JP2021505976A - Systems and methods for assessing drug efficacy - Google Patents

Abstract

In the step of inputting the genomic information of the non-trained object including the characteristics derived from the tumor sample into the trained machine learning classifier, the trained machine learning classifier is the process of inputting the characteristics of the tumor sample obtained from the trained object and the characteristics of the tumor sample. A subject is checkedpoint-inhibited using a process and a trained machine-learning classifier trained for its responsiveness to checkpoint inhibition treatment and a machine learning classifier trained to predict its responsiveness to treatment. Provided is a computer execution method including a step of generating a checkpoint inhibitory responsiveness classification that predicts a response to a response and a step of reporting the checkpoint inhibitory responsiveness classification using a graphical user interface. Also provided are a computer system for carrying out this method and a machine learning classifier trained by this method.

Description

Cross-reference to related applications This application claims the priority of US Provisional Patent Application No. 62 / 593,802 filed December 1, 2017, the entire contents of which are incorporated herein by this.

Detecting abnormal or cancerous cells in the body is an important challenge in the immune system. One mechanism involved is the immune checkpoint. For example, programmed cell death protein 1 (PD-1) and cytotoxic T lymphocyte-related protein 4 (CTLA-4) checkpoints negatively regulate immune function and prevent overreaction on T cells (ie,). , Promotes self-recognition of the immune system). However, this mechanism can be utilized by tumor cells to escape immune attack. Immunotherapy such as PD-1 inhibition (eg, anti-PD1 antibody) and CTLA-4 (eg, CTLA-4 antibody) block checkpoint activity, which allows T cells to identify the disease or tumor cells themselves. Make it easy to do.

However, while immune checkpoint therapy may be effective, responsiveness in all cancer patients is not guaranteed. Compared to traditional cancer treatments, immune checkpoint therapy has shown improved long-term survival in patients with a variety of cancers. However, only some cancer patients have currently approved checkpoint inhibitors, including anti-CTLA-4 antibody (eg, ipilimumab), and anti-PD-1 antibody (eg, nivolumab) or anti-programmed death ligand 1 Respond to treatments that target the PD-1 checkpoint pathway, such as (anti-PD-L1) antibodies (eg, atezolizumab). Therefore, it may be advantageous to be able to select patients who can respond to a particular checkpoint therapy and to be able to predict which checkpoint target may give the best results to a particular patient.

A wide variety of genomic and cellular features can contribute to the effectiveness of immunotherapy for a particular individual. For example, when the tumor mutation burden (TMB) increases, the response rate is positively affected by increased antigen presentation on tumor cells, and recognition by T cells increases when PD-1 is blocked. Can be. Leukocyte tumor infiltration expressing CD4 / CD8 / CD19 correlates with better clinical outcomes because such cells aid in the immune attack and subsequent antigen release of tumor cells. Bone marrow-derived inhibitory cells and regulatory T cells (Tregs) suppress T cell efficacy and correlate with poor survival in a variety of patients. Because these are detectable and derivable features from interacting next-generation sequencing (NGS) data, investigate their relationship to immunotherapeutic responses and for therapy such as checkpoint inhibition or other cancer treatments. Building machine learning applications that predict responsiveness and incorporate the context of many features that collaborate in a particular individual, taking into account the individual's multifactorial context, based on the responsiveness of other individuals. is important.

In addition, responsiveness predictions given the potential number of interacting genomes, cells, and other features that may interact to determine whether a particular individual responds positively to checkpoint inhibition. Need to be improved in the way it reports. For example, a wide variety of features can interact in combination in predicting responsiveness. When applying machine learning methods to assess whether a patient can respond more or less to a given checkpoint inhibition, some features may be higher or lower than others in different individuals. Determined to be of low importance, different characteristics differ to varying degrees with levels suggesting that each may affect responsiveness, and different factors determine the patient's responsiveness to different checkpoint inhibition therapies. Can indicate whether is high or low. Therefore, there is a need for a contextual report of responsiveness in a particular patient, including identification of features that are important in predicting responsiveness and directional in the effectiveness of the prediction. However, given the limited space available to present all these potential aspects of forecast reporting, current reporting methods are inadequate. Therefore, there is a need for new methods for reporting multiple elements of relevant responsiveness predictions.

Gaujoux et al. (2013) CellMix: a comprehensive toolbox for gene expression deconvolution, Bioinformatics 29:22 11-2212 F Finotello et al. (2018), Quantifying tumor-infiltrating immune cells from transcriptomics data, Cancer Immunology, Immunotherapy 67: 1031-1040 Barbie et al. (2009), Systematic RNA interference reveals that oncogenic KRAS-driven cancers require TBK1, Nature 462: 108-112 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) Van Allen et al. (2015) Genomic correlates of response to CTLA-4 blockade in metastatic melanoma. Science 350: 207-211 (doi: 10.1126 / science.aad0095)

The present disclosure is intended to overcome the above and other deficiencies in the art.

In some embodiments, the process of inputting untrained genomic information into a trained machine learning classifier, wherein the untrained genomic information includes features derived from the tumor profile obtained from the untrained subject and is a trained machine. A learning classifier is trained on the genomic information of multiple training subjects and the responsiveness of each of the multiple training subjects to treatments including checkpoint inhibition, and the genomic information of multiple training subjects is combined with multiple training subjects. Machine learning classifiers are trained to anticipate treatment responsiveness, including features of tumor samples obtained from each of the processes and checked for non-trained subjects using trained machine learning classifiers. A step of generating a point inhibition responsiveness classification, in which the checkpoint inhibition responsiveness classification predicts that the untrained subject will respond to checkpoint inhibition, and the checkpoint inhibition responsiveness of the untrained subject. Disclosure of computer execution methods, including the process of reporting classifications using a graphical user interface. In one example, at least some of the features derived from the tumor profile obtained from the untrained subject, or at least some of the features derived from the tumor profile obtained from one or more of the trained subjects, are selected from the group of features: : Total mutation frequency consisting of all mutations, total mutation frequency consisting of non-synonymous mutations, beta2 microglobulin (B2M) expression, proteasome subunit beta10 (PSMB10) expression, antigen peptide transmitter 1 (TAP1) expression, antigen peptide Transporter 2 (TAP2) expression, human leukocyte antigen A (HLA-A) expression, major tissue compatible complex class IB (HLA-B) expression, major tissue compatible complex class IC (HLA-C) expression, major tissue compatible Complex class II DQ alpha 1 (HLA-DQA1) expression, HLA class II tissue-compatible antigen DRB1 beta chain (HLA-DRB1) expression, HLA class I tissue-compatible antigen alpha chain E (HLA-E) expression, natural killer cell granules Protein 7 (NKG7) expression, chemokine-like receptor (CMKLR1) expression, tumor invasion by cells expressing surface antigen classification 8 (CD8), tumor invasion by cells expressing surface antigen classification 4 (CD4), surface antigen classification 19 Tumor invasion by cells expressing (CD19), granzyme A (GZMA) expression, perforin 1 (PRF1) expression, cytotoxic T lymphocyte-related protein 4 (CTLA4) expression, programmed cell death protein 1 (PD1) expression, program 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 and ITIM domains, surface antigen classification 276 (CD276) expression, chemokine (CC motif) ligand 5 (CCL5), CD27 expression, chemokine (CXC motif) ligand 9 (CXCL9) expression, CXC motif chemokine receptor 6 (CXCR6), indolamine 2,3-dioxygenase ( IDO) expression, signaling and transcriptional activator 1 (STAT1) expression, 3-fucosyl-N-acetyl-lactosamine (CD15) expression, interleukocyte 2 receptor alpha chain (CD25) expression, siglec-3 (CD33), Surface antigen classification 39 (CD39) expression, surface antigen classification 118 (CD118) expression, forkhead box P3 (FOXP3) expression, and any of the two or more aforementioned combinations mentioned above. Se.

In another example, at least some of the trained features, or at least some of the untrained features, include a gene set. In a further example, the gene set was selected using a single sample gene set enrichment analysis. In yet another example, the machine learning classifier is a random forest. In a further example, at least 50,000 trees are used in training machine learning classifiers. In a further example, the checkpoint inhibition responsiveness classification comprises a predicted score and one or more feature identifiers, the one or more feature identifiers being selected from the group consisting of feature value, feature importance, and feature weight. Will be done.

In another example, the graphical user interface reports the feature identifier as each side of the circular sector, the angle of the circular sector reports the feature importance, the outer radius of the circular sector reports the feature weight, and the color of the circular sector. Reports feature values. In a further example, the feature importance of a feature includes a decrease in the Gini index of the feature. In yet another example, the graphical user interface reports feature identifiers only if and only if the feature importance of the feature exceeds a threshold. In yet another example, the feature importance of a feature does not exceed a threshold if the square of the feature importance of the feature does not exceed 0.1. In a further example, each circular sector includes an inner arc, and the inner arcs of the circular sector are arranged to form a circle.

Another example further includes the step of inputting the responsiveness of the non-trained object to the treatment into the trained machine learning classifier and the step of further training the machine learning classifier, and the further training step is the trained machine learning. The classifier involves training the characteristics of the tumor sample obtained from the untrained subject and the responsiveness of the untrained subject to treatment. Yet another example further comprises selecting a treatment based on the checkpoint inhibition responsiveness classification generated.

In another aspect, the trained machine learning classifier is one or more microprocessors and a trained machine learning classifier, and one or more memories for storing untrained genomic information. , Genome information of multiple training subjects, and tumor profiles obtained from each of the multiple training subjects by being trained for the responsiveness of each of the multiple training subjects to treatments including checkpoint inhibition. A computer that includes features, a machine learning classifier is trained to predict responsiveness to treatment, and untrained genomic information contains memory and features derived from tumor profiles obtained from the untrained subject. In a system where one or more memories are run by one or more microprocessors, the computer system is untrained for checkpoint inhibition responsiveness classification using a trained machine learning classifier. To generate and report the checkpoint inhibition responsiveness classification of the untrained subject using the graphical user interface, memorize the instructions, and the checkpoint inhibition responsiveness classification of the untrained subject responds to the checkpoint inhibition. Disclose a computer system that predicts what to do.

In one example, at least some features from a tumor profile obtained from an untrained subject, or at least some features from a tumor profile obtained from one or more trained subjects, are selected from the following group: Total mutation frequency consisting of all mutations, total mutation frequency consisting of non-synonymous mutations, beta2 microglobulin (B2M) expression, proteasome subunit beta10 (PSMB10) expression, antigen peptide transmitter 1 (TAP1) expression, antigen peptide transporter 2 (TAP2) expression, human leukocyte antigen A (HLA-A) expression, major tissue compatible complex class IB (HLA-B) expression, major tissue compatible complex class IC (HLA-C) expression, major tissue compatible complex Class II DQ alpha 1 (HLA-DQA1) expression, HLA class II tissue compatible antigen DRB1 beta chain (HLA-DRB1) expression, HLA class I tissue compatible antigen alpha chain E (HLA-E) expression, natural killer cell granule protein 7 (NKG7) expression, chemokine-like receptor (CMKLR1) expression, tumor invasion by cells expressing surface antigen classification 8 (CD8), tumor invasion by cells expressing surface antigen classification 4 (CD4), surface antigen classification 19 (CD19) ), Tumor invasion by cells expressing), Granzyme A (GZMA) expression, Perforin 1 (PRF1) expression, Cytotoxic T lymphocyte-related 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 and ITIM domains, surface antigen classification 276 (CD276) ) Expression, chemokine (CC motif) ligand 5 (CCL5), CD27 expression, chemokine (CXC motif) ligand 9 (CXCL9) expression, CXC motif chemokine receptor 6 (CXCR6), indolamine 2,3-dioxygenase (IDO) Expression, signal transduction and transcriptional activator 1 (STAT1) expression, 3-fucosyl-N-acetyl-lactosamine (CD15) expression, interleukocyte 2 receptor alpha chain (CD25) expression, siglec-3 (CD33), surface antigen Classification 39 (CD39) expression, surface antigen classification 118 (CD118) expression, forkhead box P3 (FOXP3) expression, and any combination of two or more of the aforementioned.

In another example, at least some of the trained features, or at least some of the untrained features, include a gene set. In yet another example, the gene set was selected using a single sample gene set enrichment analysis. In yet another example, the machine learning classifier is a random forest. In a further example, at least 50,000 trees are used in training machine learning classifiers. In a further example, the checkpoint inhibition responsiveness classification comprises a predicted score and one or more feature identifiers, the one or more feature identifiers consisting of a group consisting of feature value, feature importance, and feature weight. Be selected. When the instruction is executed by one or more microprocessors, it causes the graphical user interface to report the feature identifier as each side of the circular sector, the angle of the circular sector reports the feature importance, and the outer radius of the circular sector. The feature weight is reported, and the circular sector color reports the feature value.

In another example, the feature importance of a feature involves a decrease in the Gini index of the feature. In yet another example, the instruction, when executed by one or more microprocessors, causes the graphical user interface to report the feature identifier only if the feature importance of the feature exceeds a threshold. In a further example, the feature importance of the feature does not exceed the threshold if the square of the feature importance of the feature does not exceed 0.1. In yet another example, the instruction, when executed by one or more microprocessors, causes the graphical user interface to report an inner arc of each circular sector and a circle containing the inner arc of the circular sector. In a further example, when the instruction is executed by one or more microprocessors, training the computer system further to train the machine learning classifier and further training the trained machine learning classifier from the untrained subject. It involves training the characteristics of the resulting tumor sample and the responsiveness of the untrained subject to treatment.

In yet another embodiment, a machine learning-based classifier for classifying immune checkpoint responsiveness that runs on a number of processors to predict the responsiveness of untrained subjects to immune checkpoint inhibition therapy. A machine learning-based classifier that is trained in a machine learning-based classifier that provides machine learning-based classifiers with genomic information for multiple training subjects and their respective responses to treatment. Trained by entering gender, machine learning-based classifiers and machine-learning-based classifiers that contain the genomic information of multiple training subjects, including tumor profile features obtained from each of the multiple training subjects, An input processor that inputs the characteristics of tumor samples obtained from untrained subjects, a machine learning-based classifier configured to generate checkpoint inhibition responsive classifications for untrained subjects. Discloses a machine learning-based classifier that includes an input processor that predicts that a subject will respond to a checkpoint inhibition treatment and an output processor that reports a checkpoint inhibition response classification. In one example, the checkpoint inhibition responsiveness classification includes a predicted score and multiple identifiers.

The above and other features, aspects, and advantages of the present disclosure will be better understood by reading the following detailed description and referring to the accompanying drawings herein.

FIG. 6 is a web diagram showing options for implementing the method according to aspects of the present disclosure. It is part of a non-limiting example of features that may be associated with predicting patient responsiveness to classifier training and treatment according to aspects of the present disclosure. It is a web diagram which shows an example of how the method of training a classifier can be carried out according to the aspect of this disclosure. It is a web diagram showing an example of how a method of predicting a subject's responsiveness to treatment using a trained classifier can be performed according to aspects of the present disclosure. It is an example of a method of reporting a subject's responsiveness to treatment when predicted according to aspects of the present disclosure by a trained machine learning based classifier. It is an example of a method of reporting a subject's responsiveness to treatment when predicted according to aspects of the present disclosure by a trained machine learning based classifier. It is an example of a method of reporting and comparing the responsiveness of a subject to various treatments when predicting according to aspects of the present disclosure by a trained machine learning based classifier. It is an example of a method of reporting and comparing the responsiveness of a subject to various treatments when predicting according to aspects of the present disclosure by a trained machine learning based classifier. It is an example of a method of reporting and comparing the responsiveness of a subject to various treatments when predicted according to aspects of the present disclosure by a trained machine learning based classifier that does or does not include a gene set as a feature. It is an example of a method of reporting a subject's responsiveness to treatment as expected according to aspects of the present disclosure by a trained machine learning based classifier. These are comparative graphs and figures demonstrating that the use of 38 features results in better classifier performance compared to the use of a single factor. These are comparative graphs and figures demonstrating that the use of 38 features results in better classifier performance compared to the use of a single factor. It is a graph and figure which compares the performance of a machine learning classifier using 38 features without a gene set or 44 features with a gene set. It is a graph and figure which compares the performance of a machine learning classifier using 38 features without a gene set or 44 features with a gene set. It is a graph and figure which compares the performance of a machine learning classifier using 38 features without a gene set or 44 features with a gene set. It is a graph and figure which compares the performance of a machine learning classifier using 38 features without a gene set or 44 features with a gene set.

The present disclosure relates to machine learning methods for obtaining predictions of an individual's ability to respond to treatment with checkpoint inhibitors or other cancer treatments. It is not known whether different individuals will respond to a given treatment. Responsiveness to treatment can depend on the presence or absence of a given feature or a quantitative amount of a given feature, respectively. Rather, a number of features may be differentially compounded among individuals to increase the likelihood of responding to a given treatment, while others may reduce it. The standard method of predicting patient responsiveness based on a single feature, or just a few features, is accurate in responsiveness in situations where multiple factors that can vary independently of each other work together in this way. Do not predict.

Machine learning methods offer new solutions to this shortcoming of common diagnostic prediction methods. In supervised machine learning, for example, a large data set that represents a large number of characteristics of many individuals and is combined with each individual's known responsiveness to a given treatment can be loaded onto computer memory storage. The storage medium of the computer instructs the computer processor to process the personal trait information and responsiveness, and issues an instruction to identify a pattern of trait information indicating the high or low likelihood that the individual is responsive to a given treatment. Including. The advantages of using machine learning methods for such analyzes are the identification of patterns of features and their adaptation to the prediction of responsiveness to treatment, which would not be possible without the high ability of computer systems to store and retrieve data. It is to enable its ability to process large amounts of information. Computer-executed machine learning systems process dozens or hundreds of features, or dozens or hundreds of objects, to identify patterns of features and their aggregated correlation with individual responsiveness. Can be done. The patterns identified in this way are otherwise undetectable and may require a computer system to perform a large amount of data processing on a set of complex information.

To determine the characteristics of an individual, tissue samples of the individual, eg, an individual with cancer, can be obtained and the characteristics of such tissue can be determined. In some examples, the tissue obtained may be a sample of cells taken from the tumor. In another example, the tissue obtained may be non-tumor tissue obtained from an individual.

As used herein, checkpoint inhibition refers to a treatment in which tumor cells activate self-recognition pathways in the immune system, thereby blocking processes that prevent immune cell attack and tumor cell lysis. Activation of such pathways by tumor cells is thought to contribute to the refractory of some patients to cancer immunotherapy, for example, when modifying immune cells to recognize and target tumor cells. An example of checkpoint inhibition is the CTLA4 pathway. CTLA4 is a protein receptor expressed on regulatory T cells. It can also be expressed on normal T cells after activation, as is often the case in cancer. When CTLA4 binds to CD80 or CD87, the protein is expressed on the surface of antigen-presenting cells, resulting in immunosuppression. In healthy cells, this mechanism promotes self-awareness and prevents the immune attack of autologous cells. However, in cancer, upregulation of this pathway helps tumor cells escape detection and attack by the immune system. Examples of checkpoint inhibitory therapies that block this pathway (eg, the anti-CTLA4 antibody ipilimumab) target and destroy tumor cells, for example, when combined with other cancer immunotherapies that stimulate anti-tumor immune responsiveness. Can promote the ability of the immune system to do so.

Similarly, another example of checkpoint inhibition is the PD-1 pathway. When PD-1 on T cells binds to the PD1-L1 receptor expressed on autologous cells, an immunosuppressive response occurs. Like the CTLA4 pathway, this pathway is also utilized by tumor cells to escape detection and attack by the immune system. Examples of checkpoint inhibition therapies that block this pathway (eg, the anti-PD-1 antibodies pembrolizumab, nivolumab, and semiprimab and the anti-PD-L1 antibodies atezolizumab, avelumab, and durvalumab) are, for example, anti-PD-1 antibodies. When combined with other cancer immunotherapies that stimulate tumor immune responsiveness, it can enhance the ability of the immune system to target and destroy tumor cells.

As used herein, terms such as checkpoint inhibitor or checkpoint inhibitory therapy refer to such therapies, as well as the interaction of CTLA4 or PD-1 with their cognate ligands or receptors, or theirs. Includes other therapies that inhibit checkpoint inhibition pathways, including treatment with other antibodies or pharmaceutical compositions that function by preventing downstream signaling sequelae or activation of cell function.

Many features may be relevant in the generation of predictions of whether an individual is responsive to a given treatment. For example, it is reflected in the gene sequence information contained in the genome of a cell obtained from an individual, the gene sequence information expressed in RNA transcribed from the genome of an individual cell, the corresponding RNA transcription in a sample, or the amount of its protein product. Expression levels of transcripts of possible genomic sequences, or various cells present in the sample. In examples where the sample contains tumor tissue or cells of individual origin, such information is the tumor cells, i.e., one compared to the non-tumor cells of the population or individual or the control genome referred to in the gene sequencing paradigm. Alternatively, it can exhibit the characteristics of cells having multiple modified genomic sequences. Tumor formation can result from one or more mutations in a nucleotide sequence and / or mutations in two or more sequences in genomic DNA. In some cases, for example, the accumulation of multiple such sequence modifications can function with each other to transform cells from non-disease cells to tumor cells. In other cases, early accumulation in one or some of the cells with such modifications may cause further accumulation of such modifications in the cells. In yet other cases, proliferation in cells with such modifications may indicate that any particular modification is not always directly involved in or responsible for the growth of the tumor. Rather, the tumorigenesis process can result in some modifications that directly induce the transformation of cells into tumor cells, but can also result in other modifications that do not directly induce the transformation of cells into tumor cells.

Thus, there are individual tumor cells that may have many such modifications in the genomic DNA sequence, and there are individual tumor cells that may have only minor such modifications. Among them, some tumor cells may produce transcripts or protein products with modified amino acid sequences as a result of genomic modification, for example, modification of the DNA sequence in genomic DNA does not result in modification. A protein or RNA molecule with a different sequence than it should have been produced may be produced. Such modifications are called non-synonymous mutations. Other modifications can be modifications of non-coding DNA or modifications of coding DNA that do not mutate the protein amino acid sequence. For example, modifications to intron sequences or non-transcribed DNA may not result in protein products that differ in amino acid sequence from amino acid sequences of proteins produced from genomes that do not retain the same modification. Such modifications are called synonymous mutations. Thus, different tumors contain different numbers of non-synonymous mutations, different numbers of synonymous mutations, or both different numbers of mutations (or contain the same total number of genomic mutations, but different numbers of synonymous mutations. And various numbers of non-synonymous mutations are not included), and various numbers of modifications can be included in genomic DNA. The number of mutations carried by a cell is called its mutation frequency or total mutation frequency, while the total number of non-synonymous mutations held by a cell is called its non-synonymous mutation frequency and is called the synonymous mutation held by that cell. The number of is called its synonymous mutation frequency.

Cell conversion from non-tumor cells to tumor cells can correspond to the accumulation of such modifications in genomic DNA. Such accumulation can be the accumulation of synonymous mutations, the accumulation of non-synonymous mutations, or both types of accumulation. In any case, the total mutation frequency can increase during non-tumor cell to tumor cell conversion, during non-tumor cell to tumor cell conversion, and after non-tumor cell to tumor cell conversion. In addition, the frequency of tumor cell mutations can affect whether checkpoint inhibition may be effective in stimulating an antitumor response by the immune system. The more mutations a tumor cell retains, the greater the chance that suppressing checkpoint inhibition will allow it to desuppress the immune system by recognizing and attacking the tumor cell as a diseased cell. In particular, the frequency of non-synonymous tumor mutations can positively correlate with the ability of checkpoint inhibition to desuppress the antitumor immune response. Proteins with mutated amino acid sequences produced as a result of non-synonymous mutations can be identified as abnormal in cells and present on the cell membrane as a sign of intracellular pathology. For example, a tumor can express a protein with a mutated amino acid sequence called a nascent antigen. Tumor cells expressing such nascent antigens can express mutant fragments of such nascent antigens on their cell membranes.

Presenting such a neogenic antigen stimulates the immune system's perception that the cell is lesioned (eg, a tumor cell) and promotes the immune system to target and destroy such cells. obtain. However, the tumor offsetting process can utilize the checkpoint pathway to escape immunodetection. Therefore, whether checkpoint pathway inhibitors can assist in enhancing cancer immunotherapy may depend on the frequency of tumor mutations. Higher frequency may correspond to higher levels of neonatal antigen presentation and increase the probability of stimulating antitumor immunogenicity when checkpoint inhibitory therapy is given. In general, a high total mutation frequency may indicate a high non-synonymous mutation frequency, and a high total mutation frequency may indicate increased expression of the nascent antigen. Furthermore, a high frequency of non-synonymous tumor mutations can also indicate increased expression of the nascent antigen, thus indicating a high likelihood that checkpoint inhibition may be effective. Response to checkpoint inhibition so that synonymous tumor mutation frequency can correlate with responsiveness to checkpoint inhibition and / or some combinations of synonymous and non-synonymous tumor mutation frequencies can be reflected in total mutation frequency. It is also possible that it can correlate with sex. Thus, total mutation frequency, non-synonymous tumor mutation frequency, synonymous tumor mutation frequency, or any combination of two or more of these can predict checkpoint responsiveness, according to the machine learning methods disclosed herein. It can be one or more features included.

In addition to whether the mutations are synonymous or non-synonymous, or in addition, there may be other or additional mutational properties, the number or type of which is synonymous with one or more mutations. As well as being or non-synonymous, it may be relevant in predicting responsiveness to treatment. For example, some mutations, called non-stop mutations, are mutations at the stop codon, where the mutant portion of the RNA transcript results in translation of the RNA product that would otherwise continue beyond termination. Another form of mutation is a frameshift mutation, which is the insertion or deletion (frameshift deletion) of a large number of adjacent nucleotides that are indivisible in triplets (eg, a single nucleotide). Insertion or deletion). This results in a shift in the codon reading sequence, thus resulting in the recruitment of various tRNA molecules in the translation of the produced RNA transcript, thus altering the amino acid sequence of the translated protein. Another mutation is a splice site mutation, which occurs at or near the splice site and thus modifies normal mRNA splicing, resulting in modified RNA transcripts. Alternatively, the mutation is a missense mutation, in which a single nucleotide is mutated, which mutates the codon containing that single nucleotide, recruiting different types of tRNAs in translation and thus different amino acid sequences. Can produce the proteins it has.

Another possible mutation is a start mutation, a mutation to the transcription start site or to the start codon, which can result in a mutation at the site that initiates transcription or translation, respectively. For example, mutations in the initiation site can prevent the initiation of transcription from that initiation site. Alternatively, the mutation may generate a transcription initiation site that did not previously exist. The mutation at the transcription initiation site is of a different length than the RNA transcript that would otherwise have been produced, but is either in-frame or out-of-frame because transcription occurs in the absence of the mutation. Obtain, transcription of RNA transcripts can occur. Similar mutations to the start codon can also occur, resulting in transcription of the RNA product, where translation initiation does not occur or translation initiation occurs from sites that have not previously initiated. Such stop codon mutations can be in-frame or out-of-frame. Alternatively, the mutation can be a nonsense mutation, i.e., a mutation that results in an RNA transcript with an immature stop codon. Any of the aforementioned mutations can be single nucleotide polymorphisms (SNPs). One or more of the various mutations described above can be features associated with a given treatment, eg, prediction of an individual's responsiveness to a given checkpoint inhibitor.

In another example, the amount of tumor infiltration by lymphocytes can predict responsiveness to checkpoint inhibitors. Tumors include not only cells transformed from non-tumor cells to tumor cells, but also other non-transformed cells. Examples include cells of the immune system that play or may play a role in stimulating an immune response against transformed cells in a tumor. Immune cells that mix with transformed cells in the tumor, especially lymphocytes, refer to tumor-infiltrating lymphocytes. Whether the subject from whom the tumor sample was taken could be responsive to checkpoint inhibition, depending on the level of tumor-infiltrating lymphocytes and tumor-infiltrating lymphocytes expressing various markers that act as identifiers for the lymphocyte phenotype. Can be predicted. Since the tumor is, for example, a heterologous mixture of tumor cells and tumor-infiltrating lymphocytes, this is a lymphocyte marker expressed on the tumor-infiltrating lymphocytes present in the sample, and in tumor samples such as transformed tumor cells. It can be an advantage for identifying lymphocyte markers that are potentially expressed on other cells.

For example, tumor-infiltrating lymphocytes may express transcripts (eg, RNA) of genes encoding surface antigen classification 8 (CD8), surface antigen classification 4 (CD4) or surface antigen classification 19 (CD19). Each of these can act as a marker of lymphocyte phenotype when expressed in cells. Thus, the expression level of CD8, CD4, CD19, or any combination of any two or more of the above can be determined from the tumor sample. Thus, for example, the amount of RNA can be determined from the sample. This is because the amount of detection of its expression is tumor invasion, as such a determination can reflect not only its expression by tumor infiltrating lymphocytes, but also other cells in the tumor sample, such as transformed tumor cells. It can be an advantage in determining to what extent it can and cannot be attributed to lymphocyte expression. To do so, a denconvolution process can be applied, which allows the level of expression by tumor-infiltrating lymphocytes to be determined relative to expression by other cells. Various options are available for performing tumor-infiltrating lymphocyte deconvolution, such as Gaujoux et al. (2013) CellMix: a comprehensive toolbox for gene expression deconvolution, Bioinformatics 29: 2211-2212, and F Finotello et al. (2018), Quantifying tumor-infiltrating immune cells from transcriptomics data, Cancer Immunology, Immunotherapy 67: 1031-1040. Deconvolution analysis can be performed in the R programming language, as a non-limiting example.

In particular, by a process called tumor infiltrating lymphocyte deconvolution, the expression levels of a given lymphocyte transcript (eg, CD8, CD4 or CD19) are used to generate CD4 or CD8 or CD19 of lymphocyte cells in a tumor sample. The rate of expression can be determined. That is, it does not merely indicate the amount of lymphocyte infiltration of the tumor, which is represented by the amount of given lymphocyte transcript that can be identified in the tumor, and the tumor infiltrating lymphocyte deconvolution depends on the type of transcript expressed. It is possible to indicate the type of lymphocyte to be identified and further provide a percentage of the total tumor infiltration. The total amount of lymphocyte infiltration in a tumor may be relevant in predicting whether an individual may be responsive to a given treatment, such as a checkpoint inhibitor, and given transcripts (eg, eg). Specific contributions of tumor lymphocyte infiltration caused by lymphocytes expressing CD4 or CD8 or CD19), but not limited to these, may also be relevant in making such predictions.

Various other features may be relevant in predicting whether an individual may be responsive to checkpoint inhibition. Such features can generally be categorized according to theoretical processes as to whether checkpoint inhibitors may be effective in promoting or enhancing the cancer immune response. For example, some features are whether tumor cells can increase or decrease the probability of an anti-tumor immune response by expressing antigens of mutant proteins on their cell surface, and which It can be about whether it can be as high or low as possible. Examples of such features already discussed include various tumor mutation frequencies, such as total tumor mutation frequency or non-synonymous tumor mutation frequency. Other examples include expression levels of transcripts of genes encoding various proteins, or proteins known to be involved in antigen presentation and stimulate an immune response. Some non-limiting examples are beta 2 microglobulin (B2M), proteasome subunit beta 10 (PSMB10), antigen peptide transmitter 1 (TAP1), antigen peptide transporter 2 (TAP2), human leukocyte antigen A ( HLA-A), major histocompatibility complex class IB (HLA-B) expression, major histocompatibility complex class IC (HLA-C), major histocompatibility complex class II DQ alpha 1 (HLA-DQA1), and HLA Class II histocompatibility complex DRB1 beta chain (HLA-DRB1) is mentioned. Such gene products are known to undergo various steps in the presentation of protein fragments, antigens, or cell surfaces and recognition by immune T cells.

The expression level of any one of such gene products, or any combination of two or more thereof, or any combination thereof in a tumor can indicate the expression level of the antigen on the surface of the tumor cell. For example, expression of such products can affect the degree of presentation of genomic DNA protein products with non-synonymous mutations on the cell surface. If expression increases antigen presentation, the likelihood of presenting a mutant antigen, which T cells recognize as indicating diseased cells and can result in an antitumor immune response, is a given checkpoint inhibition. It can increase the probability that a substance can be effective in producing a response in a subject. Therefore, if the expression level of any or a combination of two or more of the above correlates positively or negatively with the degree of antigen presentation on the cell in the tumor sample, such expression level is the subject from which the tumor was collected. Can correlate positively or negatively with the likelihood of responding to a given checkpoint inhibitor, respectively.

Another type of feature can include the expression level of T or NK cells present in the tumor sample taken from the subject, which is also associated with predicting responsiveness to treatments such as checkpoint inhibition. Can be done. For example, the level of expression of HLA Class I tissue-compatible antigen alpha chain E (HLA-E), natural killer cell granule protein 7 (NKG7), chemokine-like receptor 1 (CMKLR1), or any combination of two or more of these. Can also predict the response to checkpoint inhibitors. Thus, features may include measures of expression of one or more of such products or RNA transcripts thereof.

Another type of feature is the presence of expression of a protein or transcript by it relating to or implying an enhancement of cytolytic activity that can be promoted by an anti-tumor immune response or inhibit such activity. Or it can be about the level. The scale of tumor infiltrating lymphocyte deconvolution discussed above can be an example of such a feature (eg, surface antigen classification 8 (CD8), surface antigen classification 4 (CD4), or surface antigen classification 19 (CD19)). Deconvolution of tumor infiltration by cells expressing). Other non-limiting examples of features in this category may include Granzyme A (GZMA) or perforin 1 (PRF1) or any combination of two or more of the aforementioned, or the level of expression of RNA transcripts by them. ..

Yet other features may relate to the process or function of antitumor immune responsiveness checkpoint inhibition. Whether the level of expression of various protein products or transcripts thereof that contribute to checkpoint inhibition in a subject-derived tumor sample allows the subject to respond to treatment with cancer treatments such as checkpoint inhibition therapy. Can be relevant in the prediction of. Examples of such features are cytotoxic T lymphocyte-related protein 4 (CTLA-4), 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 and ITIM domains (TIGIT), Surface antigen classification 276 (CD276), or expression of any two or more of the above, or RNA transcription thereof. The expression of a substance can be mentioned.

Other features that may be relevant in predicting whether an individual may respond to treatment are related to interferon gamma activity, such as products whose expression is downstream of the release of interferon gamma and its activity at the receptor. Includes expression of a protein or RNA transcript thereby. Examples of this type of feature are chemokine (CC motif) ligand 5 (CCL5), CD27, chemokine (CXC motif) ligand 9 (CXCL9), CXC motif chemokine receptor 6 (CXCR6), indolamine 2,3-di Expression of oxygenases (IDOs), signaling and transcriptional activators 1 (STAT1), or any combination of two or more of the aforementioned, or RNA transcripts by them can be mentioned. Other indicators of interferon gamma activity can also predict responsiveness to treatments such as checkpoint inhibition.

Other features that may be relevant in predicting whether an individual may respond to treatment include expression of proteins or RNA transcripts by them for bone marrow-derived immunosuppressive cells (MDSCs) or regulatory T cells (Tregs). This may confer immunosuppressive effects on antitumor immune responsiveness and may blunt or hinder the effectiveness of cancer immunotherapy. Examples of such characteristics include 3-fucosyl-N-acetyl-lactosamine (CD15), interleukin-2 receptor alpha chain (CD25), siglec-3 (CD33), and surface in tumor samples derived from the target tumor. Expression of antigen classification 39 (CD39), surface antigen classification 118 (CD118), forkhead box P3 (FOXP3), or any combination of two or more of the aforementioned may be mentioned. Tumor expression levels of other types of proteins or corresponding RNA transcripts, which imply the presence or activity of such cells, are also indicated by individuals for checkpoint inhibition therapy or other therapies for cancer. It can be related to whether or not it can respond.

Any one or more of the aforementioned features will vary in degree in predicting whether an individual will be responsive to treatment with a given cancer treatment, including treatment with checkpoint inhibitors. Can be related at. Any of the features may relate to or embody genomic information about the tumor of interest that was taken and tested for trait determination. In this case, genomic terminology is used to include features that are not just information about nucleotide sequences in genomic DNA (eg, features related to mutation frequency, etc.). As used herein, genomic information represented by a measure of characteristics also includes a measure of the expression level of genomic transcripts, or various products of protein products produced from such transcripts. Thus, the expression levels of any of the various protein products described above, or other protein products involved in similar pathways, such as those identified in detail, or the expression levels of RNA transcripts by them, are described herein. It can be included in genomic information because it relates to the predictive features of the disclosure. Also, what is included in the genomic information about the features can be a measure of the features of tumor-infiltrating lymphocyte deconvolution.

In addition to measures of individual traits, patterns of correlated expression levels of traits known or thought to be related to a given pathway or function or cell type are also checkpoint inhibition or other. It can be a feature related to responsiveness to cancer treatment. For example, among the above-mentioned characteristics, groups showing some commonality of pathways, or cell or physiological responsiveness, or cell phenotype can be identified, and they can be distinguished as a group in a coordinated manner. Individual characteristics of upregulation or downregulation, or more commonly expressed, or otherwise present as a group with high or low levels of correlation in a sample from a particular subject tumor. Judgments can be made based on the measurements. In some examples, such a generalized measure of grouped features, in addition to individual features, inputs itself as a feature to train a machine learning classifier and checkpoint. The responsiveness of the subject to inhibition and / or other treatments can be predicted. In the present specification, such a feature group for obtaining additional features representing the expression level of the whole group and the like is referred to as a gene set. Thus, a gene set can include a combination of scales that represent a correlation such as the presence of genomic mutations, the expression level of a particular RNA transcript, the presence of an identified cell type, and the like.

In a non-limiting example, some of the aforementioned features relate to antigen presentation, which causes tumor-like cells to express protein fragments on their cell membranes and are monitored by the immune system. As mentioned above, antigen presentation can increase the likelihood of stimulating an antitumor immune response, for example, with checkpoint inhibitors. Some examples of such features are total mutation frequency, non-synonymous mutation frequency, or other mutation frequency (nonstop mutation frequency, frameshift mutation frequency (insertion, deletion, or either), splice. Site mutation frequency, missense mutation frequency, start mutation frequency (in-frame, out-of-frame, or either), nonsense mutation frequency, start codon mutation frequency (including start codon SNP or others), in-frame insertion mutation frequency, in-frame lack Missense mutation frequency, or other SNP mutation frequency), or any combination of the two or more mentioned above can be mentioned. Other non-limiting examples of features related to antigen presentation are beta 2 microglobulin (B2M), proteasome subunit beta 10 (PSMB10), antigen peptide transmitter 1 (TAP1), antigen peptide transporter 2 (TAP2). , Human leukocyte antigen A (HLA-A), Major histocompatibility complex class IB (HLA-B) expression, Major histocompatibility complex class IC (HLA-C), Major histocompatibility complex class II DQ alpha 1 (HLA) -DQA1) and HLA class II histocompatibility complex DRB1 beta chain (HLA-DRB1) can be mentioned. In addition to features relating to the presence or expression level of individual examples from the aforementioned features, additional features may or may not be coordinately upward or downward controlled in part or all of the aforementioned features, or otherwise subject. It may represent a high or low level of presence in the tumor (whether for training of machine learning classifiers or for prediction).

As another non-limiting example, some features relate to the expression levels of T or NK cells present in tumor samples taken from a subject, which is also responsive to treatments such as checkpoint inhibition. Can be related to the prediction of. For example, the expression level of HLA class I tissue-compatible antigen alpha chain E (HLA-E), natural killer cell granule protein 7 (NKG7), chemokine-like receptor 1 (CMKLR1), or any combination of two or more of these It can also predict the response to checkpoint inhibitors. In addition to features relating to the presence or expression level of individual examples from the aforementioned features, additional features may or may not be coordinately upward or downward controlled in part or all of the aforementioned features, or otherwise subject. It may represent a high or low level of presence in the tumor (whether for training of machine learning classifiers or for prediction).

As another non-limiting example, some features, such as when an immune response promotes cell death and lysis of cells such as tumor cells present in a tumor sample taken from a subject. With respect to indicators of cell lysis by immune stimulation, which may also be associated with predicting responsiveness to treatments such as checkpoint inhibition. For example, decombined CD8 expression, deconvoluted CD4 expression, deconvoluted CD19 expression (deconvolution representing proportionally contributing CD8, CD4, or CD19 expressing cells is tumor invasion present in the tumor sample. The expression level of RNA transcripts by (represented by the number of lymphocytes), Granzyme A (GZMA) or perforin 1 (PRF1), or any combination of two or more of the aforementioned, or by them, is also a checkpoint inhibitor. Can predict the response to. In addition to features relating to the presence or expression level of individual examples from the aforementioned features, additional features may or may not be coordinately upward or downward controlled in part or all of the aforementioned features, or otherwise subject. It may represent a high or low level of presence in the tumor (whether for training of machine learning classifiers or for prediction).

As another non-limiting example, some features relate to cellular and molecular processes involved in checkpoint inhibitory function present in tumor samples taken from a subject, which also treats such as checkpoint inhibition. May be related to predicting responsiveness to. Non-limiting examples of these characteristics include cytotoxic T lymphocyte-related 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 (TIGIT) having Ig and ITIM domains, Surface antigen classification 276 (CD276), or expression of any two or more of the above, or RNA by them. Expression of transcripts can be mentioned. In addition to features relating to the presence or expression level of individual examples from the aforementioned features, additional features may or may not be coordinately upward or downward controlled in part or all of the aforementioned features, or otherwise subject. It may represent a high or low level of presence in the tumor (whether for training of machine learning classifiers or for prediction).

As another non-limiting example, some features relate to indicators or cellular and molecular pathways involved in interferon gamma activity present in tumor samples taken from a subject, which are also like checkpoint inhibition. It may be related to the prediction of responsiveness to various treatments. Non-limiting examples of such features include chemokine (CC motif) ligand 5 (CCL5), CD27, chemokine (CXC motif) ligand 9 (CXCL9), CXC motif chemokine receptor 6 (CXCR6), indolamine 2 , 3-Dioxygenase (IDO), signal transduction and transcriptional activator 1 (STAT1), or the expression of any combination of two or more of the aforementioned, or the expression of RNA transcripts by them. In addition to features relating to the presence or expression level of individual examples from the aforementioned features, additional features may or may not be coordinately upward or downward controlled in part or all of the aforementioned features, or otherwise subject. It may represent a high or low level of presence in the tumor (whether for training of machine learning classifiers or for prediction).

As another non-limiting example, some features relate to the presence or activity of MDSCs or Tregs present in tumor samples taken from a subject, which is also responsive to treatments such as checkpoint inhibition. Can be related to the prediction of. Non-limiting examples of such characteristics include 3-fucosyl-N-acetyl-lactosamine (CD15), interleukin-2 receptor alpha chain (CD25), siglec-3 (siglec-3) in tumor samples derived from the tumor of interest. Expression of CD33), surface antigen classification 39 (CD39), surface antigen classification 118 (CD118), forkhead box P3 (FOXP3), or any combination of two or more of the aforementioned may be mentioned. In addition to features relating to the presence or expression level of individual examples from the aforementioned features, additional features may or may not be coordinately upward or downward controlled in part or all of the aforementioned features, or otherwise subject. It may represent a high or low level of presence in the tumor (whether for training of machine learning classifiers or for prediction).

Thus, in some examples, one or more gene sets can be identified, and a measure of the extent to which the up or down regulation of features of such gene sets coordinate or correlate in a tumor of interest is mechanical It may be provided as an additional feature for training a learning classifier or predicting a subject's responsiveness to checkpoint inhibition or other treatments or both. Examples of gene sets include antigen presentation, T and NK cell properties, cell lysis indicators, checkpoint inhibition, interferon gamma, and gene sets for the presence or activity of MSDC / Treg. In some cases, one or more such gene sets, along with one or more of the other individual characteristics discussed above, may be the patient's responsiveness to checkpoint inhibition or other treatments. It can be included in training or machine learning classifiers used for prediction. A generalized measure of how features in a gene set feature are coordinately up or down regulated, or otherwise coordinated or correlated at high or low levels. Various methods can be used for confirmation. One example may be an analysis called single sample gene set enrichment analysis (ssGSEA). In ssGSEA, for example, as described in Barbie et al. (2009), Systematic RNA interference reveals that oncogenic KRAS-driven cancers require TBK1, Nature 462: 108-112, using an empirical cumulative distribution function, such Confirm enrichment of the grouped gene set. ssGSEA can be implemented in the R programming language as a non-limiting example.

In the features described above, some, or any combination of any two or more, is used to train machine learning classifiers, and using trained machine learning classifiers, eg, for treatment with checkpoint inhibitors. It can be used both to predict whether the subject is responsive to it and to what extent it is likely. However, it is not necessary to train a machine learning classifier with all of the above features. Machine learning classifiers can be trained with a set of features that include all of the above, or exclude one or more of the above. All combinations and permutations suggested by this optional inclusion and exclusion are not necessarily listed explicitly and verbatim, but are incorporated by this in their entirety. One of ordinary skill in the art will be able to conceptualize subsets, combinations, partial combinations, and permutations that can occur due to the features described above. Similarly, additional features can also be included by any of all or just combinations, partial combinations, permutations, or in addition to a smaller number of other mixtures than all of the above features. All such various examples are expressly included in this disclosure.

In some examples, the characteristics of any object used in training a machine learning classifier are the same as those used in training a machine learning classifier for all other objects. However, in other examples, various features can be provided for different objects used to train machine learning classifiers. In other words, some subjects may have features included in a series of trainings for that subject that are not included in the features derived from the training set of the other subject. Similarly, in some examples, to obtain predictions about a subject's responsiveness to treatment from a machine learning classifier, the features obtained to obtain predictions from a subject-derived tumor sample are used for classifier training. Can be identical to features. That is, all the object-derived features used in the training of the machine learning classifier are the same as each other, and can be the same as the features of the object for which prediction from the machine learning classifier is required. In another example, there may be inconsistencies between the characteristics of the trained subject used to train the machine learning classifier and the characteristics of the subject for which predictions from the machine learning classifier are required. Some or all object-derived features used in training the machine learning classifier may include features for which there is no corresponding object-derived feature that is required to be predicted from the machine learning classifier.

In some examples, the subject for which prediction from the machine learning classifier is required may lack features corresponding to features from one or some or all of the objects used in training the machine learning classifier. In another example, the subject may have similar features, but not the same features, and similar features can be used in place of the same features in which the subject does not exist. For example, a machine learning classifier is trained for at least some of the training subjects for features that include one or more gene set features, eg, gene set features that can be obtained using the ssGSEA described above. You may be. For some of the training subjects that used gene sets to train machine learning classifiers, some of such sets may be derived from the same underlying personal characteristics. For example, a gene set of features of an antigen-presenting-related gene set may be derived from the same underlying features for all subjects used to train machine learning classifiers. In another example, some antigen-presenting-related gene sets for one training subject include some of these features that were not included in the identification of antigen-presenting-related gene sets from another training subject. Obtained based on the characteristics that make up. This also applies to other gene sets. In addition, gene set feature targets that require predictions from trained machine learning classifiers can be used to obtain predictions, and gene set feature values correspond to one or more trained subjects. The basis of the individual features of the subject origin, which is used to obtain the value of the features of the gene set to be used and does not contain at least one or more underlying features used for training the machine learning classifier. It may be obtained from the training set.

The characteristics can be confirmed by gene sequencing data or a known method for quantifying the expression level of a protein or RNA transcript in a biological sample. For example, the significant amount of nucleotide sequence information that can be obtained using next-generation sequencing techniques depends on the type of next-generation sequencing used to obtain a given feature, and genome-related features (eg, total). It can result in both mutation frequency) and, for example, expression levels of RNA transcripts. 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 array, or the disclosure of the present disclosure. Other relevant methods for ascertaining the presence or level or quantity of features used to train and / or acquire predictions from machine learning classifiers by aspect can be mentioned. In some examples, a series of identical techniques can be used to obtain features from all trained objects for training machine learning classifiers and from objects for which prediction is required. In other examples, methodological differences between the method of determining features or some features for various training subjects and / or the method of obtaining the features used to obtain predictions for subjects requiring prediction. Can exist.

In addition to training the machine learning classifier with features derived from the trained subject, the responsiveness of the trained subject to treatment is also loaded into the machine learning classifier. Therefore, the training object is an object that provides features and responsiveness for training a machine learning classifier. Responsiveness, for example, when the trained subject is classified as responding to treatment, did the subject exhibit a given response, including extended lifespan, tumor contraction, partial or complete remission, etc. It can be a binary classification, such as. In another example, responsiveness can be a score or value based on the degree of responsiveness obtained, rather than a binomial assessment of whether responsiveness was obtained. For example, the machine learning classifier according to the present disclosure may include classification trees and regression trees as non-limiting examples, depending on the type of prediction required.

The machine learning classifier can be any classifier suitable for computer-based machine learning. A non-limiting example is the Random Forest Machine Learning Classifier. In a random forest machine learning classifier, a decision tree based on the characteristics of the trained object and its responsiveness to treatment values is created with nodes representing classification decision points and leaves representing results based on trained inputs. Random forest classifiers can generate multiple trees using a subset of features and a subset to be trained to create a large number of trees and then aggregate them. Such multiple decision trees, including a subset of inputs, prevent overtraining and reduce errors and biases in prediction. In some examples, more decision trees can be created in the training of more accurate machine learning classifiers. In some examples, any decision tree from 5,000 to 500,000 can be created in the 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 can be run and aggregated into a random forest machine learning classifier. It is possible to run more or fewer trees than can be counted in such an exemplary possible range.

Numerous options are available for conducting random forest training and generating predictions from random forest classifiers. As a non-limiting example, the R programming language can be used. Other classifiers can also be used in accordance with aspects of the present disclosure, including neural network classifiers, support vector machines, maximum entropy classifiers, super-gradient boosting classifiers, and random fern classifiers. However, it is not limited to these.

According to the methods disclosed herein, a feature is obtained from each of a number of trained subjects as the responsiveness of each such subject. Such features and responsiveness, or inputs, enter into a computer memory storage device, such as a hard drive, server, or other memory component. Also, what is stored on the memory storage function of such a computer is an instruction included in the software, which directs one or more microprocessors. The instructions include instructions for creating a machine learning classifier using inputs from the trained subject. The trained machine learning classifier can then be stored in one or more computer memories and subsequently performed for subject-derived features for which predictions about responsiveness are required.

As a non-limiting example, an instruction directs one or more microprocessors to perform random forest training and create a decision tree from the input to be trained, with or without the presence or absence of various features. It can be confirmed whether the level etc. is responsive to treatment and is likely or low to aggregate a large number of decision trees into a trained random forest machine learning classifier. The machine learning classifier trained by this example, based on the aggregation of decision trees generated by one or more microprocessors when processing features and responsiveness according to instructions, is then stored in one or more memories. can do. Then, whether the untrained subject (ie, the subject who did not use the feature values to train the trained machine learning classifier) can be responsive to treatments such as certain checkpoint inhibitors. The features obtained from this untrained tumor sample can be loaded into one or more memories when predicting. One or more microprocessors process instructions and analyze untrained features with a trained machine learning classifier, accessed from one or more memories by one or more microprocessors. You can report predictions about the responsiveness of the subject.

In such an example, a machine learning classifier was trained and trained on the characteristics of tumor samples obtained from each of the multiple training subjects, and the responsiveness of each of the multiple training subjects to treatment, including checkpoint inhibition. A machine learning classifier, which was trained to predict responsiveness to treatment. In addition, a set of untrained subject genomic information, or feature values, including untrained features derived from the subject tumor profile can be entered into a trained machine learning classifier for treatment responsiveness classification for the untrained subject, eg. , Generated classifications or scores that predict how untrained subjects may respond to treatment. Treatment can be checkpoint inhibition.

The checkpoint inhibition responsiveness for untrained subjects generated by the trained machine learning classifier can be reported to the user. The report can include a classification of subjects that indicates in a binomial assessment whether untrained subjects are expected to respond to treatment. In another example, the numerical score by grade may indicate the probability of responsiveness in addition to, or not to, the binary classification of whether the untrained subject is expected to respond. In another example, a particular degree of responsiveness can be reported. For example, reports can show high likelihood responsiveness, but responsiveness can be limited in terms of duration or degree. In another example, the report can show a relatively low likelihood prediction of responsiveness, but if such an untrained subject is expected to respond, it will have a longer duration or degree of responsiveness. Can show predictions.

Prediction reports can be reported via a graphical user interface (GUI) as responsive scores or binary predictions, as opposed to those that are unlikely to respond. For example, a computer or computer system that is connected to one or more memories and one or more microprocessors, inputs an untrained profile of feature values, and analyzes it with a trained machine learning classifier. It can be further connected to a display device where predictions are visually reported. The GUI can take any of many forms. For example, the GUI has a high degree of importance or weight in the generation of predictions, a subset of features or features, and each such factor has a high or low likelihood of the untrained subject responding to treatment. The various aspects of what has been shown (ie, the value of the feature) can be tabulated, taking into account the value of the feature for the untrained subject, as further described below. Alternatively, the report may include various shape, shadow, or color designs to report such information.

Predictive classification reports that use the set of features disclosed herein distinguish the subject matter disclosed herein from conventional methods for predicting responsiveness to treatment. Unlike conventional treatment prediction methods, methods that utilize combined methods are disclosed herein, in which case, in some cases, one or only a few may be separated in the generation of predictions. It is possible to search for a large number of features that are not interrelated. Beyond conventional methods, such properties disclosed herein are like checkpoint inhibitors, which were previously highly limited in accuracy and generalizability and were based on a limited number of features. It has advantages in previous attempts to predict whether an individual can respond to a given treatment. As disclosed herein, new machine learning processes for assessing the contributions of multiple factors, and how each influences prediction independently and in collaboration with others, are conventional in prediction. Overcome such limitations of the attempt.

Conventional methods for confirming responsiveness predictions have involved the identification of features or a limited number of features that can indicate whether the likelihood of responsiveness to treatment is high or low. In contrast, this specification discloses a novel method in which machine learning classifiers are used to assess the relative contributions of a large number of potential features, or to make more co-related predictions with each other. With the new application of machine learning classifiers, such multifactorial methods offer significant advantages over currently available methods.

A useful feature of some GUI examples for reporting responsive predictions is that they display multifaceted information about the features related to the reported predictions in a relatively small space to provide useful information to the user. It is possible that it can be confirmed compactly. In particular, the advantages of certain examples of GUIs used in accordance with this disclosure are for displays of limited or miniaturized size, or for displays that are required or desirable to display a significant amount of other information as well. It can be its sizing and placement for. The GUI for reporting predictive classifications according to the present disclosure provides a large number of facets, and / or features related to their generation, in a display of limited display size or proportion, which makes it easy to recognize and understand predictive classifications. Can function to fit in. Such a compact reporting feature is when the display is the screen of a portable electronic device such as a telephone or other wireless communication device (eg, a tablet computer or telephone or other portable wired or wireless device). With the computer system that provides the report, such as, in that this facilitates the reception and interpretation of the report and secures the display space that may be needed for other purposes or its limited capabilities. Improve connectivity. In the example showing a report by GUI, which condenses a large amount of characteristics and aspects of a feature into a confined space, and retains the ability to quickly convey a large amount of information that remains quickly and easily identifiable, more The usefulness of computer systems is enhanced in that more display space is still available at the same time for additional purposes, or is still available for use on smaller displays.

The report can include feature identifiers, or aspects or properties of some or all features used in the generation of responsiveness classifications. For example, the report can include an indication of the value of the feature, i.e., whether that value in the user's profile indicates an increase or decrease in the likelihood that the subject can respond to treatment. That is, in a non-trained profile, the feature can have a positive or negative value. The positive value tends to positively correlate the value of the feature in the trained subject with the responsiveness to treatment, and the value of the non-trained subject for the feature was high, or the value of the feature in the trained subject was the responsiveness to treatment It tends to correlate negatively, which may mean that the untrained value for that feature was low.

Another identifier can be an indication of the importance of the feature in generating predictions for a given machine learning classifier. The importance or feature can indicate that it is likely to drive the prediction in either direction relative to the other features used for training. For example, in some cases, the gini reduction index can be identified for one or more features during training. The Gini Decrease Index indicates the importance of a feature in that it represents the proportion of the effectiveness of the feature to the functionality of the classifier to the degree of influence that the other feature has on driving the generation of forecasts. .. The Gini Decrease Index can be determined using various software packages, for example by using the R programming language. In some examples, the gini reduction index of feature importance can be used to determine which feature identifiers are included in the report, either separately from or in addition to the predicted score. For example, in a given report in the form of a GUI, the GUI may only display features whose importance meets or exceeds a given minimum importance threshold. For example, a report may include only feature identifiers where the importance of a feature is expressed numerically as a gini reduction index and squared to greater than 0.1. The stringent minimum importance threshold can be set high or low in place of or in place of a given report, depending on the degree of information that should be included with the predicted score. The higher the minimum importance threshold, the fewer identifiers of features that can be included in the report, and vice versa. For example, the minimum importance threshold can be a threshold at which the square of the numerical importance (eg, the Gini reduction index) exceeds any of 0.01 to 0.5. Other minimum importance thresholds can be selected either during or outside this range.

Another example of a feature identifier that can be included in a report is a feature weight. The weight of a feature is a measure that the value of that feature of the subject can directly suggest whether the subject can respond to treatment. For example, for each feature, a single factor decision boundary can be determined. The single factor determination boundary is a value for that characteristic that best distinguishes whether a trained subject responds to treatment. For example, if all trained subjects responding to treatment had values for features above a given amount, but all non-responding treatment subjects had values less than that amount, the amount was a single factor. It can be a decision boundary. In some other examples, some responsive training subjects may have values of features that exceed some non-responsive training subjects, while other responsive training subjects may be less than that training subject. It may also have feature values. Thus, in some examples, there may be a bright line for the value of a feature that is clearly distinguished between the respondent and the non-responder, while in others, the responder and the non-responder. There may be more duplication at the boundary of feature values with a person. In the latter example, the single factor determination boundary can be selected as the value that results in the largest possible difference between responders and non-responders in the set of training subjects.

A feature weight is a measure or indication of how different a feature's value for an untrained subject is from the single feature determination boundary for that feature. The more the value of the feature for the untrained subject differs from the single-factor determination boundary for the feature based on the trained subject, the greater the weight that the feature can have in determining the responsive classification prediction. For example, a feature may have a negative value, which is a low value for a feature in which the untrained subject positively correlates with responsiveness in the trained subject, or a feature that negatively correlates with responsiveness in the trained subject. It means having a high value. If the untrained value for the feature is substantially different from the single-factor determination boundary for the feature, the weight of the feature can be high. However, another feature is high importance and positive value (ie, high value for non-trained subjects and positive correlation with responsiveness in the trained subject, or low value in the non-trained subject and responsiveness in the trained subject. This effect on the responsive predictive classification can be proportionally stronger if it has a negative correlation of), even if its value is not very different (ie, less weighted) from the single factor determination boundary.

There can be many possible ways to indicate feature identifiers as components of responsiveness prediction reports and GUIs. Some specific examples are given in a little more detail herein. However, one of ordinary skill in the art is not limited to the fact that there may be many other possible ways for reporting the value, weight, and importance of a feature in a GUI report, and the examples provided herein. You will understand that, or as they are, they are clearly essential in any way.

The GUI can show tabular features with rows and columns. Features can be represented, for example, by rows, and various properties of a given feature can be represented by multiple columns. For example, the various columns can indicate the importance of the feature, its value, the single-factor determination boundary of the feature, the value of the feature of the untrained subject, and optionally the value of the feature of the untrained subject. Visually indicates how different from the single-factor determination boundary of the feature and whether the untrained subject could be predicted to respond if the prediction was based solely on the feature. The tabular GUI can contain any combination of the two or more mentioned above. A tabular GUI report can also include an overall predicted score.

The GUI can also be a histogram. For example, a column may indicate the value of a given feature for an untrained subject, and a line may also indicate a single factor determination boundary for that feature. The difference between the value of a feature for an untrained subject and the line indicating the single-factor determination boundary is an identifier for the weight of that feature. The line can also be drawn between the height of the column and the line indicating the single factor determination boundary. The length of the line between the two is also a weight identifier. The value may be symbolized below the pillar. For example, positive or negative symbols can indicate positive or negative values, respectively. The other pair can include arrows pointing up and down, triangles pointing up and down, etc., where one direction shows the positive value of the feature and the other direction shows the negative value. .. The valence is also indicated by the color or shadow of the pillar, a pillar with a certain color or shadow pattern may show a certain valence (positive or negative), and a pillar with a different color or shadow pattern may show the opposite valence. The color or shadow of the line between the value of the untrained feature in the histogram and the single factor determination boundary of that feature can also indicate valence. For example, if the value of a feature negatively correlates with the responsiveness of the trained subject and the value of the untrained feature is less than the single-factor determination boundary, then the value of the untrained target in the bar of the histogram report for that feature. The line connecting and the single factor determination boundary can be a color or shadow showing a positive value. On the other hand, if the value of the feature positively correlates with the responsiveness of the trained subject and the value of the feature of the untrained subject is less than the single-factor determination boundary, then the value of the untrained subject in the histogram report bar of the feature The line connecting the value and the single factor determination boundary can be a color or shadow showing a negative value. In both cases, the value of the feature of the untrained subject is less than the single-factor determination boundary of the feature, but the feature correlates with the responsiveness of the trained subject as negative (positive value) or positive (negative value). Note that the value will vary depending on whether or not it has been done. The reverse is also true (ie, whether the single-factor determination boundary of a feature is less than the value of that feature in the untrained subject, the feature correlates positively or negatively with the responsiveness of the trained subject, respectively. Can have a positive or negative value, depending on whether it is prone).

The importance of a feature can be indicated by a symbol or other indicator near, within, or below the column for that feature in the histogram. For example, the size of the symbol below the bar for a feature can indicate its importance. Alternatively, the importance can be color coded by a bar or an accompanying symbol thus colored or shaded to indicate the degree of importance. When entered, it can be associated with a histogram showing a spectrum with colors or shadows, with high importance being indicated by colors or shadows that are more similar to the spectrum at one end, and less important being the other. It is indicated by a color or shadow pattern that is more similar to the edge spectrum. In some examples, for example, the symbol placed below the bar in a histogram GUI report on untrained features indicates whether the feature values were negatively or positively correlated with responsiveness in the trained subject. obtain. For example, positive and negative signs, up and down arrows, triangles pointing up and triangles pointing down, or other pairs show features that correlate positively and negatively. In such cases, the relative size of such symbols may indicate their relative importance.

In another example, the GUI report contains a shape that conveys a feature identifier. For example, each feature that includes an identifier in a GUI report can be represented by a shape in which dimensions, colors, shadows, or other aspects can represent different identifiers. For example, each feature is represented by a quadrangle, where width represents importance and height represents weight. The color of the rectangle, the pattern of shadows or contours can indicate the value. Alternatively, the feature can be represented by a triangle, where the base represents importance, the height represents weight, and vice versa. Triangles can highlight features with a positive value and vice versa. Alternatively, the color of the triangle, or the pattern of shadows, or the pattern of lines that outline them, can indicate a value.

In another example, the feature identifier can be indicated in the form of an annular sector or an annular sector. The angle of the circular sector can indicate importance, its outer radius can indicate its weight, or vice versa. The value can be indicated by the color of the circular sector, the pattern that casts a shadow, or the pattern that outlines the pattern. In other examples, the sector, or sector, can represent an identifier of the feature, the angle and radius represent importance and weight, or vice versa, and the color, shadow, etc., for example, the circular fan, described above. Represents the value of. When the feature identifier is represented by a circular sector, in some examples all such circular sectors can be drawn to have the same inner radius, which together form an inner circle with their inner arcs. Can be arranged to do so. In the inscribed circle itself, a prediction score, other overall overview, or responsiveness prediction or classification instructions can be provided. The color or shadow of this inner circle can indicate whether the responsive classification prediction predicts whether the untrained subject is likely to respond. For example, if the prediction says that the untrained subject may respond, the inner circle may have a pattern of color or shadow, or be drawn with a line of a pattern, while the prediction is untrained. The inner circle can have a different color or shadow pattern if the subject may not respond. In other examples, instead of the inner circle, another shape, such as a square, star, triangle, pentagon or other shape, may be present inside. The size of the inner shape indicates the strength of the prediction, and the larger the inner shape, the more reliable the prediction, and vice versa.

GUI reports can also provide users with the opportunity to request more information or launch additional software applications depending on the interests of the particular features reported in the GUI. For example, the GUI allows the user to hover over pointers or other elements that can be controlled by typing into the device, such as by moving the mouse over the feature identifier, or by touching the touch screen. Can be configured. Orienting the element or touching the features of the display will open a drop-down menu with more selectable options. For example, a drop-down menu may respond to an aspect of a feature that is specific to an untrained subject, such as a feature value, or a cohort range of features, a range of values that exist in the training data, or if the untrained subject responds. Percentages of untrained feature values represented, or single-factor determination boundaries of features, or single-factor determination boundaries of features, compared only to training data for trained subjects that responded in a predicted manner, or for other trained subjects. The importance of the features or their correlation with their responsiveness to other treatments, or any combination of the two or more mentioned above can be displayed. Drop-down menus can also show links to other programs accessible by one or more microprocessors to further evaluate predictive scores for features or untrained subjects, eg, different machine learning classifiers. Can be executed. By compressing such interactivity into a GUI report, a considerable amount of space and computational resources can be secured, and the interactivity with the user's computer system is greatly enhanced. For example, the display space required for simultaneous continuous display of GUI reports while accessing the drop-down menu option is reduced. In addition, time and computational resources can be reserved and, for example, interactivity can allow access to multiple computer functions without switching between display screens or between applications.

In some examples, the trained machine learning classifier can be further trained. For example, when the responsiveness of an untrained subject to treatment is confirmed, the machine learning classifier will use the feature values of the trained subject, the responsiveness of the first training in addition to the feature values, and the untrained subject that provided the prediction. Can be retrained for features including responsiveness and responsiveness obtained. In another example, the trained machine learning classifier can be retrained by the value and responsiveness of additional features of the untrained subject that could not be predicted from the trained machine learning classifier.

In some examples, when obtaining a response prediction classification for a subject, a decision may be made as to whether or not to apply a given checkpoint inhibition therapy. If the predicted score for a particular checkpoint treatment obtained from a trained machine learning classifier as disclosed herein is high, the machine learning classifier is trained to treat the subject with a treatment that predicts responsiveness. It can be a judgment. Alternatively, if the score is low, it may be determined that such treatment is not applied. Responsively likely response predictive classifications or scores obtained according to the methods, systems, or machine learning classifiers disclosed herein, treating subjects as a result of such response predictive classifications or scores. The person who receives the instruction can treat the subject by the treatment. Treat cancer in a subject by performing checkpoint inhibitory therapy in response to obtaining a response predictive classification or predictive score generated in accordance with the present disclosure, indicating that the subject can respond to such treatment. This disclosure includes doing so, or giving such treatment based on instructions to those who have obtained such response prediction classifications or prediction scores.

The following examples are intended to illustrate certain embodiments of the present disclosure, but are by no means intended to limit their scope.

FIG. 1 is a web diagram showing options for implementing the method according to aspects of the present disclosure. Sources of feature values, such as the cancer genome atlas (TCGA), or clinical trials (eg, treatment with anti-PD1 treatment, anti-CTLA4 treatment or other checkpoint inhibition treatment, or trials of other cancer treatments. ) Shows potential non-limiting examples of derived data. Some non-limiting examples of assays that can be used to obtain feature information to be used to determine feature values are also shown, for example RNAseq and whole exome sequences, such as two non-limiting examples. There is a sing (WES). Various non-limiting examples of various features are also shown, eg, those of the example to be tested, or those in which the assay can provide a suitable scale for confirming the value of such features. Examples include HLA, gene expression, ssGSEA, tumor mutation frequency, cytolytic invasion such as by tumor infiltrating lymphocytes (after deconvolution), CGA, nascent antigens, clones and / or subclonal mutations (ie, given). Examples include the presence of a mutation present in the progeny of the cell that first exhibits the mutation, and other mutations of the progeny of the cell derived from the first mutated cell that later obtained another mutation. Features are processed by a machine learning (ML) model, which trains the machine learning classifier, inputs the values of the untrained feature into the trained machine learning classifier, and then obtains the patient response. Mark untrained patients as having the potential to respond to treatment.

FIG. 2 shows some non-limiting examples of features that may be associated with predicting patient responsiveness to classifier training and treatment according to aspects of the present disclosure. As can be understood by those skilled in the art, the features set forth in FIG. 2 are non-limiting examples. These are also not all needed. Other features than those shown in FIG. 2 may be utilized, but some of which may be omitted in the practice of the method according to aspects of the present disclosure. The features in the present specification are shown as grouped according to function, cell or molecular pathway, response, and the like. Examples of such groups include antigen presentation, T or NK cell properties, immune-mediated cell lysis properties, checkpoint pathway-related substances, interferon gamma pathway-related substances, and MDSC / Treg properties. Other groups representing other functions or cell or molecular processes, etc. may be included in addition to, or instead of, as non-limiting examples of any of those presented herein.

FIG. 3 is a web diagram showing an example of how a method of training a classifier can be performed in accordance with aspects of the present disclosure. As will be appreciated by those skilled in the art, the features set forth in FIG. 3 are non-limiting examples. These are also not all needed. Other features than those shown in FIG. 3 can be utilized, but some of which may be omitted in the practice of the method according to aspects of the present disclosure. The features in the present specification are shown as grouped according to function, cell or molecular pathway, response, and the like. Examples of such groups include antigen presentation, T or NK cell properties, immune-mediated cell lysis properties, checkpoint pathway-related substances, interferon gamma pathway-related substances, and MDSC / Treg properties. In the training, as shown in FIG. 3, the characteristics derived from the training target, for example, those shown here are in a machine learning classifier such as a random forest classifier, and the trained subject is a place such as checkpoint inhibition treatment. Enter as a marker corresponding to how you responded to your treatment. Train the classifier in this way. Other machine learning classifiers can also be used as disclosed.

FIG. 4 is an extension of FIG. 3 showing an example of how a trained classifier can be used to predict a subject's responsiveness to treatment in accordance with aspects of the present disclosure. is there. A machine learning classifier (in this non-limiting example, a random forest classifier) trained on a feature (an example of which is shown herein as an example) receives additional input from an untrained subject. Receive. In particular, input the value of the feature derived from the non-trained object into the machine learning classifier. The trained machine learning classifier then generates a response predictive classification and / or feature identifier in the prediction report that may include a score indicating the likelihood of responsiveness (indicated herein as an immune score).

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) Obtained from /j.cell.2016.02.065). In this study, total exome sequencing and RNAseq data from tumor samples from 26 patients with melanoma were treated with inhibitors of the PD-1 checkpoint pathway (anti-PD-1 antibody therapy (nivolumab)). ) Or anti-PD-L1 antibody treatment (pembrolizumab)). Raw data is publicly available and has been accessed for the examples herein. Transcriptome data obtained by RNAsep (including expression levels of transcripts in samples) is available from the National Center for Biotechnology Information (NCBI) Gene Expression Omnibus. It was available online at GSE78220 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE78220). All exome sequencing data obtained by the NGS method is available from the NCBI Sequence Read Archive (https://www.ncbi.nlm.nih.gov/sra) with accession numbers SRA: SRP067938 and SRA: SRP090294. Was available online by. The responsiveness of patients for whom such data were available was also obtained from the results of published trials. Data were selected from such sources to train machine learning classifiers to create features for training subjects and responsiveness for training subjects to predict responsiveness to anti-PD1 checkpoint pathway inhibitors. .. Feature data for obtaining predictions from trained machine learning classifiers were also obtained from such sources.

Characteristics and responsiveness to CTLA-4 inhibition were obtained from Van Allen et al. (2015) Genomic correlates of response to CTLA-4 blockade in metastatic melanoma. Science 350: 207-211 (doi: 10.1126 / science.aad0095). In this study, total exome sequencing and RNAseq data from tumor samples from 30 patients with melanoma were obtained before and after treatment with a CTLA-4 inhibitor (ipilimumab). Raw data is publicly available and has been accessed for the examples herein. All exome sequencing and transcriptome data obtained by the NGS method are available online at accession number phs000452.v2.pl from the NCBI database of Genotypes and Phenotypes (dbGaP) for genotypes and phenotypes. It was (https://www.ncbi.nlm.nih.gov/gap/?term=phs000452.v2.p1). The responsiveness of patients for whom such data were available was also derived from the results of published trials. Data were selected from such sources to train machine learning classifiers to create features for training subjects and responsiveness for training subjects to predict responsiveness to anti-CTLA4 checkpoint pathway inhibitors. .. Feature data for obtaining predictions from trained machine learning classifiers were also obtained from such sources.

From the data obtained from both tests, the data was preprocessed to create feature inputs to the machine learning classifier for training. Random forest machine learning classifiers that received such inputs were then trained on the training data. Predictions were then generated from one or both trained classifiers using features from some subjects, and responsive predictive classifications, including predictive scores and feature identifiers, were reported via GUI. FIG. 1 outlines an example of some of the methods used in such an example, as described above. Figure 2 shows the 38 features selected for input to the machine learning classifier. The features were selected based on the current understanding of the role of such features and the expectation that they could be involved in predicting whether the subject would respond to checkpoint inhibition. allMut refers to the frequency of tumor mutations that all mutate, and nonSynMut refers to the frequency of tumor mutations that can result from non-synonymous mutations. cd8_dec, cd4_dec and cd19_dec refer to the tumor infiltrating lymphocyte deconvolution scale for CD8, CD4 and CD19, respectively. The remaining features defined in Figure 2 were taken from RNAseq data and included a measure of relative expression levels. In some examples, ssGSEA was also obtained for a range of features (antigen presentation, T cell / NK cell properties, immune cell lysis properties, checkpoint pathway, interferon gamma and MDSC / Treg properties). In such an example, some machine learning classifiers trained only on the original 38 features, while other machine learning classifiers added the gene set obtained by ssGSEA analysis to the original 38 features. As a result, a total of 44 features were trained.

The target data per value of a feature is normal as the ratio of each value per feature to the range of values for that feature in the set, depending on the order in which those values are requested. It became. The responsiveness for training in such cases is a binomial assessment and the subject is responsive (both partial and full responsive, or long-term treatment as reported in each basic study. Labeled as either effective (including efficacy) or non-responsive (including subjects with treatment-responsive disease progression, reported in basic trials, and reported as ineffective treatment).

Features have been entered into a computer system that includes one or more memory storage devices and one or more microprocessors. One or more memory storage devices included instructions to train a random forest machine learning classifier using the R programming language when executed by one or more microprocessors. The features of the subject were stored on one or more memory storage devices and analyzed according to instructions by one or more microprocessors. In such an example, we used 50,000 trees. After training, enter the features of interest into the trained machine learning classifier so that they are stored on one or more memory storage devices in the computer system to generate a response prediction classification score and a GUI on the computer display. Reported by. An example of training is shown in Fig. 3, and an example of predictive generation is shown in Fig. 4.

The final expected score was generated by the probability of classification and adjusted to a score of 0 to 10. As an example, for illustration purposes, the 0.75 probability of "responding" to a given immunotherapy was replaced with a predictive responsiveness score of 7.5. In addition, for each feature, the single factor determination boundary was determined by a value that maximized classification accuracy. The direction of the correlation between the classification when the feature was responsive and whether it was positively or negatively correlated was determined by Spearman's correlation between the feature and the response. The above analysis was performed using all samples without separating the trained and untrained subjects. Performance analysis was performed separately. Area under the curve (AUC) plot, R programming language packages "caret" and "cvAUC" were used with default features and parameters.

Figure 5 shows an example of a GUI500 report reporting response classification prediction scores and feature identifiers. In such an example, the feature identifier was shown only if the square of the feature's importance (the importance determined by the decrease in the feature's Gini index) was greater than 0.1. Figure 5 shows a GUI for subjects for which data were obtained for anti-PD1 trials performed on machine learning classifiers trained to predict responsiveness to anti-PD1 treatment (eg, anti-PD1 or PD1-L1 antibodies). Show the report. Fifteen features with importance above the minimum importance threshold are shown in FEATURES column 510. The importance of the features is shown in IMP. Column 520. The groups whose features are related to those in Figure 2 are shown in GROUP column 530. Positive or negative correlation with feature responsiveness is shown in CORR column 540. In this example, triangles are used to indicate the direction of correlation, upward-pointing triangles show a positive correlation between feature values and responsiveness, and downward-pointing triangles show feature values. Shows a negative correlation between and responsiveness. The single-factor determination boundaries of the features are shown in column 1FDB column 550. In this example, the single factor determination boundary provided the highest difference obtained between responders and non-responders, above and below the subject for that sample (ie, responsive and non-responsive). In the distinction between objects, the accuracy of the ratio is generally low, and it is numerically determined as the ratio in the range of upper or lower values). Single-factor determination boundaries are also indicated by shadowing from left to right within the cells of each feature in the 1FDB column, representing the proportions indicated by the single-factor determination boundaries (ie, proportions in proportion to the value of the boundary). If is low, the shadow is light, and if the ratio is high, the shadow is dark). The values of interest for each feature are shown in INPUT column 560 (for this PT5 INPUT, the patient reporting the prediction here is identified as patient number 5 or PT5). The number in column 560 of PT5 INPUT indicates the value of the object for a given feature, and the shadow indicates the order of the ratio of the value of the feature of that object to the range of values for the training object.

Report in column 1F.PRED column 570 whether a subject could be predicted to respond to treatment based solely on the value of a given feature. Thus, for positively correlated features, 1F.PRED570 indicates YES if the value of PT5 Input560 exceeds the value of 1FDB550 (the feature can be predicted that the subject may respond to treatment). Means). For positively correlated features, 1F.PRED570 indicates NO if the value of PT5 Input560 is less than the value of 1FDB550 (the feature can be predicted that the subject may not respond to treatment). means). For negatively correlated features, 1F.PRED570 indicates NO if the value of PT5 Input560 exceeds the value of 1FDB550 (meaning that the feature can predict that the subject may not respond to treatment). To do). Also, for negatively correlated features, 1F.PRED570 indicates YES if the value of PT5 Input560 is less than the value of 1FDB550 (features can be predicted that the subject may respond to treatment). Means that). In this example, the cells of 1F.PRED570 can also be color coded depending on whether only the features can predict responsiveness. The YES cell can be colored green, but the NO cell can be colored red (R), for example (R). The last line, FULL MODEL 580, reports a response classification score of 5.5 in this case. The cutoff can be determined by a value above or below and can predict whether the treatment is effective or not. For example, a score below 5.0 can be considered to predict that treatment may not affect this patient, and a score above 5.0 may be considered to indicate that treatment may affect this patient. it can. Thus, if a score greater than 5 is considered to indicate that treatment may affect this patient, the usefulness of training and using the machine learning classifiers disclosed herein is an indicator in the GUI report. The basis for prediction is significantly improved over what can be basically predicted based on only one feature shown, but is non-responsive with only some features (including the most important features). However, it can be understood that the machine learning classifier brings about it in terms of overall predicting that the patient can respond.

In this example, the response prediction score is reported by the GUI as an indicator of each feature contained in the GUI report, weight (by presenting a comparison between PT5 INPUT560 and 1FDB550), importance 520, and 1F. Includes PRED value. In some examples, the user touches, for example, a part of a touch screen display that corresponds to a part of a GUI report, or a graphic element such as a cursor is featured or associated with a device such as a mouse. Additional information that can be moved over the score to access additional information or to be executed by various programming instructions stored on one or more memory storage devices for one or more microprocessors. You may have the option to access the drop-down menu from various aspects of the GUI by choosing from the analysis of.

Another GUI report 610 for this subject is shown in Figure 6. Combine multiple GUI reports into a single report in the example shown in Figure 6. The upper part of FIG. 6 shows the annular fan-shaped 610 ring, which corresponds to each feature whose importance exceeds the minimum importance threshold, as described in FIG. The corresponding features of each annular sector are also indicated by description. For example, the annular fan shape corresponding to feature HLA.B is indicated by 630. The annular sector has an angle, an outer radius, an inner radius, and an inner arc, respectively. In this example, the angle corresponds to the importance of the feature. In this example, the angles of the features are proportional to each other, allowing direct visual comparison. Also, in this example, the difference between the outer and inner radii corresponds to its weight (ie, the difference between the value of the feature of interest and the single factor determination boundary). The feature valence in this example is also reported by the type of line that outlines the feature circular sector. Features in which the circular sector has a positive valence for the subject (eg, for total tumor mutation frequency 640) are outlined with a solid line, while features in which the circular sector has a negative valence for the subject (eg, for HLA.B630) are dotted. Draw a contour with.

In this example, the inner arcs are arranged to form an inner circle. Also, in this example, in the inner circle, the responsive predictive classification score for this patient is reported to be 5.5 in this case. Also, in this example, the overall prediction is indicated by a solid line forming an inner circle, meaning that the response prediction classification score of the subject exceeds a predetermined level that distinguishes between responsive and non-responsive predictions. .. In another example, the dotted line may form an inscribed circle if the response prediction classification score is below such a predetermined score threshold. In another example, the difference in color or shadow pattern in the circular sector and / or the inner circle, and / or the difference in the coloring of the contour of the circular sector or the inner circle may indicate a value.

In this example, the top 610 of the GUI report shown in FIG. 6 of 600 includes a report of response prediction classification scores, showing the importance, value, and weight of features. Another example of such a GUI report 1010 is shown in Figure 10. The annular fan shape for features is shown individually, not with its inner arc forming a circle. Examples of outer arc 1002, inner arc 1001, and difference 1003 between outer and inner radii are shown for the annular sector of CD15, and the angle of the annular sector 1004 is below the annular sector of HLA.B 1030 for illustration. Shown. The outer radius of the feature (or the difference between the inner radius and the outer radius) reports the weight, the angle reports the importance, and the pattern of lines that outline the circular sector is positive (solid line, eg, solid line). , Total tumor mutation frequency all_tmb 1040 circular sector) or negative (dotted line, eg circular sector for CD15) valence. Response prediction classification scores are not shown here, but are also optionally included in such GUI report 1010. Patterns of different colors or shadows, rather than contours of different patterns, can be used to indicate the value of the feature. Shapes other than the annular fan shape can also be used. For example, features are reported as rectangles and importance and weight are represented by width and height, or, for example, as triangles, the base width reflects importance and height is. It represents a weight, and the directionality can represent a value. One of ordinary skill in the art will appreciate that a number of possible methods are applicable for reporting multiple indicators of multiple features of the reporting GUI according to aspects of the present disclosure.

Returning to the GUI report 600 shown in FIG. 6, below the top of the report containing the circular sector 610 is the histogram portion of the GUI report 620. GUI reports may or may not have both or one of these parts. Histogram 620 shows the pillars of each feature whose values reported for the subject exceed the minimum importance threshold set for that feature. The scale 650 on the left shows the order of the percentage of the value of interest for each feature. The pillar of each feature represents the value of the target feature for that feature as a percentage of the range of values for the value of the training target. As an example of the target value, feature CD15 660 is shown. Also, for each column, a horizontal line is shown in this case as a single feature determination boundary for that feature. As an example of a single feature determination boundary, feature CD15 670 is shown. The triangle below each pillar indicates whether the feature correlates with responsiveness positively (upward pointing triangle) or negative (downward pointing triangle). As an example, it refers to 680 of CD15. The triangles are also assigned a large triangle representing high importance and a small triangle representing low importance to reflect the relative importance of each feature. The line between the value of the feature of interest and the single-factor determination boundary for that feature indicates the weight of that feature. As an example of a weight report for a feature, feature CD15 690 is shown. The value for the feature is indicated by whether the weight line is a solid line (positive value) or a dotted line (negative value). As will be appreciated by those skilled in the art, each such specific example for reporting different instructions for different features may be omitted or replaced with different graphic representations. Colors and shadows represent valences and / or correlations for features, arrows or other directional shapes represent valences or correlations, and importance can be represented by a graduated color-coded design or the like.

In some examples, gene sets were used to train machine learning classifiers and generate predictions from the trained machine learning classifiers. Figure 7 shows an example of the gene set produced by ssGSEA. Six sets are shown grouped according to the characteristics of the individual used to determine the gene set. Examples include those relating to antigen processing pathways, namely antigen presentation (710), T cell and NK cell properties 720, cell lysis properties 730, checkpoint pathway 740, interferon gamma 750, and MSDC / Treg properties 760. Correlation and importance for each feature, including cell set, is shown in 770 and 780 when used to train PD1 and CTLA4 machine learning classifiers, respectively. In some examples, a dashed box outlines a gene set that has either been highly correlated or of greater importance than any individual feature when the gene set value was determined using ssGSEA. Is emphasized in the cell in which is drawn, and the value including the gene set is shown as a feature. The usefulness when ssGSEA is included is also shown in FIG. Figure 9 shows two GUI reports on responsiveness prediction for the same subject using two different trained machine learning classifiers. The prediction 910 on the left was obtained by generating a prediction without using the ssGSEA gene set as a feature in training or prediction. Prediction 920 on the right was obtained by generating a prediction using the ssGSEA gene set as a feature in training or prediction. When including the gene set by ssGSEA, it includes some gene sets (by definition, those not included in the prediction 910 obtained without the gene set), and the minimum threshold boundary is set without giving up the overall prediction. Fewer features were exceeded (11:15) (eg, in both cases where the prediction exceeded the score 5.0 set as the minimum response prediction classification score, which initiates the classification as the respondent of the subject).

Figure 8 shows two GUI reports obtained from the same patient who generated predictions using two different machine learning classifiers, one 810 trained to predict responsiveness to anti-CTLA4 treatment. The other 820 was trained to predict responsiveness to anti-PD1 and both used the features shown in Figure 2. The anti-CTLA4 machine learning classifier 810 generated a response predictive classification score of 3.8 and predicted that the subject might not respond to anti-CTLA4 treatment (using the response predictive classification score threshold of 5.0). .. In the case of this 810, the value (non-responsiveness) of the response prediction classification score is indicated by a dotted line around the inner circle 815 indicating the response prediction classification score. The accuracy of such predictions is due to clinical trial sources (Van Allen et al.) Showing that the patient was categorized as "disease progression" and that the subject was non-responsive to anti-CTLA4 treatment. Demonstrated. However, the anti-PD1 machine learning classifier 820 produces a response-predictive classification score of 6.7, predicting that a subject may respond to anti-PD1 (eg, anti-PD1 or anti-PD-L1 antibody) treatment. (Again using the response prediction classification score threshold of 5.0). In the case of this 820, the response prediction classification score (responsiveness) is indicated by a solid line around the inner circle 817 indicating the response prediction classification score.

The various response prediction classification scores (and the corresponding differences in valence, weight, and importance of various features) depending on the machine learning classifier used reflect the capabilities of the methods disclosed herein. For example, index 830, which reports HLA.A, indicates that it has a negative value for predicting the patient's responsiveness to anti-CTLA4 treatment, but has a positive value for predicting the subject's responsiveness to anti-PD1 treatment. Furthermore, the feature values for non-synonymous tumor mutation frequency and total tumor mutation frequency did not exceed the minimum importance threshold in the anti-CTLA4 machine learning classifier, whereas in the anti-PD1 machine learning classifier 840, the minimum importance threshold. Beyond.

Similar to the example shown in FIG. 5, the exemplary GUI reports shown in FIGS. 6, 8, 9 and 10 may include user interactivity. Thus, in some examples, the user touches, for example, a part of the touch screen display that corresponds to a part of the GUI report, or features a graphic element such as a cursor in a device such as a mouse (eg, for example). Various programming stored on one or more memory storage devices for access to additional information or for one or more microprocessors by moving over a circular fan or column or other indicator of the histogram. You may have the option to access the drop-down menu from various aspects of the GUI by choosing from additional parsings that can be performed by instructions.

Figures 11A to 11D demonstrate that the use of all 38 features shown in Figure 2 results in classifier performance over that of using a single factor. This was true for machine learning classifiers, regardless of whether they were trained to predict responsiveness to anti-PD1 or anti-CTLA4 treatment. Figure 11A trains and tests machine learning classifiers and overtrains machine learning classifiers without cross-validation (CV), for example, using all 38 features to predict responsiveness to anti-PD1 treatment. In this case, an AUC of 1.00 false positive vs. true positive receiver operating characteristic (auROC, Receiver Operator Characteristic) was obtained, reaching 1.00 auROC vs. 0.64 auROC for a single factor mean, as shown in Figure 11B. Is shown. Figures 11C and 11D further show that using the 38 features shown in Figure 2, they are more accurate than all three of the top three single-factor features, HLA-B, nonSyn_tmb and all_tmb.

12A-12D are graphs showing the effectiveness of using the gene set obtained by ssGSEA with classifier performance. The performance of ssGSEA is comparable when using the 38 features shown in Figure 2. 12A, 12B and 12C use the 38 features shown in Figure 2 to predict responsiveness to anti-PD1 (Figure 12A) or anti-CTLA4 (Figure 12B) treatment, or in addition to the 38 features. We report a 3-fold cross-validation au ROC for predicting responsiveness to anti-PD1 therapy (Figure 12C) using the six gene sets shown in Figure 7. When including the 6 gene sets obtained by ssGSEA, performance was reduced to 0.69 to 0.64 for anti-PD1 treatment prediction (comparing Figures 12A to 12C), but with importance above the minimum importance threshold. The number of features decreased from 15 to 11 (see, for example, Figure 8 as discussed above). Figure 12D shows the results of comparable t-tests for false-positive and true-positive results from the machine learning classifier's response prediction classification score (without cross-validation). Thus, in some cases, the gene set included as a feature may improve the overall stability of the classifier, help avoid overtraining, and thereby allow a clearer interpretation of key features and indicators. ..

The data used in such examples are from subjects enrolled in a study that tests the efficacy of the checkpoint inhibitors anti-CTLA4 antibody, anti-PD1 antibody, and anti-PD1-L1 antibody in patients with melanoma. there were. However, as will be appreciated by those skilled in the art, a role based on checkpoint inhibition is known to act in slowing the efficacy of cancer immunotherapy for other cancers, checkpoint pathway function, herein. The role of features in the methods, systems, and classifiers disclosed herein, such as those included herein, is to generate predictions of a subject's responsiveness to such checkpoint inhibitors in breast cancer, gastrointestinal tract. Equally useful and effective in other cancers, including systemic cancer, liver cancer, bladder cancer, lymphoma, leukemia, bone tissue cancer, nervous system cancer, lung cancer, pancreatic cancer, or others Can be. Moreover, as will also be appreciated by those skilled in the art, responsiveness to checkpoint inhibitors is also a method disclosed herein, in addition to those specifically used in the aforementioned examples disclosed herein. It can be predicted using systems and classifiers, and these are treated merely as non-limiting examples of their adaptability.

The pitfall of this performance analysis is the small sample size. Overtraining is inevitable, even with many features. On the other hand, the results of cross-validation are very unstable and show the discretization of auROC points rather than a continuous curve. However, in the limitations of the present inventors, the analysis by all models clearly outperforms the analysis by a single factor.

Although preferred embodiments are presented and described in detail herein, various modifications, additions, substitutions, etc. can be made without departing from the spirit of the present disclosure, and thus the following claims. As defined in, it will be apparent to those skilled in the art concerned that these are considered to be within the scope of the present disclosure.

All combinations of the aforementioned and additional concepts discussed in more detail herein (unless such concepts contradict each other) are part of the subject matter of the invention disclosed herein. It should be understood that there is. In particular, all combinations of claimed subject matter appearing at the end of this disclosure are considered to be part of the subject matter of the present invention disclosed herein.

GSE 78220
SRA SRP067938
SRA SRP090294
dbGaP phs000452.v2.pl

Claims (29)

In the process of inputting genomic information of a non-trained subject into a trained machine learning classifier, the genomic information of the non-trained subject contains features derived from a tumor profile obtained from the non-trained subject.
A trained machine learning classifier is trained on the genomic information of multiple training subjects and the responsiveness of each of the training subjects to treatments including checkpoint inhibition, and the genomic information of the multiple training subjects is the multiple training subjects. Machine learning classifiers are trained to predict responsiveness to treatment, including features of the tumor profile obtained from each of the processes and
In the process of generating a checkpoint inhibition responsiveness classification for a non-trained subject using a trained machine learning classifier, the checkpoint inhibition responsiveness classification is that the untrained subject responds to checkpoint inhibition. Predict the process and
A computer execution method that includes the process of reporting an untrained checkpoint inhibition responsiveness classification using a graphical user interface. At least some of the features derived from the tumor profile obtained from the untrained subject, or at least some of the features derived from the tumor profile obtained from one or more of the trained subjects, consist of all mutations. Total mutation frequency consisting of, beta 2 microglobulin (B2M) expression, proteasome subunit beta 10 (PSMB10) expression, antigen peptide transmitter 1 (TAP1) expression, antigen peptide transporter 2 (TAP2) expression, human leukocyte antigen A (HLA) -A) expression, major histocompatibility complex class IB (HLA-B) expression, major histocompatibility complex class IC (HLA-C) expression, major histocompatibility complex class II DQ alpha 1 (HLA-DQA1) expression, HLA class II histocompatibility antigen DRB1 beta chain (HLA-DRB1) expression, HLA class I histocompatibility antigen alpha chain E (HLA-E) expression, natural killer cell granule protein 7 (NKG7) expression, chemokine-like receptor (CMKLR1) Expression, tumor invasion by cells expressing surface antigen classification 8 (CD8), tumor invasion by cells expressing surface antigen classification 4 (CD4), tumor invasion by cells expressing surface antigen classification 19 (CD19), granzyme A ( GZMA) expression, perforin 1 (PRF1) expression, cytotoxic T lymphocyte-related 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 and ITIM domains, surface antigen classification 276 (CD276) expression, chemokine (CC motif) ligand 5 (CCL5) ), CD27 expression, chemokine (CXC motif) ligand 9 (CXCL9) expression, CXC motif chemokine receptor 6 (CXCR6), indolamine 2,3-dioxygenase (IDO) expression, signaling and transcriptional activator 1 (STAT1) ) Expression, 3-fucosyl-N-acetyl-lactosamine (CD15) expression, interleukocyte 2 receptor alpha chain (CD25) expression, siglec-3 (CD33), surface antigen classification 39 (CD39) expression, surface antigen classification 118 ( The first aspect of the present invention, which is selected from the group consisting of CD118) expression, forkhead box P3 (FOXP3) expression, and any combination of two or more described above. How to put. The method of claim 1, wherein at least a portion of the trained features, or at least a portion of the non-trained features, comprises a gene set. The method of claim 3, wherein the gene set was selected using enrichment analysis of a single sample gene set. The method of claim 1, wherein the machine learning classifier is a random forest. The method of claim 5, wherein at least 50,000 trees are used in the training of machine learning classifiers. Claim that the checkpoint inhibition responsiveness classification comprises a predicted score and one or more feature identifiers, the one or more feature identifiers being selected from the group consisting of feature value, feature importance, and feature weight. The method described in 1. The graphical user interface reports the feature identifier as each side of the circular sector, the angle of the circular sector reports the importance of the feature, the outer radius of the circular sector reports the feature weight, and the color of the circular sector reports the feature value. The method according to claim 7. The method of claim 8, wherein the feature importance of the feature comprises a decrease in the Gini index of the feature. The method of claim 9, wherein the graphical user interface reports the feature identifier only if the feature importance of the feature exceeds a threshold. The method according to claim 10, wherein the feature importance of the feature does not exceed the threshold value when the square of the feature importance of the feature does not exceed 0.1. The method of claim 10, wherein each circular sector includes an inner arc and the inner arcs of the circular fan are arranged to form a circle. The step of inputting the responsiveness of the non-trained object to the treatment into the trained machine learning classifier and the step of further training the machine learning classifier are further included, and the step of further training is the trained machine learning classifier. The method of claim 1, comprising training the characteristics of the tumor sample obtained from the non-trained subject and the responsiveness of the non-trained subject to treatment. The method of claim 1, further comprising selecting a treatment based on the generated checkpoint inhibitory responsiveness classification. With one or more microprocessors,
A trained machine learning classifier and one or more memories for storing non-trained genomic information, and the trained machine learning classifier can perform multiple trained genomic information and checkpoint inhibition. Trained for the responsiveness of each of the multiple training subjects to the including treatment, the genomic information of the multiple training subjects contains the features of the tumor profile obtained from each of the multiple training subjects, and the machine learning classifier responds to the treatment. A computer system in which the genomic information of a non-trained subject trained to predict sex contains memory and features derived from a tumor profile obtained from the non-trained subject.
When one or more memories are run by one or more microprocessors, the computer system is forced to generate a checkpoint inhibition responsive classification for the untrained subject using a trained machine learning classifier, and is non-trained. The checkpoint inhibition responsiveness classification of the trained subject is reported using a graphical user interface, the instruction is memorized, and the checkpoint inhibition responsiveness classification predicts that the non-trained subject responds to the checkpoint inhibition. Computer system. At least some of the features derived from the tumor profile obtained from the untrained subject, or at least some of the features derived from the tumor profile obtained from one or more of the trained subjects, consist of all mutations. Total mutation frequency consisting of, beta 2 microglobulin (B2M) expression, proteasome subunit beta 10 (PSMB10) expression, antigen peptide transmitter 1 (TAP1) expression, antigen peptide transporter 2 (TAP2) expression, human leukocyte antigen A (HLA) -A) expression, major histocompatibility complex class IB (HLA-B) expression, major histocompatibility complex class IC (HLA-C) expression, major histocompatibility complex class II DQ alpha 1 (HLA-DQA1) expression, HLA class II histocompatibility antigen DRB1 beta chain (HLA-DRB1) expression, HLA class I histocompatibility antigen alpha chain E (HLA-E) expression, natural killer cell granule protein 7 (NKG7) expression, chemokine-like receptor (CMKLR1) Expression, tumor invasion by cells expressing surface antigen classification 8 (CD8), tumor invasion by cells expressing surface antigen classification 4 (CD4), tumor invasion by cells expressing surface antigen classification 19 (CD19), granzyme A ( GZMA) expression, perforin 1 (PRF1) expression, cytotoxic T lymphocyte-related 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 and ITIM domains, surface antigen classification 276 (CD276) expression, chemokine (CC motif) ligand 5 (CCL5) ), CD27 expression, chemokine (CXC motif) ligand 9 (CXCL9) expression, CXC motif chemokine receptor 6 (CXCR6), indolamine 2,3-dioxygenase (IDO) expression, signaling and transcriptional activator 1 (STAT1) ) Expression, 3-fucosyl-N-acetyl-lactosamine (CD15) expression, interleukocyte 2 receptor alpha chain (CD25) expression, siglec-3 (CD33), surface antigen classification 39 (CD39) expression, surface antigen classification 118 ( Claim 1 selected from the group consisting of CD118) expression, forkhead box P3 (FOXP3) expression, and any combination of two or more described above. The system described in 5. The system of claim 15, wherein at least some of the trained features, or at least some of the non-trained features, contain a gene set. 17. The system of claim 17, wherein the gene set was selected using a single sample gene set enrichment analysis. The system of claim 15, wherein the machine learning classifier is a random forest. The system of claim 19, wherein at least 50,000 trees are used in the training of machine learning classifiers. Claim that the checkpoint inhibition responsiveness classification comprises a predicted score and one or more feature identifiers, the one or more feature identifiers being selected from the group consisting of feature value, feature importance, and feature weight. The system described in 15. When the instruction is executed by one or more microprocessors, have the graphical user interface report the feature identifier as each side of the circular sector, the angle of the circular sector reports the feature importance, and the outer radius of the circular sector. 21. The method of claim 21, wherein the feature weights are reported and the circular sector color reports the feature values. 22. The system of claim 22, wherein the feature importance of the feature comprises a decrease in the Gini index of the feature. 23. Claim 23, wherein the instruction is executed by one or more microprocessors and causes the graphical user interface to report the feature identifier if and only if the feature feature importance exceeds the threshold. system. 24. The system of claim 24, wherein the feature importance of the feature does not exceed a threshold if the square of the feature importance of the feature does not exceed 0.1. 24. The system of claim 24, wherein when the instruction is executed by one or more microprocessors, the graphical user interface is made to report an inner arc of each circular sector and a circle containing the inner arc of the circular sector. When the instruction is executed by one or more microprocessors, the computer system may be further trained and further trained to train the machine learning classifier, a tumor sample obtained from an untrained subject. 15. The system of claim 15, comprising training the characteristics of the non-trained subject to treatment and responsiveness to the treatment. A machine learning-based classifier for the classification of immune checkpoint responsiveness
A machine learning-based classifier that runs on a number of processors and is trained to predict the responsiveness of untrained subjects to immune checkpoint inhibition therapy.
A machine learning-based classifier is trained by inputting the genomic information of multiple training targets and the responsiveness of each of the multiple training targets to treatment into the machine learning-based classifier, and the genomes of multiple training targets. With a machine learning-based classifier, the information contains features of the tumor profile obtained from each of the multiple training subjects.
A machine learning-based classifier that is an input processor that inputs untrained genomic information into a machine-learning-based classifier, in which the non-trained genomic information contains features derived from tumor profiles obtained from the non-trained subject. But,
With an input processor, which is configured to generate a checkpoint inhibitory responsiveness classification for untrained subjects, the checkpoint inhibitory responsiveness classification predicts that the subject will respond to checkpoint inhibitory therapy.
A machine learning-based classifier that includes an output processor that reports checkpoint inhibition responsiveness classification. The machine learning based classifier according to claim 28, wherein the checkpoint inhibition responsiveness classification comprises a predicted score and a plurality of identifiers.