CN110678930A

CN110678930A - Systems and methods for assessing drug efficacy

Info

Publication number: CN110678930A
Application number: CN201880036325.1A
Authority: CN
Inventors: S·张; M·王; A·怀斯; H·康; V·F·奥努席克; K·克鲁格利亚克
Original assignee: Irumina Co Ltd
Current assignee: Irumina Co Ltd
Priority date: 2017-12-01
Filing date: 2018-12-03
Publication date: 2020-01-10
Anticipated expiration: 2038-12-03
Also published as: KR20200093437A; AU2018375214A1; CN110678930B; JP7429539B2; MX2019014285A; RU2019135697A; NZ759413A; WO2019109089A1; SG11201910041WA; BR112019024537A2; CA3061736A1; JP2021505976A; ZA201907298B; EP3616205A1

Abstract

The invention provides a computer-implemented method comprising inputting non-trained subject genomic information comprising features from tumor samples to a trained machine learning classifier, wherein the trained machine learning classifier is trained on features of tumor samples obtained from training subjects and responsiveness of the training subjects to checkpoint inhibition therapy, and the machine learning classifier is trained to predict responsiveness to the therapy; and generating a checkpoint inhibition reactivity classification that predicts the subject's reactivity to checkpoint inhibition using the trained machine learning classifier and reporting the checkpoint inhibition reactivity classification using a graphical user interface. The invention also provides a computer system for implementing the method and a machine learning classifier trained by the method.

Description

Systems and methods for assessing drug efficacy

Cross Reference to Related Applications

This application claims priority to U.S. provisional patent application No. 62/593,802 filed on 12/1/2017, the entire contents of which are hereby incorporated by reference herein.

Background

The detection of abnormal or cancerous cells in the body is an important task of the immune system. One mechanism involved is immune checkpoints. For example, programmed cell death protein 1(PD-1) and cytotoxic T lymphocyte-associated protein 4(CTLA-4) checkpoints on T cells negatively regulate immune function and prevent overresponses (i.e., promote immune system self-recognition). However, this mechanism can be exploited by tumor cells to escape immune attack. Immunotherapy such as PD-1 inhibition (e.g., anti-PD 1 antibodies) and CTLA-4 inhibition (e.g., CTLA-4 antibodies) can block checkpoint activity, thus facilitating T cell recognition of the disease or tumor cell itself.

However, although immune checkpoint therapy may be effective, responsiveness cannot be guaranteed in all cancer patients. Immune checkpoint therapy has been shown to improve long-term survival of patients with various cancers compared to traditional cancer therapies. However, only a subset of cancer patients respond to currently approved checkpoint inhibitor drugs, including anti-CTLA-4 antibodies (e.g., ipilimumab) and therapies that target the PD-1 checkpoint pathway, such as anti-PD-1 antibodies (e.g., nivolumab) or anti-programmed death-ligand 1 (anti-PD-L1) antibodies (e.g., atezolizumab (atezolizumab))). Therefore, it would be advantageous to be able to select patients that will respond to a particular checkpoint therapy, and to be able to predict which checkpoint targets will allow the best outcome in a given patient.

Many different genomic and cellular characteristics may contribute to the effectiveness of immunotherapy for a given individual. For example, a higher Tumor Mutation Burden (TMB) may positively affect response rates by increasing antigen presentation on tumor cells, leading to increased T cell recognition when PD-1 is blocked. Leukocyte tumor infiltration with expression of CD4/CD8/CD19 correlates with better clinical outcomes because such cells contribute to immune attack and subsequent antigen release by tumor cells. Myeloid-derived suppressor and regulatory T cells (Tregs) block T cell availability and correlate with poorer survival in various patients. Since these are features that can be detected and derived from Next Generation Sequencing (NGS) data that interact with each other, it is important to establish machine learning applications that can explore the relationship of these features to immunotherapy responsiveness and can generate predictions of treatment (such as checkpoint inhibition or other cancer treatments) responsiveness, where the association (context) of multiple features that act synergistically in a given individual is integrated based on others' responsiveness and taking into account their individual multi-factor association (multigenic context).

In addition, given the potentially large number of interacting genomic, cellular, and other features that may be present that may interact to determine whether a given individual will respond positively to checkpoint inhibition, improved methods are needed to report responsiveness predictions. For example, in the prediction of reactivity, many different features may interact in combination. Some features may be determined to be of greater or lesser importance than others when applying machine learning methods to assess whether a patient is likely to be more or less responsive to a given checkpoint inhibition; different features may have different degrees and levels of predictive effect each may have on reactivity; and different factors may signal in different individuals that the patient will have greater or lesser responsiveness to different checkpoint inhibition treatments. Therefore, for a given patient's responsiveness, a relevance report (contextual report) is required, including identifying features that are meaningful in predicting responsiveness and directional in their predictive effect. However, current reporting methods have deficiencies in view of the limited space available to present all of these potential aspects of predictive reporting. Therefore, a new approach is needed to report multiple elements relevant to reactivity prediction.

The present disclosure is directed to overcoming these and other deficiencies in the art.

Summary of The Invention

In one aspect, the invention discloses a computer-implemented method comprising: inputting genomic information of a non-training subject (non-training subject) to a trained machine learning classifier, the genomic information of the non-training subject comprising features from a tumor profile (tumor profile) obtained from a non-training subject, wherein the trained machine learning classifier is trained on genomic information of a plurality of training subjects (training subjects) and responsiveness of each of the plurality of training subjects to a treatment (including checkpoint inhibition), the genomic information of the plurality of training subjects comprising features of tumor samples obtained from each of the plurality of training subjects, wherein the machine learning classifier is trained to predict responsiveness to the treatment; generating a checkpoint inhibition responsiveness classification of a non-training subject using the trained machine learning classifier, the checkpoint inhibition responsiveness classification predicting responsiveness of the non-training subject to checkpoint inhibition; and reporting the checkpoint inhibition responsiveness classification of the non-training subject using a graphical user interface. In one embodiment, at least some features of the tumor profile from a non-training subject, or at least some features of the tumor profile from one or more training subjects, are selected from the group of features consisting of: total mutation load consisting of all mutations, total mutation load consisting of non-synonymous mutations, expression of β 2 microglobulin (B2M), expression of proteasome subunit β 10(PSMB10), expression of antigenic peptide transporter 1(antigen peptide transporter 1, TAP1), expression of antigenic peptide transporter 2(antigen peptide transporter 2, TAP2), expression of human leukocyte antigen a (HLA-a), expression of major histocompatibility complex class I B (HLA-B), expression of major histocompatibility complex class I C (HLA-C), expression of major histocompatibility complex class II DQ α 1(HLA-DQA1), expression of HLA class II histocompatibility antigen DRB1 β chain (HLA-DRB1), expression of HLA class I histocompatibility antigen α chain E (HLA-E), expression of natural killer cell granule protein 7(NKG7), expression of chemokine-like receptor 1 (klr 1), cmcmcmcmcmr 3 (HLA-E) expression of HLA class I), expression of natural killer cell granule protein 7(NKG7), expression of chemokine-like receptor 1 (kl 1), Differentiation antigen cluster 8(CD8) expressing cells to tumor infiltration, differentiation antigen cluster 4(CD4) expressing cells to tumor infiltration, differentiation antigen cluster 19(CD19) expressing cells to tumor infiltration, granzyme a (gzma) expression, perforin-1 (PRF1) expression, cytotoxic T-lymphocyte-associated protein 4(CTLA4) expression, programmed cell death protein 1(PD1) expression, programmed death-ligand 1(PDL1) expression, programmed cell death-1 ligand 2(PDL2) expression, lymphocyte activation gene 3(LAG3) expression, T-cell immunoreceptor (TIGIT) expression with Ig domain and ITIM domain, differentiation antigen cluster 276(CD276) expression, chemokine (C-C motif) ligand 5(CCL5), CD27 expression, chemokine (C-X-C motif) ligand 9(CXCL9) expression, C-X-C motif receptor chemokine (CXCR6) chemokine receptor 9(CXCL9) expression, Indoleamine 2, 3-dioxygenase (IDO) expression, signal transducer and activator of transcription 1(STAT1) expression, 3-fucosyl-N-acetyl-lactosamine (CD15) expression, interleukin-2 receptor alpha chain (CD25) expression, siglec-3(CD33), differentiation antigen cluster 39(CD39) expression, differentiation antigen cluster (CD118) expression, forkhead box P3(FOXP3) expression, and any combination of two or more of the foregoing.

In another embodiment, at least some of the training features or at least some of the non-training features comprise a gene set. In yet another embodiment, the gene set is selected using a single sample gene set enrichment assay. In yet another embodiment, the machine learning classifier is a random forest. In yet another embodiment, at least 50,000 trees are used in training the machine learning classifier. In yet another embodiment, the checkpoint inhibition reactivity classification comprises a predictive score and one or more characteristic identifiers (identifiers), and the one or more characteristic identifiers are selected from the group consisting of: characteristic titer, characteristic importance and characteristic weight.

In another embodiment, the graphical user interface reports the feature identifiers as characteristics (aspects) of an annulus sector (annular sector), wherein an angle of the annulus sector reports the importance of the feature, an outer radius of the annulus sector reports the weight of the feature, and a color of the annulus sector reports the valence (value) of the feature. In yet another embodiment, the feature importance of a feature comprises a reduction in the Gini index of the feature. In yet another embodiment, the graphical user interface reports an identifier of a feature if and only if the feature importance of the feature is above a threshold. In yet another embodiment, the feature importance of the feature is not above the threshold if the square of the feature importance of the feature is not above 0.1. In yet another embodiment, each annulus sector includes an inner arc, and the inner arcs of the annulus sectors are arranged to form a circle.

Another embodiment further comprises: inputting the responsiveness of the non-trained subject to the treatment to a trained machine learning classifier, and further training the machine learning classifier, wherein further training comprises training the trained machine learning classifier as a function of features of a tumor sample obtained from the non-trained subject and the responsiveness of the non-trained subject to the treatment. Yet another embodiment further comprises selecting a therapy based on the generated checkpoint inhibition responsiveness classification.

In another aspect, a computer system is disclosed, comprising: one or more microprocessors and one or more memories for storing a trained machine learning classifier and genomic information of non-trained subjects, wherein the trained machine learning classifier is trained on genomic information of a plurality of training subjects and responsiveness of each of the plurality of training subjects to a treatment (including checkpoint inhibition), the genomic information of the plurality of training subjects including features of a tumor profile obtained from each of the plurality of training subjects, the machine learning classifier being trained to predict responsiveness to the treatment; wherein the genomic information of the non-training subject comprises features from a tumor profile obtained from the non-training subject; wherein the one or more memories store instructions that, when executed by the one or more microprocessors, cause the computer system to generate a checkpoint inhibition responsiveness classification for a non-training subject using the trained machine learning classifier and report the checkpoint inhibition responsiveness classification for the non-training subject using a graphical user interface, wherein the checkpoint inhibition responsiveness classification predicts responsiveness of the non-training subject to checkpoint inhibition.

In one embodiment, at least some features of the tumor profile from a non-training subject or at least some features of the tumor profile from one or more training subjects are selected from the group of: total mutation load consisting of all mutations, total mutation load consisting of non-synonymous mutations, expression of β 2 microglobulin (B2M), expression of proteasome subunit β 10(PSMB10), expression of antigenic peptide transporter 1(TAP1), expression of antigenic peptide transporter 2(TAP2), expression of human leukocyte antigen a (HLA-a), expression of major histocompatibility complex class I B (HLA-B), expression of major histocompatibility complex class I C (HLA-C), expression of major histocompatibility complex class II DQ α 1(HLA-DQA1), expression of HLA class II histocompatibility antigen DRB1 β chain (HLA-DRB1), expression of HLA class I histocompatibility antigen α chain E (HLA-E), expression of natural cell granule protein 7(NKG7), expression of differentiation factor-like receptor 1 (klcmr 1), expression of antigen cluster 8(CD8) -expressing cells tumor infiltration, Tumor infiltration of differentiation antigen cluster 4(CD4) -expressing cells, tumor infiltration of differentiation antigen cluster 19(CD19) -expressing cells, granzyme A (GZMA) expression, perforin-1 (PRF1) expression, cytotoxic T-lymphocyte-associated protein 4(CTLA4) expression, programmed cell death protein 1(PD1) expression, programmed death-ligand 1(PDL1) expression, programmed cell death 1 ligand 2(PDL2) expression, lymphocyte activation gene 3(LAG3) expression, T cell immunoreceptor (TIGIT) expression with Ig domain and ITIM domain, differentiation antigen cluster 276(CD276) expression, chemokine (C-C motif) ligand 5(CCL5), CD27 expression, chemokine (C-X-C motif) ligand 9(CXCL9) expression, C-X-C motif chemokine receptor 6(CXCR6), indoleamine 2, 3-dioxygenase (IDO) expression, signal transduction and activator of transcription 1(STAT1) expression, 3-fucosyl-N-acetyl-lactosamine (CD15) expression, interleukin-2 receptor alpha chain (CD25) expression, siglec-3(CD33), differentiation antigen cluster 39(CD39) expression, differentiation antigen cluster (CD118) expression, forkhead box P3(FOXP3) expression, and any combination of two or more of the foregoing.

In another embodiment, at least some of the training features or at least some of the non-training features comprise a gene set. In yet another embodiment, the gene set is selected using a single sample gene set enrichment assay. In another embodiment, the machine learning classifier is a random forest. In yet another embodiment, at least 50,000 trees are used in training the machine learning classifier. In yet another embodiment, the checkpoint inhibition reactivity classification comprises a predictive score and one or more feature identifiers, and the one or more feature identifiers are selected from the group consisting of: feature valence, feature importance, and feature weight, wherein when executed by the one or more microprocessors, the instructions cause the graphical user interface to report the feature identifier as a characteristic of an annulus sector, wherein an angle of the annulus sector reports the importance of the feature, an outer radius of the annulus sector reports the weight of the feature, and a color of the annulus sector reports the valence of the feature.

In another embodiment, the feature importance of a feature comprises a reduction in the Gini index of the feature. In yet another embodiment, the instructions, when executed by the one or more microprocessors, cause the graphical user interface to report the identifier of the feature if and only if the feature importance of the feature is above a threshold. In yet another embodiment, the feature importance of the feature is not above the threshold if the square of the feature importance of the feature is not above 0.1. In yet another embodiment, the instructions, when executed by the one or more microprocessors, cause the graphical user interface to report the inner arc of each annulus sector and a circle comprised of the inner arcs of the annulus sectors. In yet another embodiment, the instructions, when executed by the one or more microprocessors, cause the computer system to further train the machine learning classifier, wherein further training comprises training the trained machine learning classifier in accordance with features of a tumor sample obtained from a non-training subject and responsiveness of the non-training subject to treatment.

In yet another aspect, a machine learning based classifier for immune checkpoint responsiveness classification is disclosed, the machine learning based classifier comprising: a machine learning-based classifier, running on a plurality of processors, trained to predict responsiveness of non-training subjects to immune checkpoint inhibition therapy, wherein the machine learning-based classifier is trained by inputting to the machine learning-based classifier genomic information of a plurality of training subjects and responsiveness of each of the plurality of training subjects to therapy, wherein the genomic information of the plurality of training subjects comprises features of a tumor profile obtained from each of the plurality of training subjects; inputting features of a tumor sample obtained from a non-training subject into an input processor in the machine learning based classifier, wherein the machine learning classifier is configured to generate a checkpoint inhibition responsiveness classification for the non-training subject that predicts responsiveness of the subject to a checkpoint inhibition therapy; and an output processor that reports the checkpoint inhibition reactivity classification. In one embodiment, the checkpoint inhibition reactivity classification includes a predictive score and a plurality of identifiers.

Brief Description of Drawings

These and other features, aspects, and advantages of the present disclosure will become better understood when the following detailed description is read with reference to the accompanying drawings, wherein:

the network diagram of fig. 1 shows options for performing a method according to aspects of the present disclosure.

Fig. 2 shows some non-limiting examples of features that may be relevant in training a classifier and predicting a patient's treatment responsiveness, according to aspects of the present disclosure.

The network diagram of FIG. 3 provides an example illustrating how the method of training a classifier may be implemented in accordance with aspects of the present disclosure.

The network diagram of fig. 4 gives an example illustrating how a method of predicting treatment responsiveness of a subject using a trained classifier may be implemented in accordance with aspects of the present disclosure.

Fig. 5 presents an example illustrating a method of reporting a subject's responsiveness to treatment predicted by a trained machine-learning based classifier in accordance with aspects of the present disclosure.

Fig. 6 presents an example illustrating a method of reporting a subject's responsiveness to treatment predicted by a trained machine-learning based classifier in accordance with aspects of the present disclosure.

Fig. 7 presents an example illustrating a method of reporting and comparing the responsiveness of a subject to different treatments as predicted by a trained machine-learning based classifier, in accordance with aspects of the present disclosure.

Fig. 8 presents an example illustrating a method of reporting and comparing the responsiveness of a subject to different treatments as predicted by a trained machine-learning based classifier, in accordance with aspects of the present disclosure.

Fig. 9 presents an example illustrating a method of reporting and comparing subject treatment responsiveness predicted by a trained machine-learning based classifier, wherein the prediction is performed when a gene set is not included or included as a feature, according to aspects of the present disclosure.

Fig. 10 presents an example illustrating a method of reporting a subject's responsiveness to treatment predicted by a trained machine-learning based classifier in accordance with aspects of the present disclosure.

The comparison curves and graphs of fig. 11A-11D show that using 38 features can yield superior classifier performance compared to using a single factor.

The graphs and figures of fig. 12A-12D compare the performance of a machine-learned classifier using 38 features that do not include a gene set or 44 features that include a gene set.

Detailed Description

The present disclosure relates to machine learning methods for obtaining a prediction of whether an individual may respond to checkpoint inhibitor therapy or other cancer therapy. Different individuals may or may not respond to a given treatment. Therapeutic responsiveness may not depend solely on the presence or absence of a given characteristic, or the amount of a given characteristic, in isolation. Conversely, multiple features may be combined differently between individuals such that some individuals are more likely to respond to a given treatment, while others have less likelihood of responding. Standard methods for predicting patient responsiveness based on a single feature, or even several features, will not accurately predict responsiveness in the following situations: in such a situation, a number of factors that may vary independently of one another act synergistically in some manner.

In response to this shortcoming of the mainstream diagnostic prediction approach, the machine learning approach constitutes a new solution. In supervised machine learning, for example, large data sets may be loaded into a computerMemory (memory storage)The data set above represents a number of characteristics of many individuals paired with each individual's known responsiveness to a given treatment. The storage medium of the computer contains instructions that direct the computer processor to process the characteristic information and the responsiveness of the individual to identify patterns in the characteristic information that may be predictive of a high or low likelihood that the individual will respond to a given treatment. An advantage of a machine learning method for such analysis is that it allows to recognize patterns in features and their correlation with predicted treatment responsiveness, which is not possible without the high data storage and retrieval capabilities of computer systems and their high amount of information processing capabilities. A computer-implemented machine learning system may process tens or hundreds of features or more of tens or hundreds of subjects or more to identify patterns of features and their integrated correlations with individual reactivity (aggregatate correlation). Patterns so recognized may otherwise be undetectable, requiring complex pairs of features as they are implementedA large amount of data in the information set of (2) is processed.

To determine characteristics of an individual, a tissue sample of the individual (e.g., an individual with cancer) can be obtained and characteristics of the tissue determined. In some embodiments, the obtained tissue may be a cell sample taken from a tumor. In other embodiments, the obtained tissue may be non-tumor tissue obtained from an individual.

Herein, checkpoint inhibition refers to a treatment that blocks the process by which tumor cells activate self-recognition pathways in the immune system and thereby prevent the tumor cells from being attacked and lysed by immune cells. It is believed that the activation of such pathways by tumor cells is a factor contributing to the refractoriness (responsiveness) of some patients to immunooncology, such as immunooncology in which immune cells are engineered to recognize and target tumor cells. Examples of checkpoint inhibition include the CTLA4 pathway. CTLA4 is a protein receptor expressed on regulatory T cells. It can also be expressed on activated conventional T cells, as is often seen in cancer. When CTLA4 binds to CD80 or CD87 (proteins expressed on the surface of antigen presenting cells), immunosuppression occurs. In healthy cells, this mechanism promotes self-recognition and prevents immune attack on the own cells. However, in cancer, upregulation of this pathway helps tumor cells escape detection and attack by the immune system. For example, an example of checkpoint inhibitor therapy that inhibits this pathway (e.g., anti-CTLA 4-antibody ipilimumab) may facilitate the ability of the immune system to target and destroy tumor cells when used in conjunction with other immunooncology therapies that stimulate anti-tumor immune reactivity.

Similarly, another example of checkpoint inhibition is the PD-1 pathway. The PD1-L1 receptor expressed on self cells binds to PD-1 on T cells, causing an immunosuppressive response. Like the CTLA4 pathway, this pathway can also be used by tumor cells to escape detection and attack by the immune system. Examples of checkpoint inhibitor treatments that inhibit this pathway, such as anti-PD-1 antibodies, pembrolizumab, nivolumab and cemiplimab, and anti-PD-L1 antibodies, atezolizumab, avelumab and devaluzumab, may promote the ability of the immune system to target and destroy tumor cells when used, for example, in conjunction with other immunooncology treatments that stimulate anti-tumor immunoreactivity.

As used herein, the term "checkpoint inhibitor" or "checkpoint inhibitory therapy" and the like includes these therapies as well as other therapies that inhibit checkpoint inhibitor pathways, including therapies that use other antibodies or pharmaceutical compounds that can act by preventing CTLA4 or PD-1 from interacting with its associated ligand or receptor or preventing its downstream signaling of subsequent events or activation of cellular function.

A number of characteristics may be relevant in predicting whether an individual is responsive to a given treatment. Examples may include: genetic sequence information contained in the genome of cells taken from the individual, genetic sequence information expressed in RNA transcribed from the genome of cells of the individual, the amount of expression of a transcript of the genomic sequence (which may be reflected in the amount of the corresponding RNA transcript in the sample or the amount of its protein product), or the type of cells present in the sample. In one embodiment where the sample comprises tumor tissue or cells from an individual, such information may be indicative of the identity of the tumor cells (i.e., cells having a modified genomic sequence(s) compared to the general population or non-tumor cells of the individual or a reference genome mentioned in the classic paradigm of genetic sequencing (paramigm)). Neoplasia may be caused by a change or changes in one nucleotide sequence in genomic DNA and/or by a change or changes in more than one sequence. In some cases, for example, accumulation of multiple such sequence modifications can work together to transform a cell from a non-diseased cell to a tumor cell. In other cases, initial accumulation of one or more such modifications in a cell can result in the cell having a susceptibility to accumulate other such modifications. In still other cases, expansion of such modifications in cells may indicate that not any particular modification is directly involved in or responsible for tumor formation. In contrast, the process of tumor formation may result in some modifications that may directly induce cell transformation into tumor cells, but may also result in other modifications that do not have this effect.

Thus, some individuals may have tumor cells with numerous such genomic DNA sequence modifications, while other individuals may have fewer. Among these modifications may be modifications that result in the transcript or protein product having an altered amino acid sequence as a result of a genomic modification, e.g., a modification of a DNA sequence in genomic DNA may result in the production of a protein or RNA molecule having a different sequence than would be produced if such modification had not occurred. Such modifications are referred to as non-synonymous mutations. Other modifications may be modifications of non-coding DNA, or may be modifications of coding DNA that do not alter the amino acid sequence of the protein. For example, modification of an intron sequence or non-transcribed DNA may not produce a protein product that differs in amino acid sequence from a protein produced from a genome that does not carry the same modification. Such modifications are referred to as synonymous mutations. Thus, different tumors can contain a different total number of genomic DNA modifications, including a different total number of non-synonymous mutations, a different number of synonymous mutations, or a different number of two types of mutations (or, the same total number of genomic mutations, but a different number of synonymous mutations and a different number of non-synonymous mutations). The number of mutations carried by a cell is referred to as the mutation load or total mutation load of the cell, wherein the total number of non-synonymous mutations carried by a cell is referred to as its non-synonymous mutation load and the total number of synonymous mutations carried by a cell is referred to as its synonymous mutation load.

Transformation of cells from non-tumor cells to tumor cells may correspond to the accumulation of such genomic DNA modifications. This accumulation can be a synonymous mutation accumulation, a non-synonymous mutation accumulation or both types of accumulation. In either case, the total mutation load may increase before, during, and after transformation of non-tumor cells into tumor cells. Furthermore, whether checkpoint inhibitors may be effective in stimulating the anti-tumor response of the immune system, the mutational load of tumor cells may have an impact. The more mutations a tumor cell carries, the more likely it will be that the suppressive checkpoint inhibition will also release the suppression of the immune system to recognize the tumor cell as a diseased cell and initiate an attack thereon. In particular, nonsynonymous mutational burden of a tumor may be positively correlated with the ability of checkpoint inhibition to abrogate the anti-tumor immune response. Proteins having mutated amino acid sequences, which are produced by non-synonymous mutations, can be recognized as abnormal in cells and presented on cell membranes as signals for the occurrence of disease states inside the cells. For example, tumors may express proteins with mutated amino acid sequences (called neoantigens). Tumor cells expressing such neoantigens may express mutated fragments of the neoantigen on their cell membranes.

This neoantigen presentation can stimulate the immune system to recognize diseased cells (e.g., tumor cells) and promote targeted destruction of such cells by the immune system. However, the counter-processes in tumors can escape immune probing using checkpoint pathways. Thus, whether a checkpoint inhibitor can aid in enhancing immunooncology therapy may depend on the mutational burden of the tumor. Higher loads may correspond to higher levels of neoantigen presentation, thereby increasing the probability of stimulating anti-tumor immunogenicity when checkpoint inhibition therapy is administered. A higher total mutation load may mean more neoantigen expression, since a higher total mutation load may mean a higher non-synonymous mutation load on average. In addition, a higher tumor non-synonymous mutation load may also mean more neoantigen expression, and thus a greater likelihood that checkpoint inhibition will be effective. It is also possible that the synonymous mutation burden of the tumor may be correlated with responsiveness to checkpoint inhibition, and/or that certain combinations of synonymous and non-synonymous mutation burden of the tumor (e.g. as may be reflected in the overall mutation burden) may be correlated with responsiveness to checkpoint inhibition. Thus, total mutational burden, tumor non-synonymous mutational burden, tumor synonymous mutational burden, or any combination of two or more thereof, may predict checkpoint responsiveness and may be the feature(s) included in the machine learning methods as disclosed herein.

In addition to whether the mutation is synonymous or non-synonymous, other characteristics or additional properties of the mutation may also be present, and its number or type, as with the mutation being synonymous or non-synonymous, may be of interest in predicting therapeutic responsiveness. For example, some mutations, called non-stop mutations (non-stop mutations), are mutations within the stop codon that result in translation of the resulting RNA product, since the mutated portion of the RNA transcript will continue to proceed past the position where it would otherwise stop. Another form of mutation is a frameshift mutation, comprising a plurality of consecutive nucleotides that are not divisible by three (e.g., single nucleotide insertions or deletions) that result in a shift in codon reading sequence, and thus will result in the recruitment of a different tRNA molecule during translation of the resulting RNA transcript and thus alter the amino acid sequence of the translated protein. Other mutations may be splice site mutations that occur at or near the splice site, thereby altering normal mRNA splicing and producing a modified RNA transcript. Alternatively, the mutation may be a missense mutation, in which a single nucleotide is altered to alter the codon containing it, whereby a different species of tRNA will be recruited during translation and thus produce a protein with a different amino acid sequence.

Another possible mutation may be an initiation mutation, which is a mutation to the transcription start site or to the initiation codon, resulting in a change in the position at which transcription or translation, respectively, starts. For example, a start site mutation may prevent transcription from initiating from the start site. Alternatively, the mutation may create a transcription initiation site that was not previously present. A mutation in the transcription start site can result in transcription of an RNA transcript, which, although of a different length than the transcript it originally produced, can be in-frame or out-of-frame relative to the transcript produced in the absence of the mutation. Similar start codon mutations may also occur, resulting in transcription of an RNA product from which no translation initiation occurs, or from a position previously non-initiation. Such stop codon mutations may also be in-frame or out-of-frame. Alternatively, the mutation may be a nonsense mutation, i.e., the mutation results in an RNA transcript having a premature stop codon. Any of the foregoing mutations may be Single Nucleotide Polymorphisms (SNPs). Any one or more of the foregoing types of mutations can be a characteristic that correlates with predicting an individual's responsiveness to a given treatment (e.g., a given checkpoint inhibitor).

In another embodiment, responsiveness to a checkpoint inhibitor can be predicted by the amount of tumor infiltrated by lymphocytes. Tumors contain not only cells that have been transformed from non-tumor cells to tumor cells, but also other non-transformed cells. Examples include cells of the immune system within a tumor that do or may have the effect of stimulating an immune response against transformed cells. Immune cells, especially lymphocytes, which are mixed with transformed cells in tumors are called tumor infiltrating lymphocytes. The level of tumor-infiltrating lymphocytes, and the level of tumor-infiltrating lymphocytes expressing different markers as markers of lymphocyte phenotype, can predict whether a subject from which a tumor sample is taken will respond to checkpoint inhibition. Since the tumor is, for example, a heterogeneous mixture of tumor cells and tumor-infiltrating lymphocytes, it may be advantageous to distinguish between lymphocyte markers expressed on tumor-infiltrating lymphocytes present in the sample and lymphocyte markers potentially expressed on other cells (e.g., transformed tumor cells) in the tumor sample.

For example, tumor infiltrating lymphocytes may express transcripts (e.g., RNA) of genes encoding differentiation antigen cluster 8(CD8), differentiation antigen cluster 4(CD4), or differentiation antigen cluster 19(CD19), each of which may serve as markers of lymphocyte phenotype when expressed in the cell. Thus, the expression level of CD8, CD4, CD19, or any combination of any two or more of the foregoing may be determined from a tumor sample. For example, the amount of RNA can be determined from the sample. Since this determination may reflect not only the expression by tumor-infiltrating lymphocytes, but also by other cells in the tumor sample (e.g., transformed tumor cells), it may be advantageous to determine how much of the detected expression is due to, and not how much is due to, the expression of tumor-infiltrating lymphocytes. For this purpose, a deconvolution (deconvolution) method can be applied, whereby the expression level by tumor-infiltrating lymphocytes can be determined relative to the expression by other cells. Various alternatives are available for effecting deconvolution of tumor infiltrating lymphocytes, including, for example, Gaujoux et al (2013) CellMix: aerobic reactive decoding. bioinformatics 29: 2211-; and Finotello et al (2018), Quantifying tumor-encapsulating animal cells from transcriptics data, Cancer Immunology, Immunology 67: 1031-1040. As a non-limiting embodiment, the deconvolution analysis may be performed in the R programming language.

In particular, the expression level of a given lymphocyte transcript (e.g., CD8, CD4, or CD19) can be used to determine what percentage of lymphocytes in a tumor sample are CD 4-expressing or CD 8-expressing or CD 19-expressing lymphocytes by a method known as tumor infiltrating lymphocyte deconvolution. That is, rather than merely indicating the amount of lymphocyte infiltration by a tumor (as represented by the amount of a given lymphocyte transcript that can be recognized in a tumor), deconvolution of tumor-infiltrating lymphocytes can also indicate the percentage of lymphocyte types identified by the type of transcript expressed in the total infiltration of a tumor. The total amount of lymphocyte infiltration by a tumor may be relevant in predicting whether an individual will respond to a given treatment, such as a checkpoint inhibitor, and the amount of tumor lymphocyte infiltration specifically contributed by lymphocytes expressing a given transcript (e.g., without limitation, CD4 or CD8 or CD19) may also be relevant in such a prediction.

Various other features may be of interest in predicting whether an individual will respond to checkpoint inhibition. These characteristics can generally be classified according to a process that is theoretically related to whether checkpoint inhibitors will be effective in promoting or enhancing tumor immune responses. For example, some features may be associated with: whether and to what extent tumor cells are likely to express mutein antigens on their cell surface and thus increase the chance of an anti-tumor immune response. Examples of such features that have been discussed include types of tumor mutational burden, such as total tumor mutational burden or tumor non-synonymous mutational burden. Other examples include the expression levels of different proteins or gene transcripts encoding proteins known to be involved in stimulating antigen presentation of an immune response. Some non-limiting examples include β 2 microglobulin (B2M), proteasomal subunit β 10(PSMB10), antigenic peptide transporter 1(TAP1), antigenic peptide transporter 2(TAP2), human leukocyte antigen a (HLA-a), major histocompatibility complex class I B (HLA-B) expression, major histocompatibility complex class I C (HLA-C), major histocompatibility complex class II DQ α 1(HLA-DQA1), and HLA class II histocompatibility antigen DRB1 β chain (HLA-DRB 1). These gene products are known to play multiple roles in presenting protein fragments or antigens on the cell surface and for immune T cell recognition.

Expression levels of any one of these gene products, or any combination of two or more, or any of these in a tumor, can be indicative of the level of antigen expression on the surface of the tumor cells. For example, expression of such products can affect the extent to which protein products of genomic DNA carrying non-synonymous mutations are presented on the cell surface. In the case of expression of a mutant antigen that increases antigen presentation, the likelihood of presenting the mutant antigen for which T cells may recognize as a signal of diseased cells and thus trigger an anti-tumor immune response may increase the likelihood that a given checkpoint inhibitor may effectively elicit a response in a subject. Thus, where the expression level of any one or a combination of two or more of the foregoing is directly or negatively correlated with the extent of antigen presentation on cells in a tumor sample, such expression level can be directly or negatively correlated with the likelihood, respectively, that a subject from which the tumor sample will respond to a given checkpoint inhibitor.

Another type of characteristic may include the expression level of T cells or NK cells present in a tumor sample taken from the subject, which may also be correlated with a prediction of responsiveness to a treatment (e.g., checkpoint inhibition). For example, the expression level of HLA class I histocompatibility antigen alpha chain E (HLA-E), natural killer cell particle protein 7(NKG7), chemokine-like receptor 1(CMKLR1), or any combination of two or more thereof, may also be predictive of a response to a checkpoint inhibitor. Thus, a feature may include an amount of expression of one or more of these products or their RNA transcripts.

Another type of feature may be related to the presence or expression level of a protein or transcript thereof associated with or indicative of enhanced lytic activity (e.g., that may be elicited by an anti-tumor immune response), or that the protein or transcript thereof may inhibit such activity. The tumor-infiltrating lymphocyte deconvolution magnitudes discussed above may be examples of such features (e.g., tumor-infiltrating deconvolution of differentiation-antigen cluster 8(CD8), differentiation-antigen cluster 4(CD4), or differentiation-antigen cluster 19(CD19) -expressing cells). Other non-limiting examples of characteristics of this class may include the expression level of granzyme a (gzma) or perforin-1 (PRF1) or any combination of two or more of the foregoing, or RNA transcripts thereof.

In addition, other features may be associated with processes or functions in checkpoint inhibition of anti-tumor immune reactivity. The expression levels of various protein products, or transcripts thereof, that are involved in checkpoint inhibition in a tumor sample from a subject can be of interest in predicting whether a subject will respond to treatment with a cancer therapy (e.g., checkpoint inhibition therapy). Examples of such features may include expression of cytotoxic T lymphocyte-associated protein 4(CTLA4), programmed cell death protein 1(PD1), programmed death-ligand 1(PDL1), programmed cell death 1-ligand 2(PDL2), lymphocyte activation gene 3(LAG3), T cell immunoreceptor with Ig domain and ITIM domain (TIGIT), differentiation antigen cluster 276(CD276), or expression of two or more of any of the foregoing or expression of RNA transcript thereof.

Other features that may be of interest in predicting whether an individual will respond to treatment include expression of proteins or RNA transcripts thereof, wherein the proteins are associated with interferon gamma activity, e.g., products expressed downstream of interferon gamma release and activity at a receptor. Examples of this type of feature may include expression of chemokine (C-C motif) ligand 5(CCL5), CD27, chemokine (C-X-C motif) ligand 9(CXCL9), C-X-C motif chemokine receptor 6(CXCR6), indoleamine 2, 3-dioxygenase (IDO), signal transducer and activator of transcription 1(STAT1), or any combination of two or more of the foregoing, or expression of an RNA transcript thereof. Other indicators of interferon gamma activity may also be predictive of therapeutic (e.g., checkpoint inhibition) responsiveness.

Other features that may be of interest in predicting whether an individual will respond to treatment include expression of proteins, or their RNA transcripts, associated with Myeloid Derived Suppressor Cells (MDSCs) or regulatory T cells (tregs), which can confer an immunosuppressive effect on anti-tumor immune reactivity and can impair or hinder the effectiveness of immunooncological treatments. Examples of such features may include: expression of 3-fucosyl-N-acetyl-lactosamine (CD15), interleukin-2 receptor alpha chain (CD25), siglec-3(CD33), cluster of differentiation antigen 39(CD39), cluster of differentiation antigen (CD118) expression, forkhead box P3(FOXP3), or any combination of two or more of the foregoing in a tumor sample from a tumor of a subject. The tumor expression level of other species of proteins or corresponding RNA transcripts, indicative of the presence or activity of such cells, may also be correlated with whether the individual will respond to checkpoint inhibitor therapy or other cancer therapies.

Any of the foregoing characteristics of any one or more may have varying degrees of relevance for predicting whether an individual will respond to a given cancer therapy treatment, including checkpoint inhibitor treatment. When a tumor of a subject is sampled and tested to determine characteristics, any of these characteristics may relate to or embody genomic information relating to the tumor of the subject. In this context, the term "genomic" is intended to include not only information about nucleotide sequences in genomic DNA (e.g., features associated with mutation load). In this context, the genomic information represented by the characteristic magnitudes can also include expression level magnitudes for a plurality of genomic transcripts or protein products produced from such transcripts. Thus, the expression level of any of the different protein products described above, or the expression level of any of the other protein products involved in similar pathways, or the expression level of an RNA transcript thereof, may be included in the genomic information relating to the predictive features herein. The genomic information associated with the features may also include magnitudes of tumor infiltrating lymphocyte deconvolution features.

In addition to the magnitude of a single feature, the pattern of correlated expression levels of multiple features known or believed to be associated with a given pathway or function or cell type may also be a feature that correlates with checkpoint inhibition or responsiveness to other cancer treatments. For example, among the aforementioned features, a cohort of features can be identified that have commonality in pathway or cellular or physiological responsiveness or in indicative cell phenotype, and whether the features are synergistically up-or down-regulated as a cohort, or more generally, expressed or present at related high or low levels as a cohort, in a sample from a tumor of a given subject can be determined based on the magnitude of the individual features. In some embodiments, the magnitude of such a generalized measure of packet characteristics may itself be used as a feature input, in addition to individual characteristics, for training a machine learning classifier, or predicting a subject's responsiveness to checkpoint inhibition or other treatment, or both. Such a grouping of features is referred to herein as a gene set (gene set), and additional features may be obtained to represent the level of expression of the grouping as a whole, etc. Thus, a gene set can include a combined quantity value that indicates a correlation of the presence of a genomic mutation, the expression level of a particular RNA transcript, the presence of an identified cell type, and the like.

As a non-limiting example, some of the aforementioned characteristics are associated with antigen presentation by which cells, such as tumors, express protein fragments on their cell membranes for monitoring by the immune system. As described above, antigen presentation may increase the likelihood of stimulating an anti-tumor immune response, for example with checkpoint inhibitors. Some examples of such features may include total, non-synonymous or other mutant loads (no termination, frameshift, or deletion, or either), splice site, missense, initiation (in-frame, out-of-frame, or either), nonsense, initiation codon, including initiation codon SNPs or others, in-frame insertion, in-frame deletion, or other SNP, or any combination of two or more of the foregoing Major histocompatibility complex class I B (HLA-B) expression, major histocompatibility complex class I C (HLA-C), major histocompatibility complex class II DQ α 1(HLA-DQA1), and HLA class II histocompatibility antigen DRB1 β chain (HLA-DRB 1). In addition to features related to the presence or expression levels, etc., associated with individual instances of the aforementioned features, additional features may represent the degree to which some or all of the aforementioned features are synergistically up-or down-regulated or present at high or low levels in a subject's tumor (for machine learning classifier training or prediction).

As another non-limiting example, some characteristics are associated with the expression levels of T cells or NK cells present in a tumor sample of a subject, which may also be correlated with predicting treatment (e.g., checkpoint inhibition) responsiveness. For example, the expression level of HLA class I histocompatibility antigen alpha chain E (HLA-E), natural killer cell particle protein 7(NKG7), chemokine-like receptor 1(CMKLR1), or any combination of two or more thereof, may also be predictive of a response to a checkpoint inhibitor. In addition to features related to the presence or expression levels, etc., associated with individual instances of the aforementioned features, additional features may represent the degree to which some or all of the aforementioned features are synergistically up-or down-regulated or present at high or low levels in a subject's tumor (for machine learning classifier training or prediction).

As another non-limiting example, some characteristics relate to an indicator of lysis of an immune stimulus present in a tumor sample of a subject (e.g., when an immune response promotes, e.g., cell death and cell lysis of tumor cells), and may also be correlated with predictive therapy (e.g., checkpoint inhibition) responsiveness. For example, deconvolved CD8 expression, deconvolved CD4 expression, deconvolved CD19 expression (deconvolution means the proportion of contribution given by cells expressing CD8, CD4 or CD19 relative to the number of tumor-infiltrating lymphocytes present in the tumor sample), the level of expression of granzyme a (gzma) or perforin-1 (PRF1), or any combination of two or more of the foregoing, or the level of expression of an RNA transcript thereof, may also be predictive of response to a checkpoint inhibitor. In addition to features related to the presence or expression levels, etc., associated with individual instances from the aforementioned features, additional features may represent the degree to which some or all of the aforementioned features are synergistically up-or down-regulated or present at high or low levels in a subject's tumor (for machine learning classifier training or prediction).

As another non-limiting embodiment, certain features are associated with cellular and molecular processes involved in checkpoint inhibitory function that are present in a tumor sample taken from a subject, and these features may also be correlated with predicting responsiveness to a treatment (e.g., checkpoint inhibition). Non-limiting examples of such features may include expression of cytotoxic T lymphocyte-associated protein 4(CTLA4), programmed cell death protein 1(PD1), programmed death-ligand 1(PDL1), programmed cell death 1 ligand 2(PDL2), lymphocyte activation gene 3(LAG3), T cell immunoreceptor with Ig domain and ITIM domain (TIGIT), differentiation antigen cluster 276(CD276), or any two or more of the foregoing, or expression of RNA transcript thereof. In addition to features related to the presence or expression levels, etc., associated with individual instances from the aforementioned features, additional features may represent the degree to which some or all of the aforementioned features are synergistically up-or down-regulated or present at high or low levels in a subject's tumor (for machine learning classifier training or prediction).

As another non-limiting embodiment, some features are associated with an indicator of interferon gamma activity or cellular and molecular pathways present in a tumor sample taken from the subject, which may also be correlated with predicting responsiveness to treatment (e.g., checkpoint inhibition). Non-limiting examples of such features can include expression of chemokine (C-C motif) ligand 5(CCL5), CD27, chemokine (C-X-C motif) ligand 9(CXCL9), C-X-C motif chemokine receptor 6(CXCR6), indoleamine 2, 3-dioxygenase (IDO), signal transducer and activator of transcription 1(STAT1), or any combination of two or more of the foregoing, or expression of an RNA transcript thereof. In addition to features related to the presence or expression levels, etc., associated with individual instances from the aforementioned features, additional features may represent the degree to which some or all of the aforementioned features are synergistically up-or down-regulated or present at high or low levels in a subject's tumor (for machine learning classifier training or prediction).

As another non-limiting embodiment, some features are associated with the presence or activity of MDSCs or tregs present in a tumor sample taken from the subject, which may also be correlated with predicting responsiveness to a treatment (e.g., checkpoint inhibition). Non-limiting examples of such features may include expression of 3-fucosyl-N-acetyl-lactosamine (CD15), interleukin-2 receptor alpha chain (CD25), siglec-3(CD33), cluster of differentiation antigens 39(CD39), cluster of differentiation antigens (CD118) expression, forkhead box P3(FOXP3), or any combination of two or more of the foregoing, in a tumor sample from a tumor of a subject. In addition to features related to the presence or expression levels, etc., associated with individual instances from the aforementioned features, additional features may represent the degree to which some or all of the aforementioned features are synergistically up-or down-regulated or present at high or low levels in a subject's tumor (for machine learning classifier training or prediction).

Thus, in some embodiments, one or more gene sets may be identified, and a measure of the cooperativity or degree of cooperativity of features related to the gene sets up or down-regulated in a tumor of a subject may be provided as an additional feature for training a machine learning classifier or predicting responsiveness of a subject to checkpoint inhibition or other therapy, or both. Examples of gene sets include those associated with antigen presentation, T cell and NK cell signature (signature), indices of lysis, checkpoint inhibition, interferon gamma, and MSDC/Treg presence or activity. In some cases, one or more such gene sets may be included in training the machine learning classifier, along with any other single feature or features discussed above, and used to predict patient responsiveness to checkpoint inhibition or other therapy. Various methods can be used to determine a generalized measure that can measure how features within the gene set are synergistically up-or down-regulated, or synergistically expressed or present at high or low levels or in a correlated manner. One embodiment may include an assay known as single sample gene set enrichment assay (ssGSEA). ssGSEA uses an empirical cumulative distribution function to determine this grouping enrichment of a gene set, as described, for example, in Barbie et al (2009), Systematic RNAIntermediate references present at genetic KRAS-drive recipients require TBK1, Nature462: 108-. As a non-limiting example, ssGSEA may be implemented in the R programming language.

For the foregoing features, some or any combination of any two or more features may be used to train a machine learning classifier and to use the trained machine learning classifier to predict whether or not a subject will respond to a treatment, such as a checkpoint inhibitor treatment, or how likely the subject will respond to a treatment, such as a checkpoint inhibitor treatment. However, it is not necessary that all of the aforementioned features be used to train the machine learning classifier. The machine learning classifier may be trained with a feature set that includes all of the foregoing features or excludes any one or more of the foregoing features. All combinations and permutations presented in accordance with this optional inclusion and exclusion are herein incorporated in their entirety without being explicitly listed one by one. The skilled person will be able to conceive of possible subsets, combinations, sub-combinations and permutations in which the aforementioned features are employed. Likewise, where all of the aforementioned features are used, or where combinations, subcombinations, permutations, or other mixtures of features are used with less than all of the aforementioned features, additional features may be further included. The present disclosure expressly incorporates all of these various embodiments.

In some embodiments, the features of any subject used to train the machine-learned classifier may be the same as the features of any and all other subjects used to train the machine-learned classifier. However, in other embodiments, different features may be provided for different subjects used to train the machine learning classifier. In other words, some subjects may have features that are included in the training set of the subject, but not in the features of the training set from other subjects. Similarly, in some embodiments, to obtain a prediction from a machine learning classifier relating to the responsiveness to treatment of a subject, the features obtained from a tumor sample of the subject for such prediction may be the same as the features used to train the classifier. That is, the features of all subjects used to train the machine-learned classifier may all be the same as each other, and also the same as the features of the subject used to obtain the prediction results from the machine-learned classifier. In other embodiments, there may be a mismatch between the features of the training subject used to train the machine-learned classifier and the features of the subject used to obtain the prediction from the machine-learned classifier. The features of some or all of the subjects used to train the machine-learned classifier may include features where the corresponding features are not present in the subject features used to obtain the prediction from the machine-learned classifier.

In some embodiments, for subjects using a machine-learned classifier to obtain a prediction, features corresponding to features from one or some or all of the subjects used to train the machine-learned classifier may be absent. In other embodiments, the subject may have similar characteristics, but not identical characteristics, and similar characteristics may be used to replace identical characteristics that the subject lacks. For example, the machine learning classifier may have been trained on features that include one or more gene set features, such as may be obtained using ssGSEA as described above, for at least some of the training subjects. For some training subjects whose gene sets were used to train machine learning classifiers, some of the gene sets may be obtained from the same underlying single features. For example, a gene set of antigen presentation-related gene set features may be obtained from the same underlying features of all subjects used to train the machine learning classifier. In other embodiments, the antigen presentation related gene set of one training subject may be based on a basic characteristic comprising some features that are not included in determining the antigen presentation related gene set of another training subject. For other gene sets, the same may apply. Additionally, for subjects seeking a prediction result using a trained machine learning classifier, the prediction may be obtained using the subject's gene set features, wherein the feature values of the gene set may be obtained from a basic single feature set from the subject that does not include at least one or more of the basic features used to obtain the corresponding gene set feature values in one or more training subjects and to train the machine learning classifier.

Genetic sequencing data or the level of expression of a protein or RNA transcript can be determined in a biological sample by known methods to determine a characteristic. For example, significant amounts of nucleotide sequence information can be obtained using next generation sequencing techniques, thereby providing genome-related features (e.g., total mutation load, etc.) as well as expression levels (e.g., of RNA transcripts), depending on the type of next generation sequencing used to obtain a given feature. Examples of suitable methods include whole genome sequencing, whole exome sequencing, whole transcriptome sequencing, mRNA sequencing, gene array analysis, RNA array analysis, protein analysis such as protein arrays, or other related methods that may be used to determine the presence or level or amount of features that may be used to train a machine learning classifier and/or obtain predictions therefrom according to aspects of the present disclosure. In some embodiments, the same set of techniques may be used to obtain features from all training subjects training the machine learning classifier and from subjects seeking predictive outcomes. In other embodiments, the method for determining the feature(s) in different training subjects, and/or the method for determining the features for use in the prediction in subjects receiving the prediction, may be different.

In addition to training the machine-learned classifier with features from the training subjects, the responsiveness of the training subjects to the treatment is loaded into the machine-learned classifier. Thus, a training subject is a subject who provides its features and responsiveness to train a machine learning classifier. Responsiveness can be a binary classification, e.g., a training subject is classified as responsive to treatment when the subject exhibits a predetermined response (including prolonged lifespan, shrinkage of tumor, partial or complete remission, etc.). In other embodiments, reactivity may be a score or value based on the degree of reactivity obtained, rather than a binary assessment of the presence or absence of reactivity. For example, as a non-limiting embodiment, the machine learning classifier of the present disclosure may include classification and regression trees, depending on the type of prediction sought.

The machine learning classifier may be any classifier suitable for computer-based machine learning. Non-limiting examples include random forest machine learning classifiers. In a random forest machine learning classifier, a decision tree is created based on training subjects' features and treatment responsiveness values, with nodes representing classification decision points, and leaves representing decision results based on training inputs. The random forest classifier may generate a plurality of trees using the feature subset and the training subject subset to create a multitude of trees that may be subsequently integrated (aggregated). Such multiple decision trees with input subsets may prevent overfitting and reduce errors and bias in prediction. In some embodiments, the more decision trees created during training, the more accurate the machine-learned classifier can be derived. In some embodiments, any number of decision trees from 5,000 to 500,000 may be generated during training. For example, 5,000, 10,000, 15,000, 20,000, 25,000, 30,000, 50,000, 75,000, 90,000, 100,000, 125,000, 150,000, 175,000, 200,000, 225,000, 250,000, 275,000, 300,000, 325,000, 350,000, 375,000, 400,000, 425,000, 450,000, 475,000, or 500,000 decision trees may be run and integrated in a random forest machine learning classifier. Alternatively, more or fewer decision trees may be run, as many as between these exemplary possibilities.

Numerous alternative methods are available for performing random forest training and generating predictions from random forest classifiers. As a non-limiting example, the R programming language may be used. Other classifiers may also be used in accordance with aspects of the present disclosure, including, without limitation, neural network classifiers, support vector machines, maximum entropy classifiers, extreme gradient boosting classifiers, and random fern classifiers.

According to the methods disclosed herein, features are obtained from each of a number of training subjects, and the responsiveness of each training subject is obtained. These characteristics and reactivities, or entries, are entered into a computer memory (memory store), such as a hard drive, server, or other storage device. Also stored on the computer's memory are instructions, which may be embodied in software, that instruct one or more microprocessors. The instructions include instructions to create a machine-learned classifier using input from a training subject. The trained machine-learned classifier is then stored in one or more computer memories, and may then be run on the features of the subject for which a responsiveness prediction is to be made.

As one non-limiting example, the instructions may instruct the one or more microprocessors to perform random forest training, create decision trees from training subject inputs to determine whether the presence, absence, level, etc. of the plurality of features is more or less likely to indicate treatment responsiveness, and integrate numerous decision trees into a trained random forest machine learning classifier. According to this embodiment, the trained machine learning classifier, based on the integration of one or more decision trees generated by the microprocessor in processing features and reactivities according to instructions, may then be stored in one or more memories. Subsequently, when predicting whether a non-trained subject (i.e., a subject whose feature values are not used to train the trained machine learning classifier) will respond to a therapy (e.g., a particular checkpoint inhibitor), the features of the tumor sample obtained from the non-trained subject can be loaded into one or more memories. The one or more microprocessors can process instructions such that features of a non-trained subject can be analyzed by the trained machine learning classifier, accessed from one or more memories by the one or more microprocessors, and a prediction related to subject responsiveness reported.

In this embodiment, the machine learning classifier is a trained machine learning classifier that has been trained based on features of tumor samples obtained from each of a plurality of training subjects and responsiveness of each of the plurality of training subjects to therapy (including checkpoint inhibition), wherein the machine learning classifier is trained to predict responsiveness to therapy. Genomic information of the non-trained subject (including non-trained features, or a set of feature values, from the subject's tumor profile) is input to the trained machine learning classifier to produce a therapeutic responsiveness classification of the non-trained subject, e.g., a classification or score that can be indicative of a therapeutic responsiveness prediction outcome of the non-trained subject. The treatment may be checkpoint inhibition.

Checkpoint inhibition reactivities of non-trained subjects generated by the trained machine learning classifier can be reported to a user. A report may include a classification of the subject that indicates in a binary manner whether an untrained subject is predicted to respond to treatment. In other embodiments, the likelihood of a response may be indicated by a numerical score in addition to or instead of a binary classification of whether the non-training subject is predicted to respond. In other embodiments, a specific degree of reactivity may be reported. For example, the report may indicate a high reaction probability, but the reactivity may be described in terms of duration or degree. In other examples, where non-training subjects are predicted to respond, the report may give a predicted outcome with a relatively low likelihood of responding in a greater duration or degree.

The prediction report (in the form of a score or in the form of a binary prediction that may or may not react) may be reported through a Graphical User Interface (GUI). For example, a computer or computer system connected to one or more memories and one or more microprocessors, on which a signature spectrum of a non-training subject is input and analyzed by a trained machine learning classifier, may also be connected to a display device, on which the prediction is visually reported. The GUI may take any of a number of forms. For example, the GUI may be tabulated, showing aspects of the feature or subset of features that have a higher importance or weight in generating the predicted outcome, and whether these factors indicate a greater or lesser likelihood, respectively, of non-training subject therapeutic responsiveness (i.e., the potency of the feature) based on the feature value of the non-training subject, as further explained below. Alternatively, the report may include different shapes, shading, or color schemes for reporting such information.

The subject matter disclosed herein is distinguished from conventional methods for predicting treatment responsiveness using a predictive classification report of clustering of features (constellations) as disclosed herein. Unlike conventional therapy prediction methods, disclosed herein is a method of using a combination approach in which, in some embodiments, a large number of features may be queried in association with each other in generating a prediction, rather than querying one or just a few features in isolation. This feature of the present disclosure, which is different from conventional methods, is advantageous because efforts to predict whether an individual will respond to a given treatment (e.g., checkpoint inhibitor) based on a limited number of features have heretofore had very limited accuracy and universality. As disclosed herein, the non-conventional machine learning approach may overcome this limitation of conventional prediction approaches by evaluating the contributions of numerous factors and how each factor may influence the prediction independently and in concert with other factors.

Conventional methods for determining responsiveness prediction results involve identifying one feature or a limited number of features that may indicate a higher or lower likelihood of a therapeutic response. In contrast, disclosed herein is an unconventional approach in which the relative contributions of perhaps tens of features or more are simultaneously evaluated in the correlation of features to one another using machine-learned classifiers to form predictions. Through the application of non-conventional machine learning classifiers, this multi-factor approach provides significant benefits over currently available approaches.

One advantageous feature of some GUI embodiments for reporting reactive predictors may be to present multifaceted information (which relates to features related to the reported predictor) in a relatively tight or small space so that a user can determine meaningful information in a compact manner. In particular, one advantage of certain GUI embodiments used in accordance with the present disclosure may be their size and layout on displays of limited or reduced size or on displays that also require or desire to present significant amounts of other information. The GUI for reporting the predictive classification results of the present disclosure may function as follows: small aspects of the large number of easily identifiable and recognizable categories of prediction classifications and/or features relevant to their generation are compressed in a limited display size or limited portion of the display. Such a compressed report may improve the user's interface with the computer system that gave the report, expedite the receipt and interpretation of the report, and save display space that may be needed for other purposes or may have limited availability, for example, where the display is a screen of a portable electronic device such as a telephone or other wireless communication device (e.g., a tablet computer or telephone or other portable wired or wireless device). In one embodiment where the report is presented by a GUI that compresses a large number of features and aspects of the features into compacted space, yet retains the ability to quickly transfer a large amount of information that can still be quickly and easily determined, the usability of the computer system is improved because more display space can be simultaneously left for others or a smaller display can be used.

The report may include feature identifiers, or aspects or characteristics of some or all of the features used in generating the reactivity classification. For example, the report may include an indication of the potency of the feature, i.e., whether its value reflects an increased or decreased likelihood that the subject will produce a therapeutic response in the user's profile. In other words, for profile of a non-training subject, one feature may have a positive or negative titer. A positive titer can refer to a value of the feature in the training subject that tends to correlate positively with the treatment responsiveness and a value of the feature in the non-training subject that is high, or a value of the feature in the training subject that tends to correlate negatively with the treatment responsiveness and a value of the feature in the non-training subject that is low.

Another identifier may indicate the importance of a feature in predictive outcome generation for a given machine-learned classifier. The importance of one feature may indicate that it is more likely to push the prediction in one direction or another relative to other features used in training. For example, in some embodiments, the Gini reduction index for one or more features may be determined during training. The Gini reduction index may indicate the importance of a feature because it explains the impact that a feature has on the classifier function relative to the amount of impact other features have in driving the generation of predictions. The Gini reduction index may be determined using a variety of software packages, such as by using the R programming language. In some embodiments, the importance of a feature may be used to determine which identifiers of features are included in the report, either independently of the predicted score or in addition to the predicted score. For example, for a given report reported in the form of a GUI, the GUI may display only identifiers for those features whose importance meets or exceeds a predetermined minimum importance threshold. For example, a report may contain an identifier of only features whose significance, when expressed in numerical form as Gini reduction indices, is squared to more than 0.1. Alternatively, a more or less stringent minimum importance threshold may be set, or changed for a given report (which may depend on the amount of information desired to be included in the report along with the prediction score). The higher the minimum importance threshold, the fewer feature identifiers that may be included in the report, and vice versa. For example, the minimum importance threshold may be one in which the square of the numerical importance (e.g., Gini reduction index) is higher than any value between 0.01 and 0.5. Other minimum importance thresholds may be selected between or outside of this range.

Another example of a feature identifier that may be included in a report is a weight of a feature. The weight of a feature is a quantity that measures how well the value of the feature, for a subject, will suggest to the subject itself that the subject will respond to treatment or will not respond to treatment. For example, for each feature, a single-factor decision boundary may be determined. The one-way decision boundary is the value of a feature that can best distinguish between training subjects that respond to treatment and training subjects that do not. For example, when all training subjects who respond to treatment have a characteristic value above a given amount, and all subjects who do not respond to treatment have a characteristic value below that amount, then that amount may be a one-way decision boundary. In some other embodiments, some responsive training subjects may have higher feature values than some non-responsive training subjects, while other responsive training subjects may have lower feature values than these training subjects. Thus, in some embodiments, for a certain characteristic value, there may be a clear line that clearly distinguishes between a responder and a non-responder; while for other eigenvalues, there may be more overlap of eigenvalues between reactors and non-reactors at the boundary. For the latter type of example, the one-way decision boundary may be selected to be the value that provides the greatest possible distinction between responders and non-responders in the training subject set.

The weight of a feature may measure or indicate: how far the feature value of the non-training subject is from the single feature decision boundary of the feature. The greater the difference between the feature value of the non-training subject and the one-way decision boundary based on the feature of the training subject, the greater the weight that the feature may have in determining the reactive class prediction. For example, one feature may have a negative titer, meaning that the non-training subject has a low value on the feature that is positively correlated with reactivity in the training subject, or a high value on the feature that is negatively correlated with reactivity in the training subject. A feature may have a high weight if its untrained subject value is significantly different from the single-factor decision boundary of the feature. However, if another feature has a higher importance and positive valence (i.e., the feature has a high value in non-trained subjects and is positively correlated with responsiveness in trained subjects, or the feature has a low value in non-trained subjects and is negatively correlated with responsiveness in trained subjects), then the feature may have a proportionally stronger impact on the responsiveness prediction classification even though its value is less different (i.e., has less weight) from the one-way decision boundary.

There are a number of possible ways in which the characteristic identifier may be exposed as an element of the responsiveness prediction report and GUI. Several specific embodiments are shown herein with certain details. However, the skilled person will appreciate that there are many other possible scenarios for reporting the potency and weight and importance of a feature in a GUI report, and the embodiments presented herein are not limiting or are not particularly required in any way.

The GUI may present the tabulated features in rows and columns. The features may be shown, for example, in rows, and different characteristics of a given feature may be shown in multiple columns. For example, different columns may indicate the importance of a feature, its potency, a one-way decision boundary for the feature, a non-training subject value for the feature, optionally and visually indicating a difference between the non-training subject value for the feature and the one-way decision boundary for the feature, and whether the non-training subject would be predicted to be a responder when performing predictions based on the feature alone. The tabular GUI may include any combination of two or more of the foregoing. The tabular GUI report may also include an overall predicted score.

The GUI may also be a histogram. For example, a bar may indicate the value of a non-trained subject for a given feature, while a line may indicate a one-factor decision boundary for that feature. The difference between the feature value of the non-training subject and the line indicating the one-factor decision boundary is an identifier of the weight of the feature. A line may also be drawn between the column height and the line indicating the single factor decision boundary. The length of the line between the two is also an indicator of the weight. Titers may be indicated by symbols below the bars. For example, a plus sign or a minus sign may indicate a positive titer or a negative titer, respectively. Other pairing schemes may include up and down pointing arrows, positive and inverted triangles, etc., where one direction indicates a positive valence of the feature and the other direction indicates a negative valence. Valence may also be indicated by the color or shading of the bars, one color or shading pattern of bars indicating one valence (positive or negative) and a different color or shading pattern of bars indicating a relative valence. The color or shading of the line between the value of the feature of the non-training subject in the histogram and the one-factor decision boundary for that feature may also indicate valence. For example, if a certain feature value is negatively correlated with reactivity in a training subject and the non-training subject value of the feature is below a single-feature decision boundary, the line in the histogram report connecting the non-training subject value of the feature and the single-factor decision boundary may have a color or shading indicating a positive valence. Whereas if a certain feature value is positively correlated with reactivity in a training subject and the non-training subject value of the feature is below the single-feature decision boundary, the line in the histogram report connecting the non-training subject value of the feature and the single-factor decision boundary may have a color or shading indicating a negative valence. Note that in both cases, the non-training subject value for a feature is less than the one-way decision boundary for that feature, but its potency is different depending on whether the feature is negatively (positive potency) or positively (negative potency) related to reactivity in the training subjects. The reverse is true (i.e., when the one-way decision boundary of a feature is less than the non-training subject value of the feature, it may have a positive or negative valence depending on whether the feature tends to correlate positively or negatively with reactivity in the training subject).

The importance of a feature may be indicated by a symbol or other indicator in the histogram that is located near, inside, or below the feature bin. For example, the importance of a feature may be indicated by the size of a symbol located below the feature pillar. Alternatively, the importance may be color coded, with the columns or associated symbols colored or shaded in this manner to indicate the degree of importance. An index table (key) may accompany the histogram indicating a color or shading band spectrum, where colors or shading more similar to one end of the band spectrum indicate higher importance and colors or shading more similar to the other end of the band spectrum indicate lower importance. In some embodiments, a symbol, such as a symbol in a histogram GUI report placed below the feature bar of a non-training subject, may indicate whether the value of the feature is negatively or positively correlated with reactivity in the training subject. For example, plus and minus signs, up and down arrows, positive and inverted triangles, or other pairing symbols may be used to indicate positive and negative correlation characteristics. In this case, the relative size of such symbols may indicate their relative importance.

In another embodiment, the GUI report includes an identifier of the shape used to convey the feature. For example, for each feature whose identifier is incorporated into the GUI report, it may be represented by a shape, where a dimension, color, shading, or other aspect of the shape may represent a different identifier. For example, each feature may be represented by a rectangle, respectively, where width represents importance and height represents weight. A color or shading or contour pattern of the rectangle may indicate the valence. Alternatively, features may be represented by triangles, where the base represents importance and the height represents weight, or vice versa. For features of positive valence, the orientation of the triangle may be upward, and vice versa. Alternatively, the color or shading pattern of the triangle or the pattern of lines outlining the triangle may indicate the valence.

In another embodiment, the identifier of a feature may be indicated by the shape of the sector of the annulus or the sector of the annulus. The angle of the annulus sector may indicate importance and its outer radius indicates the weight of the feature, or vice versa. The valence may be indicated by the color of the annulus sector, or the shading pattern of the annulus sector, or the pattern of lines drawn from the annulus sector. In other embodiments, sectors of a circle, or circular sectors, may represent identifiers of features, where angles and radii represent importance and weight or vice versa, and colors or shading, etc., represent valence, as in the embodiments of the annulus sectors given above. Where the feature identifier is represented by an annulus sector, in some embodiments, all such annulus sectors may be drawn with the same inner radius, and the annulus sectors may be arranged such that their inner arcs collectively form an inner circle. Within the inner circle itself, a prediction score or other general summary or indication of the responsiveness prediction or classification may be displayed. The color or shading of the inner circle may indicate whether the responsiveness classification predicts whether the non-training subject is predicted to be likely or unlikely to respond. For example, where the predicted outcome is that a non-trained subject is likely to respond, the inner circle may have a color or shading pattern, or be drawn with lines of a pattern; while the inner circle may have a different color or shading pattern when the predicted outcome is less likely to be responsive to an untrained subject. In other embodiments, instead of an inner circle, there may be other inner shapes such as a square or star or triangle or pentagon or other shapes. The size of the inner shape may indicate the strength of the prediction, with larger inner shapes indicating higher confidence in the prediction, and vice versa.

GUI reports may also provide the user with the opportunity to seek more information or launch additional software applications, depending on the interest in the particular features of the report in the GUI. For example, the GUI may be configured such that a user may controllably hover a cursor or other element over an identifier of a feature via an input device (e.g., a mouse), or by touching a touch screen. By the orientation of the element or the display position where the touch feature is located, a pull-down menu can be opened with options that can be further selected. For example, a drop down menu may display aspects of the feature that are specific to non-training subjects, such as the value of the feature; or a queue range for the feature; a percentage of the non-training subject value representation of the feature (relative to a range of values present in the training data, or compared to only training data of training subjects who responded in the same manner as predicted for non-training subjects, or compared to only training data of other training subjects); or a single factor decision boundary for the feature; or the importance of the characteristic or the relevance of the responsiveness to other treatments; or any combination of two or more of the foregoing. The drop down menu may also show links to other programs that may be accessed by one or more microprocessors to further evaluate the characteristic or untrained subject prediction score, such as running a different machine learning classifier. By compressing this interactivity in one GUI report, significant space and computing resources can be saved and user interactivity with the computer system can be significantly enhanced. For example, less display space will be required for continued display of GUI reports while accessing the drop down menu options. In addition, time and computing resources will be saved because the interactivity will allow access to a variety of computer functions without switching between displays or applications.

In some embodiments, the trained machine learning classifier may be further trained. For example, after determining responsiveness of a non-trained subject to treatment, the machine learning classifier can be retrained based on features including the training subject feature values and reactivities originally used to train the machine learning classifier, and the feature values and reactivities of the non-trained subjects that achieved the prediction and obtained responsiveness. In other embodiments, the trained machine-learned classifier may be retrained with other feature values and reactivities from non-trained subjects from which no prediction was obtained for the non-trained subjects.

In some embodiments, upon obtaining a predictive classification of a subject's response, a decision can be made as to whether to use a given checkpoint inhibition therapy. For a particular checkpoint treatment, when a machine-learned classifier is trained to predict the responsiveness of that therapy, upon obtaining a high prediction score with a machine-learned classifier trained in accordance with the disclosure herein, a decision to treat the subject with the therapy can be made. Alternatively, after a low score is obtained, a decision may be made to not treat with such therapy. Once a person has obtained a response prediction classification or score according to the methods, systems, or machine learning classifiers disclosed herein suggesting a suitably high likelihood of responsiveness, or has obtained instructions to treat a subject based on such a response prediction classification or score that has been obtained, the subject may be treated with the therapy. The present disclosure encompasses treating cancer in a subject by: administering a checkpoint inhibition therapy in response to a response prediction classification or prediction score obtained according to the present disclosure, wherein the response prediction classification or prediction score indicates that the subject will respond to the treatment; or administering the treatment according to instructions from the individual who obtained such a response prediction classification or prediction score.

Examples

The following examples are intended to illustrate specific embodiments of the present disclosure, but are not intended to limit the scope thereof.

The network diagram of fig. 1 shows options for performing a method according to aspects of the present disclosure. Some possible non-limiting examples of the source of the characteristic values are shown, such as cancer genome atlas (TCGA), or clinical trial data (such as treatment trials with anti-PD 1 therapy, anti-CTLA 4 therapy, or other checkpoint inhibitor therapy, or other cancer therapy). Some non-limiting examples of assays that can be used to obtain characteristic information for determining characteristic values are also shown, such as RNAseq and Whole Exome Sequencing (WES), as two non-limiting examples. A number of different non-limiting examples of types of features are also shown, as are examples of which assay or assays may provide relevant quantities to determine the value of such features. Examples include HLA, gene expression, ssGSEA, tumor mutation burden, cytolytic infiltrates such as infiltration of tumor infiltrating lymphocytes (deconvolution), CGA, neoantigens, the presence of clonal and/or subclonal mutations (i.e., mutations present in progeny of cells initially exhibiting a given mutation, and other mutations in progeny of cells from such initially mutated cells that have subsequently acquired other mutations), and the like. The features are processed by a Machine Learning (ML) model, wherein a machine learning classifier is trained and feature values of non-trained subjects are input to the trained machine learning classifier, thereby obtaining responsiveness of a patient who is tagged as likely or unlikely to respond to therapy.

Fig. 2 shows some non-limiting examples of features that may be relevant in training a classifier and predicting a patient's responsiveness to treatment, according to aspects of the present disclosure. As the skilled person will appreciate, the features identified in fig. 2 are non-limiting examples. Neither of these features is required. Other features not shown in fig. 2 may be used and some of the features shown may be omitted in performing methods according to aspects of the present disclosure. Features are shown in this grouping, which may be based on function, cellular or molecular pathway or reaction, etc. Examples of such groupings include antigen presentation, T cell or NK cell tags (signatures), tags for immune-mediated lysis, checkpoint pathway participants, interferon gamma pathway participants, and MDSC/Treg tags. In addition to any groupings shown as non-limiting examples, other groupings representing, for example, other functions or cellular or molecular processes may additionally or alternatively be included.

The network diagram of fig. 3 shows an example of how the method of training a classifier may be implemented in accordance with aspects of the present disclosure. As the skilled person will appreciate, the features identified in fig. 3 are non-limiting examples. Neither of these features is required. Other features not shown in fig. 3 may be used and some of the features shown may be omitted in performing methods according to aspects of the present disclosure. Features are shown in this grouping, which may be based on function, cellular or molecular pathway or reaction, etc. Examples of such groupings include antigen presentation, T cell or NK cell tags (signatures), tags for immune-mediated lysis, checkpoint pathway participants, interferon gamma pathway participants, and MDSC/Treg tags. For training, as shown in fig. 3, features from the training subjects, such as those shown, are input to a machine learning classifier, such as a random forest classifier, as well as labels (labels) corresponding to the responsiveness of the training subjects to a given treatment, such as checkpoint inhibition treatment. In this way, the classifier is trained. Other machine learning classifiers may also be used, as disclosed.

Fig. 4 is an expanded view of fig. 3, the network diagram shown showing one example of how a prediction method according to aspects of the present disclosure may be implemented, wherein the method uses a trained classifier to predict a subject's responsiveness to treatment. A machine-learned classifier (in this non-limiting example, a random forest classifier) that has been trained on features (e.g., the non-limiting feature example shown here) accepts other inputs obtained from non-trained subjects. In particular, feature values from untrained subjects are input to a machine learning classifier. The trained machine-learned classifier then generates a response prediction classification when reporting the prediction results, which may include a score (here, an immune score) indicating the likelihood of responsiveness and/or an identifier of the feature.

From Hugo et al (2016) Genomic and Transcriptomic Features of Response to anti-PD-1 Therapy in Metastatic Melanoma.Cell.2016; 165(1) 35-44(doi:10.1016/j. cell.2016.02.065) obtain the characteristics and reactivity against the inhibition of PD-1. In this study, whole exome sequencing data and RNAseq data were obtained from tumor samples of 26 melanoma patients before and after treatment with PD-1 checkpoint pathway inhibitors (anti-PD-1 antibody treatment (nivolumab) or anti-PD-L1 antibody treatment (pembrolizumab)). The raw data is publicly available and accessed for the examples herein. Transcriptome data obtained by RNAseq (including expression levels of transcripts in samples) were obtained online from the american National Center for Biotechnology Information (NCBI) gene expression compilation under accession number GSE78220(https:// www.ncbi.nlm.nih.gov/geo/query/acc. cgi. According to the accession number SRA: SRP067938 and SRA: SRP090294, the files read from the NCBI sequence (https:// www.ncbi.nlm.nih.gov/sra) Whole exome sequencing data obtained by the NGS method is obtained online. For patients for whom such data is available, the responsiveness of the patient is also obtained from the results of published studies. Data is selected from these sources, creating features and reactivities of training subjects to train machine learning classifiers to predict reactivities to the anti-PD 1 checkpoint pathway inhibitors. Feature data is also obtained from these sources for obtaining predictions from the trained machine-learned classifier.

Characterization and reactivity against CTLA-4 inhibition was obtained from Van Allen et al (2015) Genomic peptides of stress to CTLA-4blockade in metastic melanoma.science 350: 207-. In this study, whole exome sequencing data and RNAseq data were obtained from tumor samples of 30 melanoma patients before and after treatment with CTLA-4 inhibitor (ipilimumab). Raw data is publicly available and accessed for the examples herein. Whole exome sequencing data and transcriptome data obtained by NGS method were obtained online from the genotype and phenotype database of NCBI (dbGaP), accession number phs000452.v2.p1(https:// www.ncbi.nlm.nih.gov/gap/. For patients for whom such data is available, the responsiveness of the patient is also obtained from the results of published studies. Data was selected from these sources, creating features and reactivities of training subjects to train machine learning classifiers to predict reactivities to the anti-CTLA 4 checkpoint pathway inhibitors. Feature data is also obtained from these sources for obtaining predictions from the trained machine-learned classifier.

Starting from data obtained from both studies, the data is preprocessed to create feature entries that are input into a machine learning classifier for training. The random forest machine learning classifier that received the input is then trained based on the training data. Features from some subjects are then used to generate predictions from one or both trained classifiers, and the responsiveness prediction classifications are reported through the GUI, including prediction scores and identifiers of the features. Fig. 1, as described above, illustrates an overview of some examples of methods used in these examples. Fig. 2 shows the selected 38 features input to the machine learning classifier. Features are selected based on the following expectations: based on the current understanding of the effects of such features, these features may be relevant to predicting whether a subject is likely to respond to checkpoint inhibition. allMut refers to the tumor mutational burden of all mutations, nosynumut refers to the tumor mutational burden due to non-synonymous mutations. CD8_ dec, CD4_ dec and CD19_ dec refer to tumor infiltrating lymphocyte deconvolution magnitudes for respective CD8, CD4 and CD 19. The remaining features identified in figure 2 were obtained from RNAseq data and included the magnitude of the relative expression levels. For some examples, ssGSEA was also obtained for the feature set (antigen presentation, T cell/NK cell signature, signature of immune cytolytic cells, checkpoint pathway, interferon gamma, and MDSC/Treg signature). Some machine learning classifiers were trained on only the original 38 features, while others were trained on the 38 original features plus the gene set obtained by ssGSEA analysis (resulting in a total of 44 features for this class of examples).

By ranking the values of the subjects, the data for each feature value of the subjects is normalized to the percentage of the range of values for each value of each feature relative to the total subject values for that feature in the set. The responsiveness used for training in these examples was binary, i.e., subjects were marked as either responsive (including partial responsiveness and full responsiveness reported in the corresponding basic study or long-term benefit from treatment) or non-responsive (including progression of the condition in response to treatment and lack of therapeutic benefit in subjects reported in the basic study).

The features are input into a computer system that includes one or more memories and one or more microprocessors. The one or more memories contain instructions, wherein the instructions, when executed by the one or more microprocessors, train a random forest machine learning classifier using an R programming language. Subject characteristics are stored on the one or more memories and analyzed by the one or more microprocessors according to the instructions. In these examples, 50,000 trees are used. After training, the subject features (stored on the one or more memories of the computer system) are input into the trained machine learning classifier and a reaction prediction classification score is generated and reported by the GUI on the computer display. An example of training is shown in fig. 3 and an example of generating a prediction is shown in fig. 4.

The final prediction score is generated by the classification probability, and is scored from 0 to 10. As an example to explain, a classification probability of 0.75 in the category of "response" to a given immunotherapy translates to a predictive score for reactivity of 7.5. In addition, for each feature, a one-way decision boundary is determined at a value that can maximize classification accuracy. With respect to whether a feature is classified as positively or negatively correlated with reactivity, the direction of this correlation is determined by the Spearman correlation between the feature and reactivity. The aforementioned analysis was performed using all samples without separating the training subjects and the non-training subjects. Performance analysis was performed separately. For 3-fold cross validation and Area Under Curve (AUC) plotting, the R programming language packages "caret" and "cvAUC" were used, with default functions and parameters.

FIG. 5 shows an example report of a GUI 500 in which the reactive taxonomy prediction scores and feature identifiers are reported. For these examples, only the identifiers of features whose importance (the importance is determined by the Gini exponential decline of the feature) is greater than 0.1 squared are shown. Fig. 5 shows a GUI report for a subject, where the subject's data previously obtained for the anti-PD-1 trial was run on a trained machine learning classifier, wherein the machine learning classifier was trained to predict responsiveness to an anti-PD 1 treatment (e.g., anti-PD 1 or anti-PD 1-L1 antibody). In the FEATURES column 510 are shown 15 FEATURES whose importance exceeds a minimum importance threshold. The importance of the features is shown in imp. column 520. The GROUPs associated with features according to fig. 2 are shown in column GROUP 530. Whether a feature is positively or negatively correlated with reactivity is shown in CORR column 540. In this example, triangles are used to indicate the direction of correlation, where a positive triangle pointing upwards represents a positive correlation between the eigenvalue and reactivity and an inverted triangle pointing downwards represents a negative correlation. The single factor decision boundaries for the features are shown in column 1FDB 550. In this example, the single-factor decision boundary is numerically determined as a percentage in the subject value range of the sample that, by being above and below the percentage, provides the highest degree of achievable discrimination between responders and non-responders (i.e., values above or below the percentage are generally less accurate in distinguishing between responders and non-responders). The single factor decision boundary is also indicated by shading left to right within the cell of each feature in the 1FDB column, which represents the percentage indicated by the single factor decision boundary (i.e., low percentages have less shading, higher percentages have more shading, proportional to the value of the boundary). The subject value for each feature (in the case of PT5INPUT, the patient for which the prediction report here is directed, identified as patient number 5 or PT5) is shown in INPUT column 560. The values in PT5INPUT column 560 indicate the values of a given characteristic of the subject, and the shading indicates the percentage ranking of the characteristic values of the subject relative to the range of values for the training subject.

Pred 570, based on the values of a given feature only, whether the subject will be predicted to respond to treatment is reported in column 1f. Thus, for a positive correlation signature, if the value in PT5INPUT 560 exceeds the value in 1FDB 550, then 1f.pred 570 gives YES (meaning that the signature will predict that the subject will respond to treatment). For a positive correlation signature, if the value in PT5INPUT 560 is lower than the value in 1FDB 550, then 1f.pred 570 gives NO (meaning that the signature will predict that the subject will not respond to treatment). For a negatively correlated feature, if the value in PT5INPUT 560 exceeds the value in 1FDB 550, then 1f.pred 570 gives NO (meaning that the feature would predict that the subject would not respond to treatment). And for a negatively correlated feature, if the value in PT5INPUT 560 is lower than the value in 1FDB 550, then 1f.pred 570 gives YES (meaning that the feature will predict that the subject will respond to treatment). In this example, the cells in 1f.pred 570 can also be color coded according to whether the features alone can predict reactivity. YES cells may be colored green, e.g., (G), while NO cells may be colored red (R). The last FULL MODEL 580 reports a reactive classification score, which in this example is 5.5. A threshold value may be determined, wherein above or below said value, the treatment may be predicted to be effective or ineffective. For example, a score below 5.0 may be considered to predict that the treatment will be ineffective for that patient, while a score above 5.0 may be considered to indicate that the treatment will be effective for that patient. In this example, where a score above 5 is considered to indicate that the treatment will be effective for this patient, it can be seen that training and using the values obtained by the machine learning classifier as disclosed herein can provide a significantly improved basis for prediction, compared to a prediction based on only one of the features that exhibit the index in the GUI report, where some features alone predict unresponsiveness (including features with the highest importance), while the machine learning classifier overall predicts that the patient will respond.

In this example, the responsiveness prediction score is reported by the GUI, which also incorporates indicators for each feature, including weight (by providing a comparison between PT5INPUT 560 and 1FDB 550), importance 520, and potency 1 f.pred. In some examples, a user may have the option to access a drop down menu from different aspects of the GUI, such as by touching a portion of a touch screen display corresponding to a portion of a GUI report, or moving a graphical element (such as a cursor) over a feature or related indicator or its score with a device (such as a mouse) to access other information or select other analyses, which may be run on one or more microprocessors with different programming instructions stored on one or more memories.

Another GUI report 610 for this subject is shown in fig. 6. In the example illustrated in FIG. 6, multiple GUI reports are consolidated into a single report. The upper part of fig. 6 shows a ring of annular sectors 610, each corresponding respectively to the feature explained in fig. 5 whose importance exceeds the minimum importance threshold. The features corresponding to each zone sector are also given in text. For example, the zone sector corresponding to characteristic hla.b is indicated by 630. Each annulus sector has an angle, an outer radius and an inner radius, and an inner arc. In this example, the angle corresponds to the importance of the feature. In this example, the angles of the features are directly proportional to each other to allow direct visual comparison. Also in this example, the difference between the outer radius and the inner radius corresponds to the weight of the feature (i.e., the difference in the subject's feature value from the one-factor decision boundary). In this example, the valence of the feature is also reported by the style of the line used to outline the annulus sector of the feature. The zonal sectors characteristic of positive titers (e.g., total tumor mutation burden 640) to subjects are outlined with solid lines, while the zonal sectors characteristic of negative titers (e.g., hla.b 630) to subjects are outlined with dotted lines.

In this example, the inner arcs are arranged to form an inner circle. Also in this example, inside the inner circle, the responsiveness prediction classification score for this patient is reported, here 5.5. Also in this example, the overall prediction is indicated by the solid line forming the inner circle, meaning that the subject's responsiveness prediction classification score exceeds a predetermined level that distinguishes between a responsive and non-responsive prediction. In other examples, if the responsiveness prediction classification score is below this predetermined score threshold, a dotted line may be used to form an inner circle. In other examples, a color or differential shading pattern of the inside of the annulus sector and/or circle, and/or a differential coloration of the outline of the annulus sector or circle, may indicate valence.

In this example, the top half 610 of the GUI report 600 illustrated in FIG. 6 includes a reaction prediction classification score report and shows the importance, valence, and weight of the features. Another example of such a GUI report 1010 is shown in fig. 10. The annulus sectors of the features are shown one by one rather than as inner arcs forming a circle. For purposes of illustration, for the annulus sector of CD15, the outer arc 1002, the inner arc 1001, and the difference between the outer and inner radii 1003 are shown for exemplary purposes; for hla.b1030, the angle of the zone sector 1004 is shown below the zone sector. The outer radius of a feature (or the difference between the inner and outer radii) reports the weight, the angle reports the importance, and the pattern of the contour lines of the annulus sector reports whether the feature has a positive valence (solid line, e.g., annulus sector of total tumor abrupt burden, all _ tmb 1040) or a negative valence (dotted line, e.g., annulus sector of CD 15). Not shown in such GUI report 1010, but optionally can include, a reaction prediction classification score. Instead of different patterns of contours, different colors or different patterns of shading may be used to indicate the valence of the feature. Shapes other than the annulus sectors may also be used. For example, features may be reported as rectangles, where importance and weight are represented by, for example, width and height; or as triangles, where base width reflects importance, height represents weight, and direction represents valence. The skilled person will appreciate that numerous possibilities may be applicable for the purpose of reporting multiple indicators of multiple features in a GUI report according to aspects of the present disclosure.

Returning to the GUI report 600 depicted in FIG. 6, below the upper half of the report, which includes the annulus sector 610, is the histogram portion 620 of the GUI report. The GUI report may have two such portions or only one portion or neither. Histogram 620 shows a bar for each feature whose value exceeds the minimum importance threshold set for that feature for the subject being reported. The scale 650 on the left indicates the percentage ranking of the subject values for each feature. The bars for each feature represent the subject feature values for that feature, expressed as a percentage of the range of values for the training subject values. The subject value 660 of characteristic CD15 is an example. For each bin, the single feature decision boundary for that feature is also shown, in this case as a horizontal line. The single feature decision boundary 670 of feature CD15 is an example. The triangle below each bar indicates whether the feature is positively correlated (upward pointing triangle) or negatively correlated (downward pointing triangle) with reactivity. CD 15680 is an example. The triangles are also scaled to reflect the relative importance of each feature, with larger triangles representing higher importance and smaller triangles representing lower importance. A line between the subject feature value and the one-factor decision boundary for a feature indicates the weight of the feature. Feature CD 15690 is an example of a feature weight report. The titer of a feature is indicated by whether the weight line is solid (positive titer) or dotted (negative titer). As the skilled person will appreciate, these specific examples of different indicators for reporting different characteristics, each may be omitted or replaced with a different illustration. Color and shading may represent valence and/or correlation of features, arrows or other directional shapes may represent valence or correlation, importance may be represented by a color code scale scheme, or the like.

In some examples, the gene set is used to train a machine learning classifier and generate a prediction from the trained machine learning classifier. An example of a gene set generated from ssGSEA is shown in fig. 7. Six sets are shown, with individual features grouped for use in determining gene sets. Examples include antigen processing pathways (i.e., associated with antigen presentation) (710), T cell and NK cell tags 720, lysis tags 730, checkpoint pathways 740, interferon gamma 750, and MSDC/Treg tags 760. The relevance and importance of each feature, including the set of cells, when used to train PD1 and CTLA4 machine learning classifiers is indicated at 770 and 780, respectively. In some cases, the gene set highlighted in the cell outlined with the dashed box provides a higher magnitude of correlation or greater significance relative to any single feature upon which the value of the gene set was determined using ssGSEA, thereby indicating the value of incorporating the gene set as a feature. The efficacy of incorporating ssGSEA is also shown in figure 9. Figure 9 shows two GUI reports reporting the responsiveness predictions obtained for the same subject using two different trained machine-learned classifiers. By generating predictions without using the ssGSEA gene set as a feature during training or prediction, the left prediction 910 is obtained. The prediction on the right is obtained 920 by generating a prediction using the ssGSEA gene set as a feature during training and prediction. When included in the ssGSEA derived gene set, fewer features exceed the minimum threshold boundary (11 versus 15), including some gene sets (by definition, these gene sets are not included in the prediction 910, where the prediction 910 was obtained without using a gene set), but without sacrificing the overall prediction (e.g., in both cases the prediction outcome exceeded a score of 5.0, set as the minimum response prediction classification score used to trigger classification of the subject as a responder).

Figure 8 shows two GUI reports obtained from the same patient in which predictions for the patient were generated using two different machine learning classifiers, one 810 trained to predict responsiveness to anti-CTLA 4 treatment and the other 820 trained to predict responsiveness to anti-PD 1, both using the features described in figure 2. anti-CTLA 4 machine learning classifier 810 produced a response predicted classification score of 3.8, predicting that the subject will not respond to anti-CTLA 4 treatment (using a response predicted classification score threshold of 5.0). In this example 810, the response prediction classification score is given in the inner circle 815, while the dotted line around the inner circle 815 indicates the titer of the response prediction classification score (no response). The accuracy of the prediction was confirmed by the patient's response classification in the originating clinical study (Van Allen et al) in which the patient's response was classified as "disease progression" (indicating that the subject was not responsive to anti-CTLA 4 therapy). However, the anti-PD 1 machine-learned classifier 820 generated a response prediction classification score of 6.7, predicting that the subject will respond to treatment with anti-PD 1 (e.g., anti-PD 1 or anti-PD-L1 antibody) (again using a response prediction classification score threshold of 5.0). In this example 820, the response prediction classification score is depicted in the inner circle 817, while the solid line surrounding the inner circle 817 indicates the valence (response) of the response prediction classification score.

The response prediction classification scores (and corresponding differences in potency, weight, and importance of multiple features) that differ depending on the machine learning classifier used reflect the efficacy of the methods as disclosed herein. For example, the index reported for hla.a830 shows that it has a negative titer for predicting the patient's response to anti-CTLA 4 therapy; but it has a positive potency for predicting the patient's responsiveness to anti-PD 1 treatment. In addition, the eigenvalues of tumor non-synonymous and total tumor mutation burden did not exceed the minimum importance threshold in the anti-CTLA 4 machine learning classifier, but exceeded the minimum importance threshold in the anti-PD 1 machine learning classifier 840.

As with the example described in fig. 5, the exemplary GUI reports in fig. 6, 8, 9, and 10 may include user interactions. Thus, for some examples, a user may have the option to access a drop down menu from different aspects of the GUI, such as by touching a portion of a touch screen display corresponding to a portion of a GUI report, or moving a graphical element (such as a cursor) over a feature (e.g., a bar or other indicator of a circle sector or histogram) with a device (such as a mouse) to access other information or select other analyses, which may be run on one or more microprocessors via different programming instructions stored on one or more memories.

11A-11D show that using all 38 features described in FIG. 2, classifier performance other than using non-single factors can be generated. This was true for both machine learning classifiers trained to predict responsiveness to anti-PD 1 or anti-CTLA 4 treatments. Fig. 11A shows that using all 38 features for training and testing a machine learning classifier to predict responsiveness to, for example, anti-PD 1 therapy, when the machine learning classifier was over-trained and no Cross Validation (CV), yielded a subject working characteristic curve (auROC) AUC value of 1.00 false positive versus true positive, achieving 1.00auROC, versus an average single factor auROC of 0.64, as shown in fig. 11B. Fig. 11C and 11D further show that using the 38 features described in fig. 2, it is more accurate than all three of the highest single factor features HLA-B, nonSyn _ tmb and all _ tmb.

The graphs of fig. 12A-12D show the effect on classifier performance using the gene set obtained from ssGSEA. The performance of ssGSEA is comparable to using the 38 features described in figure 2. Fig. 12A, 12B and 12C report a 3-fold cross validation auROC in which 38 signatures described in fig. 2 were used to predict needle responsiveness to anti-PD 1 therapy (fig. 12A) or anti-CTLA 4 therapy (fig. 12B), or 38 signatures plus 6 gene sets described in fig. 7 were used to predict needle responsiveness to anti-PD 1 therapy (fig. 12C). Although performance was reduced from 0.69 to 0.64 for anti-PD 1 treatment prediction when the 6 gene sets obtained by ssGSEA were included (compare fig. 12A with fig. 12C), the number of features with significance exceeding the minimum significance threshold was reduced from 15 to 11 (as discussed above; see, e.g., fig. 8). Figure 12D shows the comparability t-test results of false positive results and true positive results generated from the response prediction classification scores according to the machine learning classifier (no cross-validation). In some examples, the inclusion of a gene set as a feature may therefore improve the overall robustness of the classifier, help avoid over-training, and allow for clearer interpretation of important features and their indices.

The data used in these examples were from subjects enrolled in the study to test the effectiveness of checkpoint inhibitor anti-CTLA 4 antibodies, anti-PD 1 antibodies, and anti-PD 1-L1 antibodies in melanoma patients. However, as the skilled person will appreciate, based on the role of known checkpoint inhibition in impairing the effectiveness of immunooncology treatments of other cancers, and the role of features (such as those contained herein) in checkpoint pathway function, the methods, systems and classifiers as disclosed herein will be equally applicable to, and effectively achieve, predicting the responsiveness of a subject to these checkpoint inhibitors in other cancers, including breast cancer, digestive system cancer, liver cancer, bladder cancer, lymphoma, leukemia, cancer of bone tissue, nervous system cancer, lung cancer, pancreatic cancer or others. In addition, as the skilled person will also appreciate, reactivity may be predicted for checkpoint inhibitors other than those specifically used in the foregoing examples disclosed herein using methods, systems and classifiers as disclosed herein, which are merely some non-limiting examples of the applicability of the present invention.

The potential for this performance analysis is low sample size. Even with tens of features, overtraining is inevitable. Meanwhile, the fold-cross validation results are highly unstable, showing discrete auROC points, rather than continuous curves. However, in this limitation we have, the full model (full model) clearly outperforms the single factor in performance.

Although preferred embodiments have been illustrated and described in detail herein, it will be apparent to those skilled in the relevant art that various modifications, additions, substitutions and the like can be made without departing from the spirit of the disclosure and these are therefore considered to be within the scope of the disclosure as defined in the following claims.

It should be understood that all combinations of the foregoing concepts and other concepts discussed in greater detail herein (provided such concepts are not mutually inconsistent) are contemplated as being part of the inventive subject matter disclosed herein. In particular, all combinations of claimed subject matter appearing at the end of this disclosure are considered part of the inventive subject matter disclosed herein.

Claims

1. A computer-implemented method, comprising:

inputting genomic information of a non-trained subject to a trained machine learning classifier, the genomic information of the non-trained subject comprising features of a tumor profile obtained from the non-trained subject, wherein

The trained machine learning classifier is trained based on genomic information of a plurality of training subjects and responsiveness of each of the plurality of training subjects to a treatment, the treatment comprising checkpoint suppression, the genomic information of the plurality of training subjects including features of a tumor profile obtained from each of the plurality of training subjects, wherein the machine learning classifier is trained to predict responsiveness to the treatment;

generating a checkpoint inhibition responsiveness classification for the non-training subject using the trained machine learning classifier, wherein the checkpoint inhibition responsiveness classification predicts responsiveness of the non-training subject to the checkpoint inhibition; and

using a graphical user interface, checkpoint inhibition responsiveness classifications for non-training subjects are reported.

2. The method of claim 1, wherein at least some features of the tumor profile from a non-training subject, or at least some features of the tumor profile from one or more training subjects are selected from the group consisting of: a total mutation load consisting of all mutations, a total mutation load consisting of non-synonymous mutations, expression of β 2 microglobulin (B2M), expression of proteasome subunit β 10(PSMB10), expression of antigenic peptide transporter 1(TAP1), expression of antigenic peptide transporter 2(TAP2), expression of human leukocyte antigen A (HLA-A), expression of major histocompatibility complex class I B (HLA-B), expression of major histocompatibility complex class I C (HLA-C), expression of major histocompatibility complex class II DQ α 1(HLA-DQA1), expression of HLA class II histocompatibility antigen DRB1 β chain (HLA-DRB1), expression of HLA class I histocompatibility antigen α chain E (HLA-E), expression of natural cell granule protein 7(NKG7), expression of differentiation factor-like receptor 1 (KLCMR 1), expression of antigen cluster 8(CD8) expressing cells, tumor infiltration, Tumor infiltration of differentiation antigen cluster 4(CD4) -expressing cells, tumor infiltration of differentiation antigen cluster 19(CD19) -expressing cells, granzyme A (GZMA) expression, perforin-1 (PRF1) expression, cytotoxic T-lymphocyte-associated protein 4(CTLA4) expression, programmed cell death protein 1(PD1) expression, programmed death-ligand 1(PDL1) expression, programmed cell death 1 ligand 2(PDL2) expression, lymphocyte activation gene 3(LAG3) expression, T cell immunoreceptor (TIGIT) expression with Ig domain and ITIM domain, differentiation antigen cluster 276(CD276) expression, chemokine (C-C motif) ligand 5(CCL5), CD27 expression, chemokine (C-X-C motif) ligand 9(CXCL9) expression, C-X-C motif chemokine receptor 6(CXCR6), indoleamine 2, 3-dioxygenase (IDO) expression, signal transduction and activator of transcription 1(STAT1) expression, 3-fucosyl-N-acetyl-lactosamine (CD15) expression, interleukin-2 receptor alpha chain (CD25) expression, siglec-3(CD33), differentiation antigen cluster 39(CD39) expression, differentiation antigen cluster (CD118) expression, forkhead box P3(FOXP3) expression, and any combination of two or more of the foregoing.

3. The method of claim 1, wherein at least some training features or at least some non-training features comprise a gene set.

4. The method of claim 3, wherein the gene set is selected using a single sample gene set enrichment assay.

5. The method of claim 1, wherein the machine learning classifier is a random forest.

6. The method of claim 5, wherein at least 50,000 trees are used in training the machine learning classifier.

7. The method of claim 1, wherein checkpoint inhibition reactivity classification comprises a prediction score and one or more feature identifiers, and the one or more feature identifiers are selected from feature potency, feature importance, and feature weight.

8. The method of claim 7, wherein the graphical user interface reports the feature identifier as an aspect of an annulus sector, wherein an angle of the annulus sector reports a feature importance, an outer radius of the annulus sector reports a feature weight, and a color of the annulus sector reports a feature valence.

9. The method of claim 8, wherein the feature importance of a feature comprises a Gini index drop of the feature.

10. The method of claim 9, wherein the graphical user interface reports an identifier of a feature if and only if the feature importance of the feature is above a threshold.

11. The method of claim 10, wherein the feature importance of a feature is not above a threshold if the square of the feature importance is not above 0.1.

12. The method of claim 10, wherein each annulus sector comprises an inner arc, and the inner arcs of the annulus sectors are arranged to form a circle.

13. The method of claim 1, further comprising inputting the responsiveness of the non-training subject to the therapy to a trained machine learning classifier and further training the machine learning classifier, wherein further training comprises training the trained machine learning classifier based on features of a tumor sample obtained from the non-training subject and the responsiveness of the non-training subject to the therapy.

14. The method of claim 1, further comprising selecting a treatment based on the generated checkpoint inhibition responsiveness classification.

15. A computer system, comprising:

one or more microprocessors, one or more of which,

one or more memories for storing a trained machine learning classifier and genomic information of a non-trained subject, wherein the trained machine learning classifier is trained from genomic information of a plurality of trained subjects and responsiveness of each of the plurality of trained subjects to a treatment, the treatment comprising checkpoint suppression, the genomic information of the plurality of trained subjects comprising features of a tumor profile obtained from each of the plurality of trained subjects, and the machine learning classifier is trained to predict responsiveness to the treatment, wherein the genomic information of the non-trained subjects comprises features of the tumor profile from the non-trained subjects, and

one or more memories storing instructions that, when executed by the one or more microprocessors, cause a computer system to generate a checkpoint inhibition responsiveness classification for a non-training subject using the trained machine learning classifier and report the checkpoint inhibition responsiveness classification for the non-training subject using a graphical user interface, wherein the checkpoint inhibition responsiveness classification predicts responsiveness of the non-training subject to checkpoint inhibition.

16. The system of claim 15, wherein at least some features of the tumor profile from a non-training subject, or at least some features of the tumor profile from one or more training subjects are selected from the group consisting of: a total mutation load consisting of all mutations, a total mutation load consisting of non-synonymous mutations, expression of β 2 microglobulin (B2M), expression of proteasome subunit β 10(PSMB10), expression of antigenic peptide transporter 1(TAP1), expression of antigenic peptide transporter 2(TAP2), expression of human leukocyte antigen A (HLA-A), expression of major histocompatibility complex class I B (HLA-B), expression of major histocompatibility complex class I C (HLA-C), expression of major histocompatibility complex class II DQ α 1(HLA-DQA1), expression of HLA class II histocompatibility antigen DRB1 β chain (HLA-DRB1), expression of HLA class I histocompatibility antigen α chain E (HLA-E), expression of natural cell granule protein 7(NKG7), expression of differentiation factor-like receptor 1 (KLCMR 1), expression of antigen cluster 8(CD8) expressing cells, tumor infiltration, Tumor infiltration of differentiation antigen cluster 4(CD4) -expressing cells, tumor infiltration of differentiation antigen cluster 19(CD19) -expressing cells, granzyme A (GZMA) expression, perforin-1 (PRF1) expression, cytotoxic T-lymphocyte-associated protein 4(CTLA4) expression, programmed cell death protein 1(PD1) expression, programmed death-ligand 1(PDL1) expression, programmed cell death 1 ligand 2(PDL2) expression, lymphocyte activation gene 3(LAG3) expression, T cell immunoreceptor (TIGIT) expression with Ig domain and ITIM domain, differentiation antigen cluster 276(CD276) expression, chemokine (C-C motif) ligand 5(CCL5), CD27 expression, chemokine (C-X-C motif) ligand 9(CXCL9) expression, C-X-C motif chemokine receptor 6(CXCR6), indoleamine 2, 3-dioxygenase (IDO) expression, signal transduction and activator of transcription 1(STAT1) expression, 3-fucosyl-N-acetyl-lactosamine (CD15) expression, interleukin-2 receptor alpha chain (CD25) expression, siglec-3(CD33), differentiation antigen cluster 39(CD39) expression, differentiation antigen cluster (CD118) expression, forkhead box P3(FOXP3) expression, and any combination of two or more of the foregoing.

17. The system of claim 15, wherein at least some training features or at least some non-training features comprise a gene set.

18. The system of claim 17, wherein the gene set is selected using a single sample gene set enrichment assay.

19. The system of claim 15, wherein the machine learning classifier is a random forest.

20. The system of claim 19, wherein at least 50,000 trees are used in training the machine learning classifier.

21. The system of claim 15, wherein checkpoint inhibition reactivity classification comprises a prediction score and one or more feature identifiers, and the one or more feature identifiers are selected from feature potency, feature importance, and feature weight.

22. The method of claim 21, wherein the instructions, when executed by one or more microprocessors, cause the graphical user interface to report the feature identifier as an aspect of an annulus sector, wherein an angle of the annulus sector reports a feature importance, an outer radius of the annulus sector reports a feature weight, and a color of the annulus sector reports a feature valence.

23. The system of claim 22, wherein the feature importance of a feature comprises a Gini exponential decline of the feature.

24. The system of claim 23, wherein the instructions, when executed by the one or more microprocessors, cause the graphical user interface to report an identifier of a feature if and only if the feature importance of the feature is above a threshold.

25. The system of claim 24, wherein the feature importance of a feature is not above a threshold if the square of the feature importance is not above 0.1.

26. The system of claim 24, wherein the instructions, when executed by one or more microprocessors, cause the graphical user interface to report an inner arc of each annulus sector and a circle containing the inner arc of the annulus sector.

27. The system of claim 15, wherein the instructions, when executed by the one or more microprocessors, cause the computer system to further train the machine learning classifier, wherein further training comprises training the trained machine learning classifier as a function of features of a tumor sample obtained from a non-training subject and the non-training subject's responsiveness to treatment.

28. A machine learning based classifier for immune checkpoint reactivity classification, the machine learning based classifier comprising:

a machine learning based classifier, running on a plurality of processors, trained to predict responsiveness of a non-trained subject to immune checkpoint suppression therapy,

wherein the machine learning based classifier is trained by inputting to the machine learning based classifier genomic information of a plurality of training subjects including features of a tumor profile obtained from each of the plurality of training subjects and responsiveness of each of the plurality of training subjects to treatment;

an input processor that inputs genomic information of a non-trained subject into a machine learning based classifier, the genomic information of the non-trained subject including features of a tumor profile from the non-trained subject, wherein the machine learning classifier is configured to generate a checkpoint inhibition reactivity classification of the non-trained subject that predicts a subject's reactivity to checkpoint inhibition; and

an output processor that reports the checkpoint inhibition reactivity classification.

29. The machine learning based classifier of claim 28, wherein the checkpoint inhibition reactivity classification includes a predictive score and a plurality of identifiers.