Identification method for newly generated antigens for T-cell therapy

Abstract

This invention provides a method for identifying antigen-specific T cells for at least one type of nascent antigen that is likely to be presented on the surface of target tumor cells. [Solution] The peptide sequences of neoplastic antigens are obtained by sequencing the target tumor cells. These peptide sequences are input into a machine learning presentation model to generate presentation likelihoods for the neoplastic antigens, each representing the likelihood that the neoplastic antigen will be presented by an MHC allele on the surface of the target tumor cells. A subset of neoplastic antigens is selected based on the presentation likelihoods. Antigen-specific T cells are identified for at least one of the neoplastic antigens in the subset. These T cells can be proliferated for use in T cell therapy. The TCRs of these identified T cells can also be sequenced and cloned into new T cells for use in T cell therapy.

Description

Background Art

[0001] Cancer vaccines and T cell therapies based on tumor-specific neoantigens are highly promising as next-generation personalized cancer immunotherapies. 1~3 Cancers with high genetic mutation loads, such as non-small cell lung cancer (NSCLC) and melanoma, are particularly promising targets for such therapies because they are relatively likely to generate neoantigens. 4,5 Initial evidence indicates that vaccination based on neoantigens induces T cell responses 6 and that T cell therapies targeting neoantigens can cause tumor regression in selected patients. 7 Both MHC class I and MHC class II influence T cell responses. 70~71

[0002] However, the identification of neoantigens and neoantigen-recognizing T cells is a central challenge in assessing tumor responses, 77,110 examining tumor evolution, 111 and designing next-generation personalized therapies. Current neoantigen identification methods are either time-consuming and laborious 112 or not sufficiently accurate. Neoantigen-recognizing T cells are a major component of TILs 84,96 and have recently been shown to circulate in the peripheral blood of cancer patients, but current methods for identifying neoantigen-reactive T cells have some combination of the following limitations: (1) they rely on clinically difficult-to-obtain specimens such as TILs 87,91-93 or leukapheresis, (2) they require screening of impractically large peptide libraries, or (3) they rely on MHC alleles that are only practically available for a small number of MHC alleles. 84,96,113,114 107 97,98 107

[0003] Furthermore, initial methods incorporating mutation-based analysis using next-generation sequencing, RNA gene expression, and prediction of MHC binding affinity of neoantigen peptides have been proposed.8 However, these proposed methods involve many steps other than gene expression and MHC binding (e.g., TAP transport, proteasome cleavage, MHC binding, transport of peptide-MHC complexes to the cell surface, and / or recognition of MHC-I by the TCR; endocytosis or autophagy, cleavage by extracellular or lysosomal proteases (e.g., cathepsin), competition with CLIP peptides for HLA binding catalyzed by HLA-DM, transport of peptide-MHC complexes to the cell surface, and / or recognition of MHC-II by the TCR). 9 It is not possible to model the entire epitope generation process. Therefore, existing methods tend to have the problem of low positive predictive value (PPV) (Figure 1A).

[0004] In fact, analyses of peptides presented by tumor cells conducted by multiple groups have shown that less than 5% of the peptides predicted to be presented based on gene expression and MHC binding affinity are found on the MHC on the tumor surface. 10,11 (Figure 1B). This low correlation between binding prediction and MHC presentation is further indicated by the lack of improvement in the predictive accuracy of binding-limited nascent antigens for checkpoint inhibitor responses for the number of mutations alone. 12 .

[0005] Such low positive predictive values ​​(PPVs) of existing methods for predicting presentation present problems in the design of neogenic antigen-based vaccines and neogenic antigen-based T-cell therapies. When vaccines are designed using low PPV predictions, it is unlikely that the majority of patients will receive therapeutic neogenic antigens, and even fewer will receive multiple neogenic antigens (even assuming that all presented peptides are immunogenic). Similarly, when therapeutic T cells are designed based on low PPV predictions, it is unlikely that the majority of patients will receive T cells that are responsive to tumor neogenic antigens, and the time and physical resource costs of identifying predicted neogenic antigens using downstream testing methods after prediction may be unnecessarily high. Therefore, neogenic antigen vaccination and T-cell therapy using current methods are unlikely to be effective in a significant number of subjects with tumors (Figure 1C).

[0006] Furthermore, previous approaches have only used cis-acting mutations to generate candidate neogenic antigens, which occur in multiple tumor types and lead to abnormal splicing in many genes due to mutations in splicing factors. 13 In most cases, further sources of nascent ORFs, including mutations that create or remove protease cleavage sites, were not considered.

[0007] Finally, standard approaches to tumor genome and transcriptome analysis may miss somatic mutations that give rise to candidate nascent antigens due to suboptimal conditions in library construction, exome and transcriptome capture, sequencing, or data analysis. Similarly, standard tumor analysis approaches may falsely promote sequence artifacts or germline polymorphisms as nascent antigens, potentially leading to inefficient vaccine capacity utilization or an autoimmune risk, respectively. [Overview of the project]

[0008] This specification discloses optimized approaches for identifying and selecting nascent antigens for personalized cancer vaccines, T-cell therapies, or both. Firstly, we address optimized tumor exome and transcriptome analysis approaches for identifying nascent antigen candidates using next-generation antigens (NGS). These methods build upon standard approaches to tumor analysis with NGS so that the most sensitive and specific nascent antigen candidates are developed across all classes of genomic alterations. Secondly, novel approaches are provided for high-PPV nascent antigen selection to overcome specificity issues and make nascent antigens developed for vaccine addition and / or as targets for T-cell therapy more likely to induce anti-tumor immunity. Depending on the embodiment, these approaches include trained statistical regression or nonlinear deep learning models that jointly model peptide-allele mapping, as well as allele-specific motifs for multiple peptide lengths that share statistical efficacy across peptides of different lengths. In particular, nonlinear deep learning models can be designed and trained to treat different MHC alleles within the same cell as independent, thus resolving the problems associated with linear models where linear models interfere with each other. Finally, further concerns regarding the design and manufacture of personalized vaccines based on newly synthesized antigens, and in the manufacture of personalized newly synthesized antigen-specific T cells for T cell therapy, will be resolved.

[0009] The models disclosed herein outperform, by up to an order of magnitude, state-of-the-art predictive tools trained on binding affinity and earlier predictive tools based on MS peptide data. By predicting peptide presentation with greater reliability, the models enable a more time-efficient and cost-effective identification of nascent antigen-specific or tumor antigen-specific T cells for personalized therapy using a clinically applicable process that utilizes a limited amount of patient peripheral blood, screens fewer peptides per patient, and does not necessarily rely on MHC multimers. However, in another embodiment, the models disclosed herein enable a more time-efficient and cost-effective identification of tumor antigen-specific T cells using MHC multimers by reducing the number of MHC multimer-bound peptides that need to be screened to identify nascent antigen-specific or tumor antigen-specific T cells.

[0010] The predictive performance of the models disclosed herein in the TIL neonatal epitope dataset and the task of identifying expected neonatal antigen-reactive T cells demonstrates that it is now possible to obtain predictions of therapeutically useful neonatal epitopes by modeling HLA processing and presentation. In summary, this work accelerates patients' progress toward healing by enabling actionable in silico antigen identification for antigen-targeted immunotherapy. [Invention 1001] A method for identifying one or more antigen-specific T cells for at least one type of neogenic antigen that is likely to be presented on the surface of one or more target tumor cells, A step of obtaining at least one of exome, transcriptome, or whole genome nucleotide sequencing data from the target tumor cells and normal cells, wherein the nucleotide sequencing data is used to obtain data representing the respective peptide sequences of a set of nascent antigens identified by comparing the nucleotide sequencing data from the tumor cells with the nucleotide sequencing data from the normal cells, and the peptide sequence of each nascent antigen includes at least one modification that causes the peptide sequence to differ from the corresponding wild-type peptide sequence identified from the target normal cells; A step of encoding each of the peptide sequences of the nascent antigen into a corresponding numerical vector, wherein each numerical vector includes information about a plurality of amino acids constituting the peptide sequence and a set of positions of the amino acids in the peptide sequence; A step of inputting the numerical vectors into a machine learning presentation model using a computer processor in order to generate a set of presentation likelihoods for the set of neonatal antigens, wherein each presentation likelihood in the set represents the likelihood that the corresponding neonatal antigen is presented on the surface of the target tumor cell by one or more MHC alleles, and the machine learning presentation model For each of the multiple samples, a label obtained by mass spectrometry measuring the presence of a peptide bound to at least one MHC allele within the set of MHC alleles identified as being present in the sample, and For each of the aforementioned samples, a training peptide sequence encoded as a numerical vector containing information about the multiple amino acids constituting the peptide and the set of positions of the amino acids in the peptide. Several parameters identified at least based on the training dataset, A function representing the relationship between the numerical vector received as input and the suggested likelihood generated as an output based on the numerical vector and the parameters. The process including the aforementioned steps; A step of selecting a subset of the set of nascent antigens based on the set of presentation likelihoods in order to generate a selected set of nascent antigens; A step of identifying one or more T cells that are antigen-specific to at least one of the nascent antigens in the subset; and The step of returning one or more identified T cells. The method, including the method described above. [Invention 1002] The process of inputting the numerical vector into the machine learning presentation model is as follows: For each of the one or more MHC alleles, the machine learning presentation model is applied to the peptide sequence of the nascent antigen in order to generate a dependency score indicating whether the MHC allele presents the nascent antigen based on a specific amino acid at a specific position in the peptide sequence. The method of the present invention 1001, including the method of the present invention 1001. [Invention 1003] The process of inputting the numerical vector into the machine learning presentation model is as follows: For each MHC allele, the dependency score is transformed to generate a corresponding allele-specific likelihood that indicates the likelihood that the corresponding MHC allele will present the corresponding nascent antigen, and To generate the presentation likelihood of the nascent antigen, the allele-specific likelihoods are combined. The method of the present invention 1002, further comprising the above. [Invention 1004] The method of the present invention 1003, wherein converting the dependency score models the presentation of the nascent antigen as mutually exclusive across one or more MHC alleles. [Invention 1005] The process of inputting the numerical vector into the machine learning presentation model is as follows: To generate the presentation likelihood, the transformation of the combination of dependency scores, wherein the transformation of the combination of dependency scores models the presentation of the nascent antigen as interference between one or more MHC alleles. The method of the present invention 1002, further comprising the above. [Invention 1006] The set of presented likelihoods is further identified by at least one allele-non-interaction property, Based on the allele-non-interaction characteristics, the machine learning presentation model is applied to the allele-non-interaction characteristics to generate a dependency score for the allele-non-interaction characteristics, indicating whether or not the peptide sequence of the corresponding nascent antigen is presented. Any method of the present invention 1002 to 1005, further comprising the above. [Invention 1007] The dependency score for each MHC allele in the one or more MHC alleles is combined with the dependency score for the allele non-interaction properties. To generate an allele-specific likelihood for each MHC allele, which indicates the likelihood that the corresponding MHC allele presents the corresponding nascent antigen, the combined dependency score for each MHC allele is transformed, and To generate the aforementioned presentation likelihood, the allele-specific likelihoods are combined. The method of the present invention 1006, further comprising the above. [Invention 1008] Combining the dependency score for each of the MHC alleles with the dependency score for the allele non-interaction properties, and To generate the aforementioned presentation likelihood, the combined dependency scores are transformed. The method of the present invention 1006, further comprising the above. [Invention 1009] Any method 1001 to 1008 of the present invention, wherein the one or more MHC alleles comprises two or more different MHC alleles. [Invention 1010] The method according to any one of the present invention 1001 to 1009, wherein the peptide sequence includes a peptide sequence having a length other than nine amino acids. [Invention 1011] The method according to any one of the present invention 1001 to 1010, wherein the step of encoding the peptide sequence includes encoding the peptide sequence using a one-hot encoding scheme. [Invention 1012] The aforementioned multiple samples (a) One or more cell lines engineered to express a single MHC allele, (b) One or more cell lines engineered to express multiple MHC alleles, (c) One or more human cell lines obtained from or derived from multiple patients, (d) Fresh or frozen tumor samples obtained from multiple patients, and (e) Fresh or frozen tissue samples obtained from multiple patients A method of the present invention, any one of the methods described in 1001 to 1011, comprising at least one of the above. [Invention 1013] The aforementioned training dataset is (a) Data relating to the measurement of peptide-MHC binding affinity for at least one of the peptides, and (b) Data relating to the measurement of peptide-MHC binding stability for at least one of the peptides. Any method of the present invention 1001 to 1012, further comprising at least one of the above. [Invention 1014] Any method 1001 to 1013 of the present invention, wherein the set of presented likelihoods is further identified by the expression levels of at least one or more MHC alleles in the subject, as measured by RNA-seq or mass spectrometry. [Invention 1015] The aforementioned set of likelihoods is (a) the predicted affinity between the nascent antigens in the set of nascent antigens and the one or more MHC alleles, and (b) Predicted stability of the newly synthesized antigen-coding peptide-MHC complex Any method of the present invention 1001 to 1014, further specified by a property including at least one of the following. [Invention 1016] The aforementioned set of numerical likelihoods is (a) The C-terminal sequence within the source protein sequence adjacent to the nascent antigen-coding peptide sequence, and (b) The N-terminal sequence adjacent to the nascent antigen-coding peptide sequence within the source protein sequence. A method of any of the present invention 1001 to 1015, further specified by a property including at least one of the following. [Invention 1017] Any method 1001 to 1016 of the present invention, wherein the step of selecting the set of selected neonatal antigens includes selecting neonatal antigens that have a higher likelihood of being presented on the surface of tumor cells compared to non-selected neonatal antigens, based on the machine learning presentation model. [Invention 1018] The method according to any one of the present invention 1001 to 1017, wherein the step of selecting the set of selected neonatal antigens includes selecting neonatal antigens that have a higher likelihood of inducing a tumor-specific immune response in the subject compared to non-selected neonatal antigens, based on the machine learning presentation model. [Invention 1019] The method of any of the present invention 1001 to 1018, wherein the step of selecting the set of selected nascent antigens includes selecting nascent antigens that, based on the presentation model, have a higher likelihood of being presented to naive T cells by professional antigen-presenting cells (APCs) compared to nascent antigens that are not selected, and optionally the APCs are dendritic cells (DCs). [Invention 1020] The method according to any one of the 1001 to 1019 of the present invention, wherein the step of selecting the set of selected nascent antigens includes selecting nascent antigens that have a reduced likelihood of being inhibited by central or peripheral tolerance compared to nascent antigens that are not selected, based on the machine learning presentation model. [Invention 1021] Any method of the present invention 1001 to 1020, wherein the step of selecting the set of selected nascent antigens includes selecting nascent antigens that, based on the machine learning presentation model, have a reduced likelihood of inducing an autoimmune response against normal tissue in the subject compared to nascent antigens that are not selected. [Invention 1022] The method according to any one of the present invention 1001 to 1021, wherein the one or more tumor cells are selected from the group consisting of lung cancer, melanoma, breast cancer, ovarian cancer, prostate cancer, kidney cancer, gastric cancer, colon cancer, testicular cancer, head and neck cancer, pancreatic cancer, brain cancer, B-cell lymphoma, acute myeloid leukemia, chronic myeloid leukemia, chronic lymphocytic leukemia, T-cell lymphocytic leukemia, non-small cell lung cancer, and small cell lung cancer. [Invention 1023] Any method of the present invention 1001 to 1022 further comprises generating an output for constructing a personalized cancer vaccine from the selected set of neoantigens. [Invention 1024] The method of the present invention 1023, wherein the output for the personalized cancer vaccine comprises at least one peptide sequence or at least one nucleotide sequence encoding the selected set of nascent antigens. [Invention 1025] The present invention according to any of the methods 1001 to 1024, wherein the machine learning-trained presentation model is a neural network model. [Invention 1026] The method of the present invention 1025, wherein the neural network model includes multiple network models for MHC alleles, each network model being assigned to a corresponding MHC allele from among the multiple MHC alleles and including a series of nodes arranged in one or more layers. [Invention 1027] The method of the present invention 1026, wherein the neural network model is trained by updating the parameters of the neural network model, wherein the parameters of at least two network models are updated together for at least one training iteration. [Invention 1028] The method according to any one of the present invention 1025 to 1027, wherein the machine learning-processed presentation model is a deep learning model that includes one or more layers of nodes. [Invention 1029] The method of any one of the present invention 1001 to 1028, wherein the step of identifying one or more T cells comprises co-culturing the one or more T cells with one or more of the nascent antigens in the subset under conditions for proliferation of the one or more T cells. [Invention 1030] Any method of the present invention 1001 to 1029, wherein the step of identifying one or more T cells includes contacting the one or more T cells with the MHC multimer containing one or more of the nascent antigens in the subset under conditions that enable binding of the T cells to the MHC multimer. [Invention 1031] Any method of the present invention 1001 to 1030, further comprising identifying one or more T cell receptors (TCRs) of the one or more identified T cells. [Invention 1032] The method of the present invention 1031, wherein identifying one or more T cell receptors comprises sequencing the T cell receptor sequences of the one or more identified T cells. [Invention 1033] Antigen-specific isolated T cells for at least one selected nascent antigen from any subset of the present invention 1001 to 1032. [Invention 1034] A step of genetically modifying multiple T cells to express at least one of the one or more identified T cell receptors, A step of culturing the plurality of T cells under conditions that cause the plurality of T cells to proliferate, and The process of injecting the proliferated T cells into the target. The method of the present invention 1032, further comprising the above. [Invention 1035] The step of genetically modifying the plurality of T cells to express at least one of the one or more identified T cell receptors is: Cloning the T cell receptor sequence of one or more identified T cells into an expression vector, and Transfecting each of the aforementioned plurality of T cells with the expression vector. The method of the present invention 1034, including the method of the present invention. [Invention 1036] A step of culturing the one or more identified T cells under conditions that cause the one or more identified T cells to proliferate, and The process of injecting the proliferated T cells into the target. Any method of the present invention 1001 to 1035, further comprising the above. [Invention 1037] Any method 1001 to 1036 of the present invention, wherein one or more T cells that are antigen-specific to at least one of the nascent antigens in the subset are identified using 5 to 30 mL of whole blood from the subject. [Invention 1038] A method according to any one of the present invention 1001 to 1037, wherein the subset of the nascent antigens comprises up to 20 different nascent antigens, and one or more identified T cells recognize at least two of the nascent antigens in the subset of the nascent antigens. [Invention 1039] Any method of the present invention 1001 to 1038, wherein the one or more MHC alleles are class I MHC alleles. [Brief explanation of the drawing]

[0011] These features, aspects, and aspects of the present invention, as well as other features, aspects, and aspects, will be better understood with reference to the following description and accompanying drawings.

[0012] [Figure 1A] This paper outlines the current clinical approach to identifying newly synthesized antigens. [Figure 1B] This indicates that less than 5% of the predicted binding peptides are presented on tumor cells. [Figure 1C] This illustrates the impact of specificity issues in predicting newly synthesized antigens. [Figure 1D]Indicates that the binding prediction is not sufficient for the identification of neoantigens. [Figure 1E] Shows the probability of MHC-I presentation as a function of peptide length. [Figure 1F] Shows an exemplary peptide spectrum generated from Promega's dynamic range standard. FIG. 1F discloses SEQ ID NO:1. [Figure 1G] Shows how the addition of features increases the positive hit rate of the model. [Figure 2A] Schematic of an environment for identifying the likelihood of peptide presentation in a patient, according to one embodiment. [Figure 2B] Explains a method for obtaining presentation information, according to one embodiment. FIG. 2B discloses SEQ ID NO:28. [Figure 2C] Explains a method for obtaining presentation information, according to one embodiment. FIG. 2C discloses SEQ ID NOs:3 - 8 in order of appearance. [Figure 3] High-level block diagram explaining the computer logic components of a presentation identification system, according to one embodiment. [Figure 4] Explains an exemplary set of training data, according to one embodiment. FIG. 4 discloses "peptide sequences" as SEQ ID NOs:10 - 13 in order of appearance, SEQ ID NOs:15, 29 - 30, and 30 as "C-terminal flanking sequences". [Figure 5] Explains an exemplary network model related to MHC alleles. [Figure 6A] Explains an exemplary network model NNH(·) shared by MHC alleles, according to one embodiment. [Figure 6B] Explains an exemplary network model NNH(·) shared by MHC alleles, according to another embodiment. [Figure 7] Explains the generation of the presentation likelihood of peptides related to MHC alleles using an exemplary network model. [Figure 8]This paper explains the generation of presentation likelihood for MHC allele-related peptides using an exemplary network model. [Figure 9] This paper explains the generation of presentation likelihood for MHC allele-related peptides using an exemplary network model. [Figure 10] This paper explains the generation of presentation likelihood for MHC allele-related peptides using an exemplary network model. [Figure 11] This paper explains the generation of presentation likelihood for MHC allele-related peptides using an exemplary network model. [Figure 12] This paper explains the generation of presentation likelihood for MHC allele-related peptides using an exemplary network model. [Figure 13A] This shows the sample frequency distribution of mutational burden in NSCLC patients. [Figure 13B] This shows the number of nascent antigens presented in the vaccine simulated for patients selected based on selection criteria that determine whether the patient meets the minimum mutagenesis requirement, according to one embodiment. [Figure 13C] One embodiment compares the number of presented neogenic antigens in a simulated vaccine between selected patients associated with a vaccine containing a therapeutic subset identified based on a presentation model and selected patients associated with a vaccine containing a therapeutic subset identified by a prior art model. [Figure 13D] This compares the number of nascent antigens presented in a simulated vaccine between selected patients associated with a vaccine containing a therapeutic subset identified based on a single-allelic presentation model for HLA-A*02:01 and selected patients associated with a vaccine containing therapeutic subsets identified based on both HLA-A*02:01 and HLA-B*07:02. According to one embodiment, the vaccine volume is set to v=20 epitopes. [Figure 13E]This embodiment compares the number of nascent antigens presented in the simulated vaccine between patients selected based on mutational burden and patients selected based on expected utility score. [Figure 14A] This study compares the positive predictive value (PPV) at a 40% recall rate for the "complete MS model," the "peptide MS model," and the MHCFlurry1.2.0 binding affinity model, which has three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with a test set consisting of five different test samples, including excluded tumor samples, each test sample having a 1:2500 ratio of presented peptide to non-presented peptide. [Figure 14B] This study compares PPV at a 40% recall rate for the "Full MS Model," the "Peptide MS Model," and the MHCFlurry1.2.0 binding affinity model, which has three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with a test set consisting of 15 different test samples, each containing excluded peptides from a single-allelic cell line test dataset where each test sample has a 1:10,000 ratio of presented peptide to non-presented peptide. [Figure 14C] This study compared the proportion of somatic mutations recognized by T cells (e.g., pre-existing T cell responses) for the top 5, 10, and 20 ranked somatic mutations, identified by a "full MS model," a "peptide MS model," and an MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds (TPM > 0, 1, and 2), for a test set consisting of 12 different test samples, each taken from a patient exhibiting at least one pre-existing T cell response. [Figure 14D]This study compared the proportion of minimally nascent epitopes recognized by T cells (e.g., pre-existing T cell responses) for a set of 12 different test samples, each taken from a patient exhibiting at least one pre-existing T cell response. The top 5, 10, and 20 ranked minimally nascent epitopes were identified using a "complete MS model," a "peptide MS model," and an MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds (TPM > 0, 1, and 2). [Figure 15A] This shows the detection of T cell responses to patient-specific nascent antigen peptide pools in nine patients. [Figure 15B] This shows the detection of T cell responses to individual patient-specific nascent antigen peptides in four patients. [Figure 15C] An illustrative image of the ELISpot well for patient CU04 is shown. [Figure 16] This study compares the positive predictive value (PPV) at a 40% recall rate for the "complete MS model" and the "anchor residue-only MS model" when testing each model with a test set consisting of five different test samples, each of which has a 1:2500 ratio of presented peptide to non-presented peptide, including excluded tumor samples. [Figure 17A] Figure 14A shows the complete precision-recall curves for the "Full MS Model," the "Peptide MS Model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds (TPM > 0, 1, and 2) when each model is tested with test sample 0. [Figure 17B] This study compares PPV at a 40% recall rate for the "Full MS Model," the "Peptide MS Model," and the MHCFlurry1.2.0 binding affinity model, which has three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with a test set consisting of 15 different test samples, each containing excluded peptides from a single-allelic cell line test dataset where each test sample has a 1:5,000 ratio of presented peptide to non-presented peptide. [Figure 17C]The complete precision-recall curves of the "complete MS model", "peptide MS model", and MHCFlurry 1.2.0 binding affinity models with three different gene expression thresholds of TPM > 0, 1, and 2 when each model is tested with test sample 0 from FIG. 14A are shown. [Figure 17D] The complete precision-recall curves of the "complete MS model", "peptide MS model", and MHCFlurry 1.2.0 binding affinity models with three different gene expression thresholds of TPM > 0, 1, and 2 when each model is tested with test sample 1 from FIG. 14A are shown. [Figure 17E] The complete precision-recall curves of the "complete MS model", "peptide MS model", and MHCFlurry 1.2.0 binding affinity models with three different gene expression thresholds of TPM > 0, 1, and 2 when each model is tested with test sample 2 from FIG. 14A are shown. [Figure 17F] The complete precision-recall curves of the "complete MS model", "peptide MS model", and MHCFlurry 1.2.0 binding affinity models with three different gene expression thresholds of TPM > 0, 1, and 2 when each model is tested with test sample 3 from FIG. 14A are shown. [Figure 17G] The complete precision-recall curves of the "complete MS model", "peptide MS model", and MHCFlurry 1.2.0 binding affinity models with three different gene expression thresholds of TPM > 0, 1, and 2 when each model is tested with test sample 4 from FIG. 14A are shown. [Figure 17H] The complete precision-recall curves of the "complete MS model", "peptide MS model", and MHCFlurry 1.2.0 binding affinity models with three different gene expression thresholds of TPM > 0, 1, and 2 when each model is tested with the excluded peptides from the test dataset of the HLA-A*01:01 cell line in FIG. 14B having a ratio of presented peptides to non-presented peptides of 1:10,000 are shown. [Figure 17I]Figure 14B shows the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with excluded peptides from the HLA-A*02:01 cell line test dataset, which has a 1:10,000 ratio of presented peptides to non-presented peptides. [Figure 17J] Figure 14B shows the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with excluded peptides from the HLA-A*02:03 cell line test dataset, which has a 1:10,000 ratio of presented peptides to non-presented peptides. [Figure 17K] Figure 14B shows the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with excluded peptides from the HLA-A*02:07 cell line test dataset, which has a 1:10,000 ratio of presented peptides to non-presented peptides. [Figure 17L] Figure 14B shows the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with excluded peptides from the HLA-A*03:01 cell line test dataset, which has a 1:10,000 ratio of presented peptides to non-presented peptides. [Figure 17M] Figure 14B shows the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with excluded peptides from the HLA-A*24:02 cell line test dataset, which has a 1:10,000 ratio of presented peptides to non-presented peptides. [Figure 17N]Figure 14B shows the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with excluded peptides from the HLA-A*29:02 cell line test dataset, which has a 1:10,000 ratio of presented peptides to non-presented peptides. [Figure 17O] Figure 14B shows the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with excluded peptides from the HLA-A*31:01 cell line test dataset, which has a 1:10,000 ratio of presented peptides to non-presented peptides. [Figure 17P] Figure 14B shows the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with excluded peptides from the HLA-A*68:02 cell line test dataset, which has a 1:10,000 ratio of presented peptides to non-presented peptides. [Figure 17Q] Figure 14B shows the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with excluded peptides from the HLA-A*35:01 cell line test dataset, which has a 1:10,000 ratio of presented peptides to non-presented peptides. [Figure 17R] Figure 14B shows the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with excluded peptides from the HLA-A*44:02 cell line test dataset, which has a 1:10,000 ratio of presented peptides to non-presented peptides. [Figure 17S]Figure 14B shows the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with excluded peptides from the HLA-A*44:03 cell line test dataset, which has a 1:10,000 ratio of presented peptides to non-presented peptides. [Figure 17T] Figure 14B shows the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with excluded peptides from the HLA-A*51:01 cell line test dataset, which has a 1:10,000 ratio of presented peptides to non-presented peptides. [Figure 17U] Figure 14B shows the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with excluded peptides from the HLA-A*54:01 cell line test dataset, which has a 1:10,000 ratio of presented peptides to non-presented peptides. [Figure 17V] Figure 14B shows the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when testing each model with excluded peptides from the HLA-A*57:01 cell line test dataset, which has a 1:10,000 ratio of presented peptides to non-presented peptides. [Figure 18] This compares the positive predictive value (PPV) at 40% recall for different versions of the MS model and initial approach29 for modeling HLA-presented peptides in human tumors when testing each model with the test set shown in Figure 14A, which consists of five different test samples, including excluded tumor samples, each test sample having a 1:2500 ratio of presented peptide to non-presented peptide. [Figure 19A-1]The results of a controlled experiment using newly synthesized antigens in healthy HLA-matched donors are shown. [Figure 19A-2] The results of a controlled experiment using newly synthesized antigens in healthy HLA-matched donors are shown. [Figure 19B-1] The results of a control experiment using newly synthesized antigens in HLA-matched healthy donors are shown. Figure 19B displays 27, 24, 21-22, 31-36, 21, and 37-45 in order of appearance. [Figure 19B-2] The results of a control experiment using newly synthesized antigens in HLA-matched healthy donors are shown. Figure 19B displays 27, 24, 21-22, 31-36, 21, and 37-45 in order of appearance. [Figure 20] Figure 15A shows the detection of the T cell response to the PHA-positive control for each donor and each in vitro proliferation. [Figure 21A] This shows the detection of T cell responses to each individual patient-specific nascent antigen peptide in pool #2 in patient CU04. [Figure 21B] This shows the detection of T cell responses to individual patient-specific nascent antigen peptides in each of the three visits of patient CU04 and in each of the two visits of patient 1-024-002 (each visit taking place at a different time). [Figure 21C] This shows the detection of T cell responses to individual patient-specific nascent antigen peptides and to the patient-specific nascent antigen peptide pool during each of the two visits of patient CU04 and each of the two visits of patient 1-024-002 (each visit taking place at a different time). [Figure 22] Figure 15A shows the detection of two patient-specific nascent antigen peptide pools and T cell responses against DMSO-negative controls for the patient in question. [Figure 23]This study compares the predictive performance of three models in ranking peptides within the HLA-DRB1*15:01 / HLA-DRB5*01:01 study dataset: "MS model," which employs the lowest NetMHCIIpan percentile rank across HLA-DRB1*15:01 and HLA-DRB5*01:01 ("NetMHCIIpan rank": NetMHCIIpan 3.177), and "NetMHCIIpan nM," which employs the strongest affinity in nM units across HLA-DRB1*15:01 and HLA-DRB5*01:01 ("NetMHCIIpan nM": NetMHCIIpan 3.1). [Figure 24-1] This document describes a method for sequencing the TCR of newly generated antigen-specific memory T cells derived from peripheral blood of NSCLC patients. Figure 24 displays images 46-48 in order of appearance. [Figure 24-2] This document describes a method for sequencing the TCR of newly generated antigen-specific memory T cells derived from peripheral blood of NSCLC patients. Figure 24 displays images 46-48 in order of appearance. [Figure 25] This shows an exemplary embodiment of a TCR construct for introducing a TCR into recipient cells. [Figure 26-1] The nucleotide sequence of an exemplary P526 construct skeleton for cloning the TCR into an expression system for therapeutic development is shown. Figure 26 discloses SEQ ID NO:49. [Figure 26-2] The nucleotide sequence of an exemplary P526 construct skeleton for cloning the TCR into an expression system for therapeutic development is shown. Figure 26 discloses SEQ ID NO:49. [Figure 27-1] This shows an exemplary construct sequence for cloning chronotype 1 of a patient's newly generated antigen-specific TCR into an expression system for the development of therapeutic methods. Figure 27 discloses SEQ ID NO: 50. [Figure 27-2] This shows an exemplary construct sequence for cloning chronotype 1 of a patient's newly generated antigen-specific TCR into an expression system for the development of therapeutic methods. Figure 27 discloses SEQ ID NO: 50. [Figure 28-1] This shows an exemplary construct sequence for cloning chronotype 3 of a patient's newly generated antigen-specific TCR into an expression system for the development of therapeutic methods. Figure 28 discloses SEQ ID NO: 51. [Figure 28-2] This shows an exemplary construct sequence for cloning chronotype 3 of a patient's newly generated antigen-specific TCR into an expression system for the development of therapeutic methods. Figure 28 discloses SEQ ID NO: 51. [Figure 29] This is a flowchart of a method for providing personalized, neoantigen-specific therapy to a patient, according to one embodiment. [Figure 30] An exemplary computer for implementing the system shown in Figures 1 and 3 will be described. [Modes for carrying out the invention]

[0013] I. Definition In general, terms used in the claims and specification shall be interpreted as having the ordinary meaning understood by those skilled in the art. Certain terms are defined below for further clarity. In the event of any conflict between the ordinary meaning and the given definition, the given definition shall prevail.

[0014] As used herein, the term "antigen" refers to a substance that induces an immune response.

[0015] As used herein, the term “nascent antigen” refers to an antigen that has at least one change that makes it different from the corresponding wild-type parent antigen, for example, due to tumor cell mutations or tumor cell-specific post-translational modifications. Nascent antigens may include polypeptide sequences or nucleotide sequences. Mutations may include frameshift or non-frameshift insertions or deletions (indels), missense or nonsense substitutions, splice site changes, genomic rearrangements or gene fusions, or any genomic or expression changes that result in nascent ORFs. Mutations may also include splice variants. Tumor cell-specific post-translational modifications may include abnormal phosphorylation. Tumor cell-specific post-translational modifications may also include splice antigens produced by the proteasome. See Liepe et al., A large fraction of HLA class I ligands are proteasome-generated spliced ​​peptides; Science. 2016 Oct 21;354(6310):354-358.

[0016] As used herein, the term “tumor-generating antigen” refers to a neoplastic antigen that is present in the tumor cells or tissue of the subject but not in the corresponding normal cells or tissue of the subject.

[0017] As used herein, the term “naegenetic antigen-based vaccine” refers to a vaccine construct based on one or more naegenetic antigens, for example, multiple naegenetic antigens.

[0018] As used herein, the term "candidate nascent antigen" refers to a mutation or other abnormality that results in a new sequence that may represent a nascent antigen.

[0019] As used herein, the term "coding region" refers to the portion of a gene that codes for a protein.

[0020] As used herein, the term "coding mutation" refers to a mutation that occurs in the coding region.

[0021] As used herein, the term "ORF" means Open Reading Frame.

[0022] As used herein, the term “new-onset ORF” refers to tumor-specific ORFs resulting from mutations or other abnormalities such as splicing.

[0023] As used herein, the term "missense mutation" refers to a mutation that results in the substitution of one amino acid with another.

[0024] As used herein, the term "nonsense mutation" refers to a mutation that results in the substitution of an amino acid into a stop codon.

[0025] As used herein, the term "frameshift mutation" refers to a mutation that causes a change in the frame of a protein.

[0026] As used herein, the term “insertion / deletion” refers to the insertion or deletion of one or more nucleic acids.

[0027] As used herein, the term “identity” (%) in relation to the sequences of two or more nucleic acids or polypeptides means two or more sequences or subsequences that, when compared and aligned for the greatest match, have the same specific ratio (%) of nucleotides or amino acid residues, either using one of the sequence comparison algorithms described below (e.g., BLASTP and BLASTN, or other algorithms available to those skilled in the art) or by visual inspection. Depending on the application, the “identity” (%) may be present across regions of the sequences being compared, for example, across functional domains, or across the full lengths of the two sequences being compared.

[0028] In sequence comparison, typically, one sequence functions as the reference sequence compared to the test sequence. When using a sequence comparison algorithm, the test sequence and reference sequence are input into a computer, subsequence coordinates are specified if necessary, and parameters for the sequence algorithm program are specified. The sequence comparison algorithm then calculates the sequence identity percentage (%) of the test sequence to the reference sequence based on the specified program parameters. Alternatively, sequence similarity or difference can also be established by the presence or absence of specific nucleotides at selected sequence locations (e.g., sequence motifs), or, in the post-translated sequence, by the presence or absence of amino acids.

[0029] The optimal alignment of sequences for comparison can be performed, for example, by the local homology algorithm of Smith & Waterman, Adv.Appl.Math.2:482 (1981), the homology alignment algorithm of Needleman & Wunsch, J.Mol.Biol.48:443 (1970), the similarity search method of Pearson & Lipman, Proc.Nat'l.Acad.Sci.USA 85:2444 (1988), by computer execution of these algorithms (GAP, BESTFIT, FASTA, and TFASTA in the Wisconsin Genetics Software Package, Genetics Computer Group, 575 Science Dr., Madison, Wis.), or by visual inspection (generally, see Ausubel et al. below).

[0030] One example of a suitable algorithm for determining sequence identity (%) and sequence similarity (%) is the BLAST algorithm described in Altschul et al., J.Mol.Biol.215:403-410 (1990). Software for performing BLAST analysis is publicly available through the National Center for Biotechnology Information.

[0031] As used herein, the terms “nonstop” or “readthrough” refer to mutations that result in the removal of a natural stop codon.

[0032] As used herein, the term “epitope” refers to a specific portion of an antigen to which an antibody or T cell receptor commonly binds.

[0033] As used herein, the term “immunogenicity” means, for example, the ability to induce an immune response via T cells, B cells, or both.

[0034] As used herein, the terms "HLA binding affinity" and "MHC binding affinity" refer to the affinity for binding between a specific antigen and a specific MHC allele.

[0035] As used herein, the term "bait" refers to a nucleic acid probe used to concentrate a specific sequence of DNA or RNA from a sample.

[0036] As used herein, the term “mutation” refers to the difference between the nucleic acid in question and a reference human genome used as a control.

[0037] As used herein, the term "mutation call" typically refers to the algorithmic determination of the presence of a mutation from sequencing.

[0038] As used herein, the term "polymorphism" refers to germline mutations, i.e., mutations found in all DNA-containing cells of an individual.

[0039] As used herein, the term “somatic mutation” refers to a mutation that occurs in the non-germline cells of an individual.

[0040] As used herein, the term "allele" refers to one version of a gene, one version of a gene sequence, or one version of a protein.

[0041] As used herein, the term "HLA type" refers to the complement of an HLA gene allele.

[0042] As used herein, the term “nonsense mutation-dependent degradation mechanism” or “NMD” refers to the degradation of mRNA by cells due to immature stop codons.

[0043] As used herein, the term "truncal mutation" refers to a mutation that occurs early in tumor development and is present in the majority of tumor cells.

[0044] As used herein, the term “subclonal mutation” refers to a mutation that occurs later in tumor development and is present in only a subset of tumor cells.

[0045] As used herein, the term “exome” refers to a subset of the genome that codes for proteins. An exome can be a collection of exons within a genome.

[0046] As used herein, the term “logistic regression” refers to a regression model for binary data from statistics, in which the logit, with a probability of the dependent variable being equal to 1, is modeled as a linear function of the dependent variable.

[0047] As used herein, the term “neural network” refers to a machine learning model for classification or regression that consists of performing multi-layer linear transformations followed by element-wise nonlinear transformations, typically trained by stochastic gradient descent and backpropagation.

[0048] As used herein, the term “proteome” refers to the set of all proteins expressed and / or translated by a cell, a group of cells, or an organism.

[0049] As used herein, the term “peptideome” refers to the set of all peptides presented by MHC-I or MHC-II on the surface of a cell. The peptideome may also refer to the properties of a cell or a collection of cells (for example, the tumor peptideome means the union of the peptideomes of all cells, including a tumor).

[0050] As used herein, the term "ELISPOT" means an enzyme-linked immunosorbent spot assay, a common method for observing immune responses in humans and animals.

[0051] As used herein, the term "dexatormer" refers to a dextran-based peptide-MHC multimer used for antigen-specific T cell staining in flow cytometry.

[0052] As used herein, the term "MHC multimer" refers to a peptide-MHC complex consisting of multiple peptide-MHC monomer units.

[0053] As used herein, the term "MHC tetramer" refers to a peptide-MHC complex consisting of four peptide-MHC monomer units.

[0054] As used herein, the term “tolerance” or “immune tolerance” refers to a state of immune non-response to one or more antigens, such as autoantigens.

[0055] As used herein, the term “central tolerance” refers to tolerance induced in the thymus by either deleting autoreactive T cell clones or promoting the differentiation of autoreactive T cell clones into immunosuppressive regulatory T cells (Tregs).

[0056] As used herein, the term “peripheral tolerance” refers to tolerance handed down in the peripheral system by downregulating or anergizing autoreactive T cells that have survived central tolerance, or by promoting the differentiation of these T cells into Tregs.

[0057] The term "sample" may include single cells or multiple cells, or cell fragments, or aliquots of bodily fluids, taken from a subject by means including venipuncture, excretion, ejaculation, massage, biopsy, needle aspiration, lavage, scraping, surgical incision, or intervention, or other means known in the art.

[0058] The term "subject" includes cells, tissues, or organisms, whether in vivo, ex vivo, or in vitro, male or female, human or non-human. The term "subject" includes mammals, including humans.

[0059] The term “mammal” encompasses both humans and non-humans, including but not limited to humans, non-human primates, dogs, cats, mice, cattle, horses, and pigs.

[0060] The term “clinical factors” refers to a measure of the subject’s condition, such as the activity or severity of a disease. “Clinical factors” encompass all markers of the subject’s health status, including non-sample markers, and / or, non-limitingly, other characteristics of the subject, such as age and sex. Clinical factors can be scores, values, or sets of values ​​that can be obtained from assessments of a subject or a sample (or population of samples) derived from a subject under given conditions. Clinical factors can also be predicted by other parameters, such as markers and / or gene expression substitutes. Clinical factors may include tumor type, tumor subtype, and smoking history.

[0061] Abbreviations: MHC: Major Histocompatibility Complex; HLA: Human Leukocyte Antigen, or Human MHC Locus; NGS: Next-Generation Sequencing; PPV: Positive Predictive Value; TSNA: Tumor-Specific Neo-Antigenic Antigen; FFPE: Formalin-Fixed Paraffin-Embedded; NMD: Nonsense Mutation-Dependent Degradation Mechanism; NSCLC: Non-Small Cell Lung Cancer; DC: Dendritic Cell.

[0062] When used in this specification and the appended claims, the singular forms "a," "an," and "the" refer to plural nouns unless the context explicitly indicates otherwise.

[0063] Terms not directly defined herein should be understood to have the general meanings associated with them as understood within the art of the present invention. Certain terms are considered herein for the purpose of providing further guidance to practitioners in describing compositions, apparatus, methods, etc., of embodiments of the present invention, as well as their manufacture or use. It will be recognized that there may be multiple ways of saying the same thing. Accordingly, alternative words and synonyms may be used for any one or more of the terms considered herein. Emphasis should not be placed on whether a term is detailed or considered herein. Several synonyms or alternative methods, materials, etc., are provided. The listing of one or more synonyms or equivalents does not exclude the use of other synonyms or equivalents unless explicitly stated. The use of examples, including examples of terms, is for illustrative purposes only and does not limit the scope and meaning of embodiments of the invention herein.

[0064] All references, published patents, and patent applications cited herein are incorporated herein by reference in their entirety for all purposes.

[0065] II. Methods for identifying new antigens This specification discloses a method for identifying T cells that are antigen-specific to a target tumor cell-derived nascent antigen, which is likely to be presented on the surface of the tumor cell. The method includes obtaining exome, transcriptome, and / or whole-genome nucleotide sequencing data from the target tumor cell and normal cells. Using this nucleotide sequencing data, the peptide sequence of each nascent antigen in a set of nascent antigens is obtained. The set of nascent antigens is identified by comparing the nucleotide sequencing data from the tumor cell with nucleotide sequencing data from normal cells. Specifically, the peptide sequence of each nascent antigen in the set of nascent antigens includes at least one change that makes the peptide sequence different from the corresponding wild-type parent peptide sequence identified from the target normal cell. The method further includes encoding the peptide sequence of each nascent antigen in the set of nascent antigens into a corresponding numerical vector. Each numerical vector includes information describing the amino acids that make up the peptide sequence and the positions of the amino acids within the peptide sequence. The method further includes generating presentation likelihoods for each nascent antigen in the set of nascent antigens by inputting the numerical vectors into a machine learning presentation model. Each presentation likelihood represents the likelihood that the corresponding nascent antigen will be presented by an MHC allele on the surface of the target tumor cell. The machine learning presentation model includes multiple parameters and a function. The multiple parameters are identified based on the training dataset. The training dataset includes, for each of several samples, a label obtained by mass spectrometry measuring the presence of a peptide bound to at least one MHC allele from a set of MHC alleles identified as present in that sample, and a training peptide sequence encoded as a numerical vector containing information describing the amino acids that make up the peptide and / or the positions of amino acids within the peptide. The function represents the relationship between the numerical vector received as input to the machine learning presentation model and the presentation likelihood generated as output by the machine learning presentation model based on the numerical vector and the multiple parameters.The method further includes selecting a subset of a set of nascent antigens based on presentation likelihood in order to generate a selected set of nascent antigens. The method further includes identifying antigen-specific T cells for at least one of the nascent antigens in the subset and returning these identified T cells.

[0066] In some embodiments, inputting numerical vectors into a machine learning presentation model includes applying the machine learning presentation model to the peptide sequence of a nascent antigen to generate a dependency score for each MHC allele. The dependency score for a given MHC allele indicates whether that MHC allele presents the nascent antigen based on a specific amino acid at a specific position in the peptide sequence. In further embodiments, inputting numerical vectors into a machine learning presentation model further includes, for each MHC allele, transforming the dependency scores to generate a corresponding allele-by-allele likelihood, which indicates the likelihood that the corresponding MHC allele presents the corresponding nascent antigen, and combining the allele-by-allele likelihoods to generate a nascent antigen presentation likelihood. In some embodiments, transforming the dependency scores models the nascent antigen presentation as mutually exclusive across MHC alleles. In alternative embodiments, inputting numerical vectors into a machine learning presentation model further includes transforming the combination of dependency scores to generate a presentation likelihood. In such embodiments, transforming the combination of dependency scores models the nascent antigen presentation as interference between MHC alleles.

[0067] In some embodiments, the set of presentation likelihoods is further identified by one or more allele-non-interaction properties. In such embodiments, the method further includes generating a dependency score for an allele-non-interaction property by applying a machine learning-trained presentation model to the allele-non-interaction property. The dependency score indicates whether the peptide sequence of the corresponding nascent antigen is presented based on the allele-non-interaction property. In some embodiments, the method further includes combining the dependency score for each MHC allele with the dependency score for the allele-non-interaction property, transforming the combined dependency score for each MHC allele to generate an allele-by-allele likelihood for each MHC allele, and combining the allele-by-allele likelihoods to generate a presentation likelihood. The allele-by-allele likelihood for a given MHC allele indicates whether that MHC allele presents the corresponding nascent antigen. In alternative embodiments, the method further includes combining the dependency score for an MHC allele with the dependency score for the allele-non-interaction property, and transforming the combined dependency score to generate a presentation likelihood.

[0068] In some embodiments, the MHC allele comprises two or more different MHC alleles.

[0069] In some embodiments, the peptide sequence includes a peptide sequence having a length other than nine amino acids.

[0070] In some embodiments, encoding the peptide sequence includes encoding the peptide sequence using a one-hot encoding scheme.

[0071] In certain embodiments, the sample comprises at least one of the following: a cell line engineered to express a single MHC allele, a cell line engineered to express multiple MHC alleles, a human cell line obtained from or derived from multiple patients, or a fresh or frozen tissue sample obtained from multiple patients.

[0072] In some embodiments, the training dataset further includes data related to measurements of peptide-MHC binding affinity for at least one of the peptides, and at least one of data related to measurements of peptide-MHC binding stability for at least one of the peptides.

[0073] In some embodiments, the set of presentation likelihoods is further identified by the expression levels of MHC alleles in the subject, measured by RNA-seq or mass spectrometry.

[0074] In some embodiments, the set of presentation likelihoods is further identified by properties including the predicted affinity between the nascent antigen and the MHC allele in the set of nascent antigens, and the predicted stability of the nascent antigen-coding peptide-MHC complex, at least one of the above.

[0075] In some embodiments, the set of numerical likelihoods is further identified by a property that includes at least one of the C-terminal sequence adjacent to the nascent antigen-coding peptide sequence within the source protein sequence, and the N-terminal sequence adjacent to the nascent antigen-coding peptide sequence within the source protein sequence.

[0076] In some embodiments, selecting a set of chosen nascent antigens involves selecting nascent antigens that have a higher likelihood of being presented on the tumor cell surface compared to unselected nascent antigens, based on a machine learning presentation model.

[0077] In some embodiments, selecting a set of selected nascent antigens involves selecting nascent antigens that, based on a machine learning presentation model, have a higher likelihood of inducing a tumor-specific immune response in the target compared to unselected nascent antigens.

[0078] In some embodiments, selecting a set of chosen nascent antigens involves selecting nascent antigens that, based on a presentation model, have a higher likelihood of being presented to naive T cells by professional antigen-presenting cells (APCs) compared to unselected nascent antigens. In such embodiments, the APCs are optionally dendritic cells (DCs).

[0079] In some embodiments, selecting a set of chosen nascent antigens involves selecting nascent antigens that, based on a machine learning presentation model, have a reduced likelihood of being inhibited by central or peripheral tolerance compared to nascent antigens that are not selected.

[0080] In some embodiments, selecting a set of nascent antigens involves selecting, based on a machine learning presentation model, nascent antigens that have a reduced likelihood of inducing an autoimmune response against normal tissue in a subject compared to unselected nascent antigens.

[0081] In some embodiments, one or more tumor cells are selected from the group consisting of lung cancer, melanoma, breast cancer, ovarian cancer, prostate cancer, kidney cancer, gastric cancer, colon cancer, testicular cancer, head and neck cancer, pancreatic cancer, brain cancer, B-cell lymphoma, acute myeloid leukemia, chronic myeloid leukemia, chronic lymphocytic leukemia, and T-cell lymphocytic leukemia, non-small cell lung cancer, and small cell lung cancer.

[0082] In some embodiments, the method further includes generating an output for constructing a personalized cancer vaccine from a selected set of nascent antigens. In such embodiments, the output for the personalized cancer vaccine may include at least one peptide sequence or at least one nucleotide sequence encoding the selected set of nascent antigens.

[0083] In some embodiments, the machine learning presentation model is a neural network model. In such embodiments, the neural network model may include multiple network models for MHC alleles, each network model being assigned to a corresponding MHC allele from a plurality of MHC alleles, and each network model including a set of nodes arranged in one or more layers. In such embodiments, the neural network model can be trained by updating the parameters of the neural network model, where the parameters of at least two network models are updated together for at least one training iteration. In some embodiments, the machine learning presentation model may be a deep learning model including one or more layers of nodes.

[0084] In some embodiments, identifying T cells involves co-culturing T cells with one or more nascent antigens from a subset under conditions that promote T cell growth.

[0085] In some embodiments, identifying T cells involves contacting the T cells with an MHC multimer containing one or more nascent antigens from a subset, under conditions that allow binding between the T cells and the MHC multimer.

[0086] In some embodiments, the method further includes identifying the T cell receptor (TCR) of the identified T cells. In such embodiments, identifying the T cell receptor includes sequencing the T cell receptor sequence of the identified T cells. In such embodiments, the method may further include genetically engineering T cells to express at least one of one or more identified T cell receptors, culturing the T cells under conditions that allow T cells to proliferate, and injecting the proliferated T cells into a target. In such embodiments, genetically engineering T cells to express at least one of the identified T cell receptors may include cloning the T cell receptor sequence of the identified T cells into an expression vector and transfecting each T cell with the expression vector.

[0087] In some embodiments, the method may further include culturing the identified T cells under conditions that promote the growth of the identified T cells, and injecting the proliferated T cells into the target.

[0088] In some embodiments, T cells that are antigen-specific to at least one of the nascent antigens in a subset are identified using 5–30 mL of whole blood from the subject.

[0089] In some embodiments, the subset of nascent antigens includes up to 20 different nascent antigens, and the identified T cells recognize at least two of the nascent antigens in the subset.

[0090] In some embodiments, the MHC allele is a class I MHC allele.

[0091] This specification also discloses isolated T cells that are antigen-specific to at least one selected nascent antigen from the subset of nascent antigens described above.

[0092] III. Identification of tumor-specific mutations in neogeneic antigens Furthermore, methods for identifying certain mutations (e.g., mutations or alleles present in cancer cells) are also disclosed herein. In particular, these mutations may be present in the genome, transcriptome, proteome, or exome of cancer cells of a subject having cancer, but not in normal tissue from which the subject originates.

[0093] Gene mutations in tumors can be considered useful for immunological targeting of tumors if they result in changes to the amino acid sequence of proteins exclusively within the tumor. Useful mutations include: (1) non-synonymous mutations resulting in different amino acids in a protein; (2) read-through mutations with modified or deleted stop codons resulting in the translation of longer proteins with a novel tumor-specific sequence at the C-terminus; (3) splice site mutations resulting in the inclusion of introns in mature mRNA, and thus a unique tumor-specific protein sequence; (4) chromosomal rearrangements (i.e., gene fusions) resulting in chimeric proteins with tumor-specific sequences at the junction of two proteins; and (5) frameshift mutations or deletions resulting in a novel open reading frame with a novel tumor-specific protein sequence. Mutations may also include one or more non-frameshift insertions or deletions, missense or nonsense substitutions, splice site changes, genomic rearrangements or gene fusions, or any genomic or expression changes resulting in a new ORF.

[0094] For example, mutated peptides or mutated polypeptides resulting from splice site, frameshift, readthrough, or gene fusion mutations in tumor cells can be identified by sequencing DNA, RNA, or proteins in tumor versus normal cells.

[0095] Furthermore, mutations may include previously identified tumor-specific mutations. Known tumor mutations can be found in the Catalogue of Somatic Mutations in Cancer (COSMIC) database.

[0096] Various methods are available to detect the presence of specific mutations or alleles in an individual's DNA or RNA. Advances in this field have provided accurate, easy, and inexpensive large-scale SNP genotyping. For example, several techniques have been described, including dynamic allele-specific hybridization (DASH), microplate array diagonal gel electrophoresis (MADGE), pyrosequencing, oligonucleotide-specific ligation, the TaqMan system, and various DNA "chip" technologies such as the Affymetrix SNP chip. These methods typically utilize amplification of the target gene region by PCR. Further methods are based on the generation of small signal molecules by invasive cleavage and subsequent mass spectrometry, or on immobilized padlock probes and rolling circle amplification. Some of the methods known in the art for detecting specific mutations are summarized below.

[0097] PCR-based detection methods can simultaneously involve multiple amplification of numerous markers. For example, it is well known in the art to select PCR primers to produce PCR products of non-overlapping sizes that can be analyzed simultaneously. Alternatively, it is possible to amplify different markers with primers that are differentially labeled and therefore can be differentially detected. Naturally, hybridization-based detection methods enable differential detection of multiple PCR products in a sample. Other techniques that enable multiple analysis of multiple markers are known in the art.

[0098] Several methods have been developed to facilitate the analysis of single nucleotide polymorphisms in genomic DNA or cellular RNA. For example, single nucleotide polymorphisms can be detected by using specialized exonuclease-resistant nucleotides, such as those disclosed in Mundy, CR (U.S. Patent No. 4,656,127). According to this method, a primer complementary to the allele sequence immediately 3' of the polymorphic site is hybridized to a target molecule obtained from a specific animal or human. If the polymorphic site on the target molecule contains a nucleotide complementary to a specific exonuclease-resistant nucleotide derivative, that derivative is incorporated into the end of the hybridized primer. Such incorporation makes the primer resistant to the exonuclease, thereby enabling its detection. Since the identity of the exonuclease-resistant derivative of the sample is known, the finding that the primer has become resistant to the exonuclease reveals that the nucleotide present in the polymorphic site of the target molecule is complementary to that of the nucleotide derivative used in the reaction. This method has the advantage of not requiring the determination of a large amount of exogenous sequence data.

[0099] To determine the identity of nucleotides at polymorphic sites, solution-based methods can be used (Cohen, D. et al. (French Patent No. 2,650,840; PCT Application No. WO91 / 02087)). These methods utilize primers complementary to the allele sequence immediately 3' of the polymorphic site, as in Mundy's method (US Patent No. 4,656,127). This method determines the identity of nucleotides at the polymorphic site using a labeled dideoxynucleotide derivative, which, if complementary to the nucleotide at the polymorphic site, is incorporated onto the end of the primer. An alternative method, known as Genetic Bit Analysis or GBA, is described by Goelet, P. et al. (PCT Application No. 92 / 15712). Goelet, P. et al.'s method uses a mixture of a labeled terminator and a primer complementary to the 3' sequence of the polymorphic site. The method of al. uses a mixture of a labeled terminator and a primer complementary to the 3' sequence of the polymorphic site. In contrast to the method of Cohen et al. (French Patent No. 2,650,840; PCT Application No. WO91 / 02087), the method of Goelet, P. et al. can be a heterogeneous phase assay in which the primer or target molecule is immobilized on a solid phase.

[0100] Several primer guide nucleotide insertion procedures for assaying polymorphism sites in DNA are described (Komher, J. Set al., Nucl. Acids. Res. 17:7779-7784 (1989); Sokolov, BP, Nucl. Acids Res. 18:3671 (1990); Syvanen, A.-C., et al., Genomics 8:684-692 (1990); Kuppuswamy, M. Net al., Proc. Natl. Acad. Sci. (USA) 88:1143-1147 (1991); Prezant, TR et al., Hum. Mutat. 1:159-164 (1992); Ugozzoli, L. et al., GATA 9:107-112 (1992); Nyren, P. et al. (Syvanen, A.-C., et al., Amer. J. Hum. Genet. 52:46-59 (1993)). These methods differ from GBA in that they utilize the incorporation of labeled deoxynucleotides to distinguish between bases at polymorphic sites. In such a form, the signal is proportional to the number of incorporation of deoxynucleotides, so polymorphisms occurring in a run of the same nucleotide can result in a signal proportional to the length of the run (Syvanen, A.-C., et al., Amer. J. Hum. Genet. 52:46-59 (1993)).

[0101] Numerous initiatives obtain sequence information directly and in parallel from millions of individual molecules of DNA or RNA. Real-time single-molecule synthesis sequencing techniques rely on the detection of fluorescent nucleotides as they are incorporated into the nascent strand of DNA that is complementary to the template being sequenced. In one method, oligonucleotides 30–50 bases long are covalently immobilized at their 5' ends to a glass coverslip. These immobilized strands serve two functions. First, they act as capture sites on the target template strand when the template is composed of a capture tail complementary to the surface-bound oligonucleotide. They also act as primers for template-directed primer extension, forming the basis for sequence reading. The capture primers function as fixed positional sites for sequencing, using multiple cycles of synthesis, detection, and chemical cleavage of a dye-linker to remove the dye. Each cycle consists of adding a polymerase / labeled nucleotide mixture, rinsing, imaging, and dye cleavage. In an alternative method, polymerase is modified with a fluorescent donor molecule and immobilized on a glass slide, while each nucleotide is color-coded by an acceptor fluorescent moiety attached to γ-phosphate. As the nucleotides are incorporated into a new chain, the system detects the interaction between the fluorescently tagged polymerase and the fluorescently modified nucleotides. Other synthetic sequencing techniques also exist.

[0102] Any suitable synthetic sequencing platform can be used to identify mutations. As described above, four major synthetic sequencing platforms are currently available: the Genome Sequencer from Roche / 454 Life Sciences, the 1G Analyzer from Illumina / Solexa, the SOLiD system from Applied BioSystems, and the Heliscope system from Helicos Bioscience. Synthetic sequencing platforms have also been described by Pacific BioSciences and VisiGen Biotechnologies. In some embodiments, the numerous nucleic acid molecules to be sequenced are bound to a support (e.g., a solid support). To immobilize the nucleic acid on the support, capture sequences / universal priming sites can be added to the 3' and / or 5' ends of the template. The nucleic acid can be immobilized to the support by hybridizing the capture sequence to a complementary sequence covalently attached to the support. A capture sequence (also called a universal capture sequence) is a nucleic acid sequence that is complementary to a sequence attached to a support, and can function dually as a universal primer.

[0103] As an alternative to the capture sequence, members of a coupling pair (e.g., antibody / antigen, receptor / ligand, or an avidin-biotin pair, as described in U.S. Patent Application No. 2006 / 0252077, for example) can be ligated to each fragment and captured on a surface coated with the second member of the respective coupling pair.

[0104] Following capture, the sequence can be analyzed by single-molecule detection / sequencing, such as described in the Examples and U.S. Patent No. 7,283,337, including template-dependent synthetic sequencing. In synthetic sequencing, surface-bound molecules are exposed to a number of labeled nucleotide triphosphates in the presence of polymerase. The template sequence is determined by the order in which the labeled nucleotides are incorporated into the 3' end of the growing chain. This can be done in real time and in step-and-repeat mode. For real-time analysis, different optical labels can be incorporated for each nucleotide, and multiple lasers can be used to stimulate the incorporated nucleotides.

[0105] Sequencing may also include other large-scale parallel sequencing, or next-generation sequencing (NGS) techniques and platforms. Additional examples of large-scale parallel sequencing techniques and platforms include Illumina HiSeq or MiSeq, ThermoPGM or Proton, Pac Bio RS II or Sequel, Qiagen's Gene Reader, and Oxford Nanopore MinION. Additional similar current large-scale parallel sequencing technologies, and future generations of these technologies, may be used.

[0106] Nucleic acid samples for use in the methods described herein can be obtained using any cell type or tissue. For example, DNA or RNA samples can be obtained from tumors or body fluids, such as blood or saliva obtained by known techniques (e.g., venipuncture). Alternatively, nucleic acid testing can be performed on dry samples (e.g., hair or skin). In addition, samples can be obtained from tumors for sequencing, and other samples can be obtained from normal tissue for sequencing if the normal tissue is of the same tissue type as the tumor. Samples can be obtained from tumors for sequencing, and other samples can be obtained from normal tissue for sequencing if the normal sample is of a different tissue type than the tumor.

[0107] Tumors may include one or more of the following: lung cancer, melanoma, breast cancer, ovarian cancer, prostate cancer, kidney cancer, stomach cancer, colon cancer, testicular cancer, head and neck cancer, pancreatic cancer, brain cancer, B-cell lymphoma, acute myeloid leukemia, chronic myeloid leukemia, chronic lymphocytic leukemia, and T-cell lymphocytic leukemia, non-small cell lung cancer, and small cell lung cancer.

[0108] Alternatively, protein mass spectrometry can be used to identify or demonstrate the presence of mutated peptides bound to MHC proteins on tumor cells. These peptides can be acid-eluted from tumor cells or from HLA molecules immunoprecipitated from tumors, and then identified using mass spectrometry.

[0109] IV. Neoantigens A nascent antigen can contain nucleotides or polynucleotides. For example, a nascent antigen can be an RNA sequence encoding a polypeptide sequence. Nascent antigens useful in vaccines can therefore contain nucleotide sequences or polypeptide sequences.

[0110] Disclosed herein are isolated peptides containing tumor-specific mutations identified by the methods disclosed herein, peptides containing known tumor-specific mutations, and mutant polypeptides or fragments thereof identified by the methods disclosed herein. Nascent antigen peptides may be described in the context of their coding sequences if the nascent antigen includes a nucleotide sequence (e.g., DNA or RNA) encoding the polypeptide sequence to which it relates.

[0111] One or more polypeptides encoded by a nascent antigen nucleotide sequence may contain at least one of the following: binding affinity to MHC with an IC50 value less than 1000 nM; for MHC class I peptides, a length of 8 to 15 amino acids, 8, 9, 10, 11, 12, 13, 14, or 15; presence of a sequence motif within or near the peptide that promotes proteasome cleavage; and presence of a sequence motif that promotes TAP transport. For MHC class II polypeptides, a length of 6 to 30 amino acids, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, or 30 amino acids; presence of a sequence motif within or near the peptide that promotes HLA binding catalyzed by extracellular or lysosomal proteases (e.g., cathepsins) or HLA-DM.

[0112] One or more newly generated antigens can be present on the surface of the tumor.

[0113] One or more newly synthesized antigens may be immunogenic in a tumor-bearing subject and, for example, may elicit a T-cell response or a B-cell response in the subject.

[0114] One or more neogeneic antigens that induce an autoimmune response in a subject can be excluded from consideration in the context of vaccine production for subjects with tumors.

[0115] The size of at least one nascent antigenic peptide molecule is approximately 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, and 35. It may include, but is not limited to, approximately 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 60, 70, 80, 90, 100, 110, 120, or more amino acid residues, and any range derived from these ranges. In a specific embodiment, the nascent antigenic peptide molecule has 50 amino acids or less.

[0116] For MHC class I, the nascent antigenic peptides and polypeptides are 15 residues or less in length, usually consisting of about 8 to 11 residues, and particularly 9 or 10 residues; for MHC class II, they can be 6 to 30 residues.

[0117] Where desirable, longer peptides can be designed in several ways. In one example, if the likelihood of peptide presentation on an HLA allele is predicted or known, the longer peptide may consist of (1) individual presented peptides having an extension of 2-5 amino acids toward the N-terminus and C-terminus of their respective corresponding gene products; or (2) a chain of some or all of the presented peptides, each having an extended sequence. In another example, if sequencing reveals long (longer than 10 residues) nascent epitope sequences present in the tumor (e.g., by frameshift, read-through, or intron inclusion resulting in a novel peptide sequence), the longer peptide may consist of (3) the entire stretch of novel tumor-specific amino acids, thus avoiding the need for computational or in vitro test-based selection of shorter peptides to present to the strongest HLA. In either example, the use of longer peptides may enable endogenous processing by patient cells, potentially leading to more effective antigen presentation and induction of T cell responses.

[0118] The nascent antigenic peptides and polypeptides can be presented on HLA proteins. In some embodiments, the nascent antigenic peptides and polypeptides are presented on HLA proteins with stronger affinity than wild-type peptides. In some embodiments, the nascent antigenic peptides or polypeptides may have an IC50 of less than 5000 nM, less than 1000 nM, less than 500 nM, less than 250 nM, less than 200 nM, less than 150 nM, less than 100 nM, less than 50 nM, or less than these values.

[0119] In some embodiments, the newly synthesized antigenic peptides and polypeptides, when administered to a subject, do not induce an autoimmune response and / or induce immune tolerance.

[0120] The present invention also provides compositions comprising at least two or more nascent antigenic peptides. In some embodiments, the composition contains at least two different peptides. The at least two different peptides may be derived from the same polypeptide. Different polypeptides mean that the peptides differ in length, amino acid sequence, or both. The peptides are derived from any polypeptide known or found to contain tumor-specific mutations. Suitable polypeptides from which nascent antigenic peptides may be derived can be found, for example, in the COSMIC database. COSMIC maintains comprehensive information on somatic mutations in human cancers. The peptides contain tumor-specific mutations. In some embodiments, the tumor-specific mutations are driver mutations for a particular type of cancer.

[0121] Newly synthesized antigenic peptides and polypeptides possessing desirable activity or properties can be modified to enhance or substantially retain at least all of the biological activity of the unmodified peptide in binding to the desired MHC molecule and activating appropriate T cells, while conferring specific desirable attributes, such as improved pharmacological features. For example, newly synthesized antigenic peptides and polypeptides can be further subjected to various modifications, such as conservative or non-conservative substitutions, which may provide certain advantages in their use, such as improved MHC binding, stability, or presentation. A conservative substitution means replacing an amino acid residue with another that is biologically and / or chemically similar, for example, one hydrophobic residue with another hydrophobic residue, or one polar residue with another polar residue. Substitutions include combinations such as Gly, Ala; Val, Ile, Leu, Met; Asp, Glu; Asn, Gln; Ser, Thr; Lys, Arg; and Phe, Tyr. The effects of single amino acid substitutions may also be explored using D-amino acids. Such modifications can be carried out using well-known peptide synthesis procedures, for example, as described in Merrifield, Science 232:341-347 (1986), Barany & Merrifield, The Peptides, Gross & Meienhofer, eds. (NY, Academic Press), pp.1-284 (1979); and Stewart & Young, Solid Phase Peptide Synthesis, (Rockford, Ill., Pierce), 2d Ed. (1984).

[0122] Modification of peptides and polypeptides with various amino acid mimes or non-natural amino acids can be particularly useful in increasing the in vivo stability of peptides and polypeptides. Stability can be assayed in many ways. For example, peptidases, as well as various biological media such as human plasma and serum, have been used to test stability. See, for example, Verhoef et al., Eur. J. Drug Metab Pharmacokin. 11:291-302 (1986). The half-life of peptides can be conveniently determined using a 25% human serum (v / v) assay. The protocol is generally as follows: Pooled human serum (type AB, non-thermally inactivated) is degreased by centrifugation before use. The serum is then diluted to 25% in RPMI tissue culture medium and used to test peptide stability. At predetermined time intervals, small amounts of the reaction solution are taken and added to either 6% aqueous trichloroacetic acid or ethanol. The turbid reaction sample is cooled for 15 minutes (4°C), and then spun to precipitate the serum protein. The presence of the peptide is then determined by reverse-phase HPLC using stability-specific chromatography conditions.

[0123] Peptides and polypeptides can be modified to provide desirable attributes other than improved serum half-life. For example, the ability of a peptide to induce CTL activity can be enhanced by ligation to a sequence containing at least one epitope capable of inducing a T helper cell response. Immunogenic peptide / T helper conjugates can be ligated by spacer molecules. Spacers typically consist of relatively small neutral molecules, such as amino acids or amino acid mimes, which are substantially uncharged under physiological conditions. Spacers are typically selected from, for example, Ala, Gly, or other neutral spacers of nonpolar or neutral polar amino acids. It will be understood that optionally present spacers do not need to consist of the same residues and can therefore be heterooligomers or homooligomers. If present, spacers will usually consist of at least one or two residues, more typically three to six residues. Alternatively, peptides can be ligated to T helper peptides without spacers.

[0124] The newly synthesized antigenic peptide can be linked to a T helper peptide either directly or via a spacer at either the amino or carboxyl terminus of the peptide. The amino terminus of either the newly synthesized antigenic peptide or the T helper peptide can be acylated. Exemplary T helper peptides include 830-843 of tetanus toxin, 307-319 of influenza, and 382-398 and 378-389 of malaria sporozoite.

[0125] Proteins or peptides can be prepared by any technique known to those skilled in the art, including the expression of proteins, polypeptides, or peptides through standard molecular biological techniques, the isolation of proteins or peptides from natural sources, or the chemical synthesis of proteins or peptides. Nucleotide sequences and protein, polypeptide, and peptide sequences corresponding to various genes have been previously disclosed and can be found in computerized databases known to those skilled in the art. One such database is the Genbank and GenPept databases of the National Center for Biotechnology Information, located on the website of the National Institutes of Health. The coding regions of known genes can be amplified and / or expressed using the techniques disclosed herein or as known to those skilled in the art. Alternatively, various commercial preparations of proteins, polypeptides, and peptides are known to those skilled in the art.

[0126] In a further embodiment, the nascent antigen comprises a nucleic acid (e.g., polynucleotide) encoding a nascent antigenic peptide or a portion thereof. The polynucleotide can be single-stranded and / or double-stranded, in its native or stabilized form, or a combination thereof, such as DNA, cDNA, PNA, CNA, RNA (e.g., mRNA), or polynucleotides having a phosphorothioate backbone, and may or may not contain introns. Further embodiments provide an expression vector capable of expressing a polypeptide or a portion thereof. Expression vectors for various cell types are well known in the art and can be selected without excessive experimentation. Generally, DNA is inserted into an expression vector, such as a plasmid, in the correct orientation and with the correct reading frame for expression. If necessary, the DNA can be ligated to a suitable transcriptional and translational regulatory nucleotide sequence recognized by the desired host, although such regulation is generally available in the expression vector. The vector is then introduced into the host through standard techniques. A guide can be found, for example, in Sambrook et al. (1989) Molecular Cloning, A Laboratory Manual, Cold Spring Harbor Laboratory, Cold Spring Harbor, NY.

[0127] IV. Vaccine Composition Also disclosed herein are immunogenic compositions, such as vaccine compositions, that can induce specific immune responses, such as tumor-specific immune responses. Vaccine compositions typically comprise a number of nascent antigens selected, for example, using the methods described herein. Vaccine compositions may also be referred to as vaccines.

[0128] The vaccine may contain 1 to 30 different peptides, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, or 30 different peptides, 6, 7, 8, 9, 10, 11, 12, 13, or 14 different peptides, or 12, 13, or 14 different peptides. The peptides may include post-translational modifications. Vaccines contain 1 to 100 or more nucleotide sequences, such as 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, It may contain 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100 or more different nucleotide sequences, 6, 7, 8, 9, 10, 11, 12, 13, or 14 different nucleotide sequences, or 12, 13, or 14 different nucleotide sequences.The vaccine contains 1 to 30 types of new antigen sequences, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 6 It may contain 6, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100 or more different nascent antigen sequences, 6, 7, 8, 9, 10, 11, 12, 13, or 14 different nascent antigen sequences, or 12, 13, or 14 different nascent antigen sequences.

[0129] In one embodiment, different peptides and / or polypeptides, or encoding nucleotide sequences, are selected so that the peptides and / or polypeptides can bind to different MHC molecules, such as different MHC class I molecules and / or different MHC class II molecules. In some embodiments, a single vaccine composition contains an encoding sequence of a peptide and / or polypeptide that can bind to the most frequently present MHC class I molecules and / or MHC class II molecules. Thus, a vaccine composition may contain different fragments that can bind to at least two preferred, at least three preferred, or at least four preferred MHC class I molecules and / or MHC class II molecules.

[0130] The vaccine composition can induce a specific cytotoxic T cell response and / or a specific helper T cell response.

[0131] The vaccine composition may further comprise an adjuvant and / or carrier. Examples of useful adjuvants and carriers are given below in this specification. The composition may be conjugated to a carrier, such as a protein, or to an antigen-presenting cell, such as a dendritic cell (DC) capable of presenting peptides to T cells.

[0132] An adjuvant is any substance whose mixing into a vaccine composition enhances or otherwise modifies the immune response to a nascent antigen. The carrier can be a scaffold structure to which the nascent antigen can bind, such as a polypeptide or polysaccharide. Optionally, the adjuvant can be conjugated covalently or non-covalently.

[0133] The ability of adjuvants to enhance the immune response to an antigen is typically demonstrated by a significant or substantial increase in the immune-mediated response or a reduction in disease symptoms. For example, an increase in humoral immunity is typically demonstrated by a significant increase in the titer of antibodies produced against the antigen, and an increase in T cell activity is typically demonstrated by increased cell proliferation, cellular cytotoxicity, or cytokine secretion. Adjuvants can also alter the immune response, for example, by changing a primarily humoral or Th response to a primarily cellular or Th response.

[0134] Suitable adjuvants include 1018 ISS, alum, aluminum salt, Amplivax, AS15, BCG, CP-870, 893, CpG7909, CyaA, dSLIM, GM-CSF, IC30, IC31, imiquimod, ImuFact IMP321, IS Patch, ISS, ISCOMATRIX, JuvImmune, LipoVac, MF59, monophosphoryl lipid A, Montanide IMS 1312, Montanide ISA206, Montanide ISA 50V, and Montanide This includes, but is not limited to, ISA-51, OK-432, OM-174, OM-197-MP-EC, ONTAK, PepTel vector systems, PLG microparticles, reciquimod, SRL172, virosoms and other virus-like particles, YF-17D, VEGF traps, R848, β-glucan, Pam3Cys, Aquila's QS21 stimulon (Aquila Biotech, Worcester, Mass., USA) derived from saponins, mycobacterial extracts and synthetic bacterial cell wall mimics, and other proprietary adjuvants such as Ribi's Detox.Quil or Superfos. Adjuvants such as incomplete Freund's or GM-CSF are also useful. Several dendritic cell-specific immunological adjuvants (e.g., MF59) and their preparations have been previously described (Dupuis M, et al., Cell Immunol. 1998;186(1):18-27; Allison AC; Dev Biol Stand. 1998;92:3-11). Cytokines can also be used. Some cytokines are directly linked to their effects on dendritic cell migration to lymphoid tissue (e.g., TNF-α), to accelerating the maturation of dendritic cells into efficient antigen-presenting cells for T lymphocytes (e.g., GM-CSF, IL-1, and IL-4) (specifically, U.S. Patent No. 5,849,589, which is incorporated herein by reference in its entirety), and to their action as immunoadjuvants (e.g., IL-12) (Gabrilovich DI, et al., J ImmunotherEmphasis Tumor Immunol. 1996(6):414-418).

[0135] CpG immunostimulatory oligonucleotides have also been reported to enhance the adjuvant effect in vaccine settings. Other TLR-binding molecules, such as RNAs that bind to TLR 7, TLR 8, and / or TLR 9, may also be used.

[0136] Other examples of useful adjuvants include, but are not limited to, chemically modified CpG (e.g., CpR, Idera), Poly(I:C) (e.g., polyi:CI2U), non-CpG bacterial DNA or RNA, as well as immunoactive small molecules and antibodies such as cyclophosphamide, sunitinib, bevacizumab, Celebrex, NCX-4016, sildenafil, tadalafil, vardenafil, sorafinib, XL-999, CP-547632, pazopanib, ZD2171, AZD2171, ipilimumab, tremelimumab, and SC58175, which may act therapeutically and / or as adjuvants. The amount and concentration of adjuvants and additives can be readily determined by those skilled in the art without excessive experimentation. Additional adjuvants include colony-stimulating factors such as granulocyte-macrophage colony-stimulating factor (GM-CSF, salglamostim).

[0137] The vaccine composition may contain more than one different adjuvant. Furthermore, the therapeutic composition may contain any adjuvant substance, including any or a combination thereof. It is also intended that the vaccine and adjuvant may be administered together or separately in any suitable sequence.

[0138] The carrier (or excipient) can exist independently of the adjuvant. The function of the carrier may be, for example, to increase activity or immunogenicity, to provide stability, to increase biological activity, or to increase serum half-life, particularly by increasing the molecular weight of the variant. Furthermore, the carrier can assist in the presentation of peptides to T cells. The carrier can be any suitable carrier known to those skilled in the art, e.g., a protein or antigen-presenting cell. Carrier proteins may be, but are not limited to, serum proteins such as keyhole limpet hemocyanin, transferrin, bovine serum albumin, human serum albumin, thyroglobulin or ovalbumin, immunoglobulins, or hormones such as insulin, or palmitic acid. For human immunization, the carrier is generally a physiologically acceptable carrier that is tolerable and safe for humans. However, tetanus toxoid and / or diphtheria toxoid are suitable carriers. Alternatively, the carrier may be dextran, e.g., cepharose.

[0139] Cytotoxic T cells (CTLs) recognize antigens in the form of peptides bound to MHC molecules rather than the intact foreign antigen itself. The MHC molecules themselves are located on the cell surface of antigen-presenting cells. Therefore, activation of CTLs is possible when a trimer complex of peptide antigen, MHC molecule, and APC is present. Correspondingly, the immune response can be enhanced not only when the peptide is used for CTL activation, but also when APCs, each containing the respective MHC molecule, are added. Therefore, in some embodiments, the vaccine composition additionally contains at least one antigen-presenting cell.

[0140] Newly generated antigens also include, but are not limited to, vaccinia, fowlpox, self-replicating alphavirus, marabavirus, adenovirus (see, e.g., Tatsis et al., Adenoviruses, Molecular Therapy (2004) 10, 616-629), or second, third, or hybrid second / third generation lentiviruses, and recombinant lentiviruses of any generation designed to target specific cell types or receptors (e.g., Hu et al., Immunization Delivered by Lentiviral Vectors for Cancer and Infectious Diseases, Immunol Rev. (2011) 239(1):45-61, Sakuma et al., Lentiviral vectors: basic translational, Biochem J. (2012) 443(3):603-18, Cooper et al., Rescue of splicing-mediated intron loss maximizes expression in lentiviral vectors containing the human It can also be incorporated into viral vector-based vaccine platforms, such as ubiquitin C promoter, Nucl. AcidsRes. (2015) 43(1):682-690, and Zufferey et al., Self-Inactivating Lentivirus Vector for Safe and Efficient In Vivo Gene Delivery, J. Virol. (1998) 72(12):9873-9880). Depending on the packaging capabilities of the aforementioned viral vector-based vaccine platforms, this approach can deliver one or more nucleotide sequences encoding one or more nascent antigen peptides.The sequence may be adjacent to a non-mutant sequence, separated by a linker, or preceded by one or more sequences targeting an intracellular compartment (see, for example, Gros et al., Prospective identification of neoantigen-specific lymphocytes in the peripheral blood of melanoma patients, Nat Med. (2016) 22(4):433-8, Stronen et al., Targeting of cancer neoantigens with donor-derived T cell receptor repertoires, Science. (2016) 352(6291):1337-41, and Lu et al., Efficient identification of mutated cancer antigens recognized by T cells associated with durable tumor regressions, Clin Cancer Res. (2014) 20(13):3401-10). Upon introduction into the host, infected cells express the neoantigen, thereby eliciting a host immune response (e.g., CTLs) to the peptide. Useful vaccinia vectors and methods in immunization protocols are described, for example, in U.S. Patent No. 4,722,848. Another vector is BCG (Bacillus calmette-Guérin). The BCG vector is described in Stover et al. (Nature 351:456-460 (1991)). A wide variety of other vaccine vectors useful for the therapeutic administration or immunization of nascent antigens, such as Salmonella typhi vectors, will be apparent to those skilled in the art from the description herein.

[0141] IV.A. Further Considerations for Vaccine Design and Manufacturing IV.A.1. Determination of the set of peptides covering all tumor subclones Truncal peptides, meaning those presented by all or most tumor subclones, are preferred for inclusion in vaccines. 53 Optionally, if there are no truncal peptides that are likely to be presented and are predicted to be immunogenic, or if the number of truncal peptides that are likely to be presented and are predicted to be immunogenic is small enough that additional non-truncal peptides can be included in the vaccine, further peptides can be prioritized by estimating the number and identity of tumor subclones and selecting peptides to maximize the number of tumor subclones covered by the vaccine. 54 .

[0142] IV.A.2. Prioritization of newly generated antigens Even after applying all of the above neonatal antigen filters, there may still be more candidate neonatal antigens available for vaccine inclusion than the vaccine technology can accommodate. Additionally, uncertainties may remain regarding various aspects of neonatal antigen analysis, and trade-offs may exist between the various properties of candidate vaccine neonatal antigens. Therefore, instead of predetermined filters at each stage of the selection process, we can consider an integral multidimensional model that places candidate neonatal antigens in a space with at least the following axes and optimizes selection using an integral approach. 1. Risk of autoimmunity or tolerance (germline risk) (lower autoimmunity risk is typically preferable) 2. Probability of sequencing artifacts (a lower probability of artifacts is generally preferable) 3. Probability of immunogenicity (a higher probability of immunogenicity is generally preferable) 4. Probability of presentation (a higher probability of presentation is generally preferable) 5. Gene expression (higher expression is generally preferable) 6. HLA gene coverage (a greater number of HLA molecules involved in the presentation of a set of nascent antigens may reduce the tumor's chances of evading immune attack through downregulation or mutation of HLA molecules). 7. HLA class coverage (covering both HLA-I and HLA-II may increase the probability of treatment response and decrease the probability of tumor immune evasion).

[0143] V. Treatment and Manufacturing Methods Also provided is a method for inducing a tumor-specific immune response in a target, thereby vaccinating the tumor and treating and / or alleviating the symptoms of the target cancer, by administering one or more neogenic antigens, such as multiple neogenic antigens identified using the methods disclosed herein, to the target.

[0144] In some embodiments, subjects are diagnosed with cancer or are at risk of developing cancer. Subjects may be humans, dogs, cats, horses, or any animal for which a tumor-specific immune response is desirable. Tumors may be any solid tumors, such as tumors of the mammary gland, ovaries, prostate, lungs, kidneys, stomach, colon, testes, head and neck, pancreas, brain, melanoma, and other tissue organs, as well as hematological malignancies, such as lymphomas and leukemias, including acute myeloid leukemia, chronic myeloid leukemia, chronic lymphocytic leukemia, T-cell lymphocytic leukemia, and B-cell lymphoma.

[0145] The newly synthesized antigen can be administered in a sufficient amount to induce a CTL response.

[0146] The newly generated antigen can be administered alone or in combination with other therapeutic substances. These therapeutic substances may include, for example, chemotherapy agents, radiation, or immunotherapy. Any suitable therapeutic treatment for a particular cancer may be administered.

[0147] In addition, the subjects may be further administered immunosuppressive / immunostimulatory substances such as checkpoint inhibitors. For example, the subjects may be further administered anti-CTLA antibodies or anti-PD-1 or anti-PD-L1 antibodies. Blocking CTLA-4 or PD-L1 with antibodies can enhance the immune response against cancer cells in patients. In particular, CTLA-4 blockade has been shown to be effective when a vaccination protocol is adopted.

[0148] The optimal amount of each nascent antigen to be included in the vaccine composition, and the optimal dosing regimen, can be determined. For example, the nascent antigen or its variants can be prepared for intravenous (iv) injection, subcutaneous (sc) injection, intradermal (id) injection, intraperitoneal (ip) injection, and intramuscular (im) injection. Methods of injection include sc, id, ip, im, and iv. Methods of DNA or RNA injection include id, im, sc, ip, and iv. Other methods of administering the vaccine composition are known to those skilled in the art.

[0149] Vaccines can be edited so that the selection, number, and / or amount of nascent antigens present in the composition are specific to the tissue, cancer, and / or patient. For example, the precise selection of peptides may be guided by the expression pattern of the parent protein in a given tissue. The selection may depend on the specific type of cancer, the disease state, earlier treatment regimens, the patient's immune status, and, of course, the patient's HLA halotype. Furthermore, vaccines may contain individualized components according to the personal needs of a particular patient. Examples include altering the selection of nascent antigens according to the expression of nascent antigens in a particular patient, or making adjustments for secondary treatments after the first round or scheme of treatment.

[0150] With respect to compositions to be used as vaccines for cancer, neoplastic antigens having similar normal self-peptides that are expressed in large quantities in normal tissues may be avoided or present in small quantities in the compositions described herein. On the other hand, if it is known that a patient's tumor expresses a particular neoplastic antigen in large quantities, each pharmaceutical composition for the treatment of this cancer may be present in large quantities and / or may contain more than one neoplastic antigen specific to this particular neoplastic antigen or the pathway of this neoplastic antigen.

[0151] Compositions containing neoplastic antigens can be administered to individuals already suffering from cancer. In therapeutic applications, the composition is administered to the patient in an amount sufficient to induce an effective CTL response to the tumor antigen and to cure or at least partially cessate the symptoms and / or complications. The amount appropriate to achieve this is defined as the "therapeutic effective dose." The effective dose for this use will depend, for example, on the composition, the mode of administration, the stage and severity of the disease being treated, the patient's weight and overall health condition, and the judgment of the prescribing physician. It should be kept in mind that the composition is generally not for use in serious disease conditions, i.e., life-threatening or potentially life-threatening situations, especially when the cancer has metastasized. In such cases, given the minimization of exogenous substances and the relatively non-toxic nature of neoplastic antigens, it may be possible and desirable for the treating physician to administer substantially excessive amounts of these compositions.

[0152] For therapeutic use, administration can be initiated at the time of tumor detection or surgical removal. This is followed by a boost dose until the symptoms have substantially subsided, and then for a period thereafter.

[0153] Pharmaceutical compositions for therapeutic treatment (e.g., vaccine compositions) are intended for parenteral, topical, nasal, oral, or local administration. Pharmaceutical compositions can be administered parenterally, for example, intravenously, subcutaneously, intradermally, or intramuscularly. Compositions can be administered to the site of surgical excision to induce a local immune response against a tumor. A composition for parenteral administration comprising a solution of a nascent antigen is disclosed herein, and the vaccine composition is dissolved or suspended in an acceptable carrier, such as an aqueous carrier. Various aqueous carriers can be used, such as water, buffer water, 0.9% saline, 0.3% glycine, hyaluronic acid, etc. These compositions can be sterilized by conventional, well-known sterilization techniques or by sterile filtration. The resulting aqueous solution can be packaged for use as is or lyophilized, and the lyophilized preparation is combined with a sterile solution before administration. The composition may contain pharmaceutically acceptable auxiliary substances necessary to approximate physiological conditions, such as pH adjusters, buffers, isotonic agents, and wetting agents, for example, sodium acetate, sodium lactate, sodium chloride, potassium chloride, calcium chloride, sorbitan monolaurate, and triethanolamine oleate.

[0154] Newly synthesized antigens can also be administered via liposomes, targeting them to specific cell tissues such as lymphoid tissue. Liposomes are also useful for increasing half-life. Liposomes include emulsions, foams, micelles, insoluble monolayers, liquid crystals, phospholipid dispersions, lamellar layers, etc. In these preparations, the newly synthesized antigen to be delivered is incorporated either alone or as part of a liposome, together with a molecule that binds to a receptor dominant among lymphoid cells, such as a monoclonal antibody that binds to the CD45 antigen, or other therapeutic or immunogenic compositions. Thus, liposomes filled with the desired newly synthesized antigen can be directed to a site on lymphoid cells, where the liposome then delivers the selected therapeutic / immunogenic composition. Liposomes can generally be formed from standard vesicle-forming lipids, including neutral and negatively charged phospholipids and sterols such as cholesterol. Lipid selection is generally guided by considerations such as liposome size, acid instability, and liposome stability in blood flow. For example, various methods are available for preparing liposomes, as described in Szoka et al., Ann. Rev. Biophys. Bioeng. 9;467 (1980), U.S. Patents No. 4,235,871, No. 4,501,728, No. 4,501,728, No. 4,837,028, and No. 5,019,369.

[0155] For targeting immune cells, the ligands to be incorporated into the liposomes may include, for example, antibodies or fragments thereof that are specific to the cell surface determinants of desired immune system cells. The liposome suspension can be administered intravenously, topically, or locally, in doses that vary depending, among other things, on the mode of administration, the peptides delivered, and the stage of the disease being treated.

[0156] For therapeutic or immunization purposes, the peptides described herein, and optionally nucleic acids encoding one or more peptides, may also be administered to patients. Numerous methods are conveniently used to deliver nucleic acids to patients. For example, nucleic acids can be delivered directly as "naked DNA." This approach is described, for example, in Wolff et al., Science 247:1465-1468 (1990), and in U.S. Patents 5,580,859 and 5,589,466. Nucleic acids can also be administered using ballistic delivery, for example, in U.S. Patent 5,204,253. Simply a particle consisting of DNA can be administered. Alternatively, DNA can be attached to particles such as gold particles. Approaches for delivering nucleic acid sequences may include viral vectors, mRNA vectors, and DNA vectors, with or without electroporation.

[0157] Nucleic acids can also be delivered by complexing them with cationic compounds such as cationic lipids. Lipid-mediated gene delivery methods are described, for example, in 9618372WOAWO 96 / 18372;9324640WOAWO 93 / 24640;Mannino & Gould-Fogerite, BioTechniques 6(7): 682-691 (1988);U.S. Patent No. 5,279,833 Rose;9106309WOAWO 91 / 06309; and Felgner et al., Proc.Natl.Acad.Sci.USA 84: 7413-7414 (1987).

[0158] Newly generated antigens also include, but are not limited to, vaccinia, fowlpox, self-replicating alphavirus, marabavirus, adenovirus (see, e.g., Tatsis et al., Adenoviruses, Molecular Therapy (2004) 10, 616-629), or second, third, or hybrid second / third generation lentiviruses, and recombinant lentiviruses of any generation designed to target specific cell types or receptors (see, e.g., Hu et al., Immunization Delivered by Lentiviral Vectors for Cancer and Infectious Diseases, Immunol Rev. (2011) 239(1): 45-61, Sakuma et al., Lentiviral vectors: basic translational, Biochem J. (2012) 443(3):603-18, Cooper et al., Rescue of splicing-mediated intron loss maximizes expression in lentiviral vectors containing the human It can also be incorporated into viral vector-based vaccine platforms, such as ubiquitin C promoter, Nucl. Acids Res. (2015) 43 (1): 682-690, and Zufferey et al., Self-Inactivating Lentivirus Vector for Safe and Efficient In Vivo Gene Delivery, J. Virol. (1998) 72 (12): 9873-9880). Depending on the packaging capabilities of the aforementioned viral vector-based vaccine platforms, this approach can deliver one or more nucleotide sequences encoding one or more nascent antigen peptides.The sequence may be adjacent to a non-mutant sequence, separated by a linker, or preceded by one or more sequences targeting an intracellular compartment (see, for example, Gros et al., Prospective identification of neoantigen-specific lymphocytes in the peripheral blood of melanoma patients, Nat Med. (2016) 22 (4):433-8, Stronen et al., Targeting of cancer neoantigens with donor-derived T cell receptor repertoires, Science. (2016) 352 (6291):1337-41, and Lu et al., Efficient identification of mutated cancer antigens recognized by T cells associated with durable tumor regressions, Clin Cancer Res. (2014) 20 (13):3401-10). Upon introduction into the host, infected cells express the neoantigen, thereby eliciting a host immune response (e.g., CTLs) to the peptide. Useful vaccinia vectors and methods in immunization protocols are described, for example, in U.S. Patent No. 4,722,848. Another vector is BCG (Bacillus calmette-Guérin). The BCG vector is described in Stover et al. (Nature 351:456-460 (1991)). A wide variety of other vaccine vectors useful for the therapeutic administration or immunization of nascent antigens, such as Salmonella typhi vectors, will be apparent to those skilled in the art from the description herein.

[0159] The means of administering nucleic acids utilize minigene constructs encoding one or more epitopes. The amino acid sequences of the epitopes are back-translated to create DNA sequences (minigenes) encoding selected CTL epitopes for expression in human cells. Human codon frequency tables are used to guide codon selection for each amino acid. These epitope-encoding DNA sequences are directly adjacent to each other to create a continuous polypeptide sequence. Additional elements can be incorporated during minigene design to optimize expression and / or immunogenicity. Examples of amino acid sequences that can be back-translated and included in the minigene sequence include helper T lymphocyte epitopes, leader (signal) sequences, and endoplasmic reticulum retention signals. In addition, MHC presentation of CTL epitopes can be improved by including synthetic (e.g., polyalanine) or naturally occurring flanking sequences adjacent to the CTL epitopes. The minigene sequences are converted to DNA by assembling oligonucleotides encoding the positive and negative strands of the minigene. Overlapping oligonucleotides (30–100 nucleotides in length) are synthesized, phosphorylated, purified, and annealed under appropriate conditions using known techniques. The ends of the oligonucleotides are ligated using T4 DNA ligase. This synthetic minigene encoding a CTL epitope polypeptide can then be cloned into a desired expression vector.

[0160] Purified plasmid DNA can be prepared for injection using a variety of formulations. The simplest of these is the reconstitution of lyophilized DNA in sterile phosphate-buffered saline (PBS). Various methods have been described, and new techniques may become available. As mentioned above, nucleic acids are conveniently formulated with cationic lipids. In addition, glycolipids, fusion liposomes, peptides, and compounds collectively called protective, interacting, and non-condensing (PINC) compounds can also be complexed with purified plasmid DNA to influence variables such as stability, intramuscular dispersion, or transport to specific organs or cell types.

[0161] Also disclosed herein are methods for producing an oncology vaccine, which include the steps of the method disclosed herein and the steps of producing an oncology vaccine comprising a number of neogenic antigens or a subset of a number of neogenic antigens.

[0162] The nascent antigens disclosed herein can be produced using methods known in the art. For example, a method for producing a nascent antigen or vector disclosed herein (e.g., a vector comprising at least one sequence encoding one or more nascent antigens) may include the steps of culturing host cells under conditions suitable for expressing the nascent antigen or vector, wherein the host cells comprise at least one polynucleotide encoding the nascent antigen or vector, and purifying the nascent antigen or vector. Standard purification methods include chromatography, electrophoresis, immunology, precipitation, dialysis, filtration, concentration, and chromatofocusing techniques.

[0163] The host cells may include Chinese hamster ovary (CHO) cells, NS0 cells, yeast, or HEK293 cells. The host cells may be transformed with one or more polynucleotides comprising at least one nucleic acid sequence encoding the nascent antigen or vector disclosed herein, and optionally, the isolated polynucleotide further comprises a promoter sequence functionally ligated to at least one nucleic acid sequence encoding the nascent antigen or vector. In certain embodiments, the isolated polynucleotide may be a cDNA.

[0164] VI. Identification of newly synthesized antigens VI.A. Identification of new antigen candidates Research methods for NGS analysis of tumor and normal exomes and transcriptomes are described and applied to specific spaces of nascent antigens. 6,14,15The following examples consider certain optimizations for greater sensitivity and specificity in identifying nascent antigens in a clinical setting. These optimizations can be grouped into two areas: those related to laboratory processes and those related to NGS data analysis.

[0165] VI.A.1. Optimization of Laboratory Processes The process improvements presented herein are concepts developed for the reliable evaluation of cancer driver genes in targeted cancer panels. 16 By expanding this to the whole exome and whole transcriptome settings necessary for identifying nascent antigens, we address the challenges in accurately discovering nascent antigens from clinical specimens with low tumor content and small volumes. Specifically, these improvements include: 1. Targeting deep (greater than 500x) intrinsic mean coverage across tumor exomes to detect mutations present at low variant allele frequencies due to either low tumor content or subclonal status. 2. To minimize the chances of missing potential nascent antigens, the base coverage at less than 100x should be less than 5%. For example, a. Use of DNA-based capture probes with individual probe QCs. 17 b. Inclusion of additional bait for areas not adequately covered 3. Targeting with uniform coverage across normal exomes, where less than 5% of bases are covered at less than 20x, so that the chances of potential nascent antigens remaining unclassified for somatic / germline status (and therefore unusable as TSNAs) are minimized. 4. To minimize the total amount of sequencing required, sequence capture probes are designed only for the coding region of a gene, since non-coding RNA cannot generate nascent antigens. Additional optimizations include: a. Supplemental probes for HLA genes that are GC-rich and not adequately captured by standard exome sequencing. 18. b. Elimination of genes that are predicted to generate little to no candidate nascent antigens due to factors such as insufficient expression, suboptimal proteasome digestion, or unusual sequence characteristics. 5. Tumor RNA is also sequenced at high depth (greater than 100M reads) to enable mutation detection, quantification of gene and splice variant ("isoform") expression, and fusion detection. RNA from FFPE samples is probe-based enriched with the same or similar probes used to capture exomes in DNA. 19 It is extracted using this method.

[0166] VI.A.2. Optimization of NGS Data Analysis Improvements to the analytical methods address the suboptimal sensitivity and specificity of common research variant calling approaches, specifically considering the customization relevant to identifying nascent antigens in clinical settings. These include: 1. Use of the HG38 reference human genome or a later version for alignment (because it contains multiple MHC region assemblies that better reflect population polymorphisms, in contrast to earlier genome releases). 2. Various programs 5 Overcoming the limitations of a single variant caller 20 by merging results from multiple sources. a. Single nucleotide mutations and insertions / deletions are detected from tumor DNA, tumor RNA, and normal DNA using a set of tools including: Strelka 21 and Mutec t22 Programs based on the comparison of tumor and normal DNA, and which are particularly advantageous in low-purity samples. 23 Programs such as UNCeqR that incorporate tumor DNA, tumor RNA, and normal DNA. b. Insertions and deletions are Strelka and ABRA 24 This is determined by programs that perform local reassembly, such as the one described above. c. Structural reorganization is Pindel 25 or Breakseq 26This is determined using specialized tools such as those listed above. 3. To detect and prevent sample swapping, mutation calls from samples of the same patient are compared at a selected number of polymorphic sites. 4. For example, extensive filtering of artificial calls is performed as follows: a. Potentially, removal of mutations found in normal DNA, using lenient detection parameters in cases of low coverage and tolerant proximity criteria in cases of insertions and deletions. b. Removal of mutations due to poor mapping quality or poor base quality 27 . c. Removal of mutations resulting from reappearing sequencing artifacts, even if they are not observed in the corresponding normal state. 27 Examples include mutations detected primarily on a single strand. d. Removal of mutations detected in the set of unrelated controls. 27 . 5. seq2HLA 28 , ATHLATES 29 , or use one of the Optitypes, and also combine exome and RNA sequencing data. 28 Accurate HLA calling from normal exomes. Additional potential optimizations include employing dedicated assays for HLA typing, such as long-read DNA sequencing. 30 or adaptation of a method for linking RNA fragments to maintain continuity. 31 Includes. 6. Robust detection of neonatal ORFs arising from tumor-specific splice variants is possible with CLASS 32 Bayesembler 33 StringTie 34 This is done by assembling transcripts from RNA-seq data using a similar program in its reference guide mode (i.e., using known transcript structures rather than attempting to recreate their entire transcripts from each experiment). 35However, although commonly used for this purpose, it frequently produces an unbelievable number of splice variants, many of which are much shorter than the full-length gene and may not be able to recover a simple positive control. The coding sequence and potential nonsense mutation-dependent degradation mechanism leads to the reintroduction of the variant sequence, SpliceR 36 and MAMBA 37 Gene expression is determined by tools such as Cufflinks. 35 Alternatively, it is determined by tools such as Express (Roberts and Pachter, 2013). Wild-type and mutant-specific expression counts and / or relative levels are determined by ASE. 38 or HTSeq 39 These are determined by tools developed for these purposes, such as: a. Removal of candidate nascent ORFs that are thought to be insufficiently expressed. b. Removal of candidate nascent ORFs that are predicted to trigger nonsense mutation-dependent degradation mechanisms (NMD). 7. Candidate neonatal antigens observed only in RNA (e.g., neonatal ORFs) that cannot be directly verified as tumor-specific are classified as likely to be tumor-specific according to additional parameters, by considering, for example, the following: a. Presence of cis-acting frameshift or splice site mutation support in tumor DNA only. b. Confirmation of trans-acting mutations in tumor DNA only in splicing factors. For example, in three independently published experiments with the R625 mutant SF3B1, the gene exhibiting the most differential splicing was one experiment that examined uveal melanoma patients. 40 The second experiment examined uveal melanoma cell lines. 41 , and a third experiment examined breast cancer patients. 42 Nevertheless, they agreed. c. For novel splicing isoforms, confirmation of "novel" splice-junction reads in RNASeq data. d. For novel rearrangements, there is confirmation of exon-proximal reads in tumor DNA that are not present in normal DNA. e.GTEx 43 Absences from gene expression summaries such as those mentioned above (i.e., making germline origin less likely). 8. To directly avoid alignment and annotation-based errors and artifacts, complement reference genome alignment-based analysis by comparing tumor and normal reads (or k-mars derived from such reads) of assembled DNA (e.g., for somatic mutations occurring near germline mutations or repeat-context insertions / deletions).

[0167] In samples containing polyadenylated RNA, the presence of viral and microbial RNA in RNA-seq data is evaluated using RNA CoMPASS44 or a similar method to identify additional factors that may predict patient response.

[0168] VI.B. Isolation and Detection of HLA Peptides HLA peptide molecules were isolated using classical immunoprecipitation (IP) after dissolution and solubilization of tissue samples. 55~58 The clarified solution was used for HLA-specific IP.

[0169] Immunoprecipitation was performed using antibodies coupled to beads, where the antibody is specific to the HLA molecule. For pan-class I HLA immunoprecipitation, pan-class I CR antibodies were used, and for class II HLA-DR, HLA-DR antibodies were used. The antibodies were covalently attached to NHS-Sepharose beads during overnight incubation. After covalent attachment, the beads were washed and divided equally for IP. 59、60Immunoprecipitation can also be performed using antibodies that are not covalently bound to beads. Generally, this is done using Sepharose or magnetic beads coated with Protein A and / or Protein G to retain the antibody on the column. Several antibodies that can be used to selectively enrich MHC / peptide complexes are listed below. TIFF2026113718000002.tif42149

[0170] The clarified tissue lysate is added to antibody beads for immunoprecipitation. After immunoprecipitation, the beads are removed from the lysate, and the lysate is saved for additional experiments, including additional IP. Using standard techniques, the IP beads are washed to remove nonspecific binding and the HLA / peptide complex is eluted from the beads. Protein components are removed from the peptide using a molecular weight spin column or C18 fractionation. The resulting peptide is dried by SpeedVac evaporation and, in some cases, stored at -20°C before MS analysis.

[0171] The dried peptide was reconstituted in an HPLC buffer suitable for reverse-phase chromatography and loaded onto a C-18 microcapillary HPLC column for gradient elution in a Fusion Lumos mass spectrometer (Thermo). The peptide mass / charge (m / z) MS1 spectrum was acquired at high resolution using an Orbitrap detector, followed by a low-resolution MS2 scan acquired using an ion trap detector after HCD fragmentation of selected ions. Additionally, the MS2 spectrum could be acquired using CID or ETD fragmentation, or any combination of the three techniques to obtain greater amino acid coverage of the peptide. The MS2 spectrum could also be measured at high-resolution mass accuracy using an Orbitrap detector.

[0172] The MS2 spectra derived from each analysis were processed by Comet 61、62 Using this method, we searched a protein database and identified peptides using Percolator.63~65 Scoring is performed using PEAKS studio (Bioinformatics Solutions Inc.) and other search engines, or spectral matching and de novo sequencing. 75 A sequencing method including the following can be used.

[0173] VI.B.1. Study of MS detection limits for comprehensive HLA peptide sequencing Using the peptide YVYVADVAAK (SEQ ID NO:1), the detection limit was determined using various amounts of the peptide loaded onto an LC column. The amounts of peptide tested were 1 pmol, 100 fmol, 10 fmol, 1 fmol, and 100 amol (Table 1). The results are shown in Figure 1F. These results indicate that the minimum detection limit (LoD) is in the atmole range (10 -18 ) is located in the low femtomole range (10) and has a dynamic range of five orders of magnitude, and the signal-to-noise ratio is in the low femtomole range (10 -15 This indicates that it appears sufficient for sequencing.

[0174] TIFF2026113718000003.tif56128

[0175] VII. Presented Model VII.A. System Overview Figure 2A is an overview of an environment 100 for identifying the likelihood of peptide presentation in a patient, according to one embodiment. Environment 100 provides a context for introducing a presentation identification system 160, which itself includes a presentation information storage device 165.

[0176] The presentation identification system 160 is one or a computer model embodied in a computer computing system, as discussed below with respect to Figure 30, which receives peptide sequences associated with a set of MHC alleles and determines the likelihood that the peptide sequences will be presented by one or more of the associated MHC alleles. The presentation identification system 160 can be applied to both class I and class II MHC alleles. This is useful in a variety of contexts. One specific application example of the presentation identification system 160 is to receive nucleotide sequences of candidate neonatal antigens associated with a set of MHC alleles derived from tumor cells of patient 110 and determine the likelihood that the candidate neonatal antigens will be presented by one or more of the associated MHC alleles of the tumor and / or induce an immunogenic response in the immune system of patient 110. Those candidate neonatal antigens that have a high likelihood when determined by the system 160 can be selected for inclusion in vaccine 118, and such an antitumor immune response can be induced from the immune system of patient 110 providing the tumor cells. Furthermore, it is possible to create T cells with TCRs that are responsive to candidate nascent antigens with a high presentation likelihood for use in T cell therapy, thereby also inducing an antitumor immune response from the immune system of patients.

[0177] The presentation identification system 160 determines the presentation likelihood through one or more presentation models. Specifically, the presentation model generates the likelihood of whether a given peptide sequence is presented for a set of relevant MHC alleles, and the likelihood is generated based on presentation information stored in the memory device 165. For example, the presentation model may generate the likelihood of whether the peptide sequence "YVYVADVAAK (SEQ ID NO:1)" is presented for a set of alleles on the cell surface of a sample: HLA-A*02:01, HLA-A*03:01, HLA-B*07:02, HLA-B*08:03, and HLA-C*01:04. The presentation information 165 contains information about whether these peptides bind to various types of MHC alleles so that the peptides are presented by the MHC alleles, which is determined in the model according to the position of amino acids in the peptide sequence. Based on the presentation information 165, the presentation model can predict whether an unrecognized peptide sequence will be presented by binding to a relevant set of MHC alleles. As mentioned above, the presented model can be applied to both Class I and Class II MHC alleles.

[0178] VII.B. Presentation information Figure 2 illustrates a method for obtaining presentation information according to one embodiment. The presentation information 165 includes two general categories of information: allele interaction information and allele non-interaction information. Allele interaction information includes information that affects the presentation of the peptide sequence, depending on the type of MHC allele. Allele non-interaction information includes information that affects the presentation of the peptide sequence, independent of the type of MHC allele.

[0179] VII.B.1. Allele Interaction Information Allele interaction information primarily includes identified peptide sequences, known to be presented by one or more identified MHC molecules, such as those derived from humans or mice. Notably, this may or may not include data obtained from tumor samples. The presented peptide sequences may be identified from cells expressing a single MHC allele. In this example, the presented peptide sequences are generally collected from single-allelic cell lines that have been engineered to express a predetermined MHC allele and subsequently exposed to synthetic proteins. Peptides presented on the MHC allele are isolated by techniques such as acid elution and identified by mass spectrometry. Figure 2B shows an exemplary peptide presented on the predetermined MHC allele HLA-DRB1*12:01. This example shows TIFF2026113718000004.tif5128 being isolated and identified by mass spectrometry. In this situation, the direct association between the presented peptide and the MHC protein to which it is bound is critically known, since the peptide is identified through cells engineered to express a single predetermined MHC protein.

[0180] The presented peptide sequences may also be collected from cells expressing multiple MHC alleles. Typically in humans, six different types of MHCI molecules and up to twelve different types of MHCII molecules are expressed in cells. Such presented peptide sequences may be identified from multi-allelic cell lines engineered to express multiple predetermined MHC alleles. Such presented peptide sequences may also be identified from tissue samples, either normal tissue samples or tumor tissue samples. In particular, in this example, MHC molecules can be immunoprecipitated from normal or tumor tissue. Peptides presented on multiple MHC alleles can similarly be isolated by techniques such as acid elution and identified by mass spectrometry. Figure 2C shows six exemplary peptides. This example shows TIFF2026113718000005.tif21128, presented on identified class I MHC alleles HLA-A*01:01, HLA-A*02:01, HLA-B*07:02, HLA-B*08:01, and class II MHC alleles HLA-DRB1*10:01, HLA-DRB1:11:01, isolated, and identified by mass spectrometry. In contrast to single-allelic cell lines, the direct relationship between the presented peptide and the MHC protein to which it is bound may be unknown, as the bound peptide is isolated from the MHC molecule before identification.

[0181] Allele interaction information can also include mass spectrometry ion currents, which depend on both the concentration of the peptide-MHC molecular complex and the ionization efficiency of the peptide. Ionization efficiency varies from peptide to peptide in a sequence-dependent manner. Generally, ion efficiency varies by approximately two orders of magnitude from peptide to peptide, while the concentration of the peptide-MHC complex varies over an even larger range.

[0182] Allele interaction information may also include measured or predicted values ​​of the binding affinity between a given MHC allele and a given peptide. One or more affinity models can generate such predicted values ​​(72, 73, 74). For example, returning to the example shown in Figure 1D, the presented information 165 is the peptide YEMFNDKSF (SEQ ID NO: 3) and the class I allele HLA-A * This may include a predicted binding affinity of 1000 nM between 01:01. Peptides with IC50 > 1000 nm are only slightly presented by MHC, and lower IC50 values ​​increase the probability of presentation. Presentation information 165 may include a predicted binding affinity between peptide KNFLENFIESOFI (SEQ ID NO: 8) and class II allele HLA-DRB1:11:01.

[0183] Allele interaction information may also include measured or predicted values ​​of MHC complex stability. One or more stability models may generate such predictions. More stable peptide-MHC complexes (i.e., complexes with longer half-lives) are more likely to be presented at high copy numbers on tumor cells and on antigen-presenting cells encountering vaccine antigens. For example, returning to the example shown in Figure 2C, presentation information 165 may include a predicted stability value for the class I molecule HLA-A*01:01 with a half-life of 1 hour. Presentation information 165 may also include a predicted stability value for the half-life of the class II molecule HLA-DRB1:11:01.

[0184] Allele interaction information may also include the measured or predicted rate of peptide-MHC complex formation. Complexes that form at a faster rate are more likely to be presented on the cell surface at higher concentrations.

[0185] Allele interaction information can also include the sequence and length of the peptide. MHC class I molecules typically prefer to present peptides with a length of 8–15 peptide units. 60–80% of presented peptides have a length of 9 units. MHC class II molecules generally tend to present peptides with a length of 6–30 peptide units.

[0186] Allele interaction information may also include the presence of kinase sequence motifs on the nascent antigen-coding peptide, and the presence or absence of specific post-translational modifications on the nascent antigen-coding peptide. The presence of kinase motifs affects the probability of post-translational modifications that may enhance or interfere with MHC binding.

[0187] Allele interaction information may also include the expression or activity levels of proteins involved in post-translational modification processes, such as kinases, as measured or predicted by RNA sequencing, mass spectrometry, or other methods.

[0188] Allele interaction information may also include the probability of presentation of peptides with similar sequences in cells from other individuals expressing a particular MHC allele, as evaluated by mass spectrometry proteomics or other means.

[0189] Allele interaction information can also include the expression levels of specific MHC alleles in the individual in question (as measured, for example, by RNA-seq or mass spectrometry). Peptides that bind most strongly to MHC alleles expressed at high levels are more likely to be presented than peptides that bind most strongly to MHC alleles expressed at low levels.

[0190] Allele interaction information may also include the overall nascent antigen-coding peptide sequence-independent probability of presentation by a particular MHC allele in other individuals expressing that specific MHC allele.

[0191] Allele interaction information can also include the overall peptide sequence-independent probability of presentation by MHC alleles of molecules from the same family (e.g., HLA-A, HLA-B, HLA-C, HLA-DQ, HLA-DR, HLA-DP) in other individuals. For example, HLA-C molecules are typically expressed at lower levels than HLA-A or HLA-B molecules, and therefore, peptide presentation by HLA-C is a priori less likely than presentation by HLA-A or HLA-B II. As another example, since HLA-DP is generally expressed at lower levels than HLA-DR or HLA-DQ, it can be inferred that peptide presentation by HLA-DP is even less likely than presentation by HLA-DR or HLA-DQ.

[0192] Allele interaction information can also include the protein sequences of specific MHC alleles.

[0193] Any MHC allele non-interaction information listed in the section below can also be modeled as MHC allele interaction information.

[0194] VII.B.2. Allele Non-Interaction Information Allele non-interaction information can include the C-terminal sequence adjacent to the nascent antigen-coding peptide within its source protein sequence. In MHC-I, the C-terminal flanking sequence can influence the proteasome processing of the peptide. However, the C-terminal flanking sequence is cleaved from the peptide by the proteasome before the peptide is transported to the endoplasmic reticulum and encounters the MHC allele on the cell surface. As a result, the MHC molecule receives no information about the C-terminal flanking sequence, and therefore the effect of the C-terminal flanking sequence cannot vary depending on the MHC allele type. For example, returning to the example shown in Figure 2C, the presented information 165 may include the C-terminal flanking sequence FOEIFNDKSLDKFJI (SEQ ID NO: 9) of the presented peptide FJIEJFOESS (SEQ ID NO: 5), identified from the source protein of the peptide.

[0195] Allele-non-interaction information may also include mRNA quantification measurements. For example, mRNA quantification data can be obtained for the same sample that provides mass spectrometry training data. As will be described later with respect to Figure 13H, RNA expression has been identified as a strong predictor of peptide presentation. In one embodiment, mRNA quantification measurements are obtained from the software tool RSEM. A detailed execution of the RSEM software tool can be found in Bo Li and Colin N. Dewey. RSEM: accurate transcript quantification from RNA-Seq data with or without a reference genome. BMC Bioinformatics, 12:323, August 2011. In one embodiment, mRNA quantification is measured in units of fragments per kilobase (FPKM) of the transcript per million mapped reads.

[0196] Allele-non-interaction information may also include the N-terminal sequence adjacent to the peptide within its source protein sequence.

[0197] Allele-non-interaction information can also include the source genes of peptide sequences. Source genes can be defined as the Ensembl protein family of the peptide sequence. In other examples, source genes can be defined as the source DNA or source RNA of the peptide sequence. Source genes can be represented, for example, as a string of nucleotides encoding a protein, or, instead, in a more categorized form based on a named set of known DNA or RNA sequences known to encode a particular protein. In another example, allele-non-interaction information can also include a set of source transcripts or isoforms, or potential source transcripts or isoforms, of peptide sequences extracted from databases such as Ensembl or RefSeq.

[0198] Allele-non-interaction information may also include the tissue type, cell type, or tumor type of the cell from which the peptide sequence originates.

[0199] Allele-non-interaction information may also optionally include the presence of protease cleavage motifs in peptides, weighted according to the expression of the corresponding protease in tumor cells (as measured by RNA-seq or mass spectrometry). Peptides containing protease cleavage motifs are less likely to be presented because they are more readily degraded by proteases and therefore less stable within cells.

[0200] Allele non-interaction information can also include the turnover rate of the source protein when measured in the appropriate cell type. Faster turnover rates (i.e., shorter half-lives) increase the probability of presentation, but the predictive power of this property is low when measured in dissimilar cell types.

[0201] Allele-non-interaction information may also include the length of the source protein, optionally considering the specific splice variant ("isoform") most highly expressed in tumor cells, as measured by RNA-seq or proteomic mass spectrometry, or as predicted from annotations of germline or somatic splicing mutations detected in DNA or RNA sequence data.

[0202] Allele-non-interaction information may also include the expression levels of proteasomes, immunoproteasomes, thymic proteasomes, or other proteases in tumor cells (which can be measured by RNA-seq, proteome-mass spectrometry, or immunohistochemistry). Different proteasomes have different cleavage site preferences. A greater weight is given to each type of proteasome's cleavage preference, proportional to its expression level.

[0203] Allele-non-interaction information may also include the expression of the peptide's source gene (as measured, for example, by RNA-seq or mass spectrometry). Possible optimizations include adjusting the measured expression to account for the presence of stromal cells and tumor-infiltrating lymphocytes in the tumor sample. Peptides derived from genes with higher expression are more likely to be presented. Peptides derived from genes with undetectable levels of expression can be excluded from consideration.

[0204] Allele-non-interaction information can also include the probability that the source mRNA of the nascent antigen-coding peptide will be subjected to a nonsense mutation-dependent degradation mechanism, such as the one predicted by models of nonsense mutation-dependent degradation mechanisms, e.g., the model from Rivas et al, Science 2015.

[0205] Allele-non-interaction information can also include typical tissue-specific expression of peptide source genes during various stages of the cell cycle. Genes that are expressed at generally low levels (as measured by RNA-seq or sample analysis proteomics) but are known to be expressed at high levels during specific stages of the cell cycle are more likely to produce more presented peptides than genes that are stably expressed at very low levels.

[0206] Allele-non-interaction information may also include a comprehensive catalog of source protein properties, such as those provided in uniProt or PDB http: / / www.rcsb.org / pdb / home / home.do. These properties may include, among other things, the protein's secondary and tertiary structure, intracellular localization11, and gene ontology (GO) terminology. Specifically, this information may include annotations that act at the protein level, e.g., 5'UTR length, and annotations that act at the specific residue level, e.g., helix motifs of residues 300-310. These properties may also include turn motifs, sheet motifs, and disordered residues.

[0207] Allele-non-interaction information may also include characteristics describing the properties of the peptide-containing source protein domain, such as secondary or tertiary structure (e.g., α-helix vs. β-sheet); alternative splicing.

[0208] Allele-non-interaction information can also include properties that describe the presence or absence of presentation hotspots at the peptide's location in the peptide's source protein.

[0209] Allele non-interaction information may also include the probability of presentation of the peptide derived from the source protein of the peptide in question in other individuals (after adjusting for the expression levels of the source protein in those individuals and the influence of various HLA types in those individuals).

[0210] Allele non-interaction information may also include the probability that the peptide will not be detected or will be over-represented by mass spectrometry due to technical bias.

[0211] Expression of various gene modules / pathways (not requiring the inclusion of peptide source proteins) as measured by single / multiple genes representing gene modules, measured by gene expression assays such as RNASeq, microarrays, nanostrings, or other targeted panels, or by assays such as RT-PCR, providing information about the state of tumor cells, stroma, or tumor-infiltrating lymphocytes (TILs).

[0212] Allele-non-interaction information can also include the copy number of the source gene of the peptide in tumor cells. For example, a peptide derived from a gene subjected to homozygous deletion in tumor cells may be assigned a presentation probability of zero.

[0213] Allele-non-interaction information may also include the probability of a peptide binding to TAP, or the measured or predicted binding affinity of the peptide to TAP. Peptides that are more likely to bind to TAP, or that bind to TAP with higher affinity, are more likely to be presented by MHC-I.

[0214] Allele-non-interaction information can also include TAP expression levels in tumor cells (which can be measured by RNA-seq, proteome-mass spectrometry, and immunohistochemistry). In MHC-I cells, higher TAP expression levels increase the probability of presentation of all peptides.

[0215] Allele-non-interaction information may also include, but is not limited to, the presence or absence of tumor mutations: i. Driver mutations in known cancer driver genes such as EGFR, KRAS, ALK, RET, ROS1, TP53, CDKN2A, CDKN2B, NTRK1, NTRK2, and NTRK3. ii. Genes encoding proteins involved in antigen-presenting machinery (e.g., B2M, HLA-A, HLA-B, HLA-C, TAP-1, TAP-2, TAPBP, CALR, CNX, ERP57, HLA-DM, HLA-DMA, HLA-DMB, HLA-DO, HLA-DOA, HLA-DOB, HLA-DP, HLA-DPA1, HLA-DPB1, HLA-DQ, HLA-DQA1, HLA-DQA2, HLA-DQB1, HLA-DQB2, HLA-DR, HLA-DRA, HLA-DRB1, HLA-DRB3, HLA-DRB4, HLA-DRB5, or any gene encoding components of the proteasome or immunoproteasome). The probability of presentation of peptides that rely on components of antigen-presenting machinery affected by loss-of-function mutations in tumors is reduced.

[0216] The presence or absence of functional germline polymorphisms, including but not limited to the following: i. In genes encoding proteins involved in antigen presentation machinery (e.g., B2M, HLA-A, HLA-B, HLA-C, TAP-1, TAP-2, TAPBP, CALR, CNX, ERP57, HLA-DM, HLA-DMA, HLA-DMB, HLA-DO, HLA-DOA, HLA-DOBHLA-DP, HLA-DPA1, HLA-DPB1, HLA-DQ, HLA-DQA1, HLA-DQA2, HLA-DQB1, HLA-DQB2, HLA-DR, HLA-DRA, HLA-DRB1, HLA-DRB3, HLA-DRB4, HLA-DRB5, or any gene encoding a component of the proteasome or immunoproteasome).

[0217] Allele-non-interaction information may also include tumor type (e.g., NSCLC, melanoma).

[0218] Allele non-interaction information may also include known functionalities of HLA alleles, such as those reflected by HLA allele suffixes. For example, the suffix N in the allele name HLA-A*24:09N indicates a null allele that is not expressed and therefore unlikely to present an epitope; the complete nomenclature for HLA allele suffixes is described at https: / / www.ebi.ac.uk / ipd / imgt / hla / nomenclature / suffixes.html.

[0219] Allele-non-interaction information may also include clinical tumor subtypes (e.g., squamous cell lung cancer vs. non-squamous cell lung cancer).

[0220] Allele-non-interaction information may also include smoking history.

[0221] Allele-non-interaction information may also include a history of sunburn, sun exposure, or exposure to other mutagens.

[0222] Allele-non-interaction information may also include local expression of peptide source genes in relevant tumor types or clinical subtypes, optionally stratified by driver mutations. Genes typically expressed at high levels in relevant tumor types are more likely to be presented.

[0223] Allele-non-interaction information may also include the frequency of mutations in all tumors, or in tumors of the same type, or in tumors originating from individuals with at least one shared MHC allele, or in tumors of the same type in individuals with at least one shared MHC allele.

[0224] In the case of mutated tumor-specific peptides, the list of characteristics used to predict the probability of presentation may also include the annotation of the mutation (e.g., missense, read-through, frameshift, fusion, etc.) or whether the mutation is predicted to result in a nonsense mutation-dependent degradation mechanism (NMD). For example, a peptide derived from a protein segment that is not translated in tumor cells due to a homozygous early termination mutation may be assigned a presentation probability of zero. NMD results in reduced mRNA translation, which decreases the probability of presentation.

[0225] VII.C. Presentation Identification System Figure 3 is a high-level block diagram illustrating the computer logic components of a presentation identification system 160 according to one embodiment. In this exemplary embodiment, the presentation identification system 160 includes a data management module 312, a coding module 314, a training module 316, and a prediction module 320. The presentation identification system 160 also comprises a training data storage device 170 and a presentation model storage device 175. Some embodiments of the model management system 160 have modules different from those described herein. Similarly, functions may be distributed among the modules in different ways than those described herein.

[0226] VII.C.1. Data Management Module The data management module 312 generates a set of training data 170 from the presentation information 165. Each set of training data contains a number of data examples, and each data example i contains at least one peptide sequence p which is presented or not presented. i and the peptide sequence p i One or more associated MHC alleles a bound to it i The dependent variable y represents information that the presentation identification system 160 is interested in predicting a new value for the independent variable. i The independent variable z includes i It includes a set.

[0227] In one particular realization referred to throughout the remainder of this specification, the dependent variable y i is peptide p i However, one or more related MHC alleles a i This is a binary label indicating whether it was presented by [the specified method]. However, in other realizations, the dependent variable y i The presentation identification system 160 determines the independent variable z i It is recognized that this can represent any other kind of information that we are interested in predicting based on. For example, in another realization, the dependent variable y i This may also be a numerical value indicating the mass spectrometry ion current identified for the data example.

[0228] Peptide sequence p for example data i i is, k i It is a sequence of individual amino acids, k i The number of peptide sequences p in the training dataset may vary within a certain range. For example, the range may be 8-15 for MHC class I, or 6-30 for MHC class II. In one specific implementation of system 160, all peptide sequences p in the training dataset may vary within a certain range. i They can have the same length, for example, 9. The number of amino acids in the peptide sequence can vary depending on the type of MHC allele (e.g., MHC alleles in humans). MHC allele a for example data i i Which MHC allele corresponds to the peptide sequence p i This indicates whether it existed in combination with [another element].

[0229] The data management module 312 also processes the peptide sequence p contained in the training data 170. i and the bound MHC allele a i Along with, binding affinity b i and stability s i This may also include additional allele interaction variables such as predicted values ​​of peptide p. i and, a i The predicted binding affinity b between each of the bound MHC molecules shown in [the diagram] imay contain. As another example, the training data 170 may contain a i stability prediction value s for each of the MHC alleles shown in i .

[0230] The data management module 312 may also include the peptide sequence p i along with allelic non-interaction variables w such as the C-terminal flanking sequence and mRNA quantification values i .

[0231] The data management module 312 also identifies peptide sequences not presented by the MHC allele and generates the training data 170. Generally, this involves identifying the "longer" sequence of the source protein containing the peptide sequence to be presented, prior to presentation. If the presentation information contains an engineered cell line, the data management module 312 identifies a series of peptide sequences in the synthetic protein to which the cells were exposed but which were not presented on the MHC alleles of the cells. If the presentation information contains a tissue sample, the data management module 312 identifies the source protein that is the origin of the presented peptide sequence and identifies a series of peptide sequences in the source protein that were not presented on the MHC alleles of the tissue sample cells.

[0232] The data management module 312 also artificially generates peptides having random sequences of amino acids and identifies the generated sequences as peptides not presented on the MHC allele. This can be achieved by randomly generating peptide sequences and enables the data management module 312 to easily generate large amounts of synthetic data for peptides not presented on the MHC allele. In practice, since a small percentage of peptide sequences are presented by the MHC allele, synthetically generated peptide sequences are very likely not to be presented by the MHC allele, even if they are contained in proteins processed by the cells.

[0233] Figure 4 illustrates an exemplary set of training data 170A in one embodiment. Specifically, the first three data examples in training data 170A are a single-allelic cell line containing the allele HLA-C*01:03, as well as three different peptide sequences. The peptide presentation information from TIFF2026113718000006.tif9128 is shown. The fourth data example in training data 170A shows peptide information from a multi-aleletic cell line containing alleles HLA-B*07:02, HLA-C*01:03, and HLA-A*01:01, and from the peptide sequence QIEJOEIJE (SEQ ID NO:13). The first data example shows that the peptide sequence QCEIOWARE (SEQ ID NO:14) was not presented by allele HLA-DRB3:01:01. As discussed in the previous two paragraphs, negatively labeled peptide sequences may be randomly generated by the data management module 312 or may be identified from the source protein of the presented peptide. Training data 170A also includes predicted binding affinity values ​​at 1000 nM and predicted stability values ​​with a half-life of 1 hour for peptide sequence-aleletic pairs. Training data 170A also contains the C-terminal flanking sequence of peptide FJELFISBOSJFIE (SEQ ID NO: 15), and 10 2 This also includes allele-non-interaction variables such as TPM mRNA quantification values. The fourth data example shows that the peptide sequence QIEJOEIJE (SEQ ID NO:13) was presented by one of the alleles HLA-B*07:02, HLA-C*01:03, or HLA-A*01:01. Training data 170A also includes binding affinity and stability predictions for each allele, as well as the C-terminal flanking sequence of the peptide and mRNA quantification values ​​for the peptide.

[0234] VII.C.2. Coding Modules The encoding module 314 encodes the information contained in the training data 170 into a numerical representation that can be used to generate one or more presentation models. In one implementation, the encoding module 314 encodes an array (e.g., a peptide array or a C-terminal flanking array) in one-hot for a predefined 20-letter amino acid alphabet. Specifically, a peptide sequence p i having k i amino acids is represented as a row vector of 20·k i elements, and the single element in p i 20·(j-1)+1 , p i 20·(j-1)+2 ,..., p i 20·j corresponding to the amino acid at the j-th position of the peptide sequence has a value of 1. All other remaining elements have a value of 0. As an example, for a given alphabet {A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W, Y}, the peptide sequence EAF of three amino acids for data example i can be represented as a row vector of 60 elements represented by TIFF2026113718000007.tif18147. The C-terminal flanking sequence c i , as well as the protein sequence d h for the MHC allele, and other array data in the presentation information can be encoded in the same manner as described above.

[0235] If the training data 170 contains amino acid sequences of different lengths, the encoding module 314 can further encode the peptides into vectors of equal length by adding PAD characters to extend the predefined alphabet. For example, this can be done by left-padding the peptide sequence with PAD characters until the length of the peptide sequence reaches the length of the peptide sequence with the maximum length in the training data 170. Thus, if the peptide sequence with the maximum length has k 最大 amino acids, the encoding module 314 encodes each sequence into a vector of (20 + 1)·k 最大It is represented numerically as a row vector of elements. For example, the extended alphabet {PAD, A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W, Y} and k 最大 For a maximum amino acid length of 5, the same exemplary peptide sequence EAF of 3 amino acids is a row vector of 105 elements. It can be represented by TIFF2026113718000008.tif18147. C-terminal flanking sequence c i Other sequence data can be coded similarly as described above. Therefore, the peptide sequence p i or c i Each independent variable or column in the sequence represents the presence of a specific amino acid at a specific position in the sequence.

[0236] The above method for encoding sequence data was described for sequences containing amino acid sequences, but the method can be similarly extended to other types of sequence data, such as DNA or RNA sequence data.

[0237] The coding module 314 also provides one or more MHC alleles a for example data i. i This is coded into a row vector of m elements, where each element h=1,2,...,m corresponds to a unique identified MHC allele. For example data i, the element corresponding to the identified MHC allele has a value of 1. The remaining elements have a value of 0. For example, in the unique identified MHC allele type {HLA-A*01:01, HLA-C*01:08, HLA-B*07:02, HLA-DRB1*10:01} with m=4, the alleles HLA-B*07:02 and HLA-DRB1*10:01 for example data i, which corresponds to a multi-allelic cell line, are represented by a row vector a of 4 elements. i =[0 0 1 1] can be expressed as a3 i =1 and a4 i = 1. Examples with four identified MHC allele types are described herein, but the number of MHC allele types can actually be hundreds or thousands. As mentioned above, each data example i is typically a peptide sequence pi It contains up to six different MHC allergen types.

[0238] The coding module 314 also labels y for each data example i. i This is coded as a binary variable having values ​​from the set {0,1}, where a value of 1 corresponds to peptide x i However, related MHC allele a i This indicates that it was presented by one of the peptides, and a value of 0 is peptide x i However, related MHC allele a i This indicates that it was not presented by any of the following methods. Dependent variable y i However, when representing mass spectrometry ion currents, the coding module 314 can further scale the values ​​using various functions, such as a log function, which has a range of [-∞,∞] for ion current values ​​between [0,∞].

[0239] The coding module 314 is peptide p i and the allele interaction variable x for the associated MHC allele h. h i The pair can be represented as a row vector in which numerical representations of allele interaction variables are successively concatenated. For example, coding module 314 is x h i [p i ], [p i b h i ], [p i s h i ], or [p i b h i s h i ] can be expressed as a row vector equivalent to b h i This is the predicted binding affinity value for peptide pi and the associated MHC allele h, and similarly, s h i This relates to stability. Alternatively, one or more combinations of allele interaction variables may be conserved individually (for example, as individual vectors or matrices).

[0240] In one example, the coding module 314 uses the measured or predicted value for binding affinity to determine the allele interaction variable x h i By incorporating this, it represents binding affinity information.

[0241] In one example, coding module 314 takes the measured or predicted value for binding stability as an allele interaction variable x h i By incorporating this, it represents binding stability information.

[0242] In one example, the coding module 314 takes the measured or predicted value for the binding on rate as the allele interaction variable x h i By incorporating it, it represents coupled on-rate information.

[0243] In one example, for a peptide presented by a class I MHC molecule, coding module 314 uses a vector to represent the peptide length. TIFF2026113718000009.tif11128(however, TIFF2026113718000010.tif3128 is an index function, L k is peptide p k It is expressed as (meaning the length of) vector T. k The allele interaction variable x h i It can be included in another example, for a peptide presented by a class II MHC molecule, coding module 314 uses a vector to represent the peptide length. TIFF2026113718000011.tif18146 (however, TIFF2026113718000012.tif3128 is an index function, L k is peptide p k It is expressed as (meaning the length of) vector T. k The allele interaction variable x hi It can be included.

[0244] In one example, coding module 314 represents MHC allele RNA expression information by incorporating the RNA-seq-based expression level of the MHC allele into the allele interaction variable xhi.

[0245] Similarly, coding module 314 is an allele non-interacting variable w i This can be represented as a row vector in which numerical representations of allele-non-interacting variables are successively chained together. For example, w i is [c i ] or [c i m i w i It may also be a row vector equivalent to ], w i This involves measuring the C-terminal flanking sequence of peptide pi and the mRNA quantification measurement related to the peptide m i In addition, it is a row vector representing any other allele non-interacting variables. Alternatively, one or more combinations of allele non-interacting variables may be stored individually (for example, as individual vectors or matrices).

[0246] In one example, coding module 314 uses the turnover rate or half-life as an allele-non-interacting variable w i By incorporating this, it represents the metabolic turnover rate of the source protein with respect to the peptide sequence.

[0247] In one example, coding module 314 uses the protein length as an allele-non-interacting variable w i By incorporating it, it represents the length of the source protein or isoform.

[0248] In one example, the coded module 314 is β1 i , β2 i , β5 i The mean expression of immunoproteasome-specific proteasome subunits, including the subunit, is expressed as an allele-non-interacting variable w. iBy incorporating it, it represents the activation of the immunoproteasome.

[0249] In one example, coding module 314 uses the RNA-seq abundance of the source protein of a peptide, or the gene or transcript of a peptide (quantified in units of FPKM, TPM by techniques such as RSEM), and the abundance of the source protein is expressed as an allele non-interaction variable w i It is represented by incorporating it into the structure.

[0250] In one example, coding module 314 calculates the probability that the peptide origin transcript will undergo nonsense mutation-dependent degradation (NMD), as estimated by the model in, for example, Rivas et al. Science, 2015, and assigns this probability to the allele non-interaction variable w i It is represented by incorporating it into the structure.

[0251] In one example, coding module 314 expresses the activation status of a gene module or pathway evaluated via RNA-seq by, for example, quantifying the gene expression in the pathway in units of TPM using RSEM for each gene in the pathway, and then calculating a summary statistic across the genes in the pathway, such as the mean, using computer calculation. The mean is expressed as an allele non-interaction variable w i It can be incorporated into it.

[0252] In one example, coding module 314 uses the copy number of the source gene as an allele non-interaction variable w. i It is represented by incorporating it into the structure.

[0253] In one example, the coding module 314 uses the measured or predicted TAP binding affinity (e.g., in nanomolar units) as an allele non-interaction variable w i By including it, the TAP binding affinity is expressed.

[0254] In one example, coding module 314 uses an allele-non-interaction variable w to measure the TAP expression level, which is measured by RNA-seq (and quantified in units of TPM, for example, by RSEM). i By including it, the TAP expression level is represented.

[0255] In one example, coding module 314 uses tumor mutations as allele-non-interacting variable w i The vector of the indicator variable in (i.e., peptide p) k If it originates from a sample containing the KRAS G12D mutation, then d k (=1, otherwise 0)

[0256] In one example, coding module 314 uses a vector of indicator variables (i.e., peptide p) to represent germline polymorphisms in antigen-presenting genes. k If it is derived from a sample that has a germline polymorphism specific to TAP, then d k These are represented as (=1). These indicator variables are allele non-interaction variables w i It can be included.

[0257] In one example, coding module 314 represents tumor types as one-hot encoded vectors of length 1 for each alphabet of tumor type (e.g., NSCLC, melanoma, colorectal cancer, etc.). These one-hot encoded variables are then associated with allele-non-interacting variables w i It can be included.

[0258] In one example, coding module 314 represents MHC allele suffixes by processing four-digit HLA alleles with various suffixes. For example, HLA-A*24:09N is considered a different allele from HLA-A*24:09 for the purposes of the model. Alternatively, since HLA alleles ending in the N suffix are not expressed, the probability of presentation by MHC alleles ending in the N suffix can be set to zero for all peptides.

[0259] In one example, coding module 314 represents tumor subtypes as one-hot encoded vectors of length 1 for each alphabet of tumor subtype (e.g., lung adenocarcinoma, lung squamous cell carcinoma, etc.). These one-hot encoded variables are then associated with allele non-interaction variables w i It can be included.

[0260] In one example, coding module 314 uses smoking history as an allele-non-interaction variable w i Binary indicator variables that can be included (if the patient has a history of smoking, d k It is represented as 1 (=1, otherwise 0). Alternatively, smoking history can be coded as a one-hot encoded variable of length 1 for an alphabet of smoking severity. For example, smoking status can be assessed on a scale of 1 to 5, where 1 indicates a non-smoker and 5 indicates a current heavy smoker. Since smoking history is primarily associated with lung tumors, when training models for multiple tumor types, this variable can also be defined as equivalent to 1 if the patient has a history of smoking and the tumor type is a lung tumor, and zero otherwise.

[0261] In one example, coding module 314 uses the sunburn history as the allele-non-interaction variable w i A binary indicator variable that can be included (if the patient has a history of severe sunburn, d k This variable is represented as 1 (=1, otherwise 0). Since severe sunburn is primarily associated with melanoma, when training models for multiple tumor types, this variable can also be defined as equivalent to 1 if the patient has a history of severe sunburn and the tumor type is melanoma, and zero otherwise.

[0262] In one example, coding module 314 represents the distribution of expression levels of a specific gene or transcript in the human genome as summary statistics (e.g., mean, median) of the expression level distribution, using a reference database such as TCGA. Specifically, peptide p in a sample with tumor type melanoma. k Regarding peptide p k The expression level of the measured gene or transcript of the origin of the gene or transcript is an allele-non-interaction variable w i Not only can it be included, but the peptide p in melanoma, as measured by TCGA, k This may also include the mean and / or median gene or transcript expression of the gene or transcript of origin.

[0263] In one example, coding module 314 represents the variant type as a one-hot encoded variable of length 1 for the alphabet of the variant type (e.g., missense, frameshift, NMD-inducible, etc.). These one-hot encoded variables are allele non-interaction variables w i It can be included.

[0264] In one example, coding module 314 uses the allele non-interaction variable w as the value of the source protein's annotation (e.g., 5'UTR length) to represent the protein-level properties of the protein. i In another example, coding module 314 represents peptide p i Residue-level annotation of the source protein for peptide p i If it overlaps with the helix motif, it is equivalent to 1; otherwise, it is 0, or peptide p i If the peptide p is completely contained within the helix motif, the indicator variable equivalent to 1 is represented by including it in the allele non-interaction variable wi. In another example, peptide p contained within the helix motif annotation is represented by including it in the allele non-interaction variable wi. iThe allele non-interaction variable w represents the property that expresses the proportion of residues in the allele. i It can be included.

[0265] In one example, coding module 314 uses an index vector o, which has a length equivalent to the number of proteins or isoforms in the human proteome, to represent the type of protein or isoform in the human proteome. k Represented as, and corresponding element o k i is peptide p k If it originates from protein i, the value is 1; otherwise, it is 0.

[0266] In one example, coding module 314 is peptide p i Source gene G = gene(p i ) is represented as a categorical variable having L possible categories (where L represents the upper limit of the number of subscripted source genes, 1, 2, ..., L).

[0267] In one example, coding module 314 is peptide p i Tissue type, cell type, tumor type, or tumor histology type T = tissue (p i ) is represented as a categorical variable having M possible categories (where M indicates the upper limit of the number of subscripted types 1, 2, ..., M). Examples of tissue types include lung tissue, heart tissue, intestinal tissue, and nerve tissue. Examples of cell types include dendritic cells, macrophages, and CD4 T cells. Examples of lung adenocarcinoma, lung squamous cell carcinoma, melanoma, and non-Hodgkin lymphoma.

[0268] The coding module 314 also contains peptide p i and the variable z for the related MHC allele h. i The overall set of allele interaction variables x i and allele non-interacting variable w iIt can also be represented as a row vector in which the numerical representations of are successively chained together. For example, coding module 314 is z h i [x h i w i ] or [w i x h i It can be represented as a row vector equivalent to ].

[0269] VIII. Training Modules Training module 316 constructs one or more presentation models that generate the likelihood of whether a peptide sequence is presented by an MHC allele associated with the peptide sequence. Specifically, the peptide sequence p k and peptide sequence p k Related MHC allergen a k Given a set of peptide sequences, each presentation model is given the peptide sequence p k However, related MHC allele a k An estimate u that shows the likelihood that will be presented by one or more of the following. k Generates.

[0270] VIII.A. Overview The training module 316 constructs one or more presentation models based on the training dataset stored in the storage device 170, which is generated from the presentation information stored in 165. Generally, regardless of the specific type of presentation model, all presentation models capture the dependencies between the independent and dependent variables in the training data 170 so that the loss function is minimized. Specifically, the loss function TIFF2026113718000013.tif4128 contains the dependent variable y for one or more data examples S in the training data 170. i∈S The value of and the estimated likelihood u for the data example S generated by the presented model i∈S This represents a contradiction between the two. In one particular realization referred to throughout the remainder of this specification, the loss function TIFF2026113718000014.tif4128 is a negative log-likelihood function given by the following equation (1a): TIFF2026113718000015.tif10128 However, in practice, a different loss function may be used. For example, when predictions are made for mass spectrometry ion current, the loss function is the mean square loss given by equation 1b below. TIFF2026113718000016.tif10128

[0271] The presented model may be a parametric model in which one or more parameters θ mathematically specify the dependence between the independent and dependent variables. Typically, the loss function The various parameters of the parametric model that minimizes TIFF2026113718000017.tif4128 are determined through gradient-based numerical optimization algorithms, such as batch gradient algorithms or stochastic gradient algorithms. Alternatively, the model may be a non-parametric model in which the model structure is determined from the training data 170 and is not strictly based on a fixed set of parameters.

[0272] VIII.B. Allergy-Specific Models The training module 316 can construct a presentation model to predict the presentation likelihood of a peptide on an allele-by-allele basis. In this example, the training module 316 can train the presentation model based on example data S in training data 170 generated from cells expressing a single MHC allele.

[0273] In one implementation, the training module 316 is TIFF2026113718000018.tif7128 provides the estimated presentation likelihood u of peptide pk for a specific allele h. k This modeled the peptide sequence x h k is peptide p k and the coded allele interaction variable for the corresponding MHC allele h, where f(·) is an arbitrary function and, for convenience of description, is referred to as the transformation function throughout this specification. Furthermore, g h(·) is an arbitrary function, which for convenience of description will be referred to as the dependency function throughout this specification, and is a parameter θ determined for the MHC allele h. h Based on the set, allele interaction variable x h k Generate dependency scores for each MHC allele h. Parameter θ h The set of values ​​is θ h This can be determined by minimizing the loss function related to , where i is each example in a subset S of 170 training data generated from cells expressing a single MHC allele h.

[0274] Dependency function g h (x h k ;θ h The output of ) indicates that the MHC allele h has at least allele interaction properties x h k Based on, and in particular, peptide p k This represents a dependency score for MHC allele h, indicating whether it presents the corresponding nascent antigen based on the amino acid positions of the peptide sequence. For example, the dependency score for MHC allele h is: k A high value may be given when there is a high probability of presenting it, and a low value may be given when there is a low probability of presenting it. The transformation function f(·) transforms the input, and more specifically, in this example g h (x h k ;θ h The dependency score generated by ) is used for peptide p k This is converted into an appropriate value that represents the likelihood that will be presented by the MHC allele.

[0275] In one particular implementation referred to throughout the remainder of this specification, f(·) is a function having a range in [0,1] for a suitable domain range. In one example, f(·) is: This is the expit function provided by TIFF2026113718000019.tif10128. As another example, f(·) also occurs when the value of domain z is greater than or equal to 0. It can also be the hyperbolic tangent function given by TIFF2026113718000020.tif4128. Alternatively, if the prediction is made for mass spectrometry ion currents with values ​​outside the range [0,1], f(·) can be any function such as the identity function, exponential function, or log function.

[0276] Therefore, the peptide sequence p k The allele-by-allele likelihood that will be presented by the MHC allele h is the dependency function g for the MHC allele h. h (·) peptide sequence p k This can be generated by applying it to the coded version of and generating the corresponding dependency score. The dependency score is based on the peptide sequence p k It may be transformed by a transformation function f(·) such that it generates the allele-specific likelihood that will be presented by the MHC allele h.

[0277] VIII.B.1 Dependency Functions for Allele Interaction Variables In one particular implementation referred to herein, the dependency function g h (·) is x h k Each allele interaction variable in the context is determined by the parameter θ determined for the relevant MHC allele h. h Linearly combine the corresponding parameters in the set. This is an affine function given by TIFF2026113718000021.tif5128.

[0278] In other specific implementations referred to throughout this specification, the dependency function g h (·) is a network model NN having a set of nodes arranged in one or more layers. h Represented by (·), This is the network function given by TIFF2026113718000022.tif5128. The nodes are parameter θ.h A set of nodes can be connected to other nodes through connections, each having a relevant parameter. The value at a particular node can be represented as the sum of the values ​​of the nodes connected to that particular node, weighted by the relevant parameter mapped by the activation function associated with that particular node. In contrast to affine functions, network models are advantageous because the presentation model can incorporate nonlinearity and process data with amino acid sequences of different lengths. Specifically, through nonlinear modeling, network models can capture interactions between amino acids at different positions in the peptide sequence and how these interactions affect peptide presentation.

[0279] Generally speaking, network models are NN h (·) can be structured as feedforward networks such as artificial neural networks (ANNs), convolutional neural networks (CNNs), and deep neural networks (DNNs), and / or recurrent networks such as long-term short-term memory networks (LSTMs), bidirectional recurrent networks, and deep bidirectional recurrent networks.

[0280] In one example referred to throughout the remainder of this specification, each MHC allele at h=1,2,...,m relates to a separate network model, NN h (·) indicates the output from the network model related to MHC allele h.

[0281] Figure 5 illustrates an exemplary network model NN3(·) associated with an arbitrary MHC allele h=3. As shown in Figure 5, the network model NN3(·) for MHC allele h=3 includes three types of input nodes at layer l=1, four types of nodes at layer l=2, two types of nodes at layer l=3, and one type of output node at layer l=4. The network model NN3(·) is associated with a set of 10 types of parameters θ3(1), θ3(2), ..., θ3(10). The network model NN3(·) includes three types of allele interaction variables x3 for MHC allele h=3. k (1), x3 k (2), and x3 k (3) The input values ​​for (3) are received (individual data examples, including coded polypeptide sequence data and any other training data used), and the value NN3(x3 k The output is ). The network function may include one or more network models, each taking a different allele interaction variable as input.

[0282] In another example, the identified MHC alleles h=1,2,...,m are used in a single network model NN. H (•) is related to NN h (·) represents one or more outputs of a single network model associated with the MHC allele h. In such an example, the parameter θ h The set of parameters can correspond to the set of parameters for a single network model, and therefore the parameter θ h This set can be shared by all MHC alleles.

[0283] Figure 6A shows an exemplary network model NN shared by MHC alleles h=1,2,...,m. H (·) will be explained. As shown in Figure 6A, the network model NN H (·) contains m output nodes, each corresponding to an MHC allele. The network model NN3(·) has an allele interaction variable x3 for the MHC allele h=3. k It receives the value NN3(x3) corresponding to the MHC allele h=3.k Output m values, including ).

[0284] In yet another example, a single network model NN H (·) represents the allele interaction variable x of the MHC allele h. h k and encoded protein sequence d h It could be a network model that, given a parameter θ, outputs a dependency score. In such an example, the parameter θ h The set of parameters can again correspond to the set of parameters for a single network model, and therefore the parameter θ h The set of can be shared by all MHC alleles. Therefore, in such an example, NNh(·) is the input [x for a single network model. h k d h Given ], a single network model NN H This represents the output of (·). Such network models are advantageous because they can correctly predict the peptide presentation probability for MHC alleles that were unknown in the training data simply by identifying their protein sequences.

[0285] Figure 6B shows an exemplary network model NN shared by MHC alleles. H (·) will be explained. As shown in Figure 6B, the network model NN H (·) accepts the allele interaction variable and protein sequence of the MHC allele h=3 as input and performs a dependency score NN3(x3) corresponding to the MHC allele h=3. k Outputs ).

[0286] In yet another example, the dependency function g h (·)teeth, It can be represented as TIFF2026113718000023.tif5128, and in the formula, g' h (x h k ;θ' h) is an affine function, network function, etc., with a set of parameters θ'h, and represents the baseline probability of presentation for MHC allele h, with a bias parameter θ in a set of parameters for the allele interaction variable of the MHC allele. h 0 It is accompanied by.

[0287] In another implementation, the bias parameter θ h 0 This may be shared according to the gene family of the MHC allele h. That is, the bias parameter θ for the MHC allele h. h 0 is θ 遺伝子(h) 0 This can be equivalent, and gene (h) is the gene family of MHC allele h. For example, class I MHC alleles HLA-A*02:01, HLA-A*02:02, and HLA-A*02:03 may be assigned to the gene family "HLA-A", with a bias parameter θ for each of these MHC alleles. h 0 The bias parameters θ for each of these MHC alleles may be shared. As another example, assign the class II MHC alleles HLA-DRB1:10:01, HLA-DRB1:11:01, and HLA-DRB3:01:01 to the "HLA-DRB" gene family and h 0 It can be shared.

[0288] For example, returning to equation (2), the affine dependency function g h Using (·), peptide p was obtained by MHC allele h=3 among the different identified MHC alleles with m=4. k The likelihood that will be presented is, This can be generated by TIFF2026113718000024.tif5128, where x3k is the allele interaction variable identified for the MHC allele h=3, and θ3 is the set of parameters determined for the MHC allele h=3 through loss function minimization.

[0289] As another example, using separate network transformation functions gh(·), peptide p is obtained by MHC allele h=3 among different identified MHC alleles m=4. k The likelihood that will be presented is, It can be generated by TIFF2026113718000025.tif5128, in the formula, x3 k θ3 is an allele interaction variable identified for the MHC allele h=3, and θ3 is a set of parameters determined for the network model NN3(·) associated with the MHC allele h=3.

[0290] Figure 7 shows the peptide p associated with MHC allele h=3 using an exemplary network model NN3(·). k This explains the generation of the presentation likelihood. As shown in Figure 7, the network model NN3(·) uses the allele interaction variable x3 for the MHC allele h=3. k Receives the output NN3(x3 k ) generates. The output is mapped by the function f(·) to the estimated presentation likelihood u k Generates.

[0291] VIII.B.2. Allele-by-allele with allele-non-interaction variables In one implementation, the training module 316 incorporates allele-non-interaction variables, According to TIFF2026113718000026.tif7128, peptide p k The estimated presentation likelihood uk is modeled, and in the formula, w k is peptide p k This refers to the coded allele non-interaction variable about g w (·) is the parameter θ determined for the allele non-interaction variable. w Allele non-interacting variable w based on the set k This is a function of θ. Specifically, it is a parameter θ for each MHC allele h. h The set of alleles and the parameter θ for allele non-interaction variables w The set of values, θ h and θw This can be determined by minimizing the loss function related to i, where i is each example in a subset S of 170 training data generated from cells expressing a single MHC allele.

[0292] Dependency function g w (w k ;θ w The output of ) is determined by peptide p by one or more MHC alleles, based on the influence of allele non-interaction variables. k This represents a dependency score for an allele-non-interacting variable, indicating whether or not it is presented. For example, the dependency score for an allele-non-interacting variable is peptide p k The C-terminal flanking sequence and peptide p are known to have a positive effect on presentation. k If it is bound, it may have a high value, peptide p k C-terminal flanking sequences and peptide p are known to negatively affect the presentation of the peptide. k If they are combined, they may have a low value.

[0293] According to equation (8), the peptide sequence p k The allele-specific likelihood that will be presented by the MHC allele h is a function g of the MHC allele h. h (·) is the peptide sequence p k This can be generated by applying a coded version of to generate corresponding dependency scores for allele-interacting variables. The function g for allele-non-interacting variables is also generated. w (·) is also applied to the coded version of the allele non-interacting variable to generate a dependency score for the allele non-interacting variable. Both scores are combined, and the combined score is used to determine the peptide sequence p by the MHC allele h. k The transformation is performed by the transformation function f(·) to generate the allele-specific likelihoods that will be presented.

[0294] Alternatively, in equation (2), the training module 316 uses the allele non-interacting variable w. k allele interaction variable xh k By adding this, the allele-non-interacting variable wk in the prediction may be included. Therefore, the presented likelihood is, This can be given by TIFF2026113718000027.tif7128.

[0295] VIII.B.3 Dependence Functions for Allele Non-Interacting Variables Dependency function g for allele interaction variables h Similar to (·), the dependency function g for allele non-interaction variables w (·) represents an affine function, or a separate network model where the allele non-interacting variable w k This could be a related network function.

[0296] Specifically, the dependency function g w (·) is w k The allele non-interaction variable in this case is the parameter θ. w Linearly combine the corresponding parameters in the set. This is an affine function given by TIFF2026113718000028.tif5128.

[0297] Dependency function g w (·) is also the parameter θ w Network model NN with related parameters in a set w Represented by (·), This is the network function given by TIFF2026113718000029.tif5128. The network function may include one or more network models, each taking different allele non-interacting variables as input.

[0298] In another example, the dependency function g for allele-non-interacting variables w (·)teeth, The formula is given by TIFF2026113718000030.tif5128, where g' w (w k ;θ'w ) is the allele non-interaction parameter θ' w These are affine functions, network functions, etc., accompanied by a set of m k is peptide p k This is a quantitative mRNA measurement value for , where h(·) is a function that transforms the quantitative measurement value, and θ w m h(·) is a parameter in a set of parameters for allele non-interaction variables that are combined with mRNA quantification measurements to generate a dependency score for the mRNA quantification measurement. In one particular embodiment referred to throughout the remainder of this specification, h(·) is a log function, but in practice, h(·) can be any one of a variety of different functions.

[0299] In yet another example, the dependency function g for allele-non-interacting variables w (·)teeth, Given by TIFF2026113718000031.tif5128, in the formula, g' w (w k ;θ' w ) is the allele non-interaction parameter θ' w These are affine functions, network functions, etc., which involve a set of o k is peptide p k The index vectors described in Section VII.C.2 represent proteins and isoforms in the human proteome, and θ w o This is a set of parameters in the set of parameters for allele non-interacting variables combined with the index vector. In one variation, o k and parameter θ w o If the dimensionality of the set is significantly high, TIFF2026113718000032.tif4128( TIFF2026113718000033.tif4128 allows parameter regularization terms (such as L1 norm, L2 norm, and combinations) to be added to the loss function when determining parameter values. The optimal value of the hyperparameter λ can be determined through an appropriate method.

[0300] In yet another example, the dependency function g for allele-non-interacting variables w (·) is given by the following equation. That is, TIFF2026113718000034.tif13128 However, g' w (w k ;θ' w ) is the allele non-interaction parameter θ' w These include affine functions, network functions, etc., which involve a set of elements. TIFF2026113718000035.tif5128 is a peptide p k The indicator function is equal to 1 when the allele-non-interaction variable originates from the source gene l mentioned above, and θ w l θ is a parameter that indicates the "antigenicity" of the source gene l. In one variation, L is sufficiently large, and therefore the number of parameters θ w l=1, 2,...,L If it is sufficiently large, Parameter regularization terms like TIFF2026113718000036.tif5128 (where, TIFF2026113718000037.tif4128 allows the addition of L1 norm, L2 norm, combinations, etc., to the loss function when determining parameter values. The optimal value of the hyperparameter λ can be determined by an appropriate method.

[0301] In yet another example, the dependency function g for allele-non-interacting variables w (·) is given by the following equation. That is, TIFF2026113718000038.tif13146 However, g' w (w k ;θ'w ) is the allele non-interaction parameter θ' w These include affine functions, network functions, etc., which involve a set of elements. TIFF2026113718000039.tif5128 is a peptide p with respect to allele non-interaction variables as described above. k If it originates from source gene l, and peptide p k The indicator function is equal to 1 when it originates from the organization type m, and θ w lm This parameter indicates the antigenicity of the combination of source gene l and tissue type m. More specifically, the antigenicity of gene l in tissue type m may indicate a persistence tendency for cells of tissue type m to present peptides derived from gene l after regulation of RNA expression and peptide sequence context.

[0302] In one variation, L or M is sufficiently large, and therefore the number of parameters θ w lm=1, 2,...,LM If it is sufficiently large, Parameter regularization terms like TIFF2026113718000040.tif5128 (where, TIFF2026113718000041.tif4128 allows the addition of L1 norms, L2 norms, combinations, etc., to the loss function when determining parameter values. The optimal value of the hyperparameter λ can be determined by an appropriate method. In another variation, a parameter regularization term can be added to the loss function when determining parameter values ​​so that the coefficients for the same source gene do not differ significantly between tissue types. For example, the following penalty term: TIFF2026113718000042.tif18128 (however, TIFF2026113718000043.tif5128 (which is the average antigenicity of source gene l across tissue types) can be used to penalize the standard deviation of antigenicity across different tissue types in the loss function.

[0303] In practice, the dependency function g with respect to the allele non-interacting variable can be obtained by combining any of the additional terms in equations (10), (11), (12a), and (12b). w (·) can be generated. For example, by adding together the term h(·) representing the mRNA quantitative measurement in equation (10) and the term representing the antigenicity of the source gene in equation (12) along with any other affine function or network function, a dependency function for allele non-interaction variables can be generated.

[0304] For example, returning to equation (8), the affine transformation function g h (·), g w Using (·), peptide p was obtained by MHC allele h=3 among the different identified MHC alleles with m=4. k The likelihood that will be presented is, It can be generated by TIFF2026113718000044.tif5128, in the formula, w k is peptide p k The allele non-interaction variable identified for θ is θ w This is a set of parameters determined for allele non-interaction variables.

[0305] As another example, the network transformation function g h (·), g w Using (·), peptide p was obtained by MHC allele h=3 among the different identified MHC alleles with m=4. k The likelihood that will be presented is, It can be generated by TIFF2026113718000045.tif5128, in the formula, w k is peptide p k The allele interaction variable identified for θ is θ w This is a set of parameters determined for allele non-interaction variables.

[0306] Figure 8 shows exemplary network models NN3(·) and NN w (•) Peptide p related to MHC allele h=3, using (•) kThis explains the generation of the presentation likelihood. As shown in Figure 8, the network model NN3(·) uses the allele interaction variable x3 for the MHC allele h=3. k Receives the output NN3(x3 k ) generates a network model NN. w (·) represents peptide p k Allele non-interaction variable w k Received, output NN w (w k ) generates. The output is combined and mapped by the function f(·) to estimate the presented likelihood u k Generates.

[0307] VIII.C. Multiple Allergen Models The training module 316 can also construct a presentation model to predict the presentation likelihood of a peptide in a multi-allelic setting where two or more MHC alleles are present. In this example, the training module 316 can train the presentation model based on example data S in the training data 170 generated from cells expressing a single MHC allele, cells expressing multiple MHC alleles, or a combination thereof. [Examples]

[0308] VIII.C.1. Example 1: Maximum value of each allele model In one implementation, the training module 316 is a peptide p associated with a set of multiple MHC alleles H. k Estimated presentation likelihood u k The presentation likelihood determined for each MHC allele h in set H, determined based on cells expressing a single allele, as explained above along with equations (2) to (11). Model it as a function of TIFF2026113718000046.tif4128. Specifically, the presented likelihood u k teeth, It can be any function of TIFF2026113718000047.tif4128. In one realized form, as shown in equation (12), the function is a maximum function and the presented likelihood u kThis can be determined as the maximum presentation likelihood for each MHC allele h in set H. TIFF2026113718000048.tif5128

[0309] VIII.C.2. Example 2.1: Function Model of Summation In one implementation, the training module 316 uses peptide p k Estimated presentation likelihood u k of, Modeled by TIFF2026113718000049.tif13128, in the formula, element a h k The peptide sequence p k For multiple MHC alleles H related to x, the result is 1. h k is peptide p k This refers to the coded allele interaction variables for the corresponding MHC alleles. The parameter θ for each MHC allele h. h The set of values ​​is θ h This can be determined by minimizing the loss function related to g, where i is each example in a subset S of training data 170 generated from cells expressing a single MHC allele and / or cells expressing multiple MHC alleles. h This is the dependency function g introduced above in Section VIII.B.1. h It could take any of the following forms.

[0310] According to equation (13), the peptide sequence p k The likelihood of presentation by one or more MHC alleles h is given by the dependency function g h (·) represents the peptide sequence p for each MHC allele H. k This can be generated by applying it to the coded version and generating corresponding scores for allele interaction variables. The scores for each MHC allele h are combined to form the peptide sequence p k The transformation function f(·) is used to generate the presentation likelihood that will be presented by the set of MHC alleles H.

[0311] The model presented in equation (13) is that each peptide p k This differs from the allele-by-allele model in equation (2) in that the number of related alleles for a can be greater than 1. In other words, a h k More than one element in the peptide sequence p k Multiple MHC alleles H related to this can have a value of 1.

[0312] For example, the affine transformation function g h Using (·), peptide p is produced by MHC alleles h=2 and h=3 among the different identified MHC alleles of m=4. k The likelihood that will be presented is, It can be generated by TIFF2026113718000050.tif5128, in the formula, x2 k , x3 k θ is an allele interaction variable identified for MHC alleles h=2 and h=3, and θ2 and θ3 are sets of parameters determined for MHC alleles h=2 and h=3.

[0313] As another example, the network transformation function g h (·), g w Using (·), peptide p is produced by MHC alleles h=2 and h=3 among the different identified MHC alleles of m=4. k The likelihood that will be presented is, This can be generated by TIFF2026113718000051.tif5128, where NN2(·) and NN3(·) are network models identified for MHC alleles h=2 and h=3, and θ2 and θ3 are sets of parameters determined for MHC alleles h=2 and h=3.

[0314] Figure 9 shows the peptide p associated with MHC alleles h=2 and h=3, using exemplary network models NN2(·) and NN3(·). kThis explains the generation of the presentation likelihood. As shown in Figure 9, the network model NN2(·) uses the allele interaction variable x² for the MHC allele h=2. k Receives output NN2(x2 k ) generates, and the network model NN3(·) uses the allele interaction variable x3 for the MHC allele h=3. k Receives the output NN3(x3 k ) generates. The output is combined and mapped by the function f(·) to estimate the presented likelihood u k Generates.

[0315] VIII.C.3. Example 2.2: Function model of sum with allele-non-interacting variables In one implementation, the training module 316 incorporates allele-non-interaction variables, According to TIFF2026113718000052.tif13128, peptide p k Estimated presentation likelihood u k We model this, and in the formula, w k is peptide p k This refers to coded allele non-interaction variables. Specifically, it refers to the parameter θ for each MHC allele h. h The set of alleles and the parameter θ for allele non-interaction variables w The set of values, θ h and θ w This can be determined by minimizing the loss function related to g, where i is each example in a subset S of training data 170 generated from cells expressing a single MHC allele and / or cells expressing multiple MHC alleles. w This is the dependency function g introduced above in Section VIII.B.3. w It could take any of the following forms.

[0316] Therefore, according to equation (14), one or more MHC alleles H can cause the peptide sequence p k The likelihood of presentation being given is given by the function g h (·) represents the peptide sequence p for each MHC allele H. kThis can be generated by applying a coded version of to generate the corresponding dependency score for each MHC allele h, which is the allele non-interaction variable. w (·) is also applied to the coded version of the allele non-interacting variable to generate a dependency score for the allele non-interacting variable. The scores are combined, and the combined score is used by the MHC allele H to determine the peptide sequence p k It is transformed by the transformation function f(·) to generate the presentation likelihood of what will be presented.

[0317] In the model presented by equation (14), each peptide p k The number of related alleles for a can be greater than 1. In other words, a h k More than one element in the peptide sequence p k Multiple MHC alleles H related to this can have a value of 1.

[0318] For example, the affine transformation function g h (·), g w Using (·), peptide p is produced by MHC alleles h=2 and h=3 among the different identified MHC alleles of m=4. k The likelihood that will be presented is, It can be generated by TIFF2026113718000053.tif5128, in the formula, w k is peptide p k The allele non-interaction variable identified for θ is θ w This is a set of parameters determined for allele non-interaction variables.

[0319] As another example, the network transformation function g h (·), g w Using (·), peptide p is produced by MHC alleles h=2 and h=3 among the different identified MHC alleles of m=4. k The likelihood that will be presented is, It can be generated by TIFF2026113718000054.tif5128, in the formula, w k is peptide p k The allele interaction variable identified for θ is θ w This is a set of parameters determined for allele non-interaction variables.

[0320] Figure 10 shows exemplary network models NN2(·), NN3(·), and NN w Peptides p related to MHC alleles h=2 and h=3 using (·) k This explains the generation of the presentation likelihood. As shown in Figure 10, the network model NN2(·) uses the allele interaction variable x² for the MHC allele h=2. k Receives output NN2(x2 k The network model NN3(·) generates the allele interaction variable x3 for the MHC allele h=3. k Receives the output NN3(x3 k ) generates a network model NN. w (·) represents peptide p k Allele non-interaction variable w k Received, output NN w (w k ) generates. The output is combined and mapped by the function f(·) to estimate the presented likelihood u k Generates.

[0321] Alternatively, in equation (15), the training module 316 uses the allele non-interacting variable w k allele interaction variable x h k By adding this, the allele non-interaction variable w in the prediction k It may include. Therefore, the likelihood of presentation is, This can be given by TIFF2026113718000055.tif13128.

[0322] VIII.C.4. Example 3.1: Model using implicit allele-specific likelihood In another implementation, the training module 316 uses peptide p k Estimated presentation likelihood u k of, Modeled by TIFF2026113718000056.tif7128, in the formula, element a h k The peptide sequence p k For multiple MHC alleles h∈H related to u', the value is 1. k h is the implicit allele-by-allele presentation likelihood for the MHC allele h, and the vector v is the element v h However, a h k ·u' k h The vector corresponds to v, s(·) is a function that maps the elements of v, and r(·) is a clipping function that clips the input values ​​within a given range. As described in more detail below, s(·) may be a summation function or a quadratic function, but in other embodiments, it is recognized that s(·) may be any function such as a maximum function. The set of parameters θ for the implicit allele-by-allele likelihood can be determined by minimizing a loss function with respect to θ, where i is each example in a subset S of training data 170 generated from cells expressing a single MHC allele and / or cells expressing multiple MHC alleles.

[0323] In the presentation model of equation (17), the presentation likelihood is that each individual MHC allele h produces peptide p k The implicit allele-specific likelihood u' corresponds to the likelihood that will be presented. k h It is modeled as a function of . The implicit allele-by-allele likelihood differs from the allele-by-allele presentation likelihood in Section VIII.B in that the parameters for the implicit allele-by-allele likelihood can be learned from multiple allele settings where the direct relationship between the presented peptide and the corresponding MHC allele is unknown, in addition to the single allele setting. Therefore, in the multiple allele setting, the presentation model is for peptide p kNot only can we estimate whether the whole is presented by a set of MHC alleles H, but we can also estimate which MHC allele h corresponds to peptide p k Each likelihood u' indicates which is most likely to have presented the desired outcome. k h∈H It can also be provided. The advantage of this is that the presented model can generate implicit likelihood without training data for cells expressing a single MHC allele.

[0324] In one particular implementation referred to throughout the remainder of this specification, r(·) is a function having the range [0,1]. For example, r(·) is a clipping function: r(z) = min(max(z,0),1) It may also be the case that the minimum value between z and 1 is the presented likelihood u k It is selected as. In another implementation, r(·) is r(z) = tanh(z) The hyperbolic tangent function is given as such, and the value of domain z is greater than or equal to 0.

[0325] VIII.C.5. Example 3.2: Function Sum Model In one particular realization, s(·) is a summation function, and the presentation likelihood is given by summing the implicit allele-specific presentation likelihoods. TIFF2026113718000057.tif15128

[0326] In one implementation, the implicit allele-specific presentation likelihood for MHC allele h is, The presented likelihood is generated by TIFF2026113718000058.tif7128. To be estimated by TIFF2026113718000059.tif13128.

[0327] According to equation (19), one or more MHC alleles H can transfer the peptide sequence p k The likelihood of presentation being given is given by the function g h (·) represents the peptide sequence p for each MHC allele H.k This can be generated by applying a coded version of to generate corresponding dependency scores for allele interaction variables. Each dependency score is first the implicit allele-per-presentation likelihood u' k h It is transformed by the function f(·) to generate the allele-specific likelihood u'. k h The values ​​are combined, and a clipping function is applied to the combined likelihood to clip the values ​​within the range [0,1], and the peptide sequence p k The presentation likelihood that will be presented by the set of MHC alleles H can be generated. Dependency function g h This is the dependency function g introduced above in Section VIII.B.1. h It could take any of the following forms.

[0328] For example, the affine transformation function g h Using (·), peptide p is produced by MHC alleles h=2 and h=3 among the different identified MHC alleles of m=4. k The likelihood that will be presented is, It can be generated by TIFF2026113718000060.tif7128, in the formula, x2 k , x3 k θ is an allele interaction variable identified for MHC alleles h=2 and h=3, and θ2 and θ3 are sets of parameters determined for MHC alleles h=2 and h=3.

[0329] As another example, the network transformation function g h (·), g w Using (·), peptide p is produced by MHC alleles h=2 and h=3 among the different identified MHC alleles of m=4. k The likelihood that will be presented is, This can be generated by TIFF2026113718000061.tif7128, where NN2(·) and NN3(·) are network models identified for MHC alleles h=2 and h=3, and θ2 and θ3 are sets of parameters determined for MHC alleles h=2 and h=3.

[0330] Figure 11 shows peptide p related to MHC alleles h=2 and h=3, using exemplary network models NN2(·) and NN3(·). k This explains the generation of the presentation likelihood. As shown in Figure 9, the network model NN2(·) uses the allele interaction variable x² for the MHC allele h=2. k Receives output NN2(x2 k ) generates, and the network model NN3(·) uses the allele interaction variable x3 for the MHC allele h=3. k Receives the output NN3(x3 k ) is generated. Each output is mapped by the function f(·) and combined to estimate the presentation likelihood u k Generates.

[0331] In another realization, when the prediction is made for the logarithm of the mass spectrometry ion current, r(·) is a logarithmic function and f(·) is an exponential function.

[0332] VIII.C.6. Example 3.3: Sum model of functions with allele-non-interacting variables In one implementation, the implicit allele-specific presentation likelihood for MHC allele h is, The presented likelihood is generated by TIFF2026113718000062.tif7128. The effect of allele non-interaction variables is incorporated into peptide presentation, as generated by TIFF2026113718000063.tif13128.

[0333] According to equation (21), one or more MHC alleles H can transfer the peptide sequence p k The likelihood of presentation being given is given by the function g h(·) represents the peptide sequence p for each MHC allele H. k This can be generated by applying a coded version of to generate the corresponding dependency score for each MHC allele h, which is the allele non-interaction variable. w (·) is also applied to the coded version of the allele-non-interacting variable to generate a dependency score for the allele-non-interacting variable. The scores of the allele-non-interacting variables are combined with the dependency scores of each of the allele-interacting variables. Each of the combined scores is transformed by the function f(·) to generate an implicit allele-by-allele presentation likelihood. The implicit likelihoods are combined, and a clipping function is applied to the combined output to clip the values ​​within the range [0,1], so that the peptide sequence p by the MHC allele H k A presentation likelihood of what will be presented can be generated. Dependency function g w This is the dependency function g introduced above in Section VIII.B.3. w It could take any of the following forms.

[0334] For example, the affine transformation function g h (·), g w Using (·), peptide p is produced by MHC alleles h=2 and h=3 among the different identified MHC alleles of m=4. k The likelihood that will be presented is, It can be generated by TIFF2026113718000064.tif7128, in the formula, w k is peptide p k θw is an allele non-interaction variable identified for the allele non-interaction variable, and θw is a set of parameters determined for the allele non-interaction variable.

[0335] As another example, the network transformation function g h (·), g w Using (·), peptide p is produced by MHC alleles h=2 and h=3 among the different identified MHC alleles of m=4. k The likelihood that will be presented is, It can be generated by TIFF2026113718000065.tif7128, in the formula, w k is peptide p k The allele interaction variable identified for θ is θ w This is a set of parameters determined for allele non-interaction variables.

[0336] Figure 12 shows exemplary network models NN2(·), NN3(·), and NN w Peptides p related to MHC alleles h=2 and h=3 using (·) k This explains the generation of the presentation likelihood. As shown in Figure 12, the network model NN2(·) uses the allele interaction variable x² for the MHC allele h=2. k Receives output NN2(x2 k ) generates a network model NN. w (·) represents peptide p k Allele non-interaction variable w k Received, output NN w (w k The output is combined and mapped by the function f(·). The network model NN3(·) uses the allele interaction variable x3 for MHC allele h=3. k Receives the output NN3(x3 k This generates the same network model NN. w (·) Output NN w (w k ) is combined with and mapped by the function f(·). The combined output of both gives the estimated presentation likelihood u k Generates.

[0337] In another implementation, the implicit allele-specific presentation likelihood for MHC allele h is, The presented likelihood is generated by TIFF2026113718000066.tif7128. It should be generated by TIFF2026113718000067.tif13128.

[0338] VIII.C.7. Example 4: Secondary Model In one realized form, s(·) is a quadratic function, and peptide p k Estimated presentation likelihood u k teeth, Given by TIFF2026113718000068.tif13128, in the expression, element u' k h is the implicit allele-by-allegment presentation likelihood for the MHC allele h. The set of parameters θ for the implicit allele-by-allegment likelihood can be determined by minimizing the loss function with respect to θ, where i is each example in a subset S of training data 170 generated from cells expressing a single MHC allele and / or cells expressing multiple MHC alleles. The implicit allele-by-allegment presentation likelihood can take any form shown in equations (18), (20), and (22) above.

[0339] In one embodiment, the model of equation (23) is the peptide sequence p k However, there is a possibility that the presentation may occur simultaneously by two MHC alleles, which could imply that the presentation by the two HLA alleles is statistically independent.

[0340] According to equation (23), one or more MHC alleles H can transfer the peptide sequence p k The presentation likelihood that will be presented is obtained by combining the implicit allele-by-allele presentation likelihoods and by the peptide sequence p by the MHC allele H. k Each pair of MHC alleles generates a presentation likelihood that will result in the presentation of peptide p k It can be generated by subtracting the likelihood of simultaneously presenting both from the sum.

[0341] For example, the affine transformation function g h Using (·), peptide p is produced by HLA alleles h=2 and h=3 among the different identified HLA alleles of m=4. k The likelihood that will be presented is, It can be generated by TIFF2026113718000069.tif5128, in the formula, x2k , x3 k θ is an allele interaction variable identified for HLA alleles h=2 and h=3, and θ2 and θ3 are sets of parameters determined for HLA alleles h=2 and h=3.

[0342] As another example, the network transformation function g h (·), g w Using (·), peptide p is produced by HLA alleles h=2 and h=3 among the different identified HLA alleles of m=4. k The likelihood that will be presented is, This can be generated by TIFF2026113718000070.tif5128, where NN2(·) and NN3(·) are network models identified for HLA alleles h=2 and h=3, and θ2 and θ3 are sets of parameters determined for HLA alleles h=2 and h=3.

[0343] IX. Example 5: Prediction Module The prediction module 320 receives sequence data and uses a presentation model to select candidate nascent antigens from the sequence data. Specifically, the sequence data may be DNA sequences, RNA sequences, and / or protein sequences extracted from the patient's tumor tissue cells. The prediction module 320 uses the sequence data to select multiple peptide sequences having 8 to 15 amino acids for MHC-I, or 6 to 30 amino acids for MHC-II. k The process is performed as follows. For example, the prediction module 320 processes a predetermined array TIFF2026113718000071.tif5128 contains three peptide sequences, each with nine amino acids. The data can be processed as TIFF2026113718000072.tif9128. In one embodiment, the prediction module 320 can identify candidate nascent antigens, which are mutated peptide sequences, by comparing sequence data extracted from the patient's normal tissue cells with sequence data extracted from the patient's tumor tissue cells to identify regions having one or more mutations.

[0344] The prediction module 320 applies one or more presentation models to a processed peptide sequence to estimate the presentation likelihood of the peptide sequence. Specifically, the prediction module 320 can select one or more candidate nascent antigen peptide sequences that are likely to be presented on tumor HLA molecules by applying presentation models to candidate nascent antigens. In one implementation, the prediction module 320 selects candidate nascent antigen sequences that have an estimated presentation likelihood above a predetermined threshold. In another implementation, the presentation models select v candidate nascent antigen sequences with the highest estimated presentation likelihood (where v is generally the maximum number of epitopes that can be delivered in the vaccine). For a given patient, a vaccine containing the selected candidate nascent antigens can be injected into the patient to induce an immune response.

[0345] X. Example 6: Patient Selection Module The patient selection module 324 selects a subset of patients for vaccine therapy and / or T-cell therapy based on whether the patients meet selection criteria. In one embodiment, the selection criteria are determined based on the presentation likelihood of the patient's neogenic antigen candidates generated by the presentation model. By adjusting the selection criteria, the patient selection module 324 can adjust the number of patients who receive vaccine administration and / or T-cell therapy based on the presentation likelihood of the patient's neogenic antigen candidates. Specifically, strict selection criteria may result in a smaller number of patients treated with vaccine and / or T-cell therapy, but a higher proportion of patients treated with vaccine and / or T-cell therapy who receive effective treatment (e.g., one or more tumor-specific neogenic antigens (TSNAs) and / or one or more neogenic antigen-responsive T cells). Conversely, loose selection criteria may result in a larger number of patients treated with vaccine and / or T-cell therapy, but a lower proportion of patients treated with vaccine and / or T-cell therapy who receive effective treatment. The patient selection module 324 modifies the selection criteria based on a desired balance between the target proportion of patients receiving treatment and the proportion of patients receiving effective treatment.

[0346] In some embodiments, the selection criteria for patients receiving vaccine therapy are the same as those for patients receiving T-cell therapy. However, in alternative embodiments, the selection criteria for patients receiving vaccine therapy may differ from those for patients receiving T-cell therapy. Sections XA and XB below examine the selection criteria for patients receiving vaccine therapy and the selection criteria for patients receiving T-cell therapy, respectively.

[0347] Selection of patients for XA vaccine treatment In one embodiment, a patient is associated with a corresponding therapeutic subset of v nascent antigen candidates that could potentially be included in a personalized vaccine for that patient having a vaccine capacity v. In one embodiment, the therapeutic subset for a patient is the nascent antigen candidate with the highest presentation likelihood determined by the presentation model. For example, if the vaccine may contain v = 20 epitopes, the vaccine may contain the therapeutic subset for each patient with the highest presentation likelihood determined by the presentation model. However, it is recognized that in other embodiments, the therapeutic subset for a patient may also be determined based on other methods. For example, the therapeutic subset for a patient may be randomly selected from a set of nascent antigen candidates for that patient, or may be determined in part on a combination of prior art models that model the binding affinity or stability of peptide sequences, or on specific factors including presentation likelihoods obtained from presentation models and affinity or stability information regarding these peptide sequences.

[0348] In one embodiment, the patient selection module 324 determines that a patient meets the selection criteria if their tumor mutational burden is equal to or higher than the minimum mutational burden. A patient's tumor mutational burden (TMB) represents the total number of non-synonymous mutations in the tumor exome. In one embodiment, the patient selection module 324 selects patients to receive vaccine treatment if the absolute number of the patient's TMB is equal to or higher than a predetermined threshold. In another implementation, the patient selection module 324 selects patients to receive vaccine treatment if the patient's TMB falls within a threshold percentile among TMBs determined for the set of patients.

[0349] In another embodiment, the patient selection module 324 determines that a patient meets the selection criteria if the patient's utility score based on the patient's treatment subset is equal to or higher than the minimum utility score. In one embodiment, the utility score is a measure of the estimated number of presented antigens from the treatment subset.

[0350] The estimated number of presented antigens can be predicted by modeling the presentation of nascent antigens as random variables of one or more probability distributions. In one realized form, patient i's utility score is the expected number of presented nascent antigen candidates from the treatment subset, or a specific function thereof. As an example, the presentation of each nascent antigen can be modeled as a Bernoulli random variable whose probability of presentation (success) is given by the presentation likelihood of the nascent antigen candidate. In detail, each has the highest presentation likelihood u i1 u i2 , ..., u iv v types of new antigen candidates p i1 , p i2 ..., p iv Treatment subset S i Regarding the new antigen candidate p ij The presentation of the random variable A ij Given by, here, TIFF2026113718000073.tif5128 The expected number of nascent antigens presented is given by the sum of the presentation likelihoods of each nascent antigen candidate. In other words, patient i's utility score is expressed by the following formula: The patient selection module 324 in TIFF2026113718000074.tif15128 selects a subset of patients who have a utility score equal to or higher than the minimum utility score for vaccine treatment.

[0351] In another implementation, patient i's utility score is the probability that at least a threshold number of nascent antigens k are presented. In one example, the therapeutic subset S of candidate nascent antigens. i The number of antigens presented within is modeled as a Poisson binomial random variable, where the probability of presentation (success) is given by the presentation likelihood of each epitope. More specifically, the number of antigens presented by patient i is the random variable N. i It can be given by: In formula TIFF2026113718000075.tif13128, PBD(·) represents a Poisson binomial distribution. The probability that at least a threshold number of nascent antigens k are presented is given by the number of presented antigens N. i It is given by an operation with a probability that is equal to or greater than k. In other words, the utility score of patient i is expressed as follows: The patient selection module 324 in TIFF2026113718000076.tif13128 selects a subset of patients who have a utility score equal to or higher than the minimum utility score for vaccine treatment.

[0352] In another implementation, patient i's utility score is determined by a therapeutic subset S of nascent antigen candidates that have a binding affinity lower than a fixed threshold (e.g., 500 nM) or a predicted binding affinity for one or more patient HLA alleles. i This is the number of newly generated antigens. In one example, the fixed threshold is in the range of 1000 nM to 10 nM. Optionally, the utility score may count only the newly generated antigens detected as expressed by RNA-seq.

[0353] In another implementation, patient i's utility score is determined by a therapeutic subset S of nascent antigen candidates, where the binding affinity of one or more HLA alleles in that patient is below the threshold percentile of the binding affinity of a random peptide to those HLA alleles. i This is the number of newly generated antigens. In one example, the threshold percentile is in the range of the 10th percentile to the 0.1st percentile. Optionally, the utility score may count only the newly generated antigens detected as expressed by RNA-seq.

[0354] It should be noted that the examples of utility score values ​​described with respect to equations (25) and (27) are merely illustrative, and the patient selection module 324 can also generate utility score values ​​using other statistics or probability distributions.

[0355] Patient selection for XB T-cell therapy In another embodiment, a patient may receive T-cell therapy instead of, or in addition to, vaccine therapy. Similar to vaccine therapy, in embodiments where a patient receives T-cell therapy, the patient can be associated with a corresponding therapeutic subset of the v types of nascent antigen candidates described above. This therapeutic subset of v types of nascent antigen candidates can be used to in vitro identify patient-derived T cells that are responsive to one or more of the v types of nascent antigen candidates. These identified T cells can then be proliferated and injected into the patient in personalized T-cell therapy.

[0356] Patients receiving T-cell therapy can be selected at two different time points. The first time point is after a therapeutic subset of v candidate nascent antigens for the patient has been predicted using a model, but before in vitro screening of T cells specific to the predicted therapeutic subset of v candidate nascent antigens is performed. The second time point is after in vitro screening of T cells specific to the predicted therapeutic subset of v candidate nascent antigens is performed.

[0357] First, after predicting a therapeutic subset of v nascent antigen candidates for that patient, and before in vitro identification of patient-derived T cells specific to the predicted therapeutic subset of v nascent antigen candidates, patients to receive T cell therapy can be selected. In detail, since in vitro screening of patient-derived nascent antigen-specific T cells can be costly, it is desirable to select patients to screen for nascent antigen-specific T cells only if the patient is likely to have such cells. To select patients before the in vitro T cell screening step, the same criteria used to select patients for vaccine therapy can be used. In detail, in some embodiments, the patient selection module 324 can select patients to receive T cell therapy if the patient's tumor mutational burden is equal to or higher than the minimum mutational burden described above. In another embodiment, the patient selection module 324 can select patients to receive T cell therapy if the patient's utility score based on the therapeutic subset of v nascent antigen candidates for that patient is equal to or higher than the minimum utility score described above.

[0358] Secondly, in addition to selecting patients to receive T-cell therapy before in vitro identification of patient-derived T cells that are specific to the therapeutic subset of predicted v nascent antigen candidates, or alternatively, patients to receive T-cell therapy after in vitro identification of T cells that are specific to the therapeutic subset of predicted v nascent antigen candidates can be performed. More specifically, a patient may be selected to receive T-cell therapy if at least a threshold amount of nascent antigen-specific TCRs is identified for that patient in in vitro screening of the patient's T cells for nascent antigen recognition. For example, a patient may be selected to receive T-cell therapy only if at least two nascent antigen-specific TCRs are identified for that patient, or only if nascent antigen-specific TCRs are identified for two different nascent antigens.

[0359] In another embodiment, a patient may be selected to receive T-cell therapy only if a threshold amount of a nascent antigen from a therapeutic subset of v nascent antigen candidates for that patient is recognized by the patient's TCR. For example, a patient may be selected to receive T-cell therapy only if at least one nascent antigen from the therapeutic subset of v nascent antigen candidates for that patient is recognized by the patient's TCR. In a further embodiment, a patient may be selected to receive T-cell therapy only if at least a threshold amount of the TCR for that patient is identified as nascent antigen-specific for a particular HLA-restricted class of nascent antigen peptide. For example, a patient may be selected to receive T-cell therapy only if at least one TCR for that patient is identified as a nascent antigen-specific HLA class I restriction nascent antigen peptide.

[0360] In a further embodiment, a patient may be selected to receive T-cell therapy only if at least a threshold amount of a particular HLA-restrictive class of nascent antigen peptide is recognized by the patient's TCR. For example, a patient may be selected to receive T-cell therapy only if at least one HLA class I restriction nascent antigen peptide is recognized by the patient's TCR. In another example, a patient may be selected to receive T-cell therapy only if at least two HLA class II restriction nascent antigen peptides are recognized by the patient's TCR. Any combination of the above criteria may also be used to select patients to receive T-cell therapy after in vitro identification of T cells that are specific to a therapeutic subset of v predicted nascent antigen candidates for that patient.

[0361] XI. Example 7: Experimental results demonstrating exemplary patient selection performance The validity of patient selection described in Section X is validated by performing patient selection in a set of simulated patients, each associated with a test set of simulated neogeneic antigen candidates, where a subset of simulated neogeneic antigens is known to be presented in mass spectrometry data. Specifically, each simulated neogeneic antigen candidate in the test set is associated with a label indicating whether that neogeneic antigen is presented in the mass spectrometry data set of multi-allelic JY cell lines HLA-A*02:01 and HLA-B*07:02 from the Bassani-Sternberg dataset (dataset "D1") (data can be found at www.ebi.ac.uk / pride / archive / projects / PXD0000394). As described in detail below with Figure 13A, a large number of neogeneic antigen candidates for the simulated patients are sampled from the human proteome based on the known frequency distribution of mutational burden in non-small cell lung cancer (NSCLC) patients.

[0362] Allele-specific presentation models for the same HLA allele are trained using a training set that is a subset of mass spectrometry data for single alleles HLA-A*02:01 and HLA-B*07:02 from the IEDB dataset (dataset "D2") (the data can be found at http: / / www.iedb.org / doc / mhc_ligand_full.zip). In detail, the presentation model for each allele is trained using the network dependency function g, with the N-terminal and C-terminal flanking sequences as allele non-interaction variables. h (·) and g wThe allele-specific model shown in equation (8), which incorporates (·) and the expit function f(·), was adopted. The presentation model for allele HLA-A*02:01 generates the presentation likelihood for a specific peptide to be presented on allele HLA-A*02:01, given the peptide sequence as the allele interaction variable and the N-terminal and C-terminal flanking sequences as the allele non-interaction variable. The presentation model for allele HLA-B*07:02 generates the presentation likelihood for a specific peptide to be presented on allele HLA-B*07:02, given the peptide sequence as the allele interaction variable and the N-terminal and C-terminal flanking sequences as the allele non-interaction variable.

[0363] As disclosed with reference to Figures 13A–13E in the following example, different models, such as a presentation model trained for peptide bond prediction and a prior art model, are applied to test sets of nascent antigen candidates for each simulated patient to identify different therapeutic subsets for patients based on predictions. Patients who meet selection criteria for vaccine therapy are selected and associated with personalized vaccines containing epitopes in the patient therapeutic subsets. The size of the therapeutic subsets varies depending on the different vaccine doses. No overlap is introduced between the training sets used to train the presentation models and the test sets of simulated nascent antigen candidates.

[0364] The following example analyzes the proportion of selected patients who have at least a certain number of presented neonatal antigens among the epitopes included in the vaccine. This statistic demonstrates the effectiveness of the simulated vaccine in delivering potential neonatal antigens that can induce an immune response in patients. Specifically, simulated neonatal antigens in a given test set are presented if those neonatal antigens are presented in the mass spectrometry dataset D2. A high proportion of patients with presented neonatal antigens indicates the potential for therapeutic response with neonatal antigen vaccines by inducing an immune response.

[0365] XI.A. Example 7A: Frequency distribution of mutational burden in NSCLC cancer patients Figure 13A shows the sample frequency distribution of mutational burden in NSCLC patients. Mutational burden and mutations in different tumor types, including NSCLC, can be found, for example, in the Cancer Genome Atlas (TCGA) (https: / / cancergenome.nih.gov). The X-axis represents the number of non-synonymous mutations for each patient, and the Y-axis represents the proportion of sample patients with a particular number of non-synonymous mutations. The sample frequency distribution in Figure 13A ranges from 3 to 1786 mutations, with 30% of patients having fewer than 100 mutations. Although not shown in Figure 13A, studies have shown that mutational burden is higher in smokers compared to non-smokers, and that mutational burden can be a strong indicator of neo-antigen burden in patients.

[0366] As introduced at the beginning of Section XI above, each of the simulated patient numbers is associated with a test set of nascent antigen candidates. For each patient, the test set is derived from the frequency distribution shown in Figure 13A, which is used to determine the mutation load m i It is generated by sampling. For each mutation, 21-mer peptide sequences derived from the human proteome are randomly selected to represent the mutant sequence to be simulated. A test set of nascent antigen candidate sequences is generated for patient i by identifying the peptide sequences of each (8, 9, 10, 11)-mer across the mutations within the 21-mer. Each nascent antigen candidate is associated with a label indicating whether the nascent antigen candidate sequence is present in the mass spectrometry D1 dataset. For example, a nascent antigen candidate sequence present in dataset D1 can be associated with the label "1", and a sequence not present in dataset D1 can be associated with the label "0". As will be described in more detail below, Figures 13B-13E show the experimental results of patient selection based on the nascent antigens presented by patients in the test set.

[0367] XI.B. Example 7B: Percentage of selected patients with nascent antigen presentation based on mutagenicity selection criteria Figure 13B shows the number of presented neonatal antigens in the simulated vaccine for patients selected based on the selection criteria of whether the patient meets the minimum mutagenesis requirement. The proportion of selected patients with at least a certain number of presented neonatal antigens in the corresponding trial is identified.

[0368] In Figure 13B, the x-axis represents the proportion of patients excluded from vaccine therapy based on tumor mutational burden, labeled “Minimum Number of Mutations.” For example, the data point at “Minimum Number of Mutations” 200 indicates that the patient selection module 324 selected only a subset of simulated patients with a mutational burden of at least 200 mutations. As another example, the data point at “Minimum Number of Mutations” 300 indicates that the patient selection module 324 selected a lower proportion of simulated patients with at least 300 mutations. The y-axis represents the proportion of selected patients associated with at least a certain number of presented neogeneic antigens in a test set without vaccine dose v. More specifically, the top plot shows the proportion of selected patients presenting at least one neogeneic antigen, the middle plot shows the proportion of selected patients presenting at least two antigens, and the bottom plot shows the proportion of selected patients presenting at least three antigens.

[0369] As shown in Figure 13B, the proportion of patients with the presented neogeneic antigen increased significantly with increasing mutational burden. This suggests that mutational burden as a selection criterion may be effective in selecting patients who are more likely to receive an effective immune response from the neogeneic antigen vaccine.

[0370] XI.C. Example 7C: Comparison of neoantigen presentation in vaccines identified by presentation model versus conventional technology model. Figure 13C compares the number of presented neogeneic antigens in simulated vaccines between selected patients associated with a vaccine containing a therapeutic subset identified based on the presentation model and selected patients associated with a vaccine containing a therapeutic subset identified by a conventional model. The plot on the left assumes a limited vaccine volume of v=10, and the plot on the right assumes a limited vaccine volume of v=20. Patients are selected based on a utility value score indicating the expected number of presented neogeneic antigens.

[0371] In Figure 13C, the solid line shows patients associated with the vaccine containing a therapeutic subset identified based on presentation models for alleles HLA-A*02:01 and HLA-B*07:02. The therapeutic subset for each patient is identified by applying each of the presentation models to the sequences in the test set and identifying v nascent antigen candidates with the highest presentation likelihood. The dotted line shows patients associated with the vaccine containing a therapeutic subset identified based on the conventional model NETMHCpan for single allele HLA-A*02:01. Details of the implementation of NETMHCpan are available at http: / / www.cbs.dtu.dk / services / NetMHCpan. The therapeutic subset for each patient is identified by applying the NETMHCpan model to the sequences in the test set and identifying v nascent antigen candidates with the highest estimated binding affinity. The x-axis in both graphs shows the proportion of patients excluded from vaccine treatment based on expected utility score, which indicates the expected number of presented nascent antigens in the therapeutic subset identified based on the presentation model. The expected utility score is determined as described in Section X in relation to Equation (25). The y-axis represents the proportion of selected patients who present at least a certain number of neogenic antigens (1, 2, or 3 types of neogenic antigens) contained in the vaccine.

[0372] As shown in Figure 13C, patients associated with a vaccine containing a therapeutic subset based on the presentation model receive a significantly higher proportion of the vaccine containing the presented neogenic antigen compared to patients associated with a vaccine containing a therapeutic subset based on a conventional technology model. For example, as shown in the graph on the right, 80% of selected patients associated with the presentation model-based vaccine receive at least one presented neogenic antigen in the vaccine, compared to only 40% of selected patients associated with the conventional technology model-based vaccine. These results demonstrate that the presentation model described herein is effective in selecting neogenic antigen candidates for vaccines that are likely to induce an immune response to treat tumors.

[0373] XI.D. Example 7D: Effect of HLA coverage on nascent antigen presentation of vaccines identified by presentation model Figure 13D compares the number of nascent antigens presented in simulated vaccines between selected patients associated with a vaccine containing a therapeutic subset identified based on a single-allelic presentation model for HLA-A*02:01 and selected patients associated with a vaccine containing therapeutic subsets identified based on both allelelic presentation models for HLA-A*02:01 and HLA-B*07:02. The vaccine volume is set to v=20 epitopes. For each experiment, patients are selected based on expected utility scores determined based on different therapeutic subsets.

[0374] In Figure 13D, the solid line represents patients associated with a vaccine containing a treatment subset based on both presentation models for the HLA alleles HLA-A*02:01 and HLA-B*07:02. The treatment subset for each patient is identified by applying each of the presentation models to the sequences in the test set and identifying v neonatal antigen candidates with the highest presentation likelihood. The dotted line represents patients associated with a vaccine containing a treatment subset based on a single presentation model for the HLA allele HLA-A*02:01. The treatment subset for each patient is identified by applying the presentation model for only a single HLA allele to the sequences in the test set and identifying v neonatal antigen candidates with the highest presentation likelihood. In the solid line plot, the x-axis shows the proportion of patients excluded from vaccine treatment based on the expected utility score for the treatment subsets identified by both presentation models. In the dotted line plot, the x-axis shows the proportion of patients excluded from vaccine treatment based on the expected utility score for the treatment subset identified by a single presentation model. The y-axis represents the proportion of selected patients who present at least a specific number of neogenic antigens (1, 2, or 3 different neogenic antigens).

[0375] As shown in Figure 13D, patients associated with a vaccine containing a therapeutic subset identified by presentation models for both HLA alleles presented a significantly higher proportion of nascent antigens than patients associated with a vaccine containing a therapeutic subset identified by a single presentation model. These results highlight the importance of establishing presentation models with high HLA allele coverage.

[0376] XI.E. Example 7E: Comparison of neogenic antigen presentation in patients selected based on mutational load versus expected number of presented neogenic antigens. Figure 13E compares the number of nascent antigens presented in the simulated vaccine between patients selected based on mutational burden and patients selected based on expected utility score. The expected utility score is determined based on a therapeutic subset identified by a presentation model with a size of v=20 epitopes.

[0377] In Figure 13E, the solid line represents patients selected based on the expected utility score associated with the vaccine containing the treatment subset identified by the presentation model. The treatment subset for each patient is identified by applying each of the presentation models to the sequences in the test set and identifying v=20 neonatal antigen candidates with the highest presentation likelihood. The treatment utility score is determined based on the presentation likelihood of the treatment subset identified in Section X based on Equation (25). The dotted line represents patients selected based on the mutagenesis load associated with the vaccine containing the treatment subset identified by the presentation model. The x-axis shows the proportion of patients excluded from vaccine treatment based on the expected utility score in the solid line plots and the proportion of patients excluded based on the mutagenesis load in the dotted line plots. The y-axis shows the proportion of selected patients who receive a vaccine containing at least a certain number of presented neonatal antigens (1, 2, or 3 neonatal antigens).

[0378] As shown in Figure 13E, patients selected based on expected utility score receive a higher proportion of vaccines containing the presented nascent antigen than patients selected based on mutational burden. However, patients selected based on mutational burden receive a higher proportion of vaccines containing the presented nascent antigen than patients not selected. Therefore, mutational burden is an effective patient selection criterion in effective delayed nascent antigen vaccines, but the expected utility score is more effective.

[0379] XII. Example 8: Evaluation of a mass spectrometry training model for excluded (held-out) mass spectrometry data Because HLA peptide presentation by tumor cells is a major requirement in anti-tumor immunity. 91,96,97 A large (N=74 patients) integrated dataset of human tumor and normal tissue samples, including paired class I HLA peptide sequences, HLA types, and transcriptome RNA-seq (method), was used to predict antigen presentation in human cancers. This data, along with published data, was used to predict antigen presentation in human cancers. 92,98,99 We train a novel deep learning model using this method.100 The study was conducted with the objective of generating the necessary samples. Samples were selected from several tumor types targeted for immunotherapy development based on tissue availability. Mass spectrometry identified an average of 3704 peptides per sample with a peptide level FDR of less than 0.1 (ranging from 344 to 11301 peptides). These peptides ranged in length from 8 to 15 amino acids and followed a characteristic distribution of class I HLA lengths with a modal length of 9 (56% of peptides). Consistent with previous reports, the majority of peptides (median 79%) were predicted by MHCflurry to bind to at least one patient HLA allele at a standard affinity threshold of 500 nM. 90 However, considerable variability was observed between samples (for example, in one sample, the predicted affinity was over 500 nM for 33% of the peptides). 101 The 50 nM threshold for "strong binding" captured only a median of 42% of the presented peptides. Transcriptome sequencing yielded an average of 131 M unique reads per sample, and 68% of genes were expressed at least 1 TPM (transcript per million) in at least one sample, highlighting the value of large and diverse samples set up to observe the expression of the maximum number of genes. Peptide presentation by HLA was strongly correlated with mRNA expression. Significant and reproducible differences in the rate of peptide presentation were observed between genes, larger than those explained by differences in RNA expression or sequence alone. The observed HLA types were consistent with expectations for samples from a patient population with primarily European ancestry.

[0380] These, and published HLA peptide data 92,98,99A neural network (NN) model was trained to predict HLA antigen presentation using [a specific method / tool]. A novel network architecture (method) was developed that allows for the simultaneous learning of allele-peptide mapping and allele-specific presentation motifs to learn allele-specific models from tumor mass spectrometry data, where each peptide may be presented by one of six HLA alleles. For each patient, data points shown as positive were peptides detected by mass spectrometry, and data points shown as negative were peptides from the reference proteome (SwissProt) that were not detected by mass spectrometry in that sample. The data was divided into training, validation, and test sets (method). The training set consisted of 142,844 HLA-presenting peptides (FDR < approximately 0.02) obtained from 101 samples (69 newly described in this study and 32 previously published). The validation set (used for early termination) consisted of 18,004 presenting peptides from the same 101 samples. The following two mass spectrometry datasets were used in the study: (1) a tumor sample test set consisting of 571 presentation peptides obtained from five additional tumor samples (two lungs, two colons, and one ovary) excluded from the training data; and (2) a single-allelic cell line test set consisting of 2,128 presentation peptides from genomic location windows (blocks) adjacent to (but different from) the locations of single-allelic peptides included in the training data (see "Methods" for further details on training / test splitting).

[0381] The training data identified predictive models for 53 HLA alleles. 92,104Unlike previous models, these models captured the dependence of HLA presentation on each sequence position in peptides of multiple lengths. This model appropriately learned critical dependencies on gene RNA expression and gene-specific presentation tendencies, and when mRNA abundance and learned presentation tendencies per gene were independently combined, it produced up to a 60-fold difference in presentation rates between genes with the lowest expression levels and those with the highest expression levels and those with the highest presentation rates. This model also demonstrated the measured stability of HLA / peptide complexes in the IEDB, even after controlling for predicted binding affinity. 88 Further prediction was observed (p<1e-10 for 10 alleles) (p<0.05 for 8 of the 10 alleles tested). Taken together, these properties form the basis for improved prediction of immunogenic HLA class I peptides.

[0382] We evaluated the performance of this NN model as a predictive tool for HLA presentation on an excluded mass spectrometry test set, and the state-of-the-art binding affinity prediction tool MHCFlurry, a neural network tool trained on in vitro HLA binding data. 90 This was compared with (version 1.2.0, methods). Based on previous reports emphasizing the importance of mRNA levels in HLA presentation, RNA-seq 81,92,103 We incorporated an increased threshold for gene expression assayed by this method.

[0383] Figures 14A-D compare the predictive performance of the "Full MS Model," the "Peptide MS Model," and the MHCFlurry1.2.0 binding affinity model with three gene expression thresholds. Both the "Full MS Model" and the "Peptide MS Model" are neural network models trained on the mass spectrometry data described above. However, while the "Full MS Model" is trained and tested based on all characteristics of the sample, the "Peptide MS Model" is trained and tested based only on the sample's HLA type and peptide sequence. Three different versions of the MHCFlurry1.2.0 binding affinity model are tested: one with a gene expression threshold of TPM>0, one with a gene expression threshold of TPM>1, and one with a gene expression threshold of TPM>2. Both the "peptide MS model" and the MHCFlurry1.2.0 binding affinity model, which has a gene expression threshold of TPM > 1, are trained and tested based solely on the HLA type and peptide sequence of the sample, and both have the same RNA expression threshold. Therefore, by directly comparing the performance of these two models, the improvement in prediction due to the difference in peptide motifs learned from mass spectrometry data compared to the binding affinity training data can be quantified.

[0384] Referring first to Figure 14A, Figure 14A compares the positive predictive value (PPV) at 40% recall for the “complete MS model,” the “peptide MS model,” and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2 when testing each model with a test set consisting of five different test samples, including an excluded tumor sample where each test sample had a presented peptide to non-presented peptide ratio of 1:2500 (Methods). Figure 14A also shows the mean PPV at 40% recall for the “complete MS model,” the “peptide MS model,” and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2 for the five test samples. As shown in Figure 14A, the mean PPV at a 40% recall was 0.54 for the "complete MS model" and 0.076, 0.072, and 0.061 for the MHCFlurry1.2.0 binding affinity models with gene expression thresholds of TPM > 2, 1, and 0, respectively. The comparisons between the "complete MS model" and the MHCFlurry1.2.0 binding affinity models with a gene expression threshold of TPM > 0 were all statistically significant (p < 1e-6).

[0385] Next, referring to Figure 14B, Figure 14B compares the PPV at 40% recall for the “Full MS model,” the “Peptide MS model,” and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2 when testing each model with a test set consisting of 15 different test samples, each containing excluded peptides from a single-allelic cell line test dataset where each test sample has a presented peptide-to-non-presented peptide ratio of 1:10,000 (Methods). Figure 14B also shows the mean PPV at 40% recall for the “Full MS model,” the “Peptide MS model,” and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2 for the 15 different test samples. As shown in Figure 14B, the mean PPV at a 40% recall was 0.37 for the "complete MS model" and 0.094, 0.090, and 0.071 for the MHCFlurry1.2.0 binding affinity models with gene expression thresholds of TPM > 2, 1, and 0, respectively. The comparisons between the "complete MS model" and the MHCFlurry1.2.0 binding affinity models with a gene expression threshold of TPM > 0 were all statistically significant (p < 1e-6), except for the test sample containing HLA-A*01:01, where p = 1.6e-4.

[0386] Figure 16 compares the positive predictive value (PPV) at 40% recall for the "full MS model" and the "anchor residue-only MS model" when each model is tested with the test set described above with respect to Figure 14A (Methods). Figure 16 also shows the average PPV at 40% recall for the "full MS model" and the "anchor residue-only MS model" for five different test samples. Like the "full MS model," the "anchor residue-only MS model" is a neural network model trained by mass spectrometry as described above. However, instead of being trained and tested based on the entire peptide sequence in the sample, the "anchor residue-only MS model" is trained and tested based only on the "anchor" residues (first, second, and last residues) of the peptide sequence in the sample. Therefore, the results shown in Figure 16 quantify the relative importance of anchor residues and non-anchor residues to the predictive performance of the models. As shown in Figure 16, the performance of the "anchor residue-only MS model" is significantly lower compared to the full MS model. The mean PPV at 40% recall in MS models with only anchor residues is 0.13, compared to 0.50 in full MS models. Therefore, it can be inferred that training and testing models that include non-anchored sequences in the peptide sequence will lead to improved predictive capabilities.

[0387] Figure 17A shows the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when each model is tested with test sample 0 from Figure 14A (Methods). As shown in Figure 17A, the "Full MS model" and the "Peptide MS model" perform better than the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2.

[0388] Figure 17B compares the PPV at 40% recall for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds (TPM > 0, 1, and 2) when testing each model with a test set consisting of 15 different test samples, each containing excluded peptides from a single-allelic cell line test dataset where each test sample has a 1:5,000 ratio of presented peptides to non-presented peptides (Methods). Figure 17B also shows the mean PPV at 40% recall for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds (TPM > 0, 1, and 2) for the 15 different test samples. By comparing the results in Figure 14B (each test sample includes excluded peptides from a single-allelic cell line test dataset with a ratio of 1:10,000 presented peptides to non-presented peptides) with the results in Figure 17A (each test sample includes excluded peptides from a single-allelic cell line test dataset with a ratio of 1:5,000 presented peptides to non-presented peptides), it can be inferred that the abundance of peptide presentation is strongly correlated with absolute PPV. In general, the lower the abundance of a predicted event (e.g., presentation), the more difficult it is to obtain a high PPV prediction. Therefore, decreasing (increasing) the abundance in the test data will decrease (increase) the absolute PPV for all models. However, the relative differences between PPVs of different models are not affected by changes in the expected abundance in the test set.

[0389] Figures 17C-G show the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds (TPM > 0, 1, and 2) when each model is tested with test samples 0-4 from Figure 14A (Methods).

[0390] Figures 17H-V show the complete precision-recall curves for the "Full MS model," the "Peptide MS model," and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM > 0, 1, and 2, when each model is tested with 15 different test samples from Figure 14B, each test sample containing excluded peptides from a single-allelic cell line test dataset where the ratio of presented peptides to non-presented peptides is 1:10,000 (Methods).

[0391] Figure 18 shows different versions of the MS model and initial approach to modeling HLA-presenting peptides in human tumors when each model is tested with five different test samples from Figure 14A. 104 This compares the positive predictive value (PPV) at a 40% recall (Methods). Figure 18 also shows the average PPV at a 40% recall for each model for five different test samples. The models tested in Figure 18 include the "Full MS Model," "MS Model, No Flanking Sequences," "MS Model, No Flanking Sequences or Per-Genome Coefficients," "Peptide-Only MS Model, All Lengths Trained Together," "Peptide-Only MS Model, All Lengths Trained Separately," "Linear Peptide-Only MS Model," "MixMHCPred 1.1" Model, and the "Binding Affinity" Model. The "Full MS Model," "MS Model, No Flanking Sequences," "MS Model, No Flanking Sequences or Per-Genome Coefficients," "Peptide-Only MS Model, All Lengths Trained Together," "Peptide-Only MS Model, All Lengths Trained Separately," and "Linear Peptide-Only MS Model" are all neural network models trained using mass spectrometry as described above. However, each model is trained and tested using different characteristics of the sample. The "MixMHCPred 1.1" model and the "binding affinity" model are early approaches to modeling HLA-presenting peptides. 104 .

[0392] Overall, the NN model achieved significantly improved prediction of HLA peptide presentation, with PPV up to 9 times higher than standard binding affinity + gene expression in tumor test sets (Figure 14A) and up to 5 times higher in single-allelic datasets (Figure 14B). The large PPV advantage of the MS-based NN model was maintained across different recall rates (Figure 17A) and was statistically significant (p<10 for all tumor samples in Figures 14A and 14B and for single-allelic samples except HLA-A*01:01, where p=1.6e-4). -6 The positive predictive value for standard binding affinity + gene expression for HLA peptide presentation reached a low value of 6%, which was consistent with previous estimates. 87,93 However, it should be noted that since only a small percentage of peptides are detected as presented (for example, about 1 out of 2500 in the tumor MS trial dataset), this approximately 6% PPV represents an improvement of more than 100-fold compared to the baseline prevalence.

[0393] By comparing a reduced model ("Peptide MS Model," see Figures 14A-B and Methods) trained on mass spectrometry data using only HLA type and peptide sequence as input with a full MS model, it was determined that approximately 30% of the increase in PPV relative to binding affinity prediction was due to the modeling of exogenous peptide characteristics (RNA abundance, flanking sequence, gene-specific coefficients) that can be captured by mass spectrometry but not by binding affinity assays (see also Figures 14A-B, 17A, and 18). The remaining 70% of the increase was due to improved modeling of the peptide sequence (Figures 14A-B). This modeling represents an early approach to modeling HLA-presenting peptides in human tumors. 104 The improved performance was attributed not only to the properties of the training dataset (HLA-presenting peptides) but also to the overall model architecture (Figure 18). This new model architecture allows for the development of binding affinity prediction or hard clustering approaches. 104~106This enables the training of allele-specific models through an end-to-end training process that does not require prior assignment of peptides to the expected presented alleles. Importantly, the new model architecture also does not impose constraints that degrade the accuracy of allele-specific submodels as prerequisites for deconvolution, such as linearity, and does not consider each peptide length separately. 104 It wasn't even necessary. The complete model outperformed several simplified models that imposed these constraints, as well as conventionally published approaches (Figure 18).

[0394] Figure 18 shows the performance of several simplified models on the MS test set. The relative importance of modeling improvements incorporated into the full model is quantified by testing the predictive performance on the MS test set by removing the modeling improvements one at a time. Furthermore, a comparison was made between the models presented herein and a recently published approach (MixMHCPred) for modeling peptides eluted from mass spectrometry. Since MixMHCPred does not currently model peptides of lengths other than 9 and 10, only 9-mer and 10-mer peptides were used for comparison. The models are (from left to right): "Full MS Model" (the full NN model described in "Methods"), "MS Model, No Flanking Sequences" (same as the full NN model except that the flanking sequence characteristics are excluded), "MS Model, No Flanking Sequences or Gene-Specific Coefficients" (same as the full NN model except that the flanking sequence characteristics and gene-specific coefficients are excluded), "Peptide-Only MS Model, All Lengths Trained Together" (same as the full NN model except that the only characteristics used are peptide sequences and HLA types), "Peptide-Only MS Model, Each Length Trained Separately" (the model structure is the same as the peptide-only MS model except that separate models were trained for 9-mer and 10-mer), "Linear Peptide-Only MS Model (with Ensemblement)" (same as the peptide-only MS model where each peptide length is trained separately, except that instead of modeling peptide sequences using a neural network, an ensemble of linear models used in the full model and trained using the same optimization procedure as described in "Methods") is used); "MixMHCPred "1.1" is the default setting MixMHCPred; "Binding Affinity" is the same MHCflurry 1.2.0 as in the text. The last five models ("Peptide-only MS model, train all lengths together" to "Binding Affinity") have the same input of peptide sequence and HLA type only. In detail, none of the last five models use RNA abundance in making predictions.The best-performing peptide-only model ("Peptide-only MS model, all lengths trained together") yielded an average PPV of 0.41 at a 40% recall, compared to a mere 28% average PPV for the worst-performing peptide-only model trained on mass spectrometry data ("Linear peptide-only MS model (with ensemble)") (only slightly higher than MixMHCpred's average PPV of 18%), highlighting the improved neural network modeling capabilities of peptide sequences. Note that MixMHCpred is trained on different data than the linear peptide-only MS model, but shares many of the same modeling characteristics (e.g., it is a linear model where each peptide length is trained separately).

[0395] XIII. Example 9: Retrospective Model Evaluation of a Newly Generated Antigen T Cell Day Applicator The inventors evaluated whether such accurate prediction of HLA peptide presentation leads to the ability to identify epitopes (i.e., targets for immunotherapy) on human tumor CD8 T cells. A suitable test dataset for this evaluation includes peptides that are recognized by T cells and presented by HLA on the surface of tumor cells. Furthermore, a formal performance evaluation requires not only peptides that are shown as positive (i.e., recognized by T cells) but also a sufficient number of negatively labeled peptides (i.e., tested but not recognized). Mass spectrometry datasets correspond to tumor presentation but not to T cell recognition, and conversely, post-vaccination priming or T cell assays correspond to the presence of T cell progenitor cells and T cell recognition but not to tumor presentation. For example, potent HLA-binding peptides whose source genes are expressed at low levels within the tumor may produce a strong post-immunoCD8 T cell response that has no therapeutic utility because such peptides are not presented by the tumor.

[0396] To obtain a suitable dataset, we collected publicly available CD8 T cell epitopes from the following four recent studies that met the necessary criteria: Study A 96This study investigated TILs in nine patients with gastrointestinal tumors and reported the recognition of 12 T cells out of 1,053 somatic SNV mutations tested using IFN-γELISPOT with the tandem mini-gene (TMG) method in autologous dendritic cells (DCs). Study B 107 Using TMG, they also reported the recognition of 6 T cells out of 574 SNVs obtained from CD8+PD-1+ circulating lymphocytes from 4 melanoma patients. Research C 97 This study evaluated TILs obtained from three melanoma patients using pulsed peptide stimulation and found responses to 5 out of 381 tested SNV mutations. (Study D) 108 This study used a combination of TMG assays to evaluate TILs obtained from one breast cancer patient, pulsed with minimal epitope peptides, and reported recognition of 2 out of 62 SNVs. The combined dataset consisted of 2,009 assayed SNVs from 17 patients, including 26 nascent antigens exhibiting pre-existing T-cell responses. Importantly, since the dataset largely consists of recognition of nascent antigens by tumor-infiltrating lymphocytes, effective prediction is not supported by conventional literature. 81,82,97 As described, this suggests the ability to identify not only nascent antigens that can prime T cells, but more precisely, nascent antigens presented to T cells by tumors.

[0397] To simulate antigen selection for personalized immunotherapy, somatic mutations were ranked in order of presentation probability using a "full MS model," a "peptide MS model," and an MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds: TPM > 0, 1, and 2. Antigen-specific immunotherapy is technically limited in the number of specificities that can be targeted (for example, current personalized vaccines encode approximately 10-20 somatic mutations). 80~82The prediction methods were compared by counting the number of existing T cell responses in somatic mutations ranked in the top 5, 10, or 20 for each patient showing at least one existing T cell response. These results are shown in Figure 14C. In detail, Figure 14C compares the proportion of somatic mutations recognized by T cells (e.g., existing T cell responses) for the top 5, 10, and 20 somatic mutations identified by the “Full MS model,” the “Peptide MS model,” and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds of TPM>0, 1, and 2, for a test set consisting of 12 different test samples, each taken from patients showing at least one existing T cell response. The comparisons between the “Full MS model” and the MHCFlurry1.2.0 binding affinity model with a gene expression threshold of TPM>0 are both statistically significant at p<0.005, except for the top 5 somatic mutations, where p=0.056.

[0398] As expected, the binding affinity predictions included only a small fraction of the pre-existing T cell responses between the prioritized mutations, e.g., 9 out of 26 (35%) of the top 20 ranked mutations with TPM>0 (Supplementary Table 1). In contrast, the majority of pre-existing T cell responses (19 / 26, 73%) were ranked in the top 20 by the full MS model, with the mean preserved across different ranks and gene expression thresholds (Figure 14C, Supplementary Table 1). At the patient level, the full MS model showed an average of 1.54 pre-existing nascent antigen T cell responses across the top 20 predicted mutations in 13 patients with at least one pre-existing T cell response, compared to only 0.69 (p=1.4e-4) for binding affinity between TPM>0.

[0399] Next, the inventors evaluated mutations at the minimally nascent epitope level (i.e., 8–11 mers overlapping with the mutation were recognized) to see if they could be useful in identifying T cells / TCRs for cell therapy. In other words, minimally nascent epitopes were ranked in order of presentation probability using a “complete MS model,” a “peptide MS model,” and an MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds: TPM > 0, 1, and 2. As described above, antigen-specific immunotherapy is technically limited in the number of specific targets, so predictive methods were compared by counting the number of existing T cell responses in the top 5, 10, or 20 ranked minimally nascent epitopes for each patient showing at least one existing T cell response. Epitopes showing as positive were confirmed to be immunogenic minimal epitopes by peptide-based assays (in lieu of or in addition to TMG-based assays), while negative results included all epitopes not recognized by peptide-based assays and epitopes across all mutations in unrecognized minigenes. These results are shown in Figure 14D.

[0400] In detail, Figure 14D compares the proportion of minimally nascent epitopes recognized by T cells (e.g., pre-existing T cell responses) for a test set consisting of 12 different test samples, each taken from patients exhibiting at least one pre-existing T cell response, using the “complete MS model,” the “peptide MS model,” and the MHCFlurry1.2.0 binding affinity model with three different gene expression thresholds: TPM>0, 1, and 2. The comparisons between the “complete MS model” and the MHCFlurry1.2.0 binding affinity model with a TPM>0 gene expression threshold are all statistically significant (p<0.05), except for the top 5 minimally nascent epitopes (p=0.082). Error bars in all panels represent 90% confidence intervals.

[0401] As shown in Figure 14D, the advantage of the NN model for binding affinity at TPM > 0 was even more pronounced than in Figure 14C, with at least four times more nascent epitopes included among the top-ranked minimal epitopes. It is noteworthy that studies A, B, and D individually tested only peptides with strong binding affinity, thus biasing this comparison in favor of predicting binding affinity. There may be T cell-recognizing peptides with weakly predicted HLA binding affinity that would have been selected according to this model, although they were not assayed in these studies. Such peptides were observed in this study and are discussed in detail below with respect to Figure 15A and Supplementary Table 3.

[0402] Known limitations of mass spectrometry in the detection of cysteine-containing peptides 92,104 Nevertheless, the NN model outperformed the NN model in predicting binding affinity for cysteine-containing T cell-recognized epitopes, ranking 3 out of 7 of the top 5 cysteine-containing epitopes (43%) compared to 1 out of 7 of the top 5 in binding affinity with a gene expression threshold of TPM > 0. Similar to the mass spectrometry test set, additional properties that can be modeled based on mass spectrometry training data (RNA, flanking sequences, gene-specific coefficients) contributed significantly to the increased predictive performance. However, as with the mass spectrometry test data, the predictive performance of the peptide-only MS model was significantly improved compared to binding affinity prediction, and it was shown that the majority of this improvement was due to improved modeling of the peptide sequences (Figures 14C-D, compare the light blue and green bars).

[0403] Notably, this improvement was observed despite the potential increase in false negatives of test set neonatal epitopes due to the limitations of the current TIL assay (i.e., when a T cell response is not detected for a neonatal epitope presented by a tumor that is recognizable by T cells). These limitations include (a) the immunosuppressive tumor microenvironment and inefficient T cell priming, (b) neonatal epitope-responsive T cell depletion, (c) production of cytokines other than IFNg by TIL, and (d) heterogeneity of the tumor fraction used. Therefore, the absolute predictive performance based on the number of the top 5–20 immunogenic peptides described herein may be pessimistic for other situations, such as the administration of a potent neonatal antigen cancer vaccine.

[0404] XIII.A. Data The present inventors, Gros et al. 84 Tran et al 140 Stronen et al 141 Mutation calling, HLA type, and T cell recognition data were obtained from supplementary information by Zacharakis et al. Patient-specific RNA-seq data were not available. Assuming that tumor RNA expression correlates with the same tumor type in different patients, RNA-seq data from tumor-type matched patients obtained from TCGA was used as a substitute and was used in both neural network prediction and RNA expression filtering with TPM > 1 before binding affinity prediction. Predictive performance was improved by adding tumor-type matched RNA-seq data (Figure 14C-D).

[0405] In the mutation level analysis (Figure 14C), data points shown as positive in Gros et al., Tran et al., and Zacharakis et al. were mutations recognized by patient T cells in both the TMG assay and the minimal epitope peptide pulse assay. Data points shown as negative were all other mutations tested in the TMG assay. In Stronen et al., mutations shown as positive were mutations that spanned at least one recognized peptide, and negative data points were all mutations that were tested but not recognized in the tetramer assay. Since the mutated 25-mer TMG assay tests T cell recognition of all peptides spanning the mutation, for Gros, Tran, and Zacharakis data, mutations were ranked by summing the probability of presentation across all peptides spanning the mutation or by taking the minimum binding affinity. For Stronen data, mutations were ranked by summing the probability of presentation across all peptides spanning the mutation tested in the tetramer assay or by taking the minimum binding affinity. A complete list of mutations and characteristics is shown in Supplementary Table 1.

[0406] In epitope-level analysis, positive data points were defined as all minimal epitopes recognized by patient T cells in the peptide pulse assay or tetramer assay, while negative data points were defined as all minimal epitopes not recognized by T cells in the peptide pulse assay or tetramer assay, and all peptides spanning mutations from the tested TMG that were not recognized by patient T cells. In the cases of Gros et al., Tran et al., and Zacharakis et al., minimal epitope peptides spanning mutations recognized in TMG analysis that were not tested by the peptide pulse assay were excluded from the analysis because the T cell recognition status of these peptides could not be experimentally investigated.

[0407] XIV. Example 10: Identification of newly generated antigen-responsive T cells in cancer patients This example demonstrates that improved prediction enables the identification of nascent antigens from typical patient samples. To do this, archived FFPE tumor biopsies and 5–30 ml of peripheral blood were analyzed from nine patients with metastatic NSCLC receiving anti-PD(L)1 therapy (Supplementary Table 2: Patient demographics and treatment information for N=9 patients examined in Figures 15A–C. Key fields include tumor stage and subtype, anti-PD1 therapy administered, and an overview of NGS results). Whole tumor exome sequencing, tumor transcriptome sequencing, and matched normal exome sequencing yielded an average of 198 somatic mutations (SNVs and short indels) per patient, of which an average of 118 were expressed ("Methods," Supplementary Table 2). Twenty nascent epitopes per patient were prioritized for testing against existing anti-tumor T cell responses using a complete MS model. To focus the analysis on the most likely CD8 response, prioritized peptides were synthesized as minimal 8-11 M-epitopes ("Methods"), and then peripheral blood mononuclear cells (PBMCs) were cultured with the synthesized peptides in short-term in vitro stimulation (IVS) cultures to proliferate nascent antigen-responsive T cells (Supplementary Table 3). After two weeks, the presence of antigen-specific T cells was evaluated using IFN-γELISpot against the prioritized nascent epitopes. Further separate experiments were performed in seven patients with sufficient PBMCs to perform complete or partial convolution of recognized specific antigens. These results are shown in Figures 15A-C and 19A-22.

[0408] Figure 15A shows the detection of T cell responses to patient-specific nascent antigen peptide pools in nine patients. For each patient, the predicted nascent antigens were combined into two pools of 10 peptides each, according to model ranking and arbitrary sequence homology, respectively (homologous peptides were split into different pools). Then, for each patient, PBMCs grown in vitro for that patient were stimulated with the two patient-specific nascent antigen peptide pools using IFN-γELISpot. The data in Figure 15A are from 10 seeded cells with background (corresponding DMSO-negative control) subtracted. 5The values ​​are expressed as spot-forming units (SFUs) per cell. Background measurements (DMSO-negative control) are shown in Figure 22. For patients 1-038-001, 1-050-001, 1-001-002, CU04, 1-024-001, 1-024-002, and CU05, the response to congeneral peptide pools #1 and #2 is shown for single wells (patients 1-038-001, CU02, CU03, and 1-050-001) or replicates (all other patients) including mean and standard deviation. For patients CU02 and CU03, testing was only possible against specific peptide pool #1 due to cell count. Samples with an increase factor more than twice that of the background were considered positive and are indicated with an asterisk (responsive donors include patients 1-038-001, CU04, 1-024-001, 1-024-002, and CU02). Non-responsive donors include patients 1-050-001, 1-001-002, CU05, and CU03. Figure 15C shows a photograph of ELISpot wells in IFN-γELISpot containing in vitro proliferated PBMCs derived from patient CU04 stimulated with DMSO-negative control, PHA-positive control, CU04-specific nascent antigen peptide pool #1, CU04-specific peptide 1, CU04-specific peptide 6, and CU04-specific peptide 8.

[0409] Figures 19A-B show the results of control experiments using patient-generated antigens in HLA-matched healthy donors. These results indicate that the in vitro culture conditions did not enable in vitro de novo priming, but rather only proliferated existing in vivo-primed memory T cells.

[0410] Figure 20 shows the detection of the T cell response to the PHA-positive control for each donor and each in vitro proliferation shown in Figure 15A. For each donor and each in vitro proliferation in Figure 15A, patient PBMCs grown in vitro were stimulated with PHA to achieve maximum T cell activation. The data in Figure 20 are from seeded cells 10¹¹ with background (corresponding DMSO-negative control) subtracted. 5The values ​​are shown as spot-forming units (SFUs) per cell. Responses from single wells or biological replicas are shown for patients 1-038-001, 1-050-001, 1-001-002, CU04, 1-024-001, 1-024-002, CU05, and CU03. Patient CU02 was not tested by PHA. Cells from patient CU02 were included in the analysis because a positive response to peptide pool #1 (Figure 15A) indicated viable and functional T cells. As shown in Figure 15A, donors that responded to the peptide pool included patients 1-038-001, CU04, 1-024-001, and 1-024-002. As shown in Figure 15A, the donors who did not respond to the peptide pool included patients 1-050-001, 1-001-002, CU05, and CU03.

[0411] Figure 21A shows the detection of T cell responses to each individual patient-specific nascent antigen peptide in pool #2 in patient CU04. Figure 21A also shows the detection of T cell responses to a PHA-positive control in patient CU04. (This positive control data is also shown in Figure 20.) In patient CU04, the patient's in vitro-grown PBMCs were stimulated with individual patient-specific nascent antigen peptides from pool #2 for patient CU04 in an IFN-γELISpot. The patient's in vitro-grown PBMCs were also stimulated with PHA as a positive control in an IFN-γELISpot. Data are from 10 seeded cells subtracted from background (corresponding DMSO-negative control). 5 This is shown as the number of spot-forming units (SFUs) per individual.

[0412] Figure 21B shows the detection of T cell responses to individual patient-specific nascent antigen peptides in each of the three visits of patient CU04 and in each of the two visits of patient 1-024-002 (each visit taking place at a different time). In both patients, the patient's in vitro proliferated PBMCs were stimulated with individual patient-specific nascent antigen peptides in IFN-γELISpot. For each patient, the data from each visit were taken from 100% seeded cells after subtracting the background (corresponding DMSO control).5 Data is shown as the cumulative (sum) spot-forming units (SFU) per sample. For patient CU04, the data is shown as the cumulative SFU over three visits, subtracting the background. For patient CU04, the SFU, subtracted from the background, is shown for the initial visit (T0) and subsequent visits 2 months (T0+2 months) and 14 months (T0+14 months) after the initial visit (T0). For patient 1-024-002, the data is shown as the cumulative SFU over two visits, subtracting the background. For patient 1-024-002, the SFU, subtracted from the background, is shown for the initial visit (T0) and subsequent visits 1 month (T0+1 month) after the initial visit (T0). Samples with an increase factor more than twice that of the background were considered positive and were marked with an asterisk.

[0413] Figure 21C shows the detection of T cell responses to individual patient-specific nascent antigen peptides and to the patient-specific nascent antigen peptide pool during each of two visits of patient CU04 and each of two visits of patient 1-024-002 (each visit taking place at a different time). In both patients, the patient's in vitro proliferated PBMCs were stimulated with individual patient-specific nascent antigen peptides and the patient-specific nascent antigen peptide pool in IFN-γELISpot. Specifically, in patient CU04, in vitro-grown PBMCs from patient CU04 were stimulated in IFN-γELISpot with CU04-specific individual nascent antigen peptides 6 and 8, as well as a CU04-specific nascent antigen peptide pool. In patient 1-024-002, in vitro-grown PBMCs from patient 1-024-002 were stimulated in IFN-γELISpot with 1-024-002-specific individual nascent antigen peptides 16, as well as a 1-024-002-specific nascent antigen peptide pool. The data in Figure 21C show the mean and range of seeded cells 10 for each technical replicate, with background (corresponding DMSO control) subtracted. 5The data is shown as spot-forming units (SFUs) per individual. For patient CU04, the data is shown as SFUs for two visits with background subtracted. For patient CU04, the SFUs with background subtracted are shown for the initial visit (T0, technical triplicate) and the subsequent visit two months after the initial visit (T0+2 months, technical triplicate). For patient 1-024-002, the data is shown as SFUs for two visits with background subtracted. For patient 1-024-002, the SFUs with background subtracted are shown for the initial visit (T0, technical triplicate) and the subsequent visit one month after the initial visit (T0) (T0+1 month, technical duplicate excluding samples stimulated with patient 1-024-002-specific nascent antigen peptide pool).

[0414] Figure 22 shows the detection of T cell responses to two patient-specific nascent antigen peptide pools and a DMSO-negative control for the patients in Figure 15A. For each patient, PBMCs grown in vitro for that patient were stimulated with two patient-specific nascent antigen peptide pools using IFN-γELISpot. For each donor and each in vitro growth, the patient PBMCs grown in vitro were also stimulated with DMSO as a negative control using IFN-γELISpot. The data in Figure 22 show the patient-specific nascent antigen peptide pools and the corresponding DMSO control, including the background (corresponding DMSO-negative control) in a sample of 100% of seeded cells. 5The values ​​are shown as spot-forming units (SFUs) per cell. For patients 1-038-001, 1-050-001, 1-001-002, CU04, 1-024-001, 1-024-002, and CU05, the response is shown as a single well (patients 1-038-001, CU02, CU03, and 1-050-001) or the mean (all other samples) including the standard deviation of biological duplicates against congeneral peptide pools #1 and #2. For patients CU02 and CU03, testing was only possible against specific peptide pool #1 due to cell count. Samples with an increase factor more than 2 times higher than the background were considered positive and are marked with an asterisk (responsive donors include patients 1-038-001, CU04, 1-024-001, 1-024-002, and CU02). Non-responsive donors include patients 1-050-001, 1-001-002, CU05, and CU03.

[0415] As briefly mentioned above regarding Figures 19A-B, a series of control experiments were conducted using newly generated antigens in HLA-matched healthy donors to confirm that the in vitro culture conditions did not enable in vitro de novo priming, but rather proliferated only existing in vivo-primed memory T cells. The results of these experiments are shown in Figures 19A-B and Supplementary Table 5. These results confirmed that de novo priming does not occur in healthy donors using the IVS culture method, and that no detectable newly generated antigen-specific T cell response occurs.

[0416] In contrast, existing nascent antigen-responsive T cells were identified in the majority of patients (5 / 9, 56%) tested with a patient-specific peptide pool using IFN-γELISpot (Figures 15A and 20-22). Of the seven patients whose cell counts allowed for complete or partial testing of individual nascent antigen homologous peptides, four patients responded to at least one of the tested nascent antigen peptides, and all of these patients showed a response to the corresponding pool (Figure 15B). The remaining three patients tested with individual nascent antigens (patients 1-001-002, 1-050-001, and CU05) did not show a detectable response to a single peptide (data not shown), confirming that these patients did not show the response seen in the nascent antigen pool (Figure 15A). Of the four responding patients, samples were obtained from a single visit for two responding patients (patients 1-024-001 and 1-038-001), and samples were obtained from multiple visits for the remaining two responding patients (CU04 and 1-024-002). For the two patients with samples from multiple visits, the cumulative (sum) spot-forming units (SFUs) fr...

Claims

1. A method for identifying one or more antigen-specific T cells that originate from one or more target tumor cells and are presented on the surface of the tumor cells, A step of obtaining at least one of exome, transcriptome, or whole genome nucleotide sequencing data from the target tumor cells and normal cells, wherein the nucleotide sequencing data is used to obtain data representing the respective peptide sequences of a set of nascent antigens identified by comparing the nucleotide sequencing data from the tumor cells with the nucleotide sequencing data from the normal cells, and the peptide sequence of each nascent antigen includes at least one modification that causes the peptide sequence to differ from the corresponding wild-type peptide sequence identified from the target normal cells; A step of encoding each of the peptide sequences of the nascent antigen into a corresponding numerical vector, wherein each numerical vector includes information about a plurality of amino acids constituting the peptide sequence and a set of positions of the amino acids in the peptide sequence; A step of inputting the numerical vectors into a machine learning presentation model using a computer processor in order to generate a set of presentation likelihoods for the set of neonatal antigens, wherein each presentation likelihood in the set represents the likelihood that the corresponding neonatal antigen is presented on the surface of the target tumor cell by one or more MHC alleles, and the machine learning presentation model Labels obtained by mass spectrometry, which measures the presence of a training peptide sequence presented by at least one of the multiple MHC alleles identified as present in the sample, for each of the multiple samples containing cells expressing multiple MHC alleles, and For each of the aforementioned samples, a training peptide sequence encoded as a numerical vector containing information about the multiple amino acids constituting the peptide and the set of positions of the amino acids in the peptide. Multiple parameters identified at least based on the training dataset, Includes, Here, the parameters are identified by performing a deconvolution using a machine learning presentation model to determine which MHC allele among multiple MHC alleles presented which training peptide sequence, without pre-assigning which MHC allele among multiple MHC alleles presented which training peptide sequence; A step of selecting a subset of the set of nascent antigens based on the set of presentation likelihoods in order to generate a selected set of nascent antigens; and The step of identifying one or more T cells that are antigen-specific to at least one of the nascent antigens in the subset. The method, including the method described above.

2. The process of inputting the numerical vector into the machine learning presentation model is as follows: For each of the one or more MHC alleles, the machine learning presentation model is applied to the peptide sequence of the nascent antigen in order to generate a dependency score indicating whether the MHC allele presents the nascent antigen based on a specific amino acid at a specific position in the peptide sequence. The method according to claim 1, including the method described in claim 1.

3. The process of inputting the numerical vector into the machine learning presentation model is as follows: (A) For each MHC allele, the dependency score is transformed to generate a corresponding allele-specific likelihood that indicates the likelihood that the corresponding MHC allele will present the corresponding nascent antigen, and To generate the presentation likelihood of the aforementioned nascent antigen, the allele-specific likelihoods are combined. Converting the dependency score means modeling the presentation of the nascent antigen as mutually exclusive across one or more MHC alleles, or (B) Transforming the combination of dependency scores in order to generate the presentation likelihood, wherein the transformation of the combination of dependency scores models the presentation of the nascent antigen as interfering between one or more MHC alleles. The method according to claim 2, further comprising:

4. The set of presented likelihoods is further identified by at least one allele-non-interaction property, The method further includes applying the machine learning presentation model to the allele-non-interaction property to generate a dependency score for the allele-non-interaction property indicating whether the peptide sequence of the corresponding nascent antigen is presented based on the allele-non-interaction property, The method is (A) Combining the dependency score for each MHC allele in the one or more MHC alleles with the dependency score for the allele non-interaction properties, To generate an allele-specific likelihood for each MHC allele, which indicates the likelihood that the corresponding MHC allele presents the corresponding nascent antigen, the combined dependency score for each MHC allele is converted, and To generate the aforementioned presented likelihood, combine the allele-specific likelihoods, or (B) Combining the dependency score for each of the MHC alleles with the dependency score for the allele non-interaction properties, and To generate the aforementioned presentation likelihood, the combined dependency scores are transformed. The method according to claim 2 or 3, further comprising:

5. (a) The one or more MHC alleles include two or more different MHC alleles, (b) The peptide sequence includes a peptide sequence having a length other than nine amino acids, and / or (c) The step of encoding the peptide sequence includes encoding the peptide sequence using a one-hot encoding scheme, The method according to any one of claims 1 to 4.

6. The aforementioned multiple samples (a) One or more cell lines engineered to express a single MHC allele, (b) One or more cell lines engineered to express multiple MHC alleles, (c) One or more human cell lines obtained from or derived from multiple patients, (d) Fresh or frozen tumor samples obtained from multiple patients, and (e) Fresh or frozen tissue samples obtained from multiple patients The method according to any one of claims 1 to 5, comprising at least one of the following.

7. The aforementioned training dataset is (a) Data relating to the measurement of peptide-MHC binding affinity for at least one of the peptides, and (b) Data relating to the measurement of peptide-MHC binding stability for at least one of the peptides. The method according to any one of claims 1 to 6, further comprising at least one of the following.

8. (A) The set of presentation likelihoods is further identified by the expression levels of at least one of the MHC alleles in the subject, as measured by RNA-seq or mass spectrometry. (B) The set of presented likelihoods is (a) The predicted affinity between the nascent antigens in the set of nascent antigens and the one or more MHC alleles, and (b) Predicted stability of the newly synthesized antigen-coding peptide-MHC complex Further specified by a characteristic including at least one of the following, and / or (C) The set of numerical likelihoods is (a) The C-terminal sequence adjacent to the nascent antigen-coding peptide sequence within the source protein sequence, (b) The N-terminal sequence adjacent to the nascent antigen-coding peptide sequence within the source protein sequence. Further specified by a property that includes at least one of the following: The method according to any one of claims 1 to 7.

9. The step of selecting the aforementioned set of newly selected antigens is, (A) Based on the machine learning presentation model, select the neonatal antigen that has a higher likelihood of being presented on the surface of the tumor cells compared to the neonatal antigen that is not selected. (B) Based on the machine learning presentation model, select a neonatal antigen that has a higher likelihood of inducing a tumor-specific immune response in the target compared to a neonatal antigen that is not selected. (C) Based on the presentation model, select a naive antigen that has a higher likelihood of being presented to naive T cells by a professional antigen-presenting cell (APC) compared to a naive antigen that is not selected, where the APC is a dendritic cell (DC). (D) Selecting a neonatal antigen that has a reduced likelihood of being inhibited by central or peripheral tolerance compared to a neonatal antigen that is not selected, based on the machine learning presentation model, and / or (E) Based on the machine learning presentation model, select a nascent antigen that has a reduced likelihood of inducing an autoimmune response against normal tissue in the subject compared to a nascent antigen that is not selected. The method according to any one of claims 1 to 8, including the method described in any one of claims 1 to 8.

10. (A) The one or more tumor cells are selected from the group consisting of lung cancer, melanoma, breast cancer, ovarian cancer, prostate cancer, kidney cancer, gastric cancer, colon cancer, testicular cancer, head and neck cancer, pancreatic cancer, brain cancer, B-cell lymphoma, acute myeloid leukemia, chronic myeloid leukemia, chronic lymphocytic leukemia, T-cell lymphocytic leukemia, non-small cell lung cancer, and small cell lung cancer, and / or (B) The method further comprises generating an output for constructing a personalized cancer vaccine from the selected set of nascent antigens, wherein the output for the personalized cancer vaccine comprises at least one peptide sequence or at least one nucleotide sequence encoding the selected set of nascent antigens. The method according to any one of claims 1 to 9.

11. The aforementioned machine learning presentation model (for example, a deep learning model containing one or more layers of nodes) is a neural network model. (A) The neural network model includes multiple network models for MHC alleles, each network model being assigned to a corresponding MHC allele from among the multiple MHC alleles and including a series of nodes arranged in one or more layers, or (B) The neural network model includes a plurality of network models for MHC alleles, each network model including a set of nodes assigned to a corresponding MHC allele from the plurality of MHC alleles and arranged in one or more layers, the neural network model is trained by updating the parameters of the neural network model, the parameters of at least two network models are updated together for at least one training iteration, The method according to any one of claims 1 to 10.

12. The step of identifying one or more T cells is, (A) Co-culturing the one or more T cells with one or more of the newly generated antigens in the subset under conditions that cause the one or more T cells to proliferate, and / or (B) Contacting one or more T cells with the MHC multimer containing one or more of the nascent antigens in the subset under conditions that enable binding of the T cells to the MHC multimer. The method according to any one of claims 1 to 11, including the method described in any one of claims 1 to 11.

13. The method according to any one of claims 1 to 12, wherein the method further comprises identifying one or more T cell receptors (TCRs) of the one or more identified T cells, and the identification of the one or more T cell receptors comprises sequencing the T cell receptor sequences of the one or more identified T cells.

14. (A) One or more T cells that are antigen-specific to at least one of the nascent antigens in the subset are identified using 5 to 30 mL of whole blood from the subject, (B) The subset of nascent antigens comprises up to 20 different nascent antigens, and one or more identified T cells recognize at least two of the nascent antigens in the subset of nascent antigens, and / or (C) The one or more MHC alleles are Class I MHC alleles. The method according to any one of claims 1 to 13.

Images