereof. As discussed above, the received biomarker data records may include any desired biomarker data such as (i) one or more types of next generation sequencing data obtained for genomic DNA such as single variants, insertions and deletions, substitution, translocation, fusion, break, duplication, amplification, loss, copy number, repeat, TMB, MSI, (ii) one or more types of in situ hybridization data such as DNA copies, gene copies, gene translocations, (iii) one or more types of RNA data such as gene sequence, gene expression, gene fusion, TPM, TMM and / or GSEA, or (iv) one or more types of protein data such as level and / or location obtained using immunohistochemistry.

[0092] In some embodiments, the application server 240 hosting the machine learning model 370 is configured to receive data representing other useful phenotypic data 422a for a cancer described by the cancer data 420a of the subject having biomarkers represented by the received biomarker data records 320, 322, 324. Such phenotypic data can be, e.g., a primary location of the tumor and one or more secondary tumor location (e.g., metastatic lesion), if any. The phenotypic data 422a for the cancer data 420a also can be unlabeled.

[0093] In some implementations, the cancer data 420a and any additional phenotypic data 422a is provided 305 by a terminal 405 over the network 230 and the biomarker data is obtained from a second distributed computer 310. The biomarker data may be derived from laboratory machinery used to perform various assays. In other implementations, the cancer data 420a, additional phenotypic data 422a, and the biomarker data 320, 322, 324 may each be received from the terminal 405. For example, the terminal 405 may be user device of a doctor, an employee or agent of the doctor working at the doctor’s office, or other human entity that inputs data representing a cancer, data representing other characteristics of the cancer and / or subject (e.g., the additional phenotypic data), and a data representing one or more biomarkers for a subject having the disease or disorder. In some implementations, the phenotypic data 422 may include data structures structuring fields of data representing a primary tumor location described by an organ name and / or histology. In some implementations, the phenotypic data 422 mayinclude data structures structuring fields of data representing one or more secondary tumor location.

[0094] The application server 240 receives the biomarker data records 320, 322, 324, the cancer data 420, and any additional phenotypic data 422. The application server 240 provides the biomarker data records 320, 322, 324, the cancer data 420, and any additional phenotypic data 422 to an extraction unit 242 that is configured to extract (i) particular biomarker data such as DNA biomarker data 320a-l, protein expression data 322a-l, transcriptome data 324a-l (e.g., a number of transcripts for each of a set of genes or all known genes), (ii) cancer data 420a- 1, and (iii) any additional phenotypic data 422a- 1 from the fields of the biomarker data records 320, 322, 324 and the outcome data records 420, 422. In some implementations, the extracted data is stored in the memory unit 244 as a buffer, cache or the like, and then provided as an input to the vector generation unit 250 when the vector generation unit 250 has bandwidth to receive an input for processing. In other implementations, the extracted data is provided directly to a vector generation unit 250 for processing. For example, in some implementations, multiple vector generation units 250 may be employed to enable parallel processing of inputs to reduce latency.

[0095] The vector generation unit 250 can generate a data structure such as a feature vector 360 that includes a plurality of fields and includes one or more fields for each type of biomarker data and one or more fields for each type of outcome data. For example, each field of the feature vector 360 may correspond to (i) each type of extracted biomarker data that can be extracted from the biomarker data records 320, 322, 324 such as each type of next generation sequencing data, each type of in situ hybridization data, each type of RNA data, each type of immunohistochemistry data, and each type of aptamer- data and (ii) each type of outcome data that can be extracted from the outcome data records 420, 422 such as each type of cancer, each type of metastasis data, and each type of additional phenotypic details.

[0096] The vector generation unit 250 is configured to assign a weight to each field of the feature vector 360 that indicates an extent to which the extracted biomarker data 320a-l, 322a-l, 324a- 1, the extracted cancer data 420a- 1, and the extracted additional phenotypic data 422a- 1 includes the data represented by each field. In one implementation, for example, the vector generation unit 250 may assign a ‘ 1’ to each field of the feature vector 360 that corresponds to a feature found in the extracted biomarker data 320a-l, 322a-l, 324a-l, the extracted cancer 420a-1, and the extracted additional phenotypic data 422a-l . In such implementations, the vector generation unit 250 may, for example, also assign a ‘0’ to each field of the feature vector that corresponds to a feature not found in the extracted biomarker data 320a- 1, 322a- 1, 324a- 1, the extracted cancer 420a-l, and the extracted additional phenotypic data 422a- 1. The output of the vector generation unit 250 may include a data structure such as a feature vector 360 that can be provided as an input to the trained machine learning model 370.

[0097] The trained machine learning model 370 process the generated feature vector 360 based on the adjusted parameters that were determining during the training stage and described with reference to FIG. IB. The output 272 of the trained machine learning model provides an indication of the metastatic potential of the cancer 420a- 1 for the subject having biomarkers 320a-l, 322a-l, 324a-l. In some implementations, the output 272 may include a probability that is indicative of the metastatic potential of the cancer 420a-l for the subject having biomarkers 320a-l, 322a-l, 324a-l. In some embodiments, the output 272 comprises additional phenotypic data 422a- 1, such as a secondary tumor location. The output 272 may be provided 311 to the terminal 405 using the network 230. The terminal 405 may then generate output on a user interface 415 that indicates a predicted level of metastatic potential of the cancer for a person having the biomarkers represented by the feature vector 360.

[0098] In some implementations, the output 272 may be provided to a prediction unit 380 that is configured to decipher the meaning of the output 272. For example, the prediction unit 380 can be configured to map the output 272 to one or more categories of metastasis. Then, the output of the prediction unit 328 can be used as part of message 390 that is provided 311 to the terminal 305 using the network 230 for review by the subject, a guardian of the subject, a nurse, a doctor, or the like.

[0099] As noted above, the machine learning model 270 can be trained to predict a risk of metastasis for particular primary tumor locations and / or secondary tumor locations. If the model was trained using primary tumor location data, the prediction may be that of metastatic potential of tumors from that primary location. As a non-limiting example, a model trained to predict metastasis of a primary breast tumor may be used to predict metastasis of a primary breast tumor from a test subject. If the model was trained using secondary tumor location data, the prediction may be that of metastasis to one or more particular secondary location. As a non-limitingexample, a model trained to predict metastasis to the brain may be used to predict metastasis to the brain of a primary tumor from a test subject. In some embodiments, the model was trained using both primary and secondary tumor location data. As a non-limiting example, a model trained to predict metastasis of a breast cancer (primary tumor) to the brain (secondary tumor) may be used to predict metastasis of a breast cancer in a test subject to the brain. A model is trained to more generically predict metastatic potential of any number of different cancer lineages to any number of different secondary locations may be used to predict metastasis of any number of different cancer lineages in a test subject to any number of different secondary locations.

[0100] In some embodiments, a ML model can be an ensemble of multiple submodels. For example, a set of models can each contribute a vote or numerical output, from which an average or majority value can be determined. Accordingly, multiple models can be trained to perform the predict! on / classifi cation and the joint predictions can be used to make the classification. In this scenario, each model is allowed to “vote,” which can be in a weighted fashion, and the classification receiving the majority of the votes is deemed the winner. As another example, the outputs can be averaged to obtain an overall value (e.g., an averaged probability), where the highest probability can be compared to a threshold to determine if a metastatic potential for a given level can be called, as can be done in the embodiments described above.

[0101] The prediction made by the models can include one or more prediction, e.g., to characterize a phenotype. In some embodiments, the models may provide or assist in providing a diagnosis (e.g., disease or healthy), prognosis (e.g., predict a better or worse outcome), theranosis (e.g., predict or monitor therapeutic efficacy or lack thereof), in addition to other phenotypic characterization, such as metastatic potential the origin of tissue sample. As a nonlimiting example, the models may predict a metastatic potential (risk) of a breast tumor to metastasize to the brain.

[0102] In some embodiments, a particular submodel can have its contribution weighted more, e.g., based on historical accuracy. A confidence score for a submodel can be used to determine the weight. Accordingly, a confidence score indicating that a machine learning mode is historically accurate can be used to boost a value of output data generated by the machine learning model. Similarly, a confidence score indicating that a machine learning model ishistorically less accurate can be used to reduce a value of output data generated by the machine learning model. Such boosting or reducing of the value of output data generated by a machine learning model can be achieved, for example, by using the confidence score as a multiplier of less than one for reduction and more than 1 for boosting. Other operations can also be used to adjust the value of output data such as subtracting a confidence score from the value of the output data to reduce the value of the output data or adding the confidence score to the value of the output data to boost the value of the output data. Use of confidence scores to boost or reduce the value of output data generated by the machine learning models is particularly useful when the machine learning models are configured to output probabilities that will be applied to one or more thresholds to determine a risk score, such a risk of metastasis. This is because using the confidence score to adjust the output of a machine learning model can be used to move a generated output value above or below a class threshold, thereby altering a prediction by a machine learning model based on its historical accuracy.

[0103] FIG. ID is a block diagram of a system 400 using a voting unit to interpret output generated by multiple machine learning models. The system 400 is similar to the system 300 of FIG. 1C. However, instead of a single machine learning model 370, the system 400 includes multiple machine learning models 370-0, 370-1 ... 370-x, where x is any non-zero integer greater than 1. In addition, the system 400 also include a voting unit 480. As a non-limiting example, system 400 can be used for predicting metastasis of a cancer in a subject having a particular set of biomarkers.

[0104] Each machine learning model 370-0, 370-1, 370-x can include a machine learning model that has been trained to classify a particular type of input data 320-0, 320-1 ... 320-x, wherein x is any non-zero integer greater than 1 and equal to the number x of machine learning models. In some implementations, each of the machine learning models 370-0, 370-1, 370-x can be of the same type. For example, each of the machine learning models 370-0, 370-1, 370-x can be a decision tree classification algorithm, e.g., a gradient boosted tree or random forest, trained using differing parameters. In other implementations, the machine learning models 370-0, 370-1, 370-x can be of different types. For example, there can be one or more decision trees, one or more neural networks, one or more K-nearest neighbor classifiers, one or more SVM, or other types of machine learning models, or any combination thereof.

[0105] Input data such as input data-0 320-0, input data-1 320-1 , input data-x 320-x can be obtained by the application server 240. In some implementations, the input data 320-0, 320-1, 320-x is obtained across the network 230 from one or more distributed computers 315, 405. By way of example, one or more of the input data items 320-0, 320-1, 320-x can be generated by correlating data from multiple different data sources 210, 405. In such an implementation, (i) first data describing biomarkers for a subject can be obtained from the first distributed computer 315 and (ii) second data describing a cancer and metastasis thereof can be obtained from the second computer 405. The application server 240 can correlate the first data and the second data to generate an input data structure such as input data structure 320-0. This process is described in more detail in FIG. 1C. The input data items 320-0, 320-1, 320-x can be provided as respective inputs one-at-a-time, in series, for example, to the vector generation unit. The vector generation unit can generate input vectors 360-0, 360-1, 360-x that corresponding to each respective input data 320-0, 320-1, 320-x. While some implementations may generate vectors 360-0, 360-1, 360- x serially, the present disclosure need not be so limited.

[0106] Instead, in some implementations, the vector generation unit 250 can be configured to operate multiple parallel vector generation units that can parallelize the vector generation process. In such implementations, the vector generation unit 250 can receive input data 320-0, 320-1, 320-x in parallel, process the input data 320-0, 320-1, 320-x in parallel, and generate respective vectors 360-0, 360-1, 360-x that each correspond to one of the input data 320-0, 320- 1, 320-x in parallel.

[0107] In some implementations, the vectors 360-0, 360-1, 360-x can each be generated based on corresponding input data such as input data 320-0, 320-1, 320-x, respectively. That is, vector 360-0 is generated based on, and represents, input data 320-0. Similarly, vector 360-1 is generated based on, and represents, input data 320-1. Similarly, vector 360-x is generated based on, and represents, input data 320-x.

[0108] In some implementations, each input data structure 320-0, 320-1, 320-x can include data representing biomarkers of a subject, data describing a cancer associated with the subject, data describing a metastatic outcome for the subject, or any combination thereof. The data representing the biomarkers of a subject can include data describing a specific subset or panel of genes or proteins from a subject. Alternatively, in some implementations, the data representingbiomarkers of the subject can include data representing a substantially complete set of known genes, or their gene products, for a subject. In some implementations, each of the machine learning models 370-0, 370-1, 370-x are the same type machine learning model such as a decision tree trained to classify the input data vectors as corresponding to a cancer in a subject that is likely to metastasize (high metastatic potential) or likely to not metastasize (low metastatic potential) identified associated by the vector processed by the machine learning model. In such implementations, though each of the machine learning models 370-0, 370-1, 370-x is the same type of machine learning model, each of the machine learning models 370-0, 370-1, 370-x may be trained in different ways, e.g., with respect to different parameters or different input biomarkers. The machine learning models 370-1, 370-1, 370-x can generate output data 272-0, 272-1, 272-x, respectively, representing whether a cancer in a subject associated with input vectors 360-0, 360-1, 360-x is likely to metastasize or is not likely to metastasize. The input data sets, and their corresponding input vectors, can be the same - e.g., each set of input data has the same biomarkers, same cancer / s, same primary and / or secondary tumor locations, or any desired combination. Nonetheless, given the different training methods used to train each respective machine learning model 370-0, 370-1, 370-x may generate different outputs 272-0, 272-1, 272-x, respectively, based on each machine learning model 370-0, 370-1, 370-x processing the input vector 360-0, 361-1, 361-x, as shown in FIG. ID.

[0109] In some embodiments, one or more of the machine learning models 370-0, 370-1, 370- x can be a different type of machine learning model that has been trained, or otherwise configured, to classify input data as representing a cancer in a subject that is likely to metastasize or not. For example, the first machine learning model 370-0 can include a neural network, the machine learning model 370-1 can include a gradient boosted tree classification algorithm, and the machine learning model 370-x can include a K-nearest neighbor algorithm. In this example, each of these different types of machine learning models 370-0, 370-1, 370-x can be trained, or otherwise configured, to receive and process an input vector and determine whether the input vector is associated with a cancer in a subject that is likely to metastasize or not also associated with the input vector. Consider an example wherein the input data sets, and their corresponding input vectors, are the same - e.g., each set of input data has the same biomarkers, same cancer / s, same primary and / or secondary tumor locations, or any desired combination. Accordingly, the machine learning model 370-0 can be a neural network trained to process input vector 360-0 andgenerate output data 272-0 indicating whether the cancer associated with the input vector 360-0 is likely to metastasize. In addition, the machine learning model 370-1 can be a gradient boosted tree classification algorithm trained to process input vector 360-1, which for purposes of this example is the same as input vector 360-0, and generate output data 272-1 whether the cancer associated with the input vector 360-0 is likely to metastasize. This method of input vector analysis can continue for each of the x inputs, x input vectors, and x machine learning models. Continuing with this example with reference to FIG. ID, the machine learning model 370-x can be a K-nearest neighbor algorithm trained to process input vector 360-x, which for purposes of this example is the same as input vector 360-0 and 360-1, and generate output data 272-x indicating whether the cancer associated with the input vector 360-0 is likely to metastasize.

[0110] Alternatively, each of the machine learning models 370-0, 370-1, 370-x can be the same type of machine learning models or different type of machine learning models that are each configured to receive different inputs. For example, the input to the first machine learning model 370-0 can include a vector 360-0 that includes data representing a first subset or first panel of genes or gene products of a subject and then predict, based on the machine learning models 370- 0 processing of vector 360-0 whether the cancer associated with the input vector 360-0 is likely to metastasize. In addition, in this example, an input to the second machine learning model 370-1 can include a vector 360-1 that includes data representing a second subset or second panel of genes or gene products of a subject that is different than the first subset or first panel of genes or gene products. Then, the second machine learning model can generate second output data 272-1 that is indicative of whether the cancer associated with the input vector 360-1 is likely to metastasize. This method of input vector analysis can continue for each of the x inputs, x input vectors, and x machine learning models. The input to the xth machine learning model 370-x can include a vector 360-x that includes data representing an xth subset or xth panel of genes of a subject that is different than (i) at least one, (ii) two or more, or (iii) each of the other x-1 input data vectors 360-0 to 360-x-l. In some implementations, at least one of the x input data vectors can include data representing a complete set of genes from a subject. Then, the xth machine learning model 370-x can generate second output data 272-x, the second output data 272-x being indicative of whether the cancer associated with the input vector 360-x is likely to metastasize or not.

[0111] In another alternative, each of the machine learning models 370-0, 370-1 , 370-x can be the same type of machine learning models or different type of machine learning models that are each configured to receive different inputs generated from substantially similar underlying data that is processed in different manners. For example, the input to the first machine learning model 370-0 can include a vector 360-0 that includes data representing a first transformation or normalization of genes or gene products of a subject and then predict, based on the machine learning models 370-0 processing of vector 360-0 whether the cancer associated with the input vector 360-0 is likely to metastasize. In addition, in this example, an input to the second machine learning model 370-1 can include a vector 360-1 that includes data representing a second transformation or normalization of the same (or substantially same) genes or gene products of the subject. Then, the second machine learning model can generate second output data 272-1 that is indicative of whether the cancer associated with the input vector 360-1 is likely to metastasize. This method of input vector analysis can continue for each of the x inputs, x input vectors, and x machine learning models. The input to the xth machine learning model 370-x can include a vector 360-x that includes data representing an xth subset or xth transformation or normalization of genes or gene products of a subject that is different than (i) at least one, (ii) two or more, or (iii) each of the other x-1 input data vectors 360-0 to 360-x-l. In some implementations, the x input data vectors can include data representing a substantially complete set of genes from a subject. Then, the xth machine learning model 370-x can generate second output data 272-x, the second output data 272-x being indicative of whether the cancer associated with the input vector 360-x is likely to metastasize or not. As a non-limiting example, 360-0 to 360-x-l can include TPM, TMM, and GSEA (e.g., as separate vectors) each derived from a same set of RNA transcript data. See, e.g., FIG. 2A and related text. In such an example, machine learning models 370-1 to 370-3 can operate on the separate vectors. As examples, such operation can provide an output classification or provide reduced feature vectors, e.g., where voting unit 480 can be replaced by a model that operates on the reduced feature vectors as opposed to classifications that contribute to a vote. Thus, in some embodiments, voting unit 480 can be another machine learning model.

[0112] Multiple implementations of system 400 described above are not intended to be limiting, and instead, are merely examples of configurations of the multiple machine learning models 370-0, 370-1, 370-x, and their respective inputs, that can be employed using the presentdisclosure. With reference to these examples, the subject can be any human, non-human animal, plant, or other subject. As described above, the input feature vectors can be generated, based on the input data, and represent the input data. Accordingly, each input vector can represent data that includes one or more biomarkers, one or more cancers, one or more primary and / or secondary tumor locations, and a metastatic potential of the cancer in the subject having the biomarkers.

[0113] In the implementation of FIG. ID, the output data 272-0, 272-1, 272-x can be analyzed using a voting unit 480. For example, the output data 272-0, 272-1, 272-x can be input into the voting unit 480. In some implementations, the output data 272-0, 272-1, 272-x can be data indicating whether a cancer in a subject associated with the input vector processed by the machine learning model is likely to metastasize or not. Data indicating whether the subject associated with the input vector, and generated by each machine learning model, can include a “0” or a “1.” A “0,” produced by a machine learning model 370-0 based on the machine learning model’s 370-0 processing of an input vector 360-0, can indicate that the cancer in the subject associated with the input vector 360-0 is likely to have a low metastatic potential. Similarly, a “1,” produced by a machine learning model 370-0 based on the machine learning model’s 370-0 processing of an input vector 360-0, can indicate that the cancer in the subject associated with the input vector 360-0 is likely to have a high metastatic potential. Though this example uses “0” low potential and “1” as high potential, the present disclosure is not so limited. Instead, any value can be generated as output data to represent the “low potential” and “high potential” classes. For example, in some implementations “1” can be used to represent the “low potential” class and “0” to represent the “high potential” class. In yet other implementations, the output data 272-0, 272-1, 272-x can include probabilities that indicate a likelihood that the cancer in the subject associated with an input vector processed by a machine learning model is associated with a “low potential” or “high potential” class. In such implementations, for example, the generated probability can be applied to a threshold, and if the threshold is satisfied, then the subject associated with an input vector processed by the machine learning model can be determined to be in a “high potential” class.

[0114] The voting unit 480 can evaluate the received output data 270-0, 272-1, 272-x and determine whether the cancer in a subject associated with the processed input vectors 360-0,360-1 , 360-x is likely to metastasize or not. The voting unit 480 can then determine, based on the set of received output data 270-0, 272-1, 272-x, whether the cancer in the subject associated with input vectors 360-0, 360-1, 360-x is likely to metastasize. In some implementations, the voting unit 480 can apply a “majority rule.” Applying a majority rule, the voting unit 480 can tally the outputs 272-0, 272-1, and 272-x indicating that the cancer is likely to metastasize and outputs 272-0, 272-1, 272-x indicating that the cancer is not likely to metastasize. Then, the class - e.g., likely to metastasize or not likely to metastasize - having the majority predictions or votes is selected as the appropriate classification for the cancer in the subject associated with the input vector 360-0, 360-1, 360-x. This selected class can be referred to as an actual class of the entity, with each of the predictions or votes output by the machine learning models 370-0, 370-1, 370-x being referred to as initial entity classes.

[0115] Accordingly, in some implementations, determining a majority of predictions or votes can be achieved by the voting unit 480 tallying the number of occurrences of predictions or votes for each initial entity class. For example, the system 400 can determine a number of times each initial entity class is predicted or voted for by the machine learning models 370-0, 370-1, 370-x and then select the entity class that is associated with the highest number of occurrences of predictions or votes.

[0116] In some implementations, the voting unit 480 can complete a more nuanced analysis, e.g., by being another machine learning model. For example, in some implementations, the voting unit 480 can store a confidence score for each machine learning model 370-0, 370-1, 370- x. This confidence score, for each machine learning model 370-0, 370-1, 370-x, can be initially set to a default value such as 0, 1, or the like. Then, with each round of processing of input vectors, the voting unit 480, or other module of the application server 240, can adjust the confidence score for the machine learning model 370-0, 370-1, 370-x based on whether the machine learning model accurately predicted the subject classification selected by the voting unit 480 during a previous iteration. Accordingly, the stored confidence score, for each machine learning model, can provide an indication of the historical accuracy for each machine learning model.

[0117] In the more nuanced approached, the voting unit 480 can adjust output data 272-0, 272- 0, 272-x produced by each machine learning model 370-0, 370-1, 370-x, respectively, based onthe confidence score calculated for the machine learning model. Accordingly, a confidence score indicating that a machine learning mode is historically accurate can be used to boost a value of output data generated by the machine learning model. Similarly, a confidence score indicating that a machine learning model is historically inaccurate can be used to reduce a value of output data generated by the machine learning model. Such boosting or reducing of the value of output data generated by a machine learning model can be achieved, for example, by using the confidence score as a multiplier of less than one for reduction and more than 1 for boosting. Other operations can also be used to adjust the value of output data such as subtracting a confidence score from the value of the output data to reduce the value of the output data or adding the confidence score to the value of the output data to boost the value of the output data. Use of confidence scores to boost or reduce the value of output data generated by the machine learning models is particularly useful when the machine learning models are configured to output probabilities that will be applied to one or more thresholds to determine whether a cancer in a subject is more likely or less likely to metastasize. This is because using the confidence score to adjust the output of a machine learning model can be used to move a generated output value above or below a class threshold, thereby altering a prediction by a machine learning model based on its historical accuracy.

[0118] Use of the voting unit 480 to evaluate outputs of multiple machine learning models can lead to greater accuracy in prediction of the metastatic potential of a cancer having a particular set of subject biomarkers, as the consensus amongst multiple machine learning models can be evaluated instead of the output of only a single machine learning model.II. DETERMINING METASTASES

[0119] The present disclosure describes techniques to determine a risk that an identified cancer in a subject will metastasize. Such a determination can be made by assessing biomarkers, e.g., genomic DNA or mRNA transcripts, of a biopsy of a tumor. The tumor may be treated but there is a risk that the tumor already has or will metastasize. Embodiments can use transcriptome data to determine a risk that the cancer in the subject will metastasize. Such transcriptome data regarding the expression of genes can be measured using sequencing (e.g., NGS) of RNA (sometimes referred to in the art as RNA-Seq; see e.g., Kukurba KR, Montgomery SB. RNASequencing and Analysis. Cold Spring Harb Protoc. 2015 Apr 13;2015(1 l):951-69), e.g., as described above and elsewhere herein.A. Example workflow

[0120] FIG. 2A shows an overview of determining a risk score of metastatic potential of cancer in a subject having a particular set of biomarkers by using transcriptome data as an input according to embodiments described herein. As shown, three stages 501, 502, and 503 are performed to prepare a feature vectors, which may be combined, as the input data to an Al platform 540 (e.g., one or more machine learning models) that determines a risk score 545.

[0121] At stage 501, transcriptome data 505 is obtained. Example techniques include RNA sequencing, e.g., of at least 22,000, 21,000, 20,000, 19,000, 18,000, 17,000, 16,000, 15,000, 14,000, 13,000, 12,000, 11,000, 10,000, 9,000, 8,000, 7,000, 6,000, 5,000, 4,000, 3,000, 2,000, 1,000, 500, or 100 genes. In some implementations, whole transcriptome sequencing (WTS) of patient samples (e.g., biological samples) is used. WTS can query mRNA transcripts for all or substantially known genes.

[0122] To perform the sequencing, cells from a biopsy can be lysed to release inter-cellular materials such as DNA, RNA, protein, or organelles from a cell. The RNA can be isolated and fragments can be sequenced, e.g., using next-generation sequencing technology from Illumina, Inc. (San Diego, CA) or the like. The sequence reads can be aligned to a reference sequence to determine the genes for the RNA fragments. For each transcript being analyzed, a number of sequence reads for a gene can be counted. The count for each gene can be used to determine an expression level, which can be quantified in various ways (feature types). The transcriptome data can be transferred to one or more computing modules for processing.

[0123] In other embodiments, PCR (e.g., real-time or digital PCR) or targeted sequencing may be used. For real-time PCR, an intensity signal for a given gene can be determined, possibly with normalization of the intensity signal from a different gene. Digital PCR can count transcripts in a similar manner as sequencing. PCR could be used in a targeted low-density microarray, e.g., if a limited number of transcripts are of interest. .

[0124] At stage 502, values for one or more feature types are determined. Different modules determine different feature types. As shown, the transcriptome data can be processed using aTranscripts Per Million (TPM) module 510, a Trimmed Mean of the M-values (TMM) module 515, a Gene Set Enrichment Analysis (GSEA) module 520, or a combination thereof. For each of these feature types, one value can be determined per gene.

[0125] For TPM, for each gene, the number of RNA fragments that overlap the gene is counted, and then divided by total number of fragments, possibly scaled by one million to get TPM. Aany other scaling can be used. See, e.g., Zhao Y, et al. TPM, FPKM, or Normalized Counts? A Comparative Study of Quantification Measures for the Analysis of RNA-seq Data from the NCI Patient-Derived Models Repository. J Transl Med. 2021 Jun 22;19(1):269.

[0126] For TMM, the quantity can be determined in a similar manner as TPM. For a test sample, for each gene, an initial expression level (e.g., count of transcripts) can be determined. Genes that are outliers can be removed (e.g., greater than a specified number of standard deviations or a specified amount, e.g., double, triple, etc., compared to a reference or average for the sample). Genes with a low coverage (e.g., below a threshold number of transcripts) can also be excluded. A mean (sample mean) can then be determined from the expression levels of the remaining genes. A reference sample can be analyzed in a similar manner to obtain a sample mean, e.g., using the same set of genes that remained for the test sample after the exclusions. For each sample, the ratio of sample mean to the reference mean can be calculated and used as a normalization factor (norm factor). The quantity for the expression level determined for each gene can be scaled (e.g., multiplied or divided depending on how norm factor is determined) by the norm factor for that sample. See, e.g., Zhao Y, et al. (2021).

[0127] As examples, the reference sample can be from a subject without cancer or with cancer, such as prostate cancer or other cancer that may not be one in the training set. In other embodiments, a synthetic reference can be constructed, where the synthetic reference could have some measured expression for all genes, thereby ensuring all genes can be normalized.

[0128] Other normalization can be performed in addition or instead, e.g., normalizing by a total number of transcripts measured. As another example, a percentage of transcripts for a given gene in the test sample can be divided by the percentage of the transcripts for the gene for the reference sample.

[0129] For GSEA, the initial count can be determined (e.g., by either of the two techniques described above), but then the counts of the genes are ranked. GSEA can determine whether genes that are in a group have similar expression ranks. See, e.g., Subramanian A, et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc Natl Acad Sci U S A. 2005 Oct 25; 102(43): 15545-50.

[0130] In various embodiments, one, two, or all three of these feature types are determined. Additional features may also be used. For each feature type, a respective feature vector can be determined, e.g., as a multi-dimensional data point. Each feature vector can be of length N, e.g., the number of genes after any exclusions. N can be the same or different for each feature type.

[0131] At stage 503, a dimensionality reduction can be performed, e.g., via one or more dimensional reduction models, such as uniform manifold approximation and projection (UMAP) or principal component analysis. The multi-dimensional data points determined from the transcriptome data can be clustered, e.g., using unsupervised learning, supervised learning, or semi-supervised learning. For supervised aspects, the training data can include a cluster label that indicates which samples are clustered together. In this manner, one can ensure that the dimensional reduction transformation keeps points labeled the same within a same cluster, e.g., keeps them together. After transformation of a new sample, it can be determined if the new sample belongs to one of the labeled clusters. As examples, samples within a same cluster can be identified as (1) subjects with similar cancer characteristics, such as same primary cancer that metastasized in the same secondary site and / or (2) subjects with similar timing characteristics, such as within a same time range before metastasis occurred or was detected.

[0132] Each module can provide the processed transcriptome data to a respective dimensionality reduction module configured to use Al feature generation algorithms to reduce dimensionality of the processed transcriptome data. As shown, a feature vector from each module can be fed into a respective dimensionality reduction model. For example, a TPM feature vector can be fed into TPM dimensionality reduction model 525; a TMM feature vector can be fed into a TMM dimensionality reduction model 530, and a GSEA feature vector can be fed into GSEA dimensionality reduction model 535. As part of their creation or operation, the dimensionality reduction models can identify underlying clusters that can serve as inputs into secondary AI / ML models to predict metastatic potential in subjects with cancer. For example, adimensionality reduction model can receive the processed transcriptome data from a respective module and perform dimensionality reduction to determine coordinates of clusters to provide as input to Al platform 540.

[0133] When the training is performed, the dimensionality reductions (e g., transformations thereof) can be determined. These transformation parameters can be determined independently or as part of the optimization of the machine learning model. Then for a new test sample, the transformation learned when the dimensional reduction model was trained can be used to identify similar patients. Labels of the cluster can be used in the Al platform in addition to the reduced feature vector of M values.

[0134] The clusters can be used to determine optimal dimensionality reduction. The distances between the clusters and shapes of the clusters can be preserved while transforming (projecting) the data points onto a linear combination of actual feature dimensions or of fictitious dimensions. The number of reduced dimensions (M) can be a hyperparameter specified at time of training or allowed to vary to provide optimal accuracy for classifying the samples in a training set as determined using the predicted risk score and a ground truth for the risk score of the training set. Additional hyperparameters can include the number of neighbors to create a cluster and the minimum distance between samples in a cluster.

[0135] Accordingly, the dimensionality reduction can determine clusters of samples and then transform the N values into M values of a reduced feature vector, where M is smaller than N. As examples, N can be at least 50,000, 40,000, 30,000, 22,000, 21,000, 20,000, 19,000, 18,000, 17,000, 16,000, 15,000, 14,000, 13,000, 12,000, 11,000, 10,000, 9,000, 8,000, 7,000, 6,000, 5,000, 4,000, 3,000, 2,000, 1,000, 500, or 100. As examples, M can be at least or less than 40,000, 30,000, 22,000, 21,000, 20,000, 19,000, 18,000, 17,000, 16,000, 15,000, 14,000, 13,000, 12,000, 11,000, 10,000, 9,000, 8,000, 7,000, 6,000, 5,000, 4,000, 3,000, 2,000, 1,000, 500, 100, 50, 40, 30, 20, 10, 9, 8, 7, 6, 5, 4, or 3. The value of M for each feature type can be the same or different.

[0136] Al platform 540 (also referred to as a risk model or just an ML model) can receive the reduced feature vectors. In addition, one or more parameters (e.g., coordinates) of the clusters determined in models 525, 530, and 535 can be input to Al platform 540 to determine a riskscore associated with metastatic potential. As examples, the risk score can be raw values or a ranking, e.g., in one of K levels, such as 7 levels / categories.

[0137] After dimensionality reduction provides reduced feature vectors, the transcriptome data (e.g., from two or more reduced feature vectors, such as three as shown) can combined into a single final feature vector. For instance, assume that values for 50,000 genes is reduced to 1,000 dimensions for each of TPM, TMM, and GSEA. Then, the feature vector would have 3,000 values input to the risk model of Al platform 540.

[0138] Various Al platforms can be used, e.g., model types or architectures mentioned herein. The parameters of the ML model can be trained using an optimization that minimizes a cost function that includes a difference between the predicted category and the actual category, with the error being larger if the difference in categories is larger.

[0139] In some embodiments, Al platform 540 can be a multi-layer deliberation analytic Al platform comprised of decision trees, e.g., using random forest or boosting, such as XGBoost. See, e g., Chen, Tianqi; Guestrin, Carlos (2016). "XGBoost: A Scalable Tree Boosting System". In Krishnapuram, Balaji; Shah, Mohak; Smola, Alexander J.; Aggarwal, Charu C.; Shen, Dou; Rastogi, Rajeev (eds.). Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016. ACM. pp. 785-794. Boosting can increase the chances of features specific to a sample that is initially misclassified being used in subsequent trees to obtain a correct classification. Within the decision trees, each dimension of the final feature vector (or a selected subset based on importance) can be fed into a decision tree model, where difference nodes / leaves of the tree can determine whether the value for that dimension is above or below a cutoff.

[0140] A risk score 545 generated by the Al platform can be used to help determine a course of treatment for a subject, such as more aggressive treatment for a high risk of metastasis, and vice versa. Accordingly, the risk score may assist a treating physician to determine one or more therapy selections or suggest a treatment regimen for a cancer patient. The one or more therapy selections can include therapies selected using one or more other ML models using the same underlying data or other data or selected using other techniques.

[0141] As examples, a new sample can be transformed as part of the dimensional reduction into three dimensional coordinates, where the 3D coordinates can be extracted. The transformation (projection) puts similar things together, even to the point of taking 20,000 features and reducing to three dimensions. The coordinates become the features of the feature vector. Then supervised learning algorithms (e.g., XGBoost) can take the coordinates output from the transformers instead of hugely dimensional expression data and use the coordinates in the learning and inference steps.B. Examples and results

[0142] Brain metastases occur in multiple cancer types with higher prevalence in lung, breast, melanoma, and gastrointestinal cancers. See, e.g., Abdulkarim et al., Increased risk of locoregional recurrence for women with T1-2N0 triple-negative breast cancer treated with modified radical mastectomy without adjuvant radiation therapy compared with breastconserving therapy, J Clin Oncol. 2011 Jul 20;29(21):2852-8. The prognoses of patients who develop brain metastases are poor and identification of brain metastasis risk would be useful for prognostication, monitoring, and therapy selection.

[0143] Sequencing data used herein was obtained using next-generation sequencing approaches using Illumina sequencing platforms. Illumina, Inc., San Diego, CA. Whole transcriptome sequencing (WTS) was performed on RNA for over 22,000 genes. Messenger RNAs (mRNAs) were isolated from formalin-fixed paraffin-embedded (FFPE) tumor samples. Prior to sequencing, manual microdissection techniques were used to enrich tumor tissue. Matched normal tissue was not sequenced.

[0144] Sequencing data from the whole transcriptome of 220,246 tumor samples were analyzed and multiple machine learning models were trained on various molecular subtypes. The dataset was split 50% for training and the other 50% for testing. Uniform manifold approximation and projection (UMAP) was employed for dimensionality reduction. See Mclnnes, L, et al. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction, ArXiv e-prints 1802.03426, 2018. The patterns learned across the entirety of the training dataset irrespective of brain metastasis were leveraged on the testing data set.

[0145] FIGs. 2B-E respectively show representative cluster coordinates for 1000 patients, 50,000 patients, 100,000 patients, and 200,000 patients where more data from a larger patient cohort resulted in more refined cluster coordinates for GSEA. Similar patterns were observed for TPM and TMM (data not shown). The cluster coordinates were provided as input to a multi-layer Al platform trained to determine whether patients were more or less likely to develop a brain metastasis. Patients with brain metastasis were identified using the presence of ICD-10 code C79.31 (Secondary malignant neoplasm of the brain). As the absence of C79.31 could be due to the event not happening yet, patients without brain metastasis were stratified into groups based on three, four or greater than five years without a C79.31 ICD-10 code. The brain metastasis risk score was defined by empirical evaluation of the positive predictive value in seven groups of risk probabilities. The validation set contained 1,217 patients with brain metastasis and 4,631 without an observed brain metastasis within three years.

[0146] In the validation set, the prevalence of brain metastases within the seven risk score categories across all cancer types ranged from 4% with the lowest risk score to 94% with the highest risk score. Overall, 71% of cases received the lowest two risk scores, 15% received the two intermediate risk scores, and 14% received the 3 highest risk scores. For breast, lung and colon cancers, the prevalence of brain metastasis ranged from 4-10% in patients with the lowest risk scores to 92-100% in patients with the highest risk scores. The distribution of cases with each risk score was markedly different across cancer types. For example, breast cancer had 62% of cases receiving the lowest 2 risk scores versus 27% and 92% of cases receiving the lowest two risk scores in lung and colon cancer, respectively. Those cases assigned to the three highest risk scores were 18%, 42% and 2% in breast, lung and colon cancers, respectively.

[0147] FIGs. 3A-C illustrate predictions of the risk of brain metastasis on an independent set of patients. In particular, the plots of FIGs. 3A-C respectively show the relative likelihood of developing a brain metastasis with all cancers, breast cancer, and non-small cell lung cancer (NSCLC). All samples were taken at the primary tumor site. Each plot has seven different lines, each corresponding to a different risk score. The seven different categories correspond to different risk score of progressively higher likelihood (risk) of metastasis, corresponding to progressively higher metastatic potential. The horizontal axis is days until metastasis is identified. The vertical axis is the proportion of patients in the cohort that have not metastasized.Thus, category 7 is the mostly likely since it has the lowest proportion of patients that are free of metastasis. As one can see, the predictions are accurate since the higher risk scores correspond to higher proportion of patients with metastasis.

[0148] In FIG. 3A (all cancers), line 805 corresponds to a first group with 2714 subjects, line 810 corresponds to a second group with 179 subjects, line 815 corresponds to a third group with 215 subjects, line 820 corresponds to a fourth group with 171 subjects, line 825 corresponds to a fifth group with 107 subjects, line 830 corresponds to a sixth group with 65 subjects, and line 835 corresponds to a seventh group with 43 subjects. As shown in FIG. 3A, the median time to brain metastasis development for groups 1-5 was unevaluable while groups 6-7 were 424 and 244 days, respectively.

[0149] In FIG. 3B (breast cancer), line 905 corresponds to a first group with 263 subjects, line 910 corresponds to a second group with 25 subjects, line 915 corresponds to a third group with 44 subjects, line 920 corresponds to a fourth group with 30 subjects, line 925 corresponds to a fifth group with 17 subjects, line 930 corresponds to a sixth group with 9 subjects, and line 935 corresponds to a seventh group with 6 subjects.

[0150] In FIG. 3C (NSCLC), line 1005 corresponds to a first group with 179 subjects, line 1010 corresponds to a second group with 49 subjects, line 1015 corresponds to a third group with 87 subjects, line 1020 corresponds to a fourth group with 67 subjects, line 1025 corresponds to a fifth group with 53 subjects, line 1030 corresponds to a sixth group with 39 subjects, line 1035 corresponds to a seventh group with 25 subjects.

[0151] Whole transcriptome data can be leveraged by a machine learning platform that employs dimensionality reduction techniques along with transfer learning to predict the risk of brain metastasis. The transfer learning can be employed via the use of the output of dimensional reduction models by an additional model that classifies the risk of metastasis. This tool can be used to augment the clinical picture of cancer patients an unmet clinical opportunity to inform prognosis, monitoring, and therapeutic selection.C. Applicability of Metastasis Prediction

[0152] The present disclosure provides the use of a machine learning approach to analyze molecular profiling data to discover clinically relevant biomarkers and biosignatures forpredicting a cancer’s metastatic potential, i.e., whether a cancer will metastasize. Herein, we trained machine learning classification models using molecular profiling data across cancer types that metastasized or not. The prediction can be a relative indication of the metastatic potential, such as a likelihood, risk, probability or other metric whether a cancer is more likely (higher metastatic potential) or less likely (lower metastatic potential) to become metastatic. The relative indication may be assigned a score, such as a risk score. The strength and confidence in the prediction may be given by the model. The prediction may be considered by the treating physician when determining a treatment regimen for the subject with the cancer. For example, the treating physician may prefer a more aggressive course of treatment for a cancer that is predicted to metastasize, and vice versa. A more aggressive course of treatment is relative and may, comprise factors such as additional therapeutic agents, longer course of treatment, higher dosage, or any useful combination thereof. Advantageously, the molecular profiling provided herein can be used to both predict metastasis and determine one or more candidate treatment for a given patient. Thus, the systems and methods provided herein are efficient and improve precision medicine for cancer patients.

[0153] The biological sample from the subject which is profiled can be any useful biological sample. The biological sample may comprise a single sample, by way of non-limiting example a tumor biopsy, or multiple biological samples may be assessed, by way of non-limiting examples multiple biopsy cores and / or tumor tissue and blood. In some embodiments, the one or more biological sample comprises formalin-fixed paraffin-embedded (FFPE) tissue, fixed tissue, a core needle biopsy, a fine needle aspirate, unstained slides, fresh frozen (FF) tissue, formalin samples, tissue comprised in a solution that preserves nucleic acid or protein molecules, a fresh sample, a malignant fluid, a bodily fluid, a tumor sample, a tissue sample, or any combination thereof. In preferred embodiments, FFPE tissue is used. In some embodiments, the one or more biological sample is from a solid tumor. The solid tumor can be a primary tumor, e.g., the primary tumor whose metastatic potential is predicted. In some embodiments, the primary tumor is a tumor of the myeloid, breast, bile ducts, colon, rectum, female genital tract, stomach, esophagus, gastrointestinal stromal cells, small intestine, brain, mouth, sinuses, nose, throat, blood, liver, nervous system, lung, lymph, male genital tract, pleura, skin, plasma cells, neuroendocrine cells, B-cells, T-cells, ovary, pancreas, pituitary gland, spinal cord, prostate, peritoneum, large intestine, soft tissue, connective tissue, fat tissue, thymus, thyroid, or eye. Insome embodiments, the primary tumor is a tumor of the bladder, breast, colon, rectum, endometrium, uterus, ovary, female genital tract, kidney, blood, liver, lung, skin, lymph, pancreas, prostate, or thyroid. In some embodiments, the one or more biological sample comprises a bodily fluid. In some embodiments, the bodily fluid comprises a malignant fluid, a pleural fluid, a peritoneal fluid, or any combination thereof. In some embodiments, the bodily fluid comprises peripheral blood, sera, plasma, ascites, urine, cerebrospinal fluid (CSF), sputum, saliva, bone marrow, synovial fluid, aqueous humor, amniotic fluid, cerumen, breast milk, broncheoalveolar lavage fluid, semen, prostatic fluid, cowper’s fluid, pre-ejaculatory fluid, female ejaculate, sweat, fecal matter, tears, cyst fluid, pleural fluid, peritoneal fluid, pericardial fluid, lymph, chyme, chyle, bile, interstitial fluid, menses, pus, sebum, vomit, vaginal secretions, mucosal secretion, stool water, pancreatic juice, lavage fluids from sinus cavities, bronchopulmonary aspirates, blastocyst cavity fluid, or umbilical cord blood. Nucleic acids are shed from tumor cells into the circulation. Such cell-free nucleic acids may be used in the systems and methods provided herein.

[0154] The biomarker features used to build the predictive models provided herein can be any useful set of biomarkers assayed using any useful technology. For example, the set of features extracted from the obtained molecular data according to the systems and methods provided herein can comprise a presence, level, or state of an RNA transcript for each member of the plurality of biomarkers that are assayed. Numerous techniques to assess nucleic acids, and other biological entities (e.g., proteins, lipids, carbohydrates, complexes) are described herein or known in the art. The RNA nucleic acids can comprise cell free nucleic acid. The nucleic acid can consist of cell free nucleic acid.

[0155] In some embodiments, the set of features extracted from the obtained molecular data comprises additional information in addition to the biomarker assay results. For example, the features may comprise one or more clinical characteristic of the first subject, including that of the cancer. In some embodiments, such characteristics comprise the subject’s age, gender, race, year of birth, cancer stage, histology, anatomical location / s, medical history, and / or history of surgeries and any other prior treatments (including without limitation any immunotherapy and / or chemotherapy). In some embodiments, the additional information comprises the primary tumor location, one or more secondary tumor location, and any useful combination thereof. Suchinformation may be used to refine the prediction models if so desired. As a non-limiting example, the information could include one or more primary tumor location, which could allow the prediction models to be targeted to predict metastatic potential of such one or more primary tumor location. As another non-limiting example, the information could include one or more secondary tumor location, which could allow the prediction models to be targeted to predict metastasis to such one or more secondary tumor location. As still another non-limiting example, the information could include one or more primary tumor location and one or more secondary tumor location, which could allow the prediction models to be targeted to predict metastasis from the one or more primary tumor location to the one or more secondary tumor location.

[0156] As described herein, the metastatic prediction models can be trained, tested and employed to predict metastasis of any cancer of interest. In some embodiments, the cancer comprises an acute lymphoblastic leukemia; acute myeloid leukemia; adrenocortical carcinoma; AIDS-related cancer; AIDS-related lymphoma; anal cancer; appendix cancer; astrocytomas; atypical teratoid / rhabdoid tumor; basal cell carcinoma; bladder cancer; brain stem glioma; brain tumor, brain stem glioma, central nervous system atypical teratoid / rhabdoid tumor, central nervous system embryonal tumors, astrocytomas, craniopharyngioma, ependymoblastoma, ependymoma, medulloblastoma, medulloepithelioma, pineal parenchymal tumors of intermediate differentiation, supratentorial primitive neuroectodermal tumors and pineoblastoma; breast cancer; bronchial tumors; Burkitt lymphoma; cancer of unknown primary site (CUP); carcinoid tumor; carcinoma of unknown primary site; central nervous system atypical teratoid / rhabdoid tumor; central nervous system embryonal tumors; cervical cancer; childhood cancers; chordoma; chronic lymphocytic leukemia; chronic myelogenous leukemia; chronic myeloproliferative disorders; colon cancer; colorectal cancer; craniopharyngioma; cutaneous T-cell lymphoma; endocrine pancreas islet cell tumors; endometrial cancer; ependymoblastoma; ependymoma; esophageal cancer; esthesioneuroblastoma; Ewing sarcoma; extracranial germ cell tumor; extragonadal germ cell tumor; extrahepatic bile duct cancer; gallbladder cancer; gastric (stomach) cancer; gastrointestinal carcinoid tumor; gastrointestinal stromal cell tumor; gastrointestinal stromal tumor (GIST); gestational trophoblastic tumor; glioma; hairy cell leukemia; head and neck cancer; heart cancer; Hodgkin lymphoma; hypopharyngeal cancer; intraocular melanoma; islet cell tumors; Kaposi sarcoma; kidney cancer; Langerhans cell histiocytosis; laryngeal cancer; lip cancer; liver cancer; malignant fibrous histiocytoma bonecancer; medulloblastoma; medulloepithelioma; melanoma; Merkel cell carcinoma; Merkel cell skin carcinoma; mesothelioma; metastatic squamous neck cancer with occult primary; mouth cancer; multiple endocrine neoplasia syndromes; multiple myeloma; multiple myeloma / plasma cell neoplasm; mycosis fungoides; myelodysplastic syndromes; myeloproliferative neoplasms; nasal cavity cancer; nasopharyngeal cancer; neuroblastoma; Non-Hodgkin lymphoma; nonmelanoma skin cancer; non-small cell lung cancer; oral cancer; oral cavity cancer; oropharyngeal cancer; osteosarcoma; other brain and spinal cord tumors; ovarian cancer; ovarian epithelial cancer; ovarian germ cell tumor; ovarian low malignant potential tumor; pancreatic cancer; papillomatosis; paranasal sinus cancer; parathyroid cancer; pelvic cancer; penile cancer; pharyngeal cancer; pineal parenchymal tumors of intermediate differentiation; pineoblastoma; pituitary tumor; plasma cell neoplasm / multiple myeloma; pleuropulmonary blastoma; primary central nervous system (CNS) lymphoma; primary hepatocellular liver cancer; prostate cancer; rectal cancer; renal cancer; renal cell (kidney) cancer; renal cell cancer; respiratory tract cancer; retinoblastoma; rhabdomyosarcoma; salivary gland cancer; Sezary syndrome; small cell lung cancer; small intestine cancer; soft tissue sarcoma; squamous cell carcinoma; squamous neck cancer; stomach (gastric) cancer; supratentorial primitive neuroectodermal tumors; T-cell lymphoma; testicular cancer; throat cancer; thymic carcinoma; thymoma; thyroid cancer; transitional cell cancer; transitional cell cancer of the renal pelvis and ureter; trophoblastic tumor; ureter cancer; urethral cancer; uterine cancer; uterine sarcoma; vaginal cancer; vulvar cancer; Waldenstrom macroglobulinemia; or Wilm’s tumor. In some embodiments, the cancer comprises an acute myeloid leukemia (AML), breast carcinoma, cholangiocarcinoma, colorectal adenocarcinoma, extrahepatic bile duct adenocarcinoma, female genital tract malignancy, gastric adenocarcinoma, gastroesophageal adenocarcinoma, gastrointestinal stromal tumor (GIST), glioblastoma, head and neck squamous carcinoma, leukemia, liver hepatocellular carcinoma, low grade glioma, lung bronchioloalveolar carcinoma (BAC), non-small cell lung cancer (NSCLC), lung small cell cancer (SCLC), lymphoma, male genital tract malignancy, malignant solitary fibrous tumor of the pleura (MSFT), melanoma, multiple myeloma, neuroendocrine tumor, nodal diffuse large B-cell lymphoma, non-epithelial ovarian cancer (non-EOC), ovarian surface epithelial carcinoma, pancreatic adenocarcinoma, pituitary carcinomas, oligodendroglioma, prostatic adenocarcinoma, retroperitoneal or peritoneal carcinoma, retroperitoneal or peritoneal sarcoma, small intestinal malignancy, soft tissue tumor, thymic carcinoma, thyroid carcinoma, oruveal melanoma. Tn some embodiments, the cancer comprises a breast carcinoma, colorectal adenocarcinoma, female genital tract malignancy, kidney cancer, non-small cell lung cancer (NSCLC), lung small cell cancer (SCLC), melanoma, ovarian surface epithelial carcinomas, prostatic adenocarcinoma, uterine neoplasm, endometrial carcinoma, or unknown. In some embodiments, the cancer comprises a breast cancer. The breast cancer can comprise a HER2+ breast cancer. The systems and methods herein can be used to characterize or predict metastatic potential of these and other cancers. Characterizing a phenotype can be providing a diagnosis, prognosis or theranosis of the cancer.

[0157] In some embodiments, the system is further configured to determine that the cancer in the first or second subject has indeterminate likelihood of metastasis, optionally wherein indeterminate likelihood is based on one or more statistical threshold. For example, such threshold may be set such that the metastatic potential is considered indeterminate if the prediction of metastatic potential is not very strong in either direction. This may correspond to an intermediate risk score. In some embodiments, the threshold is set such that the metastatic potential is considered indeterminate if the prediction of metastasis or no metastasis are equally likely within a desired confidence interval.

[0158] In some embodiments, the metastasis comprises secondary tumors in at least one of the lymph nodes, adrenal gland, bone, brain, liver, lung, muscle, peritoneum, skin, and vagina. In some embodiments, the metastasis comprises brain metastasis. The metastasis may consist of brain metastasis.

[0159] In some embodiments, the system further comprises operations that identify, based on profiling data obtained from assaying the one or more biological sample from the first subject: (a) one or more treatment of likely benefit for treating the cancer in the subject; (b) one or more treatment of likely lack of benefit for treating the cancer in the subject; (c) one or more treatment of likely lack of benefit for treating the cancer in the subject; and / or (d) one or more clinical trial for which the subject is indicated as eligible. In some embodiments, the profiling data comprises the molecular data, e.g., WTS data. The profiling data can consist of the molecular data. Various systems and methods for molecular profiling in order to select treatment options are provided herein or described in any one of International Patent Publications WO / 2007 / 137187 (IntT Appl. No. PCT / US2007 / 069286), published November 29, 2007; WO / 2010 / 045318 (IntT Appl. No.PCT / US2009 / 060630), published April 22, 2010; WO / 2010 / 093465 (Int’l Appl. No. PCT / US2010 / 000407), published August 19, 2010; WO / 2012 / 170715 (Int’l Appl. No. PCT / US2012 / 041393), published December 13, 2012; WO / 2014 / 089241 (Int’l Appl. No. PCT / US2013 / 073184), published June 12, 2014; WO / 2011 / 056688 (Int’l Appl. No. PCT / US2010 / 054366), published May 12, 2011; WO / 2012 / 092336 (Int’l Appl. No.PCT / US2011 / 067527), published July 5, 2012; WO / 2015 / 116868 (Int’l Appl. No. PCT / US2015 / 013618), published August 6, 2015; WO / 2017 / 053915 (Int’l Appl. No. PCT / US2016 / 053614), published March 30, 2017; WO / 2016 / 141169 (Int’l Appl. No. PCT / US2016 / 020657), published September 9, 2016; and W02018175501 (Int’l Appl. No. PCT / US2018 / 023438), published September 27, 2018; each of which publications is incorporated by reference herein in its entirety. Additional systems and methods for molecular profiling that can be used are described in Int’l Patent Appl. No. PCT / US2020 / 012815, filed January 8, 2020; Int’l Patent Appl. No. PCT7US2021 / 018263, filed February 16, 2021; Int’l Patent Appl. No. PCT / US2019 / 064078, filed December 2, 2019; Infl Patent Appl. No.PCT / US2020 / 035990, filed June 3, 2020; Int’l Patent Appl. No. PCT / US2021 / 030351, filed April 30, 2021; each of which applications is incorporated by reference herein in its entirety.

[0160] Although preferred embodiments of the present invention have been shown and described herein, it will be obvious to those skilled in the art that such embodiments are provided by way of example only. Numerous variations, changes, and substitutions will occur to those skilled in the art without departing from the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in practicing the invention. It is intended that the claims below define the scope of the invention and that methods and structures within the scope of these claims and their equivalents be covered thereby.III. ADDITIONAL OPERATIONS

[0161] In addition to a determination of risk of metastasis, various embodiments can perform additional measurements and / or assessments (e.g., additional molecular profiling) or responsive actions (e.g., treatment selection, treatment, or screening).A. Molecular Profiling

[0162] Molecular profiling of disease samples is aimed at providing clinical benefit for individuals, such as identifying therapeutic regimens that provide a longer progression free survival (PFS), longer disease free survival (DFS), longer overall survival (OS) or extended lifespan. A molecular profiling approach can provide a method for selecting treatments for an individual that could favorably change the clinical course of a medical condition, including without limitation cancer. Molecular profiling can provide a personalized approach to selecting treatments that are more likely to benefit a cancer. The molecular profiling methods described herein can be used to guide treatment in any desired setting, including without limitation the front-line / standard of care setting, or for patients with poor prognosis, such as those with metastatic disease or those whose cancer has progressed on standard front line therapies, or whose cancer has progressed on previous chemotherapeutic or hormonal regimens. The cancer can be a metastatic cancer or other recurrent cancer. The treatments can be on-compendium or off-compendium treatments. The treatments may be standard of care for the type of cancer in the individual, or the treatments may be typically used for other types of cancer. Thus, profiling may expand the choice of treatments for the individual.

[0163] The systems and methods provided herein may be used to classify patients as more or less likely to benefit or respond to various treatments. Unless otherwise noted, the terms “response” or “non-response,” as used herein, refer to any appropriate indication that a treatment provides a benefit to a patient (a “responder” or “benefiter”) or has a lack of benefit to the patient (a “non-responder” or “non-benefiter”). Such an indication may be determined using accepted clinical response criteria such as the standard Response Evaluation Criteria in Solid Tumors (RECIST) criteria, or other useful patient response criteria such as progression free survival (PFS), time to progression (TTP), disease free survival (DFS), time-to-next treatment (TNT, TTNT), tumor shrinkage or disappearance, or the like. RECIST is a set of rules published by an international consortium that define when tumors improve (“respond”), stay the same (“stabilize”), or worsen (“progress”) during treatment of a cancer patient. As used herein and unless otherwise noted, a patient “benefit” from a treatment may refer to any appropriate measure of improvement, including without limitation a RECIST response or longer PFS / TTP / DFS / TNT / TTNT, whereas “lack of benefit” from a treatment may refer to anyappropriate measure of worsening disease during treatment. Generally disease stabilization is considered a benefit, although in certain circumstances, if noted herein, stabilization may be considered a lack of benefit. A predicted or indicated benefit may be described as “indeterminate” if there is not an acceptable level of prediction of benefit or lack of benefit. In some cases, benefit is considered indeterminate if it cannot be calculated, e.g., due to lack of necessary data.

[0164] Molecular profiling can be used to provide a comprehensive view of the biological state of a sample. In some embodiments, molecular profiling is used for tumor profiling. Accordingly, a number of molecular approaches are used to assess the state of a tumor. The tumor profiling can be used for selecting a candidate treatment for the patient. Molecular profiling can be used to select candidate therapeutics on any sample for any stage of a disease, such as a newly diagnosed cancer or a cancer that has already been treated, e.g., with one or more standard-of-care therapy. In some embodiments, the cancer is refractory to the prior treatment / s. For example, the cancer may be refractory to the standard of care treatments for the cancer. The cancer can be a metastatic cancer or other recurrent cancer. The treatments can be on-compendium or off- compendium treatments.

[0165] Molecular profiling can be performed by any known means for detecting a molecule in a biological sample. Molecular profiling comprises methods that include but are not limited to, nucleic acid sequencing, such as a DNA sequencing or RNA sequencing; immunohistochemistry (IHC); in situ hybridization (ISH); fluorescent in situ hybridization (FISH); chromogenic in situ hybridization (CISH); PCR amplification (e.g., qPCR or RT-PCR); various types of microarray (mRNA expression arrays, low density arrays, protein arrays, etc); various types of sequencing (Sanger, pyrosequencing, etc); comparative genomic hybridization (CGH); high throughput or next generation sequencing (NGS); Northern blot; Southern blot; immunoassay; and any other appropriate technique to assay the presence or quantity of a biological molecule of interest. In various embodiments, any one or more of these methods can be used concurrently or subsequent to each other for assessing target genes disclosed herein.

[0166] The molecular profiling provided by the disclosure is not limited to identifying candidate treatments for patients in need thereof. Indeed, the data derived from molecular profiling can be used to characterize various phenotypes of interest. For example, providedherein are systems and methods for predicting whether a primary tumor is likely to metastasize. Thus, molecular profding of a primary tumor may provide both personalized treatment options for the patient and in addition provide a metastatic potential for the tumor. The treating physician may consider the predicted metastatic potential when deciding a course of treatment for the patient , e.g., by choosing more aggressive treatment for a high-risk cancer, or less aggressive treatment for low metastatic potential.

[0167] To carry out molecular profding, genes and / or gene products, e.g., genomic DNA, RNA transcripts, and proteins, is assessed using applicable including without limitation high throughput sequencing techniques such as whole exome sequencing (WES), whole transcriptome sequencing (WTS), and / or whole genome sequencing (WGS). Therefore, any of the markers disclosed herein can be assayed by a single molecular profding technique or by multiple methods disclosed herein (e.g., a single marker is profded at the DNA and RNA level). In some embodiments, at least 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75, 80, 85, 90, 95 or at least about 100 genes or gene products are profded. In some embodiments, at least about 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, 10,000, 11,000, 12,000, 13,000, 14,000, 15,000, 16,000, 17,000, 18,000, 19,000, 20,000, 21,000, 22,000, 23,000,24,000, 25,000, 26,000, 27,000, 28,000, 29,000, 30,000, 31,000, 32,000, 33,000, 34,000, 35,000,36,000, 37,000, 38,000, 39,000, 40,000, 41,000, 42,000, 43,000, 44,000, 45,000, 46,000, 47,000,48,000, 49,000, or at least 50,000 genes or gene products are profded using various techniques.The number of markers assayed can depend on the technique used. For example, next generation sequencing may sequence all known genes and corresponding transcripts in a simultaneous assay.

[0168] As understood by those of skill in the art, genes and proteins have developed a number of alternative names and symbols in the scientific literature. Listing of gene aliases and descriptions used herein can be found using a variety of online databases, including GeneCards® (genecards.org), HUGO Gene Nomenclature (HGNC; genenames.org), Entrez Gene (ncbi.nlm.nih.gov / entrez / query.fcgi?db=gene), UniProtKB / Swiss-Prot (uniprot.org), UniProtKB / TrEMBL (uniprot.org), OMIM (ncbi.nlm.nih.gov / entrez / query.fcgi?db=OMIM), GeneLoc (genecards.weizmann.ac.il / geneloc / ), and Ensembl (ensembl.org). For example, genesymbols and names used herein can correspond to those approved by HUGO, and protein names can be those recommended by UniProtKB / Swiss-Prot. In the specification, where a protein name indicates a precursor, the mature protein is also implied. Throughout the application, gene and protein symbols may be used interchangeably and the meaning can be derived from context, e.g., NGS is used to analyze nucleic acids whereas IHC is used to analyze protein.

[0169] One of skill will appreciate that such molecular profiling is not limited to those techniques disclosed herein but comprises any methodology conventional for assessing nucleic acid or protein levels, sequence information, or both. The methods as described herein can also take advantage of any improvements to current methods or new molecular profiling techniques developed in the future. In some embodiments, a gene or gene product is assessed by a single molecular profiling technique. In other embodiments, a gene and / or gene product is assessed by multiple molecular profiling techniques. In a non-limiting example, a gene sequence can be assayed by one or more of NGS, ISH and pyrosequencing analysis, the mRNA gene product can be assayed by one or more of NGS, RT-PCR and microarray, and the protein gene product can be assayed by one or more of IHC and immunoassay. One of skill will appreciate that any combination of biomarkers and molecular profiling techniques that will benefit disease treatment are contemplated by the present methods.

[0170] Genes and gene products that are known to play a role in cancer and can be assayed by any of the molecular profiling techniques as described herein include without limitation those listed in any of International Patent Publications WO / 2007 / 137187 (Int’l Appl. No. PCT / US2007 / 069286), published November 29, 2007; WG / 2010 / 045318 (Int’l Appl. No. PCT / US2009 / 060630), published April 22, 2010; WO / 2010 / 093465 (Int’l Appl. No.PCT / US2010 / 000407), published August 19, 2010; WO / 2012 / 170715 (Int’l Appl. No. PCT / US2012 / 041393), published December 13, 2012; WO / 2014 / 089241 (Int’l Appl. No. PCT / US2013 / 073184), published June 12, 2014; WO / 2011 / 056688 (Int’l Appl. No. PCT / US2010 / 054366), published May 12, 2011; WO / 2012 / 092336 (Int’l Appl. No.PCT / US2011 / 067527), published July 5, 2012; WO / 2015 / 116868 (Int’l Appl. No. PCT / US2015 / 013618), published August 6, 2015; WO / 2017 / 053915 (Int’l Appl. No. PCT / US2016 / 053614), published March 30, 2017; WO / 2016 / 141169 (Int’l Appl. No. PCT / US2016 / 020657), published September 9, 2016; and W02018175501 (Int’l Appl. No.PCT / US2018 / 023438), published September 27, 2018; each of which publications is incorporated by reference herein in its entirety.

[0171] In embodiments, the methods as described herein are used detect gene fusions. A fusion gene is a hybrid gene created by the juxtaposition of two previously separate genes. This can occur by chromosomal translocation or inversion, deletion or via trans-splicing. The resulting fusion gene can cause abnormal temporal and spatial expression of genes, leading to abnormal expression of cell growth factors, angiogenesis factors, tumor promoters or other factors contributing to the neoplastic transformation of the cell and the creation of a tumor. For example, such fusion genes can be oncogenic due to the juxtaposition of: 1) a strong promoter region of one gene next to the coding region of a cell growth factor, tumor promoter or other gene promoting oncogenesis leading to elevated gene expression, or 2) due to the fusion of coding regions of two different genes, giving rise to a chimeric gene and thus a chimeric protein with abnormal activity. Fusion genes are characteristic of many cancers. Once a therapeutic intervention is associated with a fusion, the presence of that fusion in any type of cancer identifies the therapeutic intervention as a candidate therapy for treating the cancer. A number of recurrent fusion genes have been catalogued in the Mittleman database (cgap.nci.nih.gov / Chromosomes / Mitelman).B . Further screening modalities

[0172] Based on the risk classification, e.g., the risk of metastasis determined using the systems and methods provided herein, the subject can be referred for additional screening modalities, e.g. using chest X ray, ultrasound, computed tomography, magnetic resonance imaging, or positron emission tomography. Such screening may be used to investigate the presence of secondary (metastatic) lesions and may be performed for cancer.C. Treatment selection

[0173] A molecular profiling approach can provide a method for selecting treatments for an individual that could favorably change the clinical course of a medical condition, including without limitation cancer. Molecular profiling can provide a personalized approach to selecting treatments that are more likely to benefit a cancer. The molecular profiling methods described herein can be used to guide treatment in any desired setting, including without limitation thefront-line / standard of care setting, or for patients with poor prognosis, such as those with metastatic disease or those whose cancer has progressed on standard front line therapies, or whose cancer has progressed on previous chemotherapeutic or hormonal regimens. The cancer can be a metastatic cancer or other recurrent cancer. The treatments can be on-compendium or off-compendium treatments. The treatments may be standard of care for the type of cancer in the individual, or the treatments may be typically used for other types of cancer. Thus, profiling may expand the choice of treatments for the individual.

[0174] There are some therapeutics such as tucatinib that lower the incidence of brain metastasis in breast cancers. In such cases, a high-risk score provided by the systems and methods herein might indicate to a physician to choose such a drug. Alternately, increased frequency of brain scanning could aid in earlier detection and treatment of metastatic lesions.

[0175] In some embodiments, based on a prediction output from a model, a treatment can be provided to a subject. A treatment can be selected from the group consisting of chemotherapy, radiation therapy, targeted therapy, surgical removal of tumor tissues, immunotherapy, hormone therapy, and stem cell therapy, and various combinations thereof. The prediction output may be considered by a treating physician when determining a treatment regimen for the subject with the cancer. For example, the treating physician may prefer a more aggressive course of treatment for a cancer that is predicted to metastasize, and vice versa. A more aggressive course of treatment is relative and may, comprise factors such as additional therapeutic agents, longer course of treatment, higher dosage, or any useful combination thereof.

[0176] When multiple treatment options are revealed by molecular profding, decision rules can be put in place to prioritize the selection of a treatment regimen. For example, treatments may be prioritized based on direct results of molecular profiling, anticipated efficacy of therapeutic agent, prior history with the same or other treatments, expected side effects, availability of therapeutic agent, cost of therapeutic agent, drug-drug interactions, and other factors. Based on the recommended and prioritized therapeutic agent targets, a treating physician can decide on the course of treatment for a particular individual.

[0177] Treatments associated with one or more of the biomarkers or sample / training / reference vectors may be determined using treatment association such as in any of International Patent Publications WO / 2007 / 137187 (Inf 1 Appl. No. PCT / US2007 / 069286), published November 29,2007; W 0 / 2010 / 045318 (Int’l Appl. No. PCT / US2009 / 060630), published April 22, 2010; WO / 2010 / 093465 (Int’l Appl. No. PCT / US2010 / 000407), published August 19, 2010;WO / 2012 / 170715 (Int’l Appl. No. PCT / US2012 / 041393), published December 13, 2012;WO / 2014 / 089241 (Int’l Appl. No. PCT / US2013 / 073184), published June 12, 2014; WO / 2011 / 056688 (Int’l Appl. No. PCI7US2010 / 054366), published May 12, 2011; WO / 2012 / 092336 (Int’l Appl. No. PCT / US2011 / 067527), published July 5, 2012; WO / 2015 / 116868 (Int’l Appl. No. PCT / US2015 / 013618), published August 6, 2015; WO / 2017 / 053915 (Int’l Appl. No. PCT / US2016 / 053614), published March 30, 2017; WO / 2016 / 141169 (Int’l Appl. No. PCT / US2016 / 020657), published September 9, 2016; and W02018175501 (Int’l Appl. No. PCT / US2018 / 023438), published September 27, 2018; each of which publications is incorporated by reference herein in its entirety. Such rules can be updated as new information becomes available regarding various biomarkers, biosignatures, treatments, and the relationships thereof. The indication whether each treatment is likely to benefit the patient, not benefit the patient, or has indeterminate benefit may be weighted. For example, a likely or potential benefit may be a strong potential benefit or a lesser potential benefit. Such weighting can be based on any appropriate criteria, e.g., the strength of the evidence of the biomarker-treatment association, or the results of the profiling, e.g., a degree or level of over- or underexpression, mutation, or any other relevant state (e.g., wild type or altered). As the treating physician is ultimately responsible for treating their patient, such physician may use the report to assist in guiding their treatment recommendations.D. Types of treatments

[0178] Embodiments may further include treating the pathology in the patient after determining a classification for the subject. Treatment can be provided according to a determined level of pathology, the fractional concentration of clinically-relevant DNA, or a tissue of origin. For example, an identified mutation can be targeted with a particular drug or chemotherapy. The tissue of origin can be used to guide a surgery or any other form of treatment. And the level of the pathology can be used to determine how aggressive to be with any type of treatment, which may also be determined based on the level of pathology. A pathology (e.g., cancer) may be treated by chemotherapy, drugs, diet, therapy, and / or surgery. In some embodiments, the more the value of aparameter (e.g., amount or size) exceeds the reference value, the more aggressive the treatment may be.

[0179] Treatment may include chemotherapy, which is the use of drugs to destroy cancer cells, usually by keeping the cancer cells from growing and dividing. The drugs may involve, for example but are not limited to, mitomycin-C (available as a generic drug), gemcitabine (Gemzar), and thiotepa (Tepadina) for intravesical chemotherapy. The systemic chemotherapy may involve, for example but not limited to, cisplatin gemcitabine, methotrexate (Rheumatrex, Trexall), vinblastine (Velban), doxorubicin, and cisplatin.

[0180] In some embodiments, treatment may include immunotherapy. Immunotherapy may include immune checkpoint inhibitors that block a protein called PD-1. Inhibitors may include but are not limited to atezolizumab (Tecentriq), nivolumab (Opdivo), avelumab (Bavencio), durvalumab (Imfinzi), and pembrolizumab (Keytruda).

[0181] Treatment embodiments may also include targeted therapy. Targeted therapy is a treatment that targets the cancer’s specific genes and / or proteins that contributes to cancer growth and survival. For example, erdafitinib is a small molecule inhibitor of fibroblast growth factor receptor (FGFR) and is approved to treat people with locally advanced or metastatic urothelial carcinoma with FGFR3 or FGFR2 genetic mutations that has continued to grow or spread of cancer cells. As another example, the monoclonal antibody trastuzumab works by binding to the HER2 receptor and inhibiting cell replication.

[0182] Some treatments may include radiation therapy. Radiation therapy is the use of high- energy x-rays or other particles to destroy cancer cells. In addition to each individual treatment, combinations of these treatments described herein may be used. In some embodiments, when the value of the parameter exceeds a threshold value, which itself exceeds a reference value, a combination of the treatments may be used. Information on treatments in the references are incorporated herein by reference.E. Report

[0183] In some embodiments, the methods as described herein comprise generating a molecular profile report. The report can be delivered to the treating physician or other caregiver of the subject whose cancer has been profiled. The report can comprise multiple sections ofrelevant information, including but not limited to: 1) description of the patient and sample; 2) a complete or partial listing of the biomarkers (nucleic acids, proteins, or other biological matter of interest) in the molecular profde; 3) a description of the state of one or more of the biomarkers in the molecular profile as determined for the subject; 4) a description of one or more biological signatures as determined for the molecular profile, such as microsatellite stability, tumor mutational load / burden, tissue-of-origin, recurrence predictors, treatment response predictors; and / or metastasis predictors; 5) one or more treatment associated with one or more of the biomarkers, groups of biomarkers, and / or biological signatures determined for the molecular profile; 6) an indication whether one or more treatment is likely to benefit the patient, not benefit the patient, or has indeterminate benefit; 7) one or more clinical trials for which the patient may be eligible based on the molecular profile; 8) an indication whether the cancer is predicted to recur and / or metastasize; and / or 9) evidence relevant to the foregoing, such as literature reports and / or clinical trial results. The description of the molecular profile of the biomarkers can include such information as the laboratory technique used to assess each biomarker (e.g., RT- PCR, FISH / CISH, PCR, FA / RFLP, NGS, WGS, WES, WTS, etc), optionally including the result and any criteria used to score each technique. By way of non-limiting example, the criteria for scoring a copy number alteration (or variation, CNA or CNV) may be a presence (i.e., a copy number that is greater or lower than the “normal” copy number present in a subject who does not have cancer, or statistically identified as present in the general population, typically diploid) or absence (i.e., a copy number that is considered the same as the “normal” copy number present in a subject who does not have cancer, or statistically identified as present in the general population, typically diploid). Treatments associated with one or more of the biomarkers or biosignatures may be determined using treatment association such as in any of International Patent Publications WO / 2007 / 137187 (Int’l Appl. No. PCT / US2007 / 069286), published November 29, 2007; WO / 2010 / 045318 (Int’l Appl. No. PCT / US2009 / 060630), published April 22, 2010; WO / 2010 / 093465 (Int’l Appl. No. PCT / US2010 / 000407), published August 19, 2010; WO / 2012 / 170715 (Int’l Appl. No. PCT / US2012 / 041393), published December 13, 2012;WO / 2014 / 089241 (Int’l Appl. No. PCT / US2013 / 073184), published June 12, 2014; WO / 2011 / 056688 (Int’l Appl. No. PCT / US2010 / 054366), published May 12, 2011; WO / 2012 / 092336 (Int’l Appl. No. PCT / US2011 / 067527), published July 5, 2012; WO / 2015 / 116868 (Int’l Appl. No. PCT / US2015 / 013618), published August 6, 2015;WO / 2017 / 05.3915 (Int’l Appl. No. PCT / US2016 / 053614), published March 30, 2017; WO / 2016 / 141169 (Int’l Appl. No. PCT / US2016 / 020657), published September 9, 2016; and W02018175501 (Int’l Appl. No. PCT7US2018 / 023438), published September 27, 2018; each of which publications is incorporated by reference herein in its entirety. Such rules can be updated as new information becomes available regarding various biomarkers, biosignatures, treatments, and the relationships thereof. The indication whether each treatment is likely to benefit the patient, not benefit the patient, or has indeterminate benefit may be weighted. For example, a likely or potential benefit may be a strong potential benefit or a lesser potential benefit. Such weighting can be based on any appropriate criteria, e.g., the strength of the evidence of the biomarker-treatment association, or the results of the profiling, e.g., a degree or level of over- or underexpression, mutation, or any other relevant state (wild type or altered). As the treating physician is ultimately responsible for treating their patient, such physician may use the report to assist in guiding their treatment recommendations.

[0184] As noted, various components can be added to the report as desired. In some embodiments, the report comprises a list having an indication of whether one or more biomarkers and / or biosignatures in the molecular profile is associated with an ongoing clinical trial. The report may include identifiers for any such trials, e.g., to facilitate the treating physician’s investigation of potential enrollment of the subject in the trial. In some embodiments, the report provides a list of evidence supporting the association of the biomarkers and / or biosignatures in the molecular profile with the reported treatment. The list can contain citations to the evidentiary literature and / or an indication of the strength of the evidence for particular treatment associations. In some embodiments, the report comprises a description of various biomarkers in the molecular profile. The description of the biomarkers in the molecular profile can comprise without limitation the biological function and / or various treatment associations.

[0185] In some embodiments, the report comprises various biosignatures determined based on the molecular profiling. Such biosignatures may include signatures of tumor characteristics including without limitation microsatellite stability and / or tumor mutational load / burden. See, e.g., Int’l Patent Appl. No. PCT / US2018 / 023438, filed March 20, 2018. Such biosignatures may include signatures of clinical characteristics including without limitation tissue-of-origin. See, e.g., Int’l Patent Appl. No. PCT / US2020 / 012815, filed lanuary 8, 2020; Int’l Patent Appl. No.PCT / US2021 / 018263, filed February 16, 2021. Such biosignatures may include predictive signatures. See, e.g., Inf 1 Patent Appl. No. PCT / US2019 / 064078, filed December 2, 2019; Int’l Patent Appl. No. PCT / US2020 / 035990, filed June 3, 2020; Int’l Patent Appl. No. PCT / US2021 / 030351, filed April 30, 2021. Each of which applications in this paragraph is incorporated by reference herein in its entirety.

[0186] The molecular profiling report can be delivered to the caregiver for the subject, e.g., the oncologist or other treating physician. The caregiver can use the results of the report to guide a treatment regimen for the subject. For example, the caregiver may use one or more treatments indicated as likely benefit in the report to treat the patient. Similarly, the caregiver may avoid treating the patient with one or more treatments indicated as likely lack of benefit in the report. In some embodiments, such as when the report includes a biosignature indicating a likely recurrence or metastasis, the treating physician may choose, for example, a more aggressive treatment regimen, more frequent monitoring, or both. Such decisions are made by the caregiver with guidance from the report. In some embodiments, the methods herein comprise administering therapies to the subject. In some embodiments, the methods herein comprise not administering therapies to the subject, e.g., if a cancer is not likely to benefit from such therapies, not likely to metastasize, or both).

[0187] The report can be computer generated, and can be a printed report, a computer file or both. The report can be made accessible via a secure web portal. The report may be displayed using any desired medium. In some embodiments, the display is a print out, a computer file, including without limitation a pdf file, or may be displayed via an application on a computer display such as a computer monitor, laptop display, tablet, smartphone, or other mobile device.

[0188] In an aspect, the disclosure provides use of a reagent in carrying out the methods as described herein. In a related aspect, the disclosure provides of a reagent in the manufacture of a reagent or kit for carrying out the methods as described herein. In still another related aspect, the disclosure provides a kit comprising a reagent for carrying out the methods as described herein. The reagent can be any useful and desired reagent. In preferred embodiments, the reagent comprises at least one of a reagent for extracting nucleic acid from a sample, and a reagent for performing next-generation sequencing.

[0189] In an aspect, the disclosure provides a system for generating a molecular profding report such as described above, comprising: (a) at least one host server; (b) at least one user interface for accessing the at least one host server to access and input data; (c) at least one processor for processing the inputted data; (d) at least one memory coupled to the processor for storing the processed data and instructions for: i) accessing a biomarker status (e.g., TPM, TMM and / or GSEA for mRNA transcripts) determined by molecular profiling methodology as described herein; and ii) identifying biomarkers, biosignatures and related data and any information derived using such data (treatments, clinical trials, phenotypes, predictions, etc, as described herein); and (e) at least one display for displaying results and outcomes of the molecular profiling. In some embodiments, the system further comprises at least one memory coupled to the processor for storing the processed data and instructions for identifying, based on the generated molecular profile according to the methods above, at least one therapy with potential benefit for treatment of the cancer; and at least one display for display thereof. The system may further comprise at least one database comprising references for various biomarker states, data for drug / biomarker associations, or both. The at least one display can be a report provided by the present disclosure.IV. EXAMPLE METHOD

[0190] FIG. 4 is a flowchart 1100 illustrating a method for predicting metastatic potential of a tumor in a subject having a particular set of biomarkers according to embodiments of the present disclosure. Method 1100 and any of the methods described herein may be totally or partially performed with a computer system including one or more processors, which can be configured to perform the steps. Thus, embodiments are directed to computer systems configured to perform the steps of any of the methods described herein, potentially with different components performing a respective step or a respective group of steps.

[0191] At block 1110, molecular data is obtained by one or more computers by sequencing a panel of mRNA transcripts in a biological sample from a subject. In some examples, the panel of mRNA transcripts can include transcripts for at least 10, 20, 50, 100, 200, 500, 1000, 2000, 5000, 10000, 20000, 21000, 22000, or 23000 genes, or anywhere in between. For example, the panel of mRNA transcripts may include transcripts for at least 600 genes, 700 genes, 800 genes, 900 genes, 1,000 genes, 1,100 genes, 1,200 genes, 1,300 genes, 1,400 genes, 1,500 genes, 1,600genes, 1,700 genes, 1,800 genes, 1,900 genes, 2,000 genes, 5,000 genes, 10,000 genes, 20,000 genes, 21,000 genes, 22,000 genes, 23,000 genes, or anywhere in between. In some embodiments, the molecular data is obtained by WTS.

[0192] The sequencing may be performed in various ways. For example, embodiments may use next-generation sequencing techniques (see, e.g., Slatko BE, et al. Overview of Next- Generation Sequencing Technologies. Curr Protoc Mol Biol. 2018 Apr; 122(1 ):e59), or in some cases long read techniques, such as single molecule sequencing such as Single-molecule realtime (SMRT) sequencing and nanopore sequencing.

[0193] At block 1120, input data that includes a set of features extracted from the obtained molecular data is generated by the one or more computers. In some examples, the input data can be provided to one or more computing modules that can extract a respective set of features from the obtained molecular data to generate the input data. For example, as described herein, the one or more computers can include a Transcripts Per Million (TPM) module, a Trimmed Mean of the M-values (TMM) module, a Gene Set Enrichment Analysis (GSEA) module, or a combination thereof to process (e.g., normalize) the obtained molecular data. Additionally, the one or more computers can include a set of dimensionality reduction modules to reduce the dimensionality of output received from the TPM module, TMM module, or GSEA module to determine the set of features to provide as input to a predictive model.

[0194] At block 1130, the generated input data is provided by the one or more computers as input to the predictive model. The predictive model can include at least one machine learning model. In some examples, each particular machine learning model of the at least one machine learning model can be trained to generate output data that indicates risk that a cancer in a subject will metastasize based on the particular machine learning model processing of the set of features extracted from the molecular data of a cohort of cancer patients comprising at least 1000, 2000, 5000, 10000, 25000, 50000, 100000, 150000 or 200000 patients. For example, the cohort may include at least 11,000 patients, 12,000 patients, 13,000 patients, 14,000 patients, 15,000 patients, 20,000 patients, 25,000 patients, 50,000 patients, 100,000 patients, 150,000 patients, 200,000 patients, or anywhere in between.

[0195] In some implementations, the predictive model uses transfer learning. The at least one machine learning model can be selected from a set of three machine learning models. A firstmachine learning model can process transcript data including TPM. A second machine learning model can process transcript data including TMM. A third machine learning model can process transcript data including GSEA. In some examples, at least one of the set of three machine learning models can include at least one dimensional reduction model that is applied to the molecular data used to generate the input data. For example, each of the first machine learning model, the second machine learning model, and the third machine learning model may include a respective dimensional reduction model. In some examples, the predictive may include an additional machine learning model that can receive an aggregated feature vector. The aggregated feature vector can include a respective reduced feature vector output by at least two of the dimensional reduction models of the first machine learning model, the second machine learning model, and the third machine learning model.

[0196] At block 1140, the generated input data is processed through the at least one machine learning model. In some examples, the generated input data may be further processed through an ensemble model to generate output data indicating risk that cancer in the subject will metastasize. The ensemble model can include multiple submodels as described herein (e.g., with respect to FIG. ID) such that the output data can be generated as an aggregation of the predictions of the submodels.

[0197] At block 1150, a risk that the cancer in the subject will metastasize is determined by the one or more computers based on the generated output data. The one or more computers can assign a risk score to the subject to indicate the risk that the cancer in the subject will metastasize. In some examples, the risk score may include a numeric value (e.g., an integer, a percentage, etc ). Additionally or alternatively, the risk score can indicate low risk, intermediate risk, high risk, or indeterminate risk of the cancer in the subject metastasizing. In some cases, a confidence level can be assigned to the risk score.

[0198] In some embodiments, the metastasis comprises brain metastasis and can further consist of brain metastasis. As described herein, the metastasis can comprise secondary tumors, e.g., in at least one of the lymph nodes, adrenal gland, bone, brain, liver, lung, muscle, peritoneum, skin, and vagina.

[0199] Based on the risk (e.g., the risk score) that the cancer in the subject will metastasize, risk score, a therapy can be administered. In some examples, the therapy may be administered if the determined risk that the cancer will metastasize is intermediate, high, or is indeterminate. Additionally or alternatively, the therapy may not be administered to the subject if the determined risk that the cancer will metastasize is low or is indeterminate.

[0200] A result of method 1100 can be provided to a user, such as in a report, as described herein. For example, based on a determination of the risk that the cancer in the subject will metastasize, one or more computers can generate rendering data that, when rendered by a user device, causes the user device to display data that identifies the risk of metastasis can provide the rendered data to the user device. As another example, a report can display the output that identifies the likely metastasis, likely lack of metastasis, or indeterminate likelihood of metastasis. For instance, the display for displaying the output can comprise a printout, a file, a computer display, and any combination thereof.

[0201] Treatments can be identified based on the risk of metastasis. For instance, a method can include operations that identify, based on the molecular data and / or additional profiling data obtained from assaying the biological sample from the subject: (a) one or more treatment of likely benefit for treating the cancer in the subject; (b) one or more treatment of likely lack of benefit for treating the cancer in the subject; (c) one or more treatment of indeterminate benefit for treating the cancer in the subject; and / or (d) one or more clinical trial for which the subject is indicated as eligibleV. EXAMPLE SYSTEMS

[0202] FIG. 5 illustrates a measurement system 1200 according to an embodiment of the present disclosure. The system as shown includes a sample 1205, such as nucleic acid molecules (e.g., DNA and / or RNA) isolated from a tissue or blood sample, within an assay device 1210, where an assay 1208, e.g., NGS, can be performed on sample 1205. For example, sample 1205 can be contacted with reagents of assay 1208 to provide a signal (e.g., an intensity signal) of a physical characteristic 1215 (e.g., RNA sequence information). An example of an assay device can be a flow cell that includes probes and / or primers of an assay or a tube through which a droplet moves (with the droplet including the assay). Physical characteristic 1215 (e.g., afluorescence intensity, a voltage, or a current), from the sample is detected by detector 1220. Detector 1220 can take a measurement at intervals (e.g., periodic intervals) to obtain data points that make up a data signal. In one embodiment, an analog-to-digital converter converts an analog signal from the detector into digital form at a plurality of times.

[0203] Assay device 1210 and detector 1220 can form an assay system, e.g., a PCR system or a sequencing system (e.g., NGS sequencing system) that performs sequencing according to embodiments described herein. A data signal 1225 is sent from detector 1220 to logic system 1230. As an example, data signal 1225 can be used to determine sequences and / or locations in a reference genome of nucleic acid molecules (e.g., DNA and / or RNA). Data signal 1225 can include various measurements made at a same time, e.g., different colors of fluorescent dyes or different electrical signals for different molecule of sample 1205, and thus data signal 1225 can correspond to multiple signals. Data signal 1225 may be stored in a local memory 1235, an external memory 1240, or a storage device 1245. The assay system can be comprised of multiple assay devices and detectors.

[0204] Logic system 1230 may be, or may include, a computer system, ASIC, microprocessor, graphics processing unit (GPU), etc. It may also include or be coupled with a display (e.g., monitor, LED display, etc.) and a user input device (e.g., mouse, keyboard, buttons, etc.). Logic system 1230 and the other components may be part of a stand-alone or network connected computer system, or they may be directly attached to or incorporated in a device (e.g., a sequencing device) that includes detector 1220 and / or assay device 1210. Logic system 1230 may also include software that executes in a processor 1250. Logic system 1230 may include a computer readable medium storing instructions for controlling measurement system 1200 to perform any of the methods described herein. For example, logic system 1230 can provide commands to a system that includes assay device 1210 such that sequencing or other physical operations are performed. Such physical operations can be performed in a particular order, e.g., with reagents being added and removed in a particular order. Such physical operations may be performed by a robotics system, e.g., including a robotic arm, as may be used to obtain a sample and perform an assay.

[0205] Measurement system 1200 may also include a treatment device 1260, which can provide a treatment to the subject. Treatment device 1260 can determine a treatment and / or beused to perform a treatment. Examples of such treatment can include surgery, radiation therapy, chemotherapy, immunotherapy, targeted therapy, hormone therapy, and stem cell transplant. Logic system 1230 may be connected to treatment device 1260, e.g., to provide results of a method described herein. The treatment device may receive inputs from other devices, such as an imaging device and user inputs (e.g., to control the treatment, such as controls over a robotic system).

[0206] Measurement system 1200 may also include a reporting device 1255, which can present results of any of the methods described herein, e.g., as determined using the measurement system. Reporting device 1255 can be in communication with a reporting module within logic system 1230 that can aggregate, format, and send a report to reporting device 1255. The reporting module can present information determined using any of the method described herein. The information can be presented by reporting device 1255 in any format that can be recognized and interpreted by a user of the measurement system 1200. For example, the information can be presented by reporting device 1255 in a displayed, printed, or transmitted format, or any combination thereof.

[0207] Any of the computer systems mentioned herein may utilize any suitable number of subsystems. Examples of such subsystems are shown in FIG. 6 in computer system 10. In some embodiments, a computer system includes a single computer apparatus, where the subsystems can be the components of the computer apparatus. In other embodiments, a computer system can include multiple computer apparatuses, each being a subsystem, with internal components. A computer system can include desktop and laptop computers, tablets, mobile phones and other mobile devices.

[0208] The subsystems shown in FIG. 6 are interconnected via a system bus 75. Additional subsystems such as a printer 74, keyboard 78, storage device(s) 79, monitor 76 (e.g., a display screen, such as an LED), which is coupled to display adapter 82, and others are shown. Peripherals and input / output (I / O) devices, which couple to I / O controller 71, can be connected to the computer system by any number of means known in the art such as input / output (I / O) port 77 (e.g., USB, FireWire®). For example, I / O port 77 or external interface 81 (e.g., Ethernet, WiFi, etc.) can be used to connect computer system 10 to a wide area network such as the Internet, a mouse input device, or a scanner. The interconnection via system bus 75 allows the centralprocessor 73 to communicate with each subsystem and to control the execution of a plurality of instructions from system memory 72 or the storage device(s) 79 (e.g., a fixed disk, such as a hard drive, or optical disk), as well as the exchange of information between subsystems. The system memory 72 and / or the storage device(s) 79 may embody a computer readable medium. Another subsystem is a data collection device 85, such as a camera, microphone, accelerometer, and the like. Any of the data mentioned herein can be output from one component to another component and can be output to the user.

[0209] A computer system can include a plurality of the same components or subsystems, e.g., connected together by external interface 81, by an internal interface, or via removable storage devices that can be connected and removed from one component to another component. In some embodiments, computer systems, subsystem, or apparatuses can communicate over a network. In such instances, one computer can be considered a client and another computer a server, where each can be part of a same computer system. A client and a server can each include multiple systems, subsystems, or components. In various embodiments, methods may involve various numbers of clients and / or servers, including at least 10, 20, 50, 100, 200, 500, 1,000, or 10,000 devices. Methods can include various numbers of communication messages between devices, including at least 100, 200, 500, 1,000, 10,000, 50,000, 100,000, 500,00, or one million communication messages. Such communications can involve at least 1 MB, 10 MB, 100 MB, 1 GB, 10 GB, or 100 GB of data.

[0210] Aspects of embodiments can be implemented in the form of control logic using hardware circuitry (e.g., an application specific integrated circuit or field programmable gate array) and / or using computer software stored in a memory with a generally programmable processor in a modular or integrated manner, and thus a processor can include memory storing software instructions that configure hardware circuitry, as well as an FPGA with configuration instructions or an ASIC. As used herein, a processor can include a single-core processor, multicore processor on a same integrated chip, or multiple processing units on a single circuit board or networked, as well as dedicated hardware. Based on the disclosure and teachings provided herein, a person of ordinary skill in the art will know and appreciate other ways and / or methods to implement embodiments of the present disclosure using hardware and a combination of hardware and software.

[0211] Any of the software components or functions described in this application may be implemented as software code to be executed by a processor using any suitable computer language such as, for example, Java, C, C++, C#, Objective-C, Swift, or scripting language such as R, Perl or Python using, for example, conventional or object-oriented techniques. The software code may be stored as a series of instructions or commands on a computer readable medium for storage and / or transmission. A suitable non-transitory computer readable medium can include random access memory (RAM), a read only memory (ROM), a magnetic medium such as a hard-drive or a floppy disk, or an optical medium such as a compact disk (CD) or DVD (digital versatile disk) or Blu-ray disk, flash memory, and the like. The computer readable medium may be any combination of such devices. In addition, the order of operations may be re-arranged. A process can be terminated when its operations are completed but could have additional steps not included in a figure. A process may correspond to a method, a function, a procedure, a subroutine, a subprogram, etc. When a process corresponds to a function, its termination may correspond to a return of the function to the calling function or the main function.

[0212] Such programs may also be encoded and transmitted using carrier signals adapted for transmission via wired, optical, and / or wireless networks conforming to a variety of protocols, including the Internet. As such, a computer readable medium may be created using a data signal encoded with such programs. Computer readable media encoded with the program code may be packaged with a compatible device (e.g., as firmware) or provided separately from other devices (e.g., via Internet download). Any such computer readable medium may reside on or within a single computer product (e.g., a hard drive, a CD, or an entire computer system), and may be present on or within different computer products within a system or network. A computer system may include a monitor, printer, or other suitable display for providing any of the results mentioned herein to a user.

[0213] Any of the methods described herein may be totally or partially performed with a computer system including one or more processors, which can be configured to perform the steps. Any operations performed with a processor may be performed in real-time. The term “real-time” may refer to computing operations or processes that are completed within a certain time constraint. The time constraint may be 1 minute, 1 hour, 1 day, or 7 days. Thus,embodiments can be directed to computer systems configured to perform the steps of any of the methods described herein, potentially with different components performing a respective step or a respective group of steps. Although presented as numbered steps, steps of methods herein can be performed at a same time or at different times or in a different order. Additionally, portions of these steps may be used with portions of other steps from other methods. Also, all or portions of a step may be optional. Additionally, any of the steps of any of the methods can be performed with modules, units, circuits, or other means of a system for performing these steps.

[0214] The specific details of particular embodiments may be combined in any suitable manner without departing from the spirit and scope of embodiments of the disclosure. However, other embodiments of the disclosure may be directed to specific embodiments relating to each individual aspect, or specific combinations of these individual aspects.

[0215] The above description of example embodiments of the present disclosure has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the disclosure to the precise form described, and many modifications and variations are possible in light of the teaching above.

[0216] A recitation of "a", "an" or "the" is intended to mean "one or more" unless specifically indicated to the contrary. The use of “or” is intended to mean an “inclusive or,” and not an “exclusive or” unless specifically indicated to the contrary. Reference to a “first” component does not necessarily require that a second component be provided. Moreover, reference to a “first” or a “second” component does not limit the referenced component to a particular location unless expressly stated. The term “based on” is intended to mean “based at least in part on.”

[0217] Th e claims may be drafted to exclude any element which may be optional. As such, this statement is intended to serve as antecedent basis for use of such exclusive terminology as “solely”, “only”, and the like in connection with the recitation of claim elements, or the use of a “negative” limitation.

[0218] All patents, patent applications, publications, and descriptions mentioned herein are incorporated by reference in their entirety for all purposes. None is admitted to be prior art. Where a conflict exists between the instant application and a reference provided herein, the instant application shall dominate.

Claims

WHAT IS CLAIMED IS:

1. A system for predicting risk that a cancer in a subject will metastasize, the system comprising: one or more computers; and one or more memory devices storing instructions that, when executed by the one or more computers, cause the one or more computers to perform operations, the operations comprising: obtaining, by the one or more computers, molecular data obtained by sequencing a panel of mRNA transcripts in a biological sample from the subject, wherein the panel of mRNA transcripts comprises transcripts for at least 10, 20, 50, 100, 200, 500, 1000, 2000, 5000, 10000, 20000, 21000, 22000 or 23000 genes; generating, by the one or more computers, input data that includes a set of features extracted from the obtained molecular data; providing, by the one or more computers, the generated input data as input to a predictive model, the predictive model comprising at least one machine learning model, wherein each particular machine learning model of the at least one machine learning models was trained to generate output data that indicates risk that a cancer in a subject will metastasize based on the particular machine learning model processing of the set of features extracted from the molecular data of a cohort of cancer patients comprising at least 1000, 2000, 5000, 10000, 25000, 50000, 100000, 150000 or 200000 patients; processing, by the one or more computers, the generated input data through the at least one machine learning model to generate output data indicating risk that the cancer in the subject will metastasize; and determining, by the one or more computers and based on the generated output data, risk that the cancer in the subject will metastasize.

2. The system of claim 1, wherein the panel of mRNA transcripts comprises transcripts for at least 10,000 genes.

3. The system of claim 1 or claim 2, wherein the operations further comprise:based on a determination of the risk that the cancer in the subject will metastasize, generating, by the one or more computers, rendering data that, when rendered by a user device, causes the user device to display data that identifies the risk of metastasis; and providing, by the one or more computers, the rendered data to the user device.

4. The system of any preceding claim, wherein the predictive model uses transfer learning.

5. The system of any preceding claim, wherein the at least one machine learning model comprises at least one machine learning model selected from:(a) a first machine learning model that processes transcript data comprising Transcripts Per Million (TPM);(b) a second machine learning model that processes transcript data comprising Trimmed Mean of the M-values (TMM); and / or(c) a third machine learning model that processes transcript data comprising Gene Set Enrichment Analysis (GSEA).

6. The system of claim 5, wherein at least one of the first machine learning model, the second machine learning model, and the third machine learning model includes a dimensional reduction model that is applied to the molecular data.

7. The system of claim 6, wherein each of the first machine learning model, the second machine learning model, and the third machine learning model include a respective dimensional reduction model.

8. The system of claim 6, wherein the predictive model includes an additional machine learning model that is configured to receive an aggregated feature vector that includes respective reduced feature vectors output by at least two dimensional reduction models of the first machine learning model, the second machine learning model, and the third machine learning model.

9. The system of any preceding claim, wherein the determined risk that the cancer will metastasize is assigned a risk score, optionally wherein:(a) the risk score comprises a numeric value;(b) the risk score indicates low risk, intermediate risk, high risk, or indeterminate risk; and / or(c) a confidence level is assigned to the risk score.

10. The system of any preceding claim, wherein the biological sample comprises formalin-fixed paraffin-embedded (FFPE) tissue, fixed tissue, a core needle biopsy, a fine needle aspirate, unstained slides, fresh frozen (FF) tissue, formalin samples, tissue comprised in a solution that preserves nucleic acid or protein molecules, a fresh sample, a malignant fluid, a bodily fluid, a tumor sample, a tissue sample, or any combination thereof.

11. The system of any preceding claim, wherein the biological sample is from a solid tumor, optionally wherein the solid tumor is a primary tumor.

12. The system of claim 11, wherein the primary tumor is a tumor of the myeloid, breast, bile ducts, colon, rectum, female genital tract, stomach, esophagus, gastrointestinal stromal cells, small intestine, brain, mouth, sinuses, nose, throat, blood, liver, nervous system, lung, lymph, male genital tract, pleura, skin, plasma cells, neuroendocrine cells, B-cells, T-cells, ovary, pancreas, pituitary gland, spinal cord, prostate, peritoneum, large intestine, soft tissue, connective tissue, fat tissue, thymus, thyroid, or eye.

13. The system of claim 11 or 12, wherein the primary tumor is a tumor of the bladder, breast, colon, rectum, endometrium, uterus, ovary, female genital tract, kidney, blood, liver, lung, skin, lymph, pancreas, prostate, or thyroid.

14. The system of any preceding claim, wherein the biological sample comprises a bodily fluid.

15. The system of claim 14, wherein the bodily fluid comprises a malignant fluid, a pleural fluid, a peritoneal fluid, or any combination thereof.

16. The system of claim 14 or 15, wherein the bodily fluid comprises peripheral blood, sera, plasma, ascites, urine, cerebrospinal fluid (CSF), sputum, saliva, bone marrow, synovial fluid, aqueous humor, amniotic fluid, cerumen, breast milk, broncheoalveolar lavage fluid, semen, prostatic fluid, cowper’s fluid, pre-ejaculatory fluid, female ejaculate, sweat, fecal matter, tears, cyst fluid, pleural fluid, peritoneal fluid, pericardial fluid, lymph, chyme, chyle,bile, interstitial fluid, menses, pus, sebum, vomit, vaginal secretions, mucosal secretion, stool water, pancreatic juice, lavage fluids from sinus cavities, bronchopulmonary aspirates, blastocyst cavity fluid, or umbilical cord blood.

17. The system of any preceding claim, wherein the set of features extracted from the obtained molecular data further comprises one or more of a clinical characteristics of the subject, a primary tumor location, one or more secondary tumor location, or any combination thereof.

18. The system of any preceding claim, wherein the cancer comprises an acute myeloid leukemia (AML), breast carcinoma, cholangiocarcinoma, colorectal adenocarcinoma, extrahepatic bile duct adenocarcinoma, female genital tract malignancy, gastric adenocarcinoma, gastroesophageal adenocarcinoma, gastrointestinal stromal tumor (GIST), glioblastoma, head and neck squamous carcinoma, leukemia, liver hepatocellular carcinoma, low grade glioma, lung bronchioloalveolar carcinoma (BAC), non-small cell lung cancer (NSCLC), lung small cell cancer (SCLC), lymphoma, male genital tract malignancy, malignant solitary fibrous tumor of the pleura (MSFT), melanoma, multiple myeloma, neuroendocrine tumor, nodal diffuse large B-cell lymphoma, non-epithelial ovarian cancer (non-EOC), ovarian surface epithelial carcinoma, pancreatic adenocarcinoma, pituitary carcinomas, oligodendroglioma, prostatic adenocarcinoma, retroperitoneal or peritoneal carcinoma, retroperitoneal or peritoneal sarcoma, small intestinal malignancy, soft tissue tumor, thymic carcinoma, thyroid carcinoma, or uveal melanoma.

19. The system of any preceding claim, wherein the cancer comprises a breast carcinoma, colorectal adenocarcinoma, female genital tract malignancy, kidney cancer, non-small cell lung cancer (NSCLC), lung small cell cancer (SCLC), melanoma, ovarian surface epithelial carcinomas, prostatic adenocarcinoma, uterine neoplasm, endometrial carcinoma, or unknown.

20. The system of any preceding claim, wherein the cancer comprises a breast cancer, optionally wherein the breast cancer comprises a HER2+ breast cancer.

21. The system of any preceding claim, wherein the metastasis comprises brain metastasis, optionally wherein the metastasis consists of brain metastasis.

22. The system of claim 1 , wherein the operations further comprise processing the processed input data through an ensemble model to generate the output data indicating risk that the cancer in the subject will metastasize.

23. The system of any preceding claim, wherein the operations further comprise generating a report displaying the output that identifies the likely metastasis, likely lack of metastasis, or indeterminate likelihood of metastasis, wherein optionally the display for displaying the output comprises a printout, a file, a computer display, and any combination thereof.

24. The system of any preceding claim, wherein the metastasis comprises secondary tumors in at least one of the lymph nodes, adrenal gland, bone, brain, liver, lung, muscle, peritoneum, skin, and vagina.

25. The system of any preceding claim, wherein the metastasis comprises brain metastasis, optionally wherein the metastasis consists of brain metastasis.

26. The system of any preceding claim, wherein the system further comprises operations that identify, based on the molecular data and / or additional profiling data obtained from assaying the biological sample from the subject:(a) one or more treatment of likely benefit for treating the cancer in the subject;(b) one or more treatment of likely lack of benefit for treating the cancer in the subject;(c) one or more treatment of indeterminate benefit for treating the cancer in the subject; and / or(d) one or more clinical trial for which the subject is indicated as eligible.

27. Anon-transitory computer-readable medium storing software comprising instructions executable by one or more computers which, upon such execution, cause the one or more computers to perform the operations described with reference to any preceding claim.

28. A method comprising steps that correspond to each of the operations of any one of claims 1- 26.

29. A method for predicting risk that a cancer in a subject will metastasize, the method comprising:obtaining, by the one or more computers, molecular data obtained by sequencing a panel of mRNA transcripts in a biological sample from the subject, wherein the panel of mRNA transcripts comprises transcripts for at least 500 genes; generating, by the one or more computers, input data that includes a set of features extracted from the obtained molecular data; providing, by the one or more computers, the generated input data as input to a predictive model, the predictive model comprising at least one machine learning model, wherein each particular machine learning model of the at least one machine learning models was trained to generate output data that indicates risk that a cancer in a subject will metastasize based on the particular machine learning model processing of the set of features extracted from the molecular data of a cohort of cancer patients comprising at least 10,000 patients; processing, by the one or more computers, the generated input data through the at least one machine learning model to generate output data indicating risk that the cancer in the subject will metastasize; and determining, by the one or more computers and based on the generated output data, risk that the cancer in the subject will metastasize.

30. The method of claim 29, further comprising administering a therapy to the subject based on the determined risk that the cancer in the subject will metastasize.

31. The method of claim 30, wherein the therapy is administered to the subject if the determined risk that the cancer will metastasize is intermediate, high, or is indeterminate.

32. The method of claim 30 or 31, wherein the therapy is not administered to the subject if the determined risk that the cancer will metastasize is low or is indeterminate.

Images