Identification of neoantigens using MHC class II models

Abstract

A method for identifying T cells specific for at least one neoantigen that is likely to be presented on the surface of a subject's tumor cells by class II MHC alleles. Peptide sequences of the tumor neoantigens are obtained by sequencing the subject's tumor cells. These peptide sequences are input into a machine-learned presentation model to generate tumor neoantigen presentation likelihoods, each representing the likelihood that a neoantigen will be presented on the surface of the subject's tumor cells by class II MHC alleles. A subset of neoantigens is selected based on the presentation likelihoods. T cells specific for at least one of the neoantigens in the subset are identified. These T cells can be expanded and used 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. TIFF2022524328000074.tif186128

Description

[Background technology]

[0001] Therapeutic vaccines and T-cell therapies based on tumor-specific neoantigens hold great promise as the next generation of personalized cancer immunotherapy 1~3 . Cancers with high mutational burden, such as non-small cell lung cancer (NSCLC) and melanoma, are particularly promising targets for such therapies due to their relatively high likelihood of producing neoantigens. 4,5 . Early evidence suggests that neoantigen-based vaccination induces T cell responses. 6 , Neoantigen-targeted T-cell therapy can cause tumor regression in selected patients 7 is shown.

[0002] Specifically, the identification of MHC class II-presented neoantigens for use in neoantigen-based vaccination and neoantigen-targeted T-cell therapy suggests that up to 50% of neoantigen-responsive TILs It is a promising treatment because it contains CD4 cells that respond to neoantigens presented by MHC class II alleles. These CD4 cells have been shown to assist CD8 cells in anti-tumor responses and in some cases have been shown to directly attack tumor cells. Despite such promise of MHC class II-presented neoantigens in cancer therapy, the positive predictive value (PPV) of MHC class II-presented neoantigens is lower than that of MHC class I-presented neoantigens recognized by CD8 cells. lower than

[0003] These relatively low presentation predictions of MHC class II-presented neoantigens are believed to be due in part to the structure of MHC class II molecules compared to MHC class I molecules. Specifically, MHC class II molecules tend to have more open peptide binding grooves compared to MHC class II molecules. Because of this structural difference, MHC class I molecules tend to bind peptides 8-10 amino acids in length, whereas MHC class II molecules bind peptides of more variable length (Fig. 14F). ). Due to the variability in the length of peptides presented by MHC class II molecules, peptides presented by MHC class II molecules can be more difficult to predict compared to peptides presented by MHC class I molecules.

[0004] Therefore, identification of MHC class II-presenting neoantigens and neoantigen-recognizing T cells assesses tumor responses. 77,110 to examine tumor evolution 111 to design the next generation of personalized therapies 112 It is a central issue. Are current neoantigen identification methods time consuming and labor intensive? 84,96 , or not sufficiently accurate 87,91-93 . Furthermore, neoantigen-recognizing T cells are a major component of TILs. 84,96,113,114 , circulating in the peripheral blood of cancer patients 107 However, current methods for identifying neoantigen-reactive T cells are: (1) TIL 97,98 or leukapheresis 107 (2) requires screening of impractically large libraries of peptides; or (3) is practically available for only a few MHC alleles. It has some combination of constraints that depend on the MHC alleles obtained.

[0005] In addition, early methods have been proposed that incorporate mutation-based analysis using next-generation sequencing, RNA gene expression, and prediction of MHC binding affinities of neoantigen peptides. 8 . However, these proposed methods involve many steps other than gene expression and MHC binding (e.g., TAP transport, proteasomal cleavage, MHC binding, transport of peptide-MHC complexes to the cell surface, and / or TCR-mediated recognition of MHC; endocytosis or autophagy, cleavage by extracellular or lysosomal proteases (e.g. cathepsins), and / or competition with CLIP peptide for HLA binding catalyzed by HLA-DM). 9 The entire epitope generation process cannot be modeled. Therefore, existing methods tend to suffer from low positive predictive value (PPV) (FIG. 1A).

[0006] Indeed, analyzes of peptides presented by tumor cells, performed by multiple groups, showed that less than 5% of the peptides predicted to be presented using gene expression and MHC binding affinity were present on tumor surface MHC. indicates that it cannot be seen above 10,11 (Figure 1B). This low correlation between binding prediction and MHC presentation is further indicated by the lack of improved prediction accuracy of binding-restricted neoantigens for checkpoint inhibitor response to the number of mutations alone. ing 12 These presentation prediction failures are particularly true in the case of neoantigens presented by MHC class II alleles.

[0007] Such low positive predictive value (PPV) of existing methods for predicting presentation presents a problem in the design of neoantigen-based vaccines and in neoantigen-based T cell therapy. If vaccines were designed with low PPV predictions, most patients would be unlikely to receive therapeutic neoantigens, and even fewer would receive multiple neoantigens. (Even assuming all of the presented peptides are immunogenic). Similarly, if therapeutic T cells are designed based on a low prediction of PPV, most patients are less likely to receive T cells with reactivity to neoplastic antigens, leading to post-prediction downstream The cost of time and physical resources to identify predictive neoantigens using test methods can be unnecessarily high. Therefore, neoantigen vaccination and T-cell therapy by current methods is unlikely to work in a significant number of tumor-bearing subjects (FIG. 1C).

[0008] Moreover, previous approaches have generated candidate neoantigens using only cis-acting mutations, splicing factor mutations that occur in multiple tumor types and lead to aberrant splicing in many genes. 13 , and mutations that create or eliminate protease cleavage sites, additional sources of nascent ORFs were in most cases not considered.

[0009] Finally, standard approaches to tumor genome and transcriptome analysis rely on somatic cells that generate candidate neoantigens due to suboptimal conditions in library construction, exome and transcriptome capture, sequencing, or data analysis. Mutations may be missed. Similarly, standard tumor analysis approaches can falsely promote sequence artifacts or germline polymorphisms as neoantigens, which can lead to inefficient use of vaccine capacity or risk of autoimmunity, respectively. . [Outline of the invention]

[0010] Disclosed herein are optimized approaches for identifying and selecting neoantigens presented by MHC class II alleles for personalized cancer vaccines, T cell therapy, or both. be. First, we work on an optimized tumor exome and transcriptome analysis approach to identify neoantigen candidates using next-generation sequencing (NGS). These methods build on standard approaches for tumor analysis by NGS to develop the most sensitive and specific neoantigen candidates across all classes of genomic alteration. Second, high PPV to overcome the specificity problem and make MHC class II allele-presenting neoantigens developed for vaccination and / or as targets for T-cell therapy more likely to induce anti-tumor immunity. A novel approach to the MHC class II allele-presenting neoantigen selection of MHC class II is provided. These approaches include, depending on the embodiment, trained statistical regression or non-linear deep learning MHC class II models that jointly model peptide-MHC class II allele mapping, as well as statistical analysis over peptides of different lengths. Included are motifs for each MHC class II allele for multiple length peptides that share potency. In particular, non-linear MHC class II deep learning models can be designed and trained to treat different MHC alleles within the same cell as independent, thus solving the problem with linear models where linear models interfere with each other. Finally, further concerns regarding the design and production of personalized vaccines based on MHC class II allele-presenting neoantigens and in the production of individualized MHC class II allele-presenting neoantigen-specific T cells for T cell therapy. matter is resolved.

[0011] The models disclosed herein outperform state-of-the-art prediction tools trained on binding affinity and earlier prediction tools based on MS peptide data by up to an order of magnitude. By more reliably predicting peptide presentation by MHC class II alleles, the model uses a limited amount of patient peripheral blood, screens fewer peptides per patient, and is more sensitive to MHC multimers. Enables more time- and cost-effective identification of MHC class II allele-presenting neoantigen-specific or tumor antigen-specific T cells for personalized therapy using a clinically viable process that does not necessarily rely on do. However, in another embodiment, using the models disclosed herein, MHC multimers need to be screened to identify MHC class II allele-presenting neoantigen- or tumor antigen-specific T cells. The reduced number of bound peptides allows more time- and cost-effective identification of MHC class II allele-presenting tumor antigen-specific T cells using MHC multimers.

[0012] The predictive performance of the MHC class II models disclosed herein in the task of identifying TIL nascent epitope datasets and predicted nascent antigen-reactive T cells was determined by modeling the processing and presentation of MHC class II alleles. We show that it is now possible to obtain predictions of therapeutically useful MHC class II allele-presenting nascent epitopes. In summary, this work will accelerate the progression towards cure of patients by enabling the identification of actionable in silico MHC class II allele-presenting antigens for MHC class II allele-presenting antigen-targeted immunotherapy. is. [Brief description of the drawing]

[0013] These and other features, aspects and aspects of the present invention will become better understood with regard to the following description and accompanying drawings.

[0014] [Fig. 1A] Current clinical approaches to identifying neoantigens are shown.

[0015] [Fig. 1B] It shows that less than 5% of the predicted bound peptides are displayed on tumor cells.

[0016] [Fig. 1C] Shows the impact of specificity issues on neoantigen prediction.

[0017] [Fig. 1D] We show that binding prediction is not sufficient to identify neoantigens.

[0018] [Fig. 1E] Shows the probability of MHC-I presentation as a function of peptide length.

[0019] [Fig.1F] An exemplary peptide spectrum generated from a Promega dynamic range standard is shown.

[0020] [Fig. 1G] We show how adding features increases the positive predictive value of the model.

[0021] [Figure 2A] 1 is a schematic of an environment for identifying the likelihood of peptide presentation in a patient, according to one embodiment;

[0022] [Fig.2B] A method for obtaining presentation information, according to one embodiment, is described. [Figure 2C] A method for obtaining presentation information, according to one embodiment, is described.

[0023] [Fig.3] 1 is a high-level block diagram illustrating computer logic components of a presentation specific system, according to one embodiment; FIG.

[0024] [Figure 4] 1 illustrates an exemplary set of training data, according to one embodiment.

[0025] [Fig.5] An exemplary network model associated with MHC alleles is described.

[0026] [Fig. 6A] FIG. 4 shows an exemplary network model NNH(•) shared by MHC alleles, according to one embodiment.

[0027] [Fig.6B] FIG. 4 shows an exemplary network model NNH(·) shared by MHC alleles, according to another embodiment.

[0028] [Fig.7] Generating presentation likelihoods for peptides associated with MHC alleles using an exemplary network model is described.

[0029] [Fig.8] Generating presentation likelihoods for peptides associated with MHC alleles using an exemplary network model is described.

[0030] [Fig.9] Generating presentation likelihoods for peptides associated with MHC alleles using an exemplary network model is described.

[0031] [Fig. 10] Generating presentation likelihoods for peptides associated with MHC alleles using an exemplary network model is described.

[0032] [Fig. 11] Generating presentation likelihoods for peptides associated with MHC alleles using an exemplary network model is described.

[0033] [Fig. 12] Generating presentation likelihoods for peptides associated with MHC alleles using an exemplary network model is described.

[0034] [Fig. 13A] Sample frequency distribution of mutation burden in NSCLC patients.

[0035] [Fig. 13B] Figure 10 shows the number of presented neoantigens in a simulated vaccine for patients selected based on selection criteria whether the patient meets the minimum mutational burden, according to one embodiment.

[0036] [Fig. 13C] Selected patients associated with a vaccine including therapeutic subsets identified based on a presentation model and selected patients associated with a vaccine including therapeutic subsets identified by prior art models, according to one embodiment Comparison of the number of neoantigens presented in simulated vaccines between .

[0037] [Fig. 13D] Selected patients associated with vaccines including therapeutic subsets identified based on a single per-allele presentation model for HLA-A*02:01 and per-allele for HLA-A*02:01 Presentation Neogenesis in Simulated Vaccines Between Selected Patients Associated with Vaccines Containing Identified Treatment Subsets Based on Both the Presentation Model and the Allele-by-Allele Presentation Model for HLA-B*07:02 It compares the number of antigens. According to one embodiment, the vaccine capacity is set at v=20 epitopes.

[0038] [Fig. 13E] Figure 10 compares the number of presented neoantigens in a simulated vaccine between patients selected based on mutational load and patients selected by expected utility score, according to one embodiment.

[0039] [Fig. 14A] FIG. 11 is a histogram of peptide lengths eluted from class II MHC alleles on human tumor cells and tumor infiltrating lymphocytes (TILs) using mass spectrometry.

[0040] [Fig. 14B] Dependence of mRNA quantification and presented peptides per residue is shown for two exemplary data sets.

[0041] [Fig. 14C] Figure 3 compares the performance results of exemplary presentation models trained and tested using two exemplary datasets.

[0042] [Fig. 14D] FIG. 12 is a histogram showing the amount of peptides sequenced using mass spectrometry for each sample of a total of 73 samples containing HLA class II molecules.

[0043] [Fig. 14E] 2 is a histogram showing the amount of samples in which specific MHC class II molecule alleles were identified.

[0044] [Fig. 14F] FIG. 10 is a histogram showing the percentage of peptides presented by MHC class II molecules in a total of 73 samples for each peptide length over a range of peptide lengths.

[0045] [Fig. 14G] FIG. 4 is a line graph showing the relationship between gene expression and incidence of presentation of gene expression products by MHC class II molecules for genes present in 73 samples.

[0046] [Fig. 14H] FIG. 4 is a line graph comparing the performance of the same model with different inputs in predicting the likelihood that a peptide in a test dataset of peptides will be presented by an MHC class II molecule.

[0047] [Fig. 14I] FIG. 4 is a line graph comparing the performance of three different models in predicting the likelihood that a peptide within a test dataset of peptides will be presented by an MHC class II molecule.

[0048] [Fig. 14J] FIG. 14I shows an exemplary embodiment of the Bi-LSTM model of FIG. 14I configured to predict peptide presentation by HLA-DRB (MHC class II gene).

[0049] [Fig. 14K] 14I is a line graph showing perfect fit-recall curves for the Bi-LSTM, MLP, RNN, and binding affinity models of FIG. 14I.

[0050] [Fig. 14L] Peptides in a test dataset of peptides of the best-in-class conventional model with two different criteria and the presentation model disclosed herein with two different inputs are presented by MHC class II molecules. is a line graph comparing performance in predicting the likelihood of

[0051] [Fig. 14M] A total of 230 samples of human tumors (NSCLC, lymphoma, and ovarian cancer) and cell lines (EBV) containing HLA class II molecules were analyzed for peptides sequenced using mass spectrometry with a q value of less than 0.1 for each sample. 2 is a histogram showing quantity;

[0052] [Fig. 14N] 2 is a histogram showing the amount of samples in which specific MHC class II molecule alleles were identified.

[0053] [Fig. 14O] Peptides bound to MHC class I molecules and peptides bound to MHC class II molecules are shown.

[0054] [Fig. 14P] Figure 14C shows an exemplary embodiment of the inception neural network of the inception model of Figure 14Q configured to predict peptide presentation by MHC class II molecules.

[0055] [Fig. 14Q] Performance of the 'Bi-LSTM' and 'Inception' presentation models in predicting the likelihood that a peptide in a test dataset of peptides will be presented by at least one of the MHC class II molecules present in the test dataset. is a line graph comparing

[0056] [Fig. 15A] Figure 15 takes the lowest NetMHCIIpan percentile rank across the "MS model", HLA-DRB1*15:01 and HLA-DRB5*01:01, "NetMHCIIpan rank": NetMHCIIpan 3.177, and HLA-DRB1*15:01 and "NetMHCIIpan nM", which adopts the strongest affinities in nM across HLA-DRB5*01:01: peptides within the HLA-DRB1*15:01 / HLA-DRB5*01:01 test dataset of NetMHCIIpan 3.1 This is a comparison of prediction performance in the ranking of [Fig. 15B] See description for Figure 15A.

[0057] [Fig. 16] FIG. 4 shows exemplary embodiments of TCR constructs for introducing TCRs into recipient cells.

[0058] [Fig. 17-1] Figure 17 shows the nucleotide sequence of an exemplary P526 construct backbone for cloning the TCR into an expression system for therapeutic development. [Fig. 17-2] Shows the continuation of Figure 17-1.

[0059] [Fig. 18-1] FIG. 18 shows an exemplary construct sequence for cloning a patient's neoantigen-specific TCR clonotype 1 into an expression system for therapeutic development. [Fig. 18-2] Shows the continuation of Figure 18-1.

[0060] [Fig. 19-1] FIG. 19 shows an exemplary construct sequence for cloning a patient's neoantigen-specific TCR clonotype 3 into an expression system for therapeutic development. [Fig. 19-2] Shows a continuation of Figure 19-1.

[0061] [Fig.20] 1 is a flowchart of a method for providing individualized neoantigen-specific therapy to a patient, according to one embodiment.

[0062] [Fig.21] An exemplary computer for implementing the entities shown in FIGS. 1 and 3 is described. [Mode for carrying out the invention]

[0063] detailed description I. Definition Generally, terms used in the claims and specification are to be interpreted as having their ordinary meanings as understood by those of ordinary skill in the art. Certain terms are defined below to provide additional clarity. In the event of a conflict between the ordinary meaning and a given definition, the given definition shall apply.

[0064] As used herein, the term "antigen" refers to a substance that induces an immune response.

[0065] As used herein, the term "neoantigen" refers to an antigen that differs from the corresponding wild-type parental antigen by, for example, tumor cell mutations or tumor cell-specific post-translational modifications. An antigen with at least one change. A neoantigen may comprise a polypeptide sequence or a nucleotide sequence. Mutations can include frameshift or non-frameshift insertion deletions (indels), missense or nonsense substitutions, splice site changes, genomic rearrangements or gene fusions, or any genomic or expression alteration that results in a nascent ORF. . Mutations can also include splice variants. Tumor cell-specific post-translational modifications can include aberrant phosphorylation. Tumor cell-specific post-translational modifications can also include splice antigens produced by the proteasome. See Liepe et al., A large fraction of HLAclass I ligandsare proteasome-generated spliced ​​peptides; Science.2016 Oct 21;354(6310):354-358.

[0066] As used herein, the term "tumor neoantigen" refers to a neoantigen present in a subject's tumor cells or tissue, but not in the subject's corresponding normal cells or tissue.

[0067] As used herein, the term "neoantigen-based vaccine" refers to vaccine constructs based on one or more neoantigens, eg, multiple neoantigens.

[0068] As used herein, the term "candidate neoantigen" refers to a mutation or other abnormality that gives rise to a new sequence that may represent a neoantigen.

[0069] As used herein, the term "coding region" refers to the portion of a gene that encodes a protein.

[0070] As used herein, the term "coding mutation" refers to a mutation occurring in the coding region.

[0071] As used herein, the term "ORF" means open reading frame.

[0072] As used herein, the term "neoplastic ORF" refers to tumor-specific ORFs resulting from mutations or other abnormalities such as splicing.

[0073] As used herein, the term "missense mutation" is a mutation that causes the substitution of one amino acid for another.

[0074] As used herein, the term "nonsense mutation" is a mutation that causes a substitution of an amino acid with a stop codon.

[0075] As used herein, the term "frameshift mutation" is a mutation that causes a change in the frame of a protein.

[0076] As used herein, the term "insertion deletion" is an insertion or deletion of one or more nucleic acids.

[0077] As used herein, the term "% identity" in the context of two or more nucleic acid or polypeptide sequences is defined by the sequence comparison algorithms described below (e.g., BLASTP and BLASTN, or by those skilled in the art). specific percentage of nucleotides or amino acid residues when aligned and compared for maximum correspondence, as determined using one of the other algorithms available) or by visual inspection refers to two or more sequences or subsequences in which are the same. Depending on the application, "% identity" can be over a region of the sequences being compared, eg, over a functional domain, or over the full length of the two sequences being compared.

[0078] For sequence comparison, typically one sequence acts as a reference sequence, to which test sequences are compared. When using a sequence comparison algorithm, test and reference sequences are entered into a computer, subsequence coordinates are designated, if necessary, and sequence algorithm program parameters are designated. The sequence comparison algorithm then calculates the % sequence identities for the test sequences relative to the reference sequence, based on the specified program parameters. Alternatively, sequence similarity or difference can be established by a combination of the presence or absence of particular nucleotides at selected sequence positions (eg, sequence motifs), or of amino acids in the translated sequence.

[0079] Optimal alignment of sequences for comparison is performed, for example, by the local homology algorithm of Smith & Waterman, Adv.Appl.Math.2:482 (1981), Needleman & Wunsch, J.Mol.Biol.48:443. (1970) by the homology alignment algorithms of Pearson & Lipman, Proc. Nat'l. Acad. Sci. GAP, BESTFIT, FASTA, and TFASTA in Software Package, Genetics Computer Group, 575 Science Dr., Madison, Wis.) or by visual inspection (generally see Ausubel et al., infra). reference).

[0080] One example of a suitable algorithm for determining % sequence identity and % sequence similarity is described in Altschul et al., J. Mol. Biol. 215:403-410 (1990). There is a BLAST algorithm. Software for performing BLAST analyzes is publicly available through the National Center for Biotechnology Information.

[0081] As used herein, the term "nonstop or readthrough" refers to mutations that cause removal of the natural stop codon.

[0082] As used herein, the term "epitope" refers to the specific portion of an antigen that is commonly bound by an antibody or T-cell receptor.

[0083] As used herein, the term "immunogenicity" refers to the ability to elicit an immune response, eg, through T cells, B cells, or both.

[0084] As used herein, the terms "HLA binding affinity", "MHC binding affinity" refer to the affinity of binding between a specific antigen and a specific MHC allele.

[0085] As used herein, the term "bait" refers to a nucleic acid probe that is used to enrich a sample for specific sequences of DNA or RNA.

[0086] As used herein, the term "mutation" is a difference between a nucleic acid of interest and a reference human genome used as a control.

[0087] As used herein, the term "mutation calling" is the algorithmic determination of the existence of a mutation, typically from sequencing.

[0088] As used herein, the term "polymorphism" is a germline mutation, ie, a mutation found in all DNA-bearing cells of an individual.

[0089] As used herein, the term "somatic mutation" is a mutation that occurs in the non-germline cells of an individual.

[0090] As used herein, the term "allele" refers to a version of a gene or a version of a gene sequence or a version of a protein.

[0091] As used herein, the term "HLA type" refers to the complement of HLA gene alleles.

[0092] As used herein, the term "nonsense mutation-mediated degradation machinery" or "NMD" refers to cellular degradation of mRNA due to premature stop codons.

[0093] As used herein, the term "truncal mutation" is a mutation that occurs early in tumor development and is present in the majority of tumor cells.

[0094] As used herein, the term "subclonal mutation" is a mutation that occurs late in tumor development and is present in only some of the cells of the tumor.

[0095] As used herein, the term "exome" is a subset of the protein-encoding genome. An exome can be the collective exons of a genome.

[0096] As used herein, the term "logistic regression" is a regression model for binary data from statistics in which the logit of probability that the dependent variable equals 1 is modeled as a linear function of the dependent variable.

[0097] As used herein, the term "neural network" is a classification consisting of multi-layered linear transformations followed by element-wise nonlinear transformations typically trained by stochastic gradient descent and backpropagation. Or a machine learning model for regression.

[0098] As used herein, the term "proteome" refers to the set of all proteins expressed and / or translated by a cell, group of cells, or individual.

[0099] As used herein, the term "peptidome" refers to the set of all peptides presented by MHC-I or MHC-II on the cell surface. A peptidome may also refer to a property of a cell or collection of cells (eg, a tumor peptidome means the union of the peptidomes of all cells, including the tumor).

[0100] As used herein, the term "ELISPOT" means an enzyme-linked immunosorbent spot assay, a common method for monitoring immune responses in humans and animals.

[0101] As used herein, the term "dextramer" is a dextran-based peptide-MHC multimer used for antigen-specific T cell staining in flow cytometry.

[0102] As used herein, the term "MHC multimer" is a peptide-MHC complex composed of multiple peptide-MHC monomeric units.

[0103] As used herein, the term "MHC tetramer" is a peptide-MHC complex composed of four peptide-MHC monomeric units.

[0104] As used herein, the term "tolerance or immune tolerance" refers to a state of immune non-responsiveness to one or more antigens, eg, self-antigens.

[0105] As used herein, the term "central tolerance" refers to deletion of autoreactive T cell clones or conversion of autoreactive T cell clones to immunosuppressive regulatory T cells (Treg). Tolerance conferred in the thymus either by promoting differentiation.

[0106] As used herein, the term "peripheral tolerance" refers to downregulating or anergizing autoreactive T cells that have survived central tolerance, or promoting the differentiation of these T cells into Tregs. It is tolerance conferred in the peripheral system by doing.

[0107] The term "sample" refers to a sample obtained from a subject by means including venipuncture, voiding, ejaculation, massage, biopsy, needle aspiration, wash sample, abrasion, surgical incision, or intervention, or other means known in the art. It can include a single cell, or multiple cells, or fragments of cells, or an aliquot of bodily fluid that has been collected.

[0108] The term "subject" includes cells, tissues, or organisms, human or non-human, in vivo, ex vivo, or in vitro, either male or female. The term subject includes mammals, including humans.

[0109] The term "mammal" encompasses both humans and non-humans, including but not limited to humans, non-human primates, dogs, cats, mice, cows, horses, and pigs.

[0110] The term "clinical factor" refers to a measure of the activity or severity of a subject's condition, eg, disease. "Clinical factors" include all markers of a subject's health status, including non-sample markers, and / or other characteristics of the subject such as, but not limited to, age and gender. A clinical factor can be a score, value, or set of values ​​that can be obtained from the evaluation of a subject or a sample (or population of samples) from a subject under given conditions. Clinical factors can also be predicted by other parameters such as markers and / or gene expression surrogates. Clinical factors can include tumor type, tumor subtype, and smoking history.

[0111] Abbreviations: MHC: major histocompatibility complex; HLA: human leukocyte antigen, or human MHC locus; NGS: next-generation sequencing; PPV: positive predictive value; TSNA: tumor-specific neoantigen; NMD: nonsense mutation-dependent degradation mechanism; NSCLC: non-small cell lung cancer; DC: dendritic cells.

[0112] As used in this specification and the appended claims, the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. Please note.

[0113] Terms not directly defined herein are to be understood as having the meaning generally associated with them, as is understood within the technical field of the present invention. Certain terms are discussed herein for the purpose of providing additional guidance to the practitioner in describing the compositions, devices, methods, etc., of aspects of the present invention, as well as how to make or use them. It will be appreciated that the same thing can be said more than once. Accordingly, alternative terms and synonyms may be used for any one or more of the terms discussed herein. No weight should be placed on whether a term is recited or discussed herein. Some synonyms or substitutable methods, materials, etc. are provided. The mention 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 the inventive aspects herein.

[0114] All references, issued patents, and patent applications cited in the text of this specification are hereby incorporated by reference in their entireties for all purposes.

[0115] II. Methods of Identifying Neoantigens Disclosed herein are methods for identifying T cells that are antigen-specific for neoantigens derived from tumor cells of interest that are likely to be presented by class II MHC alleles on the surface of the tumor cells. . The method includes obtaining exome, transcriptome, and / or whole genome nucleotide sequencing data from tumor cells and normal cells of a subject. This nucleotide sequencing data is used to obtain the peptide sequence of each neoantigen in the set of neoantigens. A set of neoantigens is identified by comparing nucleotide sequencing data from tumor cells with nucleotide sequencing data from normal cells. Specifically, the peptide sequence of each neoantigen in the set of neoantigens has at least one change that renders the peptide sequence different from the corresponding wild-type parental peptide sequence identified from a normal cell of interest. include. The method further comprises encoding the peptide sequence of each neoantigen in the set of neoantigens into a corresponding numeric vector. Each numerical vector contains information describing the amino acids that make up the peptide sequence and the positions of the amino acids in the peptide sequence. The method further includes generating a presentation likelihood for each neoantigen in the set of neoantigens by inputting the numerical vectors into the machine-learned presentation model. Each presentation likelihood represents the likelihood that the corresponding neoantigen is presented by a class II MHC allele on the surface of the subject's tumor cells. A machine-learned presentation model includes multiple parameters and functions. A plurality of parameters are identified based on the training data set. The training data set includes, for each sample in the plurality of samples, the presence of peptides bound to at least one class II MHC allele in the set of class II MHC alleles identified as being present in that sample. and a training peptide sequence encoded as a numeric vector containing information describing the amino acids that make up the peptide and / or the position of the amino acids in the peptide. The function represents the relationship between the numerical vectors received as input by the machine-learned presentation model and the presentation likelihoods generated as output by the machine-learned presentation model based on the numerical vectors and multiple parameters. The method further includes generating a selected set of neoantigens by selecting subsets of the set of neoantigens based on presentation likelihood. The method further comprises identifying T cells that are antigen-specific for at least one of the neoantigens in the subset and returning these identified T cells.

[0116] In some embodiments, inputting the numerical vector into the machine-learned presentation model applies the machine-learned presentation model to the peptide sequence of the neoantigen to generate dependency scores for each of the class II MHC alleles. including doing A dependency score for a class II MHC allele indicates whether that class II MHC allele presents a neoantigen based on a particular amino acid at a particular position in the peptide sequence. In a further embodiment, inputting the numerical vector into the machine-learned presentation model transforms the dependence score such that for each class II MHC allele the corresponding class II MHC allele presents the corresponding neoantigen. Further comprising generating a corresponding per-allele likelihood indicative of the likelihood, and generating a neoantigen presentation likelihood by combining the per-allele likelihoods. In some embodiments, transforming dependency scores models neoantigen presentation as mutually exclusive across class II MHC alleles. In an alternative embodiment, inputting the numerical vectors into the machine-learned presentation model further comprises transforming the combinations of dependency scores to generate presentation likelihoods. In such embodiments, transforming the combination of dependency scores models neoantigen presentation as interference between class II MHC alleles.

[0117] In some embodiments, the set of presentation likelihoods is further specified by one or more allelic non-interacting characteristics. In such embodiments, the method further comprises generating a dependency score for the allele non-interacting feature by applying a machine-learned presentation model to the allele non-interacting feature. The dependence score indicates whether the corresponding neoantigen peptide sequence is presented based on allelic non-interacting properties. In some embodiments, the method comprises combining a dependency score for each class II MHC allele with a dependency score for allele non-interacting characteristics and a combined dependency score for each class II MHC allele. to generate a per-allele likelihood for each class II MHC allele, and combining the per-allele likelihoods to generate a presentation likelihood. The allele-by-allele likelihood for a class II MHC allele indicates the likelihood that that class II MHC allele presents the corresponding neoantigen. In an alternative embodiment, the method comprises combining a dependency score for class II MHC alleles and a dependency score for allele non-interacting characteristics and transforming the combined dependency score into a presentation likelihood and generating a

[0118] In some embodiments, a class II MHC allele comprises two or more different class II MHC alleles.

[0119] In some embodiments, at least one class II MHC allele in the set of class II MHC alleles identified as being present in the samples of the training data set comprises two or more different types of class II MHC alleles. including.

[0120] In some embodiments, peptide sequences include peptide sequences with lengths other than 9 amino acids.

[0121] In some embodiments, encoding the peptide sequence comprises encoding the peptide sequence using a one-hot encoding scheme.

[0122] In certain embodiments, the plurality of samples is a cell line engineered to express a single class II MHC allele, a cell line engineered to express multiple class II MHC alleles, a cell line engineered to express multiple class II MHC alleles, a plurality of samples from a plurality of patients. Obtained or derived human cell lines, including at least one of fresh or frozen tissue samples obtained from multiple patients.

[0123] In some embodiments, the training data set comprises data relating to measurements of peptide-MHC binding affinity for at least one of the peptides and peptide-MHC binding for at least one of the peptides. Further including at least one of the data related to stability measurements.

[0124] In some embodiments, the set of presentation likelihoods is further identified by the level of expression of class II MHC alleles within the subject as measured by RNA-seq or mass spectrometry.

[0125] In some embodiments, the set of presentation likelihoods is the predicted affinity between the neoantigen and class II MHC alleles in the set of neoantigens and the Further defined by properties including at least one of predicted stability.

[0126] In some embodiments, the set of numerical likelihoods is the C-terminal sequence flanking the neoantigen-encoding peptide sequence within the source protein sequence and the flanking neoantigen-encoding peptide sequence within the source protein sequence. Further specified by features comprising at least one of the N-terminal sequences.

[0127] In some embodiments, selecting a selected set of neoantigens is more likely to be displayed on the tumor cell surface than unselected neoantigens based on machine-learned presentation models. Including selecting neoantigens.

[0128] In some embodiments, selecting the set of selected neoantigens induces a tumor-specific immune response in the subject relative to unselected neoantigens based on a machine-learned presentation model. This includes selecting neoantigens that have a high likelihood of generating

[0129] In some embodiments, selecting a set of selected neoantigens increases the likelihood of being presented to naive T cells by professional antigen-presenting cells (APCs) relative to unselected neoantigens based on a presentation model. This includes selecting high-grade neoantigens. In such embodiments, the APCs are optionally dendritic cells (DCs).

[0130] In some embodiments, selecting a set of selected neoantigens is more likely to be inhibited by central or peripheral tolerance than unselected neoantigens based on machine-learned presentation models. Including selecting low-grade neoantigens.

[0131] In some embodiments, selecting the set of selected neoantigens induces an autoimmune response against normal tissues in the subject relative to unselected neoantigens based on a machine-learned presentation model. This includes selecting neoantigens that have a low likelihood of generating

[0132] In some embodiments, the one or more tumor cells are 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 cells Selected from the group consisting of 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.

[0133] In some embodiments, the method further comprises generating output for constructing a personalized cancer vaccine from the set of selected neoantigens. In such embodiments, the output for a personalized cancer vaccine can include at least one peptide sequence or at least one nucleotide sequence that encodes a set of selected neoantigens.

[0134] In some embodiments, the machine-learned presentation model is a neural network model. In such embodiments, the neural network model comprises a set of nodes, each network model assigned to a corresponding class II MHC allele among the class II MHC alleles, arranged in one or more layers. Multiple network models for MHC alleles may be included. In such embodiments, the neural network model may be trained by updating the parameters of the neural network model, wherein the parameters of at least two network models are updated together in at least one training iteration.

[0135] In such embodiments, each network model may further include one or more convolutional neural networks, each of which is arranged in one or more layers and has a series of filters of different sizes. contains the nodes of . Each filter of the one or more convolutional neural networks can be sized to identify amino acid positions in the peptide sequence of each neoantigen that comprise the binding core or binding anchor of the peptide sequence.

[0136] In some embodiments, the machine-learned presentation model can be a deep learning model that includes one or more layers of nodes.

[0137] In some embodiments, identifying T cells comprises co-culturing the T cells with one or more of the neoantigens in the subset under conditions conducive to T cell expansion.

[0138] In some embodiments, identifying a T cell includes binding the T cell to an MHC multimer comprising one or more of the neoantigens in the subset and under conditions that allow binding between the T cell and the MHC multimer. including contacting with

[0139] In some embodiments, the method further comprises identifying a T cell receptor (TCR) for the identified T cells. In such embodiments, identifying the T cell receptor comprises sequencing the T cell receptor sequence of the identified T cell. In such embodiments, the method comprises genetically manipulating T cells to express at least one of one or more identified T cell receptors; and injecting the expanded T cells into the subject. In such embodiments, genetically engineering a T cell to express at least one of the identified T cell receptors comprises cloning the T cell receptor sequence of the identified T cell into an expression vector. and transfecting each of the T cells with the expression vector.

[0140] In some embodiments, the method further comprises culturing the identified T cells under conditions that expand the identified T cells and infusing the expanded T cells into the subject. be able to.

[0141] Also disclosed herein are isolated T cells that are antigen-specific for at least one selected neoantigen among the subsets of neoantigens described above.

[0142] International Patent Application Publication No. WO2018 / 195357 and International Patent Application Publication No. WO2019 / 050994 are hereby incorporated by reference in their entireties. International Patent Application Publication No. WO2018 / 195357 describes methods for predicting antigen presentation by MHC class II molecules. International Patent Application Publication No. WO2019 / 050994 describes methods for identifying T cells specific for antigens presented by MHC molecules. Although these publications are referred to in this section of this specification, the disclosures presented in International Patent Application Publication Nos. WO2018 / 195357 and WO2019 / 050994 are incorporated in their entireties in all sections of this application. incorporated by reference.

[0143] III. Identification of tumor-specific mutations in neoantigens Also disclosed herein are methods for identification of certain mutations (eg, mutations or alleles present in cancer cells). In particular, these mutations may be present in the genome, transcriptome, proteome, or exome of cancer cells of a subject with cancer, but not in normal tissue from the subject.

[0144] Genetic mutations in tumors can be considered useful for immunological targeting of tumors if they lead to alterations in the amino acid sequence of proteins exclusively in tumors. Useful mutations include: (1) non-synonymous mutations that result in different amino acids in the protein; (2) modified stop codons that result in translation of longer proteins with novel tumor-specific sequences at the C-terminus; (3) splice site mutations that lead to inclusion of introns in mature mRNAs and thus unique tumor-specific protein sequences; (4) tumor-specific mutations at the junction of two proteins; (5) frameshift mutations or deletions, resulting in new open reading frames with novel tumor-specific protein sequences; Mutations can also include one or more of non-frameshift insertion deletions, missense or nonsense substitutions, splice site changes, genomic rearrangements or gene fusions, or any genomic or expression alteration that results in a nascent ORF. can.

[0145] For example, mutated peptides or mutated polypeptides resulting from splice site, frameshift, readthrough, or gene fusion mutations in tumor cells are sequenced for DNA, RNA, or proteins in tumor versus normal cells. can be identified by

[0146] Mutations can also include previously identified tumor-specific mutations. Known tumor mutations can be found in the Catalog of Somatic Mutations in Cancer (COSMIC) database.

[0147] Various methods are available for detecting the presence of particular mutations or alleles in an individual's DNA or RNA. Advances in this field have provided accurate, easy, and inexpensive large-scale SNP genotyping. Various DNA "chip" technologies such as, for example, dynamic allele-specific hybridization (DASH), microplate array diagonal gel electrophoresis (MADGE), pyrosequencing, oligonucleotide-specific ligation, TaqMan systems, and Affymetrix SNP chips. Several techniques have been described, including These methods utilize amplification of the target gene region, typically by PCR. Yet other methods are based on generation of small signal molecules by invasive cleavage followed by 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.

[0148] PCR-based detection means can involve multiplex amplification of multiple markers simultaneously. For example, it is well known in the art to select PCR primers to generate PCR products that do not overlap in size and can be analyzed simultaneously. Alternatively, it is possible to amplify the different markers with primers that are differentially labeled and thus each can be differentially detected. Of course, hybridization-based detection means allow differential detection of multiple PCR products in a sample. Other techniques are known in the art that allow multiplex analysis of multiple markers.

[0149] Several methods have been developed to facilitate 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, C.R. (US Pat. No. 4,656,127). According to this method, a primer complementary to allelic sequences immediately 3' of the polymorphic site is hybridized to a target molecule obtained from a particular animal or human. If the polymorphic site on the target molecule contains a nucleotide that is complementary to a particular exonuclease-resistant nucleotide derivative present, that derivative will be incorporated onto the end of the hybridized primer. Such incorporation renders the primer resistant to exonucleases, thereby allowing its detection. Since the identity of the exonuclease-resistant derivative of the sample is known, the knowledge that the primer has been rendered exonuclease-resistant allows the nucleotide present at the polymorphic site of the target molecule to be the nucleotide used in the reaction. It turns out to be complementary to that of the derivative. This method has the advantage of not requiring the determination of large amounts of exogenous sequence data.

[0150] Solution-based methods can be used to determine nucleotide identity at polymorphic sites (Cohen, D. et al. (French Patent No. 2,650,840; PCT Application No. WO91 / 02087). United States. A primer complementary to the allelic sequence immediately 3' of the polymorphic site is used as in the method of Mundy of Patent No. 4,656,127, which method is complementary to the nucleotide at the polymorphic site. In some cases, labeled dideoxynucleotide derivatives, which become incorporated on the ends of the primers, are used to determine the identity of the nucleotide at that site.

[0151] An alternative method known as Genetic Bit Analysis or GBA is described by Goelet, P. et al. (PCT Application No. 92 / 15712). The Goelet, P. et al. method uses a mixture of labeled terminators and primers that are complementary to sequences 3' of the polymorphic site. The Goelet, P. et al. method uses a mixture of labeled terminators and primers that are complementary to sequences 3' 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. It can be a heterogeneous phase assay.

[0152] Several primer-guided nucleotide incorporation procedures for assaying polymorphic sites in DNA have been described (Komher, J.S. et al., Nucl. Acids. Res. 17:7779-7784 (1989); Sokolov, B.P., Nucl. Acids Res. 18:3671 (1990); Syvanen, A.-C., et al., Genomics 8:684-692 (1990); Kuppuswamy, M.N. et al., Proc. (U.S.A.) 88:1143-1147 (1991); Prezant, T.R. et al., Hum. Mutat. 1:159-164 (1992); Ugozzoli, L. et al., GATA 9:107-112 (1992) ; Nyren, P. et al., Anal. Biochem. 208:171-175 (1993)). These methods differ from GBA in that they utilize the incorporation of labeled deoxynucleotides to discriminate between bases at polymorphic sites. In such a format, the signal is proportional to the number of deoxynucleotides incorporated, 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)).

[0153] Numerous initiatives obtain sequence information directly from millions of individual molecules of DNA or RNA in parallel. Real-time single-molecule sequencing-by-synthesis technologies rely on the detection of fluorescent nucleotides as they are incorporated into nascent strands of DNA that are complementary to the template being sequenced. In one method, oligonucleotides 30-50 bases in length are covalently attached at the 5' ends to glass coverslips. These anchored strands serve two functions. First, they act as capture sites for target template strands when the template is configured with capture tails complementary to surface-bound oligonucleotides. They also act as primers for template-directed primer extension, which forms the basis for sequence reading. The capture primer serves as a fixed position site for sequencing using multiple cycles of synthesis, detection, and chemical cleavage of the dye-linker to remove the dye. Each cycle consists of addition of the polymerase / labeled nucleotide mixture, rinsing, imaging, and cleavage of the dye. In an alternative method, the polymerase is modified with a fluorescent donor molecule and immobilized on a glass slide, while each nucleotide is color-coded with an acceptor fluorescent moiety attached to the γ-phosphate. As the nucleotides become incorporated into the new strand, the system detects interactions between the fluorescently tagged polymerase and the fluorescently modified nucleotides. Other sequencing-by-synthesis techniques also exist.

[0154] Any suitable sequencing-by-synthesis platform can be used to identify mutations. As noted above, four major synthetic sequencing platforms are currently available: the Genome Sequencer sold by Roche / 454 Life Sciences, the 1G Analyzer sold by Illumina / Solexa, and the 1G Analyzer sold by Applied BioSystems. and the Heliscope system sold by Helicos Bioscience. Sequencing-by-synthesis platforms are also described by Pacific BioSciences and VisiGen Biotechnologies. In some embodiments, multiple nucleic acid molecules to be sequenced are attached to a support (eg, a solid support). A capture sequence / universal priming site can be added to the 3' and / or 5' end of the template to immobilize the nucleic acid on the support. A nucleic acid can be attached to a 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 complementary to a support-attached sequence that can double as a universal primer.

[0155] As an alternative to a capture sequence, a member of a coupling pair (e.g., antibody / antigen, receptor / ligand, or avidin-biotin pair as described, for example, in US Patent Application No. 2006 / 0252077) Each fragment can be ligated and captured on a surface coated with the respective second member of the coupling pair.

[0156] Following capture, the sequence is analyzed by single molecule detection / sequencing, e.g., as described in the Examples and in U.S. Pat. No. 7,283,337, e.g., including sequencing by template-dependent synthesis. can be done. In sequencing-by-synthesis, surface-bound molecules are exposed to a large number of labeled nucleotide triphosphates in the presence of a polymerase. The sequence of the template is determined by the order of the labeled nucleotides incorporated into the 3' end of the growing strand. This can be done in real time and can be done in step-and-repeat mode. A different optical label can be incorporated for each nucleotide for real-time analysis, and multiple lasers can be utilized for stimulation of the incorporated nucleotides.

[0157] Sequencing can also include other massively parallel sequencing or next generation sequencing (NGS) techniques and platforms. Additional examples of massively parallel sequencing techniques and platforms are Illumina HiSeq or MiSeq, ThermoPGM or Proton, Pac Bio RS II or Sequel, Qiagen's Gene Reader, and Oxford Nanopore MinION. Additional similar current massively parallel sequencing technologies and future generations of these technologies can be used.

[0158] Any cell type or tissue can be utilized to obtain a nucleic acid sample for use in the methods described herein. For example, DNA or RNA samples can be obtained from tumors or body fluids such as blood obtained by known techniques (eg, venipuncture), or saliva. Alternatively, nucleic acid testing can be performed on dry samples (eg, hair or skin). Additionally, a sample can be obtained from the tumor for sequencing and another sample is obtained from the normal tissue for sequencing if the normal tissue is of the same tissue type as the tumor. can be done. A sample can be obtained from a tumor for sequencing and another sample can be obtained from normal tissue for sequencing, where the normal sample is of a tissue type distinct from the tumor. can.

[0159] Tumors include 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 myelogenous leukemia, chronic myelogenous leukemia. , chronic lymphocytic leukemia, and one or more of T-cell lymphocytic leukemia, non-small cell lung cancer, and small cell lung cancer.

[0160] Alternatively, protein mass spectrometry can be used to identify or demonstrate the presence of mutated peptides bound to MHC proteins on tumor cells. Peptides can be acid eluted from tumor cells or from HLA molecules immunoprecipitated from tumors and then identified using mass spectrometry.

[0161] IV. Neoantigens A neoantigen can comprise nucleotides or polynucleotides. For example, a neoantigen can be an RNA sequence encoding a polypeptide sequence. Neoantigens useful in vaccines may therefore include nucleotide or polypeptide sequences.

[0162] isolated peptides containing tumor-specific mutations identified by the methods disclosed herein, peptides containing known tumor-specific mutations, and mutant polypeptides identified by the methods disclosed herein, or Fragments thereof are disclosed herein. Neoantigen peptides can be described in the context of their coding sequence when the neoantigen comprises a nucleotide sequence (eg, DNA or RNA) that encodes the related polypeptide sequence.

[0163] The one or more polypeptides encoded by the neoantigen nucleotide sequences can comprise at least one of: binding affinity to MHC with an IC50 value of less than 1000 nM, amino acids for MHC class I peptides 8-15, 8, 9, 10, 11, 12, 13, 14, or 15 in length, the presence of a sequence motif within or near the peptide that promotes proteasomal cleavage, and a sequence that promotes TAP transport presence of motifs. MHC class II polypeptides amino acids 6-30, 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 long, sequences within or near peptides that promote cleavage by extracellular or lysosomal proteases (e.g., cathepsins) or HLA binding catalyzed by HLA-DM presence of motifs.

[0164] One or more neoantigens can be present on the surface of the tumor.

[0165] One or more neoantigens can be immunogenic in a tumor-bearing subject, eg, can elicit a T-cell response or a B-cell response in the subject.

[0166] One or more neoantigens that induce an autoimmune response in a subject can be excluded from consideration in the context of vaccine production for tumor-bearing subjects.

[0167] The size of the at least one neoantigen peptide molecule is about 5, about 6, about 7, about 8, about 9, about 10, about 11, about 12, about 13, about 14 , about 15, about 16, about 17, about 18, about 19, about 20, about 21, about 22, about 23, about 24, about 25, about 26, about 27, about 28, about 29, about 30, about 31, about 32, about 33, about 34, about 35, about 36, about 37, about 38, about 39 , about 40, about 41, about 42, about 43, about 44, about 45, about 46, about 47, about 48, about 49, about 50, about 60, about 70, about 80, about 90, about 100, about 110, about 120, or more amino molecular residues, and any range derived therefrom. , but not limited to them. In a specific embodiment, a neoantigen peptide molecule is 50 amino acids or less.

[0168] Neoantigen peptides and polypeptides are 15 residues or less in length for MHC class I, and usually consist of between about 8 and about 11 residues, and can be particularly 9 or 10 residues; MHC class II; can be 6 to 30 residues.

[0169] If desired, longer peptides can be designed in several ways. In one example, if the presentation likelihood of a peptide on an HLA allele is predicted or known, then the longer peptides are (1) 2- to 2-5 towards the N- and C-termini of each corresponding gene product. (2) a sequence of some or all of the displayed peptides, with an extended sequence for each; In another example, sequencing reveals long (greater than 10 residues) nascent epitope sequences present in the tumor (e.g., by frameshifting, readthrough, or inclusion of introns that lead to novel peptide sequences). In that case, the longer peptides would (3) consist of the entire stretch of novel tumor-specific amino acids, thus obviating the need for computational or in vitro test-based selection of the strongest HLA-presented shorter peptides. To avoid. In either instance, the use of longer peptides may allow endogenous processing by patient cells, resulting in more efficient antigen presentation and induction of T cell responses.

[0170] Neoantigen peptides and polypeptides can be presented on HLA proteins. In some embodiments, neoantigen peptides and polypeptides are presented on HLA proteins with greater affinity than wild-type peptides. In some embodiments, the neoantigen peptide or polypeptide is at least less than 5000 nM, at least less than 1000 nM, at least less than 500 nM, at least less than 250 nM, at least less than 200 nM, at least less than 150 nM, at least less than 100 nM, at least less than 50 nM, or less It can have a small IC50.

[0171] In some embodiments, the neoantigenic peptides and polypeptides do not induce an autoimmune response and / or induce immune tolerance when administered to a subject.

[0172] Also provided are compositions comprising at least two or more neoantigen peptides. In some embodiments, the composition contains at least two different peptides. At least two different peptides can be derived from the same polypeptide. Different polypeptides mean that the peptides differ in length, amino acid sequence, or both. Peptides are derived from any polypeptide known or found to contain a tumor-specific mutation. Suitable polypeptides from which neoantigen peptides can be derived can be found, for example, in the COSMIC database. COSMIC is a comprehensive information manager for somatic mutations in human cancers. The peptide contains tumor-specific mutations. In some embodiments, tumor-specific mutations are driver mutations for a particular cancer type.

[0173] Neoantigen peptides and polypeptides with desirable activities or properties enhance or at least retain substantially all of the biological activity of unmodified peptides that bind to desirable MHC molecules and activate appropriate T cells. However, they can be modified to confer certain desirable attributes, such as improved pharmacological characteristics. By way of example, neoantigen peptides and polypeptides can be further subjected to various modifications, such as substitutions, either conservative or non-conservative, which can result in improved MHC binding, stability, or offer certain advantages in their use, such as presentation. Conservative substitutions involve replacing an amino acid residue with another that is biologically and / or chemically similar, e.g., one hydrophobic residue for another, or one polar residue. It means replacing a group with another polar residue. Substitutions include Gly, Ala; Val, Ile, Leu, Met; Asp, Glu; Asn, Gln; Ser, Thr; Lys, Arg; The effects of single amino acid substitutions may also be probed using D-amino acids. Such modifications are described, for example, in Merrifield, Science 232:341-347 (1986), Barany & Merrifield, The Peptides, Gross & Meienhofer, eds. (N.Y., Academic Press), pp. 1-284 (1979); Well known peptide synthesis procedures can be used as described in Stewart & Young, Solid Phase Peptide Synthesis, (Rockford, Ill., Pierce), 2d Ed. (1984).

[0174] Modification of peptides and polypeptides with various amino acid mimetics or unnatural amino acids can be particularly useful in increasing the stability of peptides and polypeptides in vivo. Stability can be assayed in many ways. By way of example, peptidases and various biological media such as human plasma and serum have been used to test stability. See, eg, Verhoef et al., Eur. J. Drug Metab Pharmacokin. 11:291-302 (1986). Peptide half-lives can be conveniently determined using a 25% human serum (v / v) assay. The protocol is generally as follows. Pooled human serum (type AB, non heat-inactivated) is delipidated by centrifugation prior to use. Serum is then diluted to 25% in RPMI tissue culture medium and used to test peptide stability. At predetermined time intervals, aliquots of the reaction solution are removed and added to either 6% aqueous trichloroacetic acid or ethanol. The cloudy reaction sample is chilled (4° C.) for 15 minutes and then spun to precipitate precipitated serum proteins. The presence of the peptide is then determined by reverse phase HPLC using stability specific chromatographic conditions.

[0175] Peptides and polypeptides can be modified to provide desirable attributes other than improved serum half-life. As an example, the ability of a peptide to induce CTL activity can be enhanced by linkage to a sequence containing at least one epitope capable of inducing a T helper cell response. Immunogenic peptide / T helper conjugates can be linked by spacer molecules. Spacers are typically composed of relatively small neutral molecules, such as amino acids or amino acid mimetics, that are substantially uncharged under physiological conditions. Spacers are typically selected from, eg, Ala, Gly, or other neutral spacers of nonpolar or neutral polar amino acids. It will be appreciated that the optionally present spacers need not be composed of the same residues and thus may be hetero-oligomers or homo-oligomers. When present, spacers will usually be at least 1 or 2 residues, more usually 3-6 residues. Alternatively, peptides can be linked to T helper peptides without spacers.

[0176] The neoantigen peptide can be linked to the T helper peptide either directly or via a spacer at either the amino- or carboxy-terminus of the peptide. The amino terminus of either the neoantigen peptide or the T helper peptide can be acylated. Exemplary T helper peptides include tetanus toxoid 830-843, influenza 307-319, malaria sporozoite perimeter 382-398 and 378-389.

[0177] Proteins or peptides may be identified by those skilled in the art, including expression of the protein, polypeptide, or peptide through standard molecular biology techniques, isolation of the protein or peptide from a naturally occurring source, or chemical synthesis of the protein or peptide. can be made by any technique known to the public. The nucleotide and protein, polypeptide and peptide sequences corresponding to various genes have been previously disclosed and can be found at computerized databases known to those of skill in the art. One such database is the Genbank and GenPept databases of the National Center for Biotechnology Information, located at the National Institutes of Health website. Coding regions of known genes can be amplified and / or expressed using techniques disclosed herein or as known to those of skill in the art. Alternatively, various commercial preparations of proteins, polypeptides and peptides are known to those skilled in the art.

[0178] In further embodiments, the neoantigen comprises a nucleic acid (eg, polynucleotide) encoding a neoantigen peptide or portion thereof. Polynucleotides can be single- and / or double-stranded polynucleotides, e.g., DNA, cDNA, PNA, CNA, RNA (e.g., mRNA), e.g., polynucleotides having a phosphorothioate backbone, or natural It can be in either the form or the stabilized form, or a combination thereof, and may or may not contain an intron. A still further aspect provides an expression vector capable of expressing a polypeptide or portion thereof. Expression vectors for various cell types are well known in the art and can be selected without undue experimentation. Generally, the DNA is inserted into an expression vector, such as a plasmid, in proper orientation and correct reading frame for expression. If necessary, the DNA can be ligated to appropriate transcriptional and translational regulatory control nucleotide sequences recognized by the desired host, although such controls are generally available on the expression vector. The vector is then introduced into the host through standard techniques. Guidance can be found, for example, in Sambrook et al. (1989) Molecular Cloning, A Laboratory Manual, Cold Spring Harbor Laboratory, Cold Spring Harbor, N.Y.

[0179] IV. Vaccine Compositions Also disclosed herein are immunogenic compositions, eg, vaccine compositions, capable of generating a specific immune response, eg, a tumor-specific immune response. Vaccine compositions typically include multiple neoantigens selected using, for example, the methods described herein. A vaccine composition can also be referred to as a vaccine.

[0180] The vaccine consists of 1-30 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. Peptides can include post-translational modifications. Vaccines are 1-100 or more nucleotide 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, 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. Vaccines contain 1-30 different neoantigen sequences, 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, 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 neoantigen sequences, 6, 7, 8, 9, 10, 11, 12, 13, or 14 different neoantigen sequences, or 12, 13, Alternatively, it can contain 14 different neoantigen sequences.

[0181] In one embodiment, the different peptides and / or polypeptides, or nucleotide sequences encoding them, are such that the peptides and / or polypeptides bind to different MHC molecules, such as different MHC class I molecules and / or different MHC class II molecules. selected to be able to In some embodiments, one vaccine composition comprises coding sequences for peptides and / or polypeptides capable of binding to most frequently occurring MHC class I and / or MHC class II molecules. Thus, a vaccine composition may comprise different fragments capable of binding to at least 2 preferred, at least 3 preferred, or at least 4 preferred MHC class I and / or MHC class II molecules.

[0182] A vaccine composition can generate a specific cytotoxic T cell response and / or a specific helper T cell response.

[0183] A vaccine composition can further comprise an adjuvant and / or carrier. Examples of useful adjuvants and carriers are provided herein below. The composition can be conjugated to a carrier such as a protein or antigen presenting cells such as dendritic cells (DC) capable of presenting peptides to T cells.

[0184] An adjuvant is any substance whose incorporation into the vaccine composition enhances or otherwise modifies the immune response to the neoantigen. A carrier can be a scaffold structure, such as a polypeptide or polysaccharide, to which the neoantigen can bind. Optionally, adjuvants are covalently or non-covalently conjugated.

[0185] The ability of an adjuvant to increase the immune response to an antigen is typically manifested by a significant or substantial increase in immune-mediated reactions or a reduction in disease symptoms. For example, increased humoral immunity is typically manifested by a significant increase in the titer of antibodies raised against an antigen, and increased T cell activity is typically characterized by cell proliferation, or cell-mediated cytotoxicity. , or in increased cytokine secretion. Adjuvants can also alter immune responses, for example, by altering a predominantly humoral or Th response to a predominantly cellular or Th response.

[0186] Suitable adjuvants are 1018 ISS, Alum, Aluminum salts, 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, Montanide ISA-51, OK-432, OM-174, OM-197-MP-EC, ONTAK, PepTel Vector System, PLG Micro particles, resiquimod, SRL172, virosomes and other virus-like particles, YF-17D, VEGF trap, R848, beta-glucan, Pam3Cys, saponin-derived Aquila's QS21 stimulon (Aquila Biotech, Worcester, Mass., USA), mycobacteria Including, but not limited to, 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 useful. Several immunological adjuvants specific for dendritic cells (eg MF59) and their preparations have been previously described (Dupuis M, et al., Cell Immunol. 1998; 186(1): 18-27; Allison A C; Dev Biol Stand. 1998;92:3-11). Cytokines can also be used. Some cytokines influence migration of dendritic cells to lymphoid tissues (e.g. TNF-α), accelerate maturation of dendritic cells into efficient antigen presenting cells for T lymphocytes (e.g. GM-α). CSF, IL-1, and IL-4) (specifically U.S. Pat. No. 5,849,589, which is hereby incorporated by reference in its entirety), and acting as immune adjuvants (e.g., IL-12). (Gabrilovich D I, et al., J ImmunotherEmphasis Tumor Immunol. 1996(6):414-418).

[0187] CpG immunostimulatory oligonucleotides have also been reported to enhance adjuvant efficacy in vaccine settings. Other TLR binding molecules such as RNA that bind TLR 7, TLR 8 and / or TLR 9 may also be used.

[0188] Other examples of useful adjuvants are chemically modified CpG (e.g. CpR, Idera), Poly(I:C) (e.g. polyi:CI2U), non-CpG bacterial DNA or RNA, and therapeutically and / or may act as an adjuvant, cyclophosphamide, sunitinib, bevacizumab, Celebrex, NCX-4016, sildenafil, tadalafil, vardenafil, sorafinib, XL-999, CP-547632, pazopanib, ZD2171, AZD2171, ipilimumab, tremelimumab , and immunologically active small molecules and antibodies such as SC58175. Amounts and concentrations of adjuvants and additives can be readily determined by those skilled in the art without undue experimentation. Additional adjuvants include colony-stimulating factors such as granulocyte-macrophage colony-stimulating factor (GM-CSF, sargramostim).

[0189] A vaccine composition can contain more than one different adjuvant. Additionally, the therapeutic composition can include any adjuvant material, including any of the above or combinations thereof. It is also contemplated that vaccines and adjuvants can be administered together or separately, in any suitable sequence.

[0190] A carrier (or excipient) can be present independently of an adjuvant. The function of the carrier is, for example, to increase activity or immunogenicity, to confer stability, to increase biological activity, or to increase serum half-life, in particular to increase the molecular weight of the variant. It can be Additionally, the carrier can assist in presenting the peptide to T cells. The carrier can be any suitable carrier known to those skilled in the art, such as proteins or antigen presenting cells. Carrier proteins can be 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, although these Not limited. For immunization of humans, the carrier is generally a physiologically acceptable carrier that is acceptable and safe for humans. However, tetanus toxoid and / or diphtheria toxoid are suitable carriers. Alternatively, the carrier can be dextran, such as Sepharose.

[0191] 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 CTL is possible when trimeric complexes of peptide antigens, MHC molecules and APCs are present. Correspondingly, it can enhance the immune response not only when the peptide is used for CTL activation, but additionally when APC with the respective MHC molecule is added. Accordingly, in some embodiments, the vaccine composition additionally contains at least one antigen-presenting cell.

[0192] Neoantigens may also be vaccinia, fowlpox, self-replicating alphaviruses, marabaviruses, adenoviruses (see, e.g., Tatsis et al., Adenoviruses, Molecular Therapy (2004) 10, 616-629), or secondary, secondary 3, or hybrid 2nd / 3rd generation lentiviruses, and any generation of recombinant lentiviruses designed to target specific cell types or receptors. For example, 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 ubiquitin C promoter, Nucl. Acids Res. (2015) 43(1):682-690, Zufferey et al., Self-Inactivating Lentivirus Vector for Safe and Efficient In Vivo Gene Delivery, J. Virol. (1998) 72(12):9873-9880). can also Depending on the packaging capabilities of the viral vector-based vaccine platforms described above, this approach can deliver one or more nucleotide sequences encoding one or more neoantigen peptides. The sequences may be flanked by non-mutated sequences, separated by linkers, or preceded by one or more sequences that target intracellular compartments (e.g., Gros et al. 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, 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 (eg, CTL) response to the peptide. Vaccinia vectors and methods useful in immunization protocols are described, for example, in US Pat. No. 4,722,848. Another vector is BCG (bacillus Calmette-Guerin). BCG vectors are described in Stover et al. (Nature 351:456-460 (1991)). A wide variety of other vaccine vectors, such as Salmonella typhi vectors, useful for therapeutic administration or immunization of neoantigens will be apparent to those skilled in the art from the description herein.

[0193] IV.A. Further Considerations for Vaccine Design and Manufacturing IV.A.1. Determination of the set of peptides covering all tumor subclones A truncal peptide, meaning that presented by all or most tumor subclones, is preferred for inclusion in the vaccine. 53 . Optionally, if there are no truncated peptides predicted to be highly probable and immunogenic, or if the number of truncated peptides predicted to be highly probable and immunogenic is , to estimate the number and identity of tumor subclones, if few additional non-truncated peptides can be included in the vaccine, and to maximize the number of tumor subclones covered by the vaccine. Further peptides can be prioritized by choosing peptides in 54 .

[0194] IV.A.2. Prioritization of neoantigens After applying all of the above neoantigen filters, there may still be more candidate neoantigens available for vaccine inclusion than vaccine technology can accommodate. Additionally, there may remain uncertainties about various aspects of neoantigen analysis, and there may be trade-offs between the various properties of candidate vaccine neoantigens. Therefore, instead of pre-determined filters at each stage of the selection process, consider an integral multidimensional model that places candidate neoantigens in a space with at least the following axes and optimizes selection using an integral approach: can be done. 1. Autoimmune or tolerance risk (germline risk) (lower autoimmune risk is typically preferred) 2. Sequencing artifact probability (lower artifact probability is typically preferred) 3. Probability of immunogenicity (higher probability of immunogenicity is typically preferred) 4. Probability of presentation (higher probabilities of presentation are typically preferred) 5. Gene expression (higher expression is typically preferred) 6. HLA gene coverage (higher number of HLA molecules involved in presenting a set of neoantigens may make tumors less likely to evade immune attack through downregulation or mutation of HLA molecules) be) 7. HLA class coverage (covering both HLA-I and HLA-II may increase the probability of therapeutic response and reduce the probability of tumor immune evasion).

[0195] In addition, in some cases, vaccination of these neoantigens if it is expected that these neoantigens will be presented by HLA alleles that are lost or inactivated in all or part of the patient's tumor Can be de-prioritized (e.g. excluded). Loss of HLA alleles can result from either somatic mutation, loss of heterozygosity, or homozygous deletion of the locus. Methods for detecting somatic HLA allele mutations are well known in the art (eg, Shukla et al., 2015). Methods for detecting somatic LOH and homozygous deletions (including HLA loci) have also been described (Carter et al., 2012; McGranahan et al., 2017; Van Loo et al., 2010).

[0196] V. Treatment and Manufacturing Methods Administering to a subject one or more neoantigens, such as a plurality of neoantigens identified using the methods disclosed herein, induces a tumor-specific immune response in the subject and results in vaccination against the tumor. Also provided are methods of treating and / or alleviating symptoms of cancer in a subject.

[0197] In some embodiments, the subject has been diagnosed with cancer or is at risk of developing cancer. A subject can be a human, dog, cat, horse, or any animal for which a tumor-specific immune response is desired. The tumor may be any solid tumor such as breast, ovarian, prostate, lung, kidney, stomach, colon, testis, head and neck, pancreatic, brain, melanoma, and tumors of other tissues and organs, as well as acute myeloid leukemia. , chronic myeloid leukemia, chronic lymphocytic leukemia, T-cell lymphocytic leukemia, and lymphomas and leukemias, including B-cell lymphoma.

[0198] Neoantigens can be administered in amounts sufficient to induce a CTL response.

[0199] Neoantigens can be administered alone or in combination with other therapeutic agents. A therapeutic agent is, for example, a chemotherapeutic agent, radiation, or immunotherapy. Any suitable therapeutic treatment for a particular cancer can be administered.

[0200] Additionally, the subject can be further administered an anti-immunosuppressive / immunostimulatory agent, such as a checkpoint inhibitor. For example, the subject can be further administered an anti-CTLA antibody or anti-PD-1 or anti-PD-L1. Blocking CTLA-4 or PD-L1 with antibodies can enhance immune responses against cancerous cells in patients. In particular, CTLA-4 blockade has been shown to be effective when employing vaccination protocols.

[0201] Optimal amounts of each neoantigen to be included in the vaccine composition, as well as optimal dosing regimens, can be determined. For example, the neoantigen or variant thereof can be prepared for intravenous (i.v.) injection, subcutaneous (s.c.) injection, intradermal (i.d.) injection, intraperitoneal (i.p.) injection, intramuscular (i.m.) injection. can. Methods of injection include s.c., i.d., i.p., i.m., and i.v. Methods of DNA or RNA injection include i.d., i.m., s.c., i.p., and i.v. Other methods of administration of vaccine compositions are known to those skilled in the art.

[0202] Vaccines can be edited such that the selection, number, and / or amount of neoantigens present in the composition is tissue-, cancer-, and / or patient-specific. By way of example, the exact selection of peptides can be guided by the expression pattern of the parent protein in a given tissue. Selection may depend on the specific type of cancer, disease state, earlier treatment regimen, patient immune status, and, of course, the patient's HLA haplotype. In addition, vaccines can contain individualized components according to the individual needs of a particular patient. Examples include varying the selection of neoantigens according to neoantigen expression in a particular patient, or adjusting for secondary treatment after a first round or scheme of treatment.

[0203] For compositions to be used as vaccines for cancer, neoantigens with similar normal self-peptides that are abundantly expressed in normal tissues are avoided in the compositions described herein, or can be present in small amounts. On the other hand, if the patient's tumor is known to express large amounts of a particular neoantigen, the respective pharmaceutical composition for the treatment of this cancer can be present in large amounts, and / Alternatively, more than one neoantigen specific to this particular neoantigen or pathway of this neoantigen can be included.

[0204] Compositions containing neoantigens can be administered to individuals already with cancer. In therapeutic applications, the composition is administered to a patient in an amount sufficient to elicit an effective CTL response against tumor antigens and to cure or at least partially arrest symptoms and / or complications. be. An amount adequate to accomplish this is defined as a "therapeutically effective dose." Amounts effective for this use will depend, for example, on the composition, the mode of administration, the stage and severity of the disease being treated, the weight and general state of health of the patient, and the judgment of the prescribing physician. . It should be borne in mind that the compositions can generally be used in serious disease states, ie, life-threatening or potentially life-threatening situations, especially when the cancer has metastasized. In such instances, given the minimization of extraneous material and the relative non-toxic nature of neoantigens, it is possible to administer substantial excesses of these compositions and the treatment is effective. can feel desirable for physicians who

[0205] For therapeutic use, administration can begin upon detection or surgical removal of a tumor. This is followed by boosting doses until at least symptoms are substantially abated and for a period thereafter.

[0206] Pharmaceutical compositions for therapeutic treatment (eg, vaccine compositions) are intended for parenteral, topical, nasal, oral, or topical administration. Pharmaceutical compositions can be administered parenterally, for example, intravenously, subcutaneously, intradermally, or intramuscularly. The composition can be administered at the site of surgical resection to induce a local immune response against the tumor. Compositions for parenteral administration comprising solutions of neoantigens are disclosed herein, the vaccine compositions dissolved or suspended in an acceptable carrier, eg, an aqueous carrier. A variety of aqueous carriers can be used, such as water, buffered water, 0.9% saline, 0.3% glycine, hyaluronic acid, and the like. These compositions may be sterilized by conventional, well-known sterilization techniques, or may be sterile filtered. The resulting aqueous solution can be packaged for use as is or lyophilized, the lyophilized preparation being combined with a sterile solution prior to administration. The compositions may also contain pH adjusting and buffering agents, tonicity agents, wetting agents and the like, such as sodium acetate, sodium lactate, sodium chloride, potassium chloride, calcium chloride, sorbitan monolaurate, triethanolamine oleate, and the like. It may also contain pharmaceutically acceptable auxiliary substances as required to approximate physiological conditions.

[0207] Neoantigens can also be administered via liposomes, which target them to specific tissue 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, and the like. In these preparations, the neoantigen to be delivered is either alone or molecules that bind to receptors that are dominant among lymphoid cells, such as monoclonal antibodies that bind the CD45 antigen, or other therapeutic agents. It is incorporated as part of a liposome along with the therapeutic or immunogenic composition. Thus, liposomes loaded with desired neoantigens can be directed to the site of lymphoid cells where they then deliver the therapeutic / immunogenic composition of choice. 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 lability, and stability of the liposomes in the blood stream. 9;467 (1980), U.S. Pat. In addition, various methods are available for preparing liposomes.

[0208] For targeting to immune cells, ligands to be incorporated into liposomes can include, for example, antibodies or fragments thereof specific for cell surface determinants of desired immune system cells. Liposomal suspensions can be administered intravenously, topically, locally, etc., with doses varying according to, inter alia, the mode of administration, the peptide to be delivered, and the stage of the disease being treated.

[0209] The peptides described herein, and optionally nucleic acids encoding one or more of the peptides, can also be administered to a patient for therapeutic or immunization purposes. A number of methods are conveniently used to deliver nucleic acids to patients. As an 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 US Pat. Nos. 5,580,859 and 5,589,466. Nucleic acids can also be administered using ballistic delivery, for example, as described in US Pat. No. 5,204,253. Particles consisting solely of DNA can be administered. Alternatively, DNA can be attached to particles such as gold particles. Approaches for delivering nucleic acid sequences can include viral vectors, mRNA vectors, and DNA vectors, with or without electroporation.

[0210] Nucleic acids can also be delivered complexed to cationic compounds such as cationic lipids. Lipid-mediated gene delivery methods are described, for example, in 9618372 WOAWO 96 / 18372; 9324640 WOAWO 93 / 24640; Mannino & Gould-Fogerite, BioTechniques 6(7): 682-691 (1988); 5,279,833; 9106309 WOAWO 91 / 06309; and Felgner et al., Proc. Natl. Acad. Sci. USA 84: 7413-7414 (1987).

[0211] Neoantigens may also be vaccinia, fowlpox, self-replicating alphaviruses, marabaviruses, adenoviruses (see, for example, Tatsis et al., Adenoviruses, Molecular Therapy (2004) 10, 616-629), or secondary , third, or hybrid second / third generation lentiviruses, and any generation of recombinant lentiviruses designed to target specific cell types or receptors. Viruses (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 to 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 ubiquitin C promoter, Nucl. AcidsRes.(2015) 43 (1): 682-690 , Zufferey et al., Self-Inactivating Lentivirus Vector for Safe and Efficient In Vivo Gene Delivery, J. Virol. (1998) 72 (12): 9873-9880). can also be included. Depending on the packaging capabilities of the viral vector-based vaccine platforms described above, this approach can deliver one or more nucleotide sequences encoding one or more neoantigen peptides. The sequences may be flanked by non-mutated sequences, separated by linkers, or preceded by one or more sequences that target intracellular compartments (e.g., Gros et al. 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, 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 (eg, CTL) response to the peptide. Vaccinia vectors and methods useful in immunization protocols are described, for example, in US Pat. No. 4,722,848. Another vector is BCG (bacillus Calmette-Guerin). BCG vectors are described by Stover et al. (Nature 351:456-460 (1991)). A wide variety of other vaccine vectors, such as Salmonella typhi vectors, useful for therapeutic administration or immunization of neoantigens will be apparent to those skilled in the art from the description herein.

[0212] A means of administering nucleic acids employs minigene constructs that encode one or more epitopes. The amino acid sequence of the epitope is back-translated to generate a DNA sequence (minigene) encoding the selected CTL epitope for expression in human cells. A human codon usage table is used to guide the codon choice for each amino acid. The DNA sequences encoding these epitopes are directly juxtaposed to create a continuous polypeptide sequence. Additional elements can be incorporated into the minigene design to optimize expression and / or immunogenicity. Examples of amino acid sequences that can be reverse 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 (eg, polyalanine) or naturally occurring flanking sequences in close proximity to CTL epitopes. The minigene sequence is converted to DNA by assembling oligonucleotides encoding the plus and minus strands of the minigene. Overlapping oligonucleotides (30-100 bases long) are synthesized, phosphorylated, purified and annealed under appropriate conditions using well 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.

[0213] Purified plasmid DNA can be prepared for injection using a variety of formulations. The simplest of these is reconstitution of lyophilized DNA in sterile phosphate-buffered saline (PBS). A variety of methods have been described, and new techniques may become available. As mentioned above, nucleic acids are conveniently formulated with cationic lipids. In addition, glycolipids, fusogenic liposomes, peptides, and compounds collectively referred to as protective, interactive, non-condensing (PINCs) can also be complexed with purified plasmid DNA to enhance stability, intramuscular Variables such as dispersal, or trafficking to specific organs or cell types can be affected.

[0214] Also disclosed herein is a method of manufacturing a tumor vaccine comprising performing the steps of the methods disclosed herein; and producing a tumor vaccine comprising multiple neoantigens or subsets of multiple neoantigens. to disclose.

[0215] The neoantigens disclosed herein can be produced using methods known in the art. For example, methods of producing neoantigens or vectors disclosed herein (eg, vectors comprising at least one sequence encoding one or more neoantigens) are suitable for expressing neoantigens or vectors. Culturing the host cell under conditions, wherein the host cell comprises at least one polynucleotide encoding the neoantigen or vector; and purifying the neoantigen or vector. Standard purification methods include chromatographic, electrophoretic, immunological, precipitation, dialysis, filtration, concentration, and chromatofocusing techniques.

[0216] Host cells can include Chinese Hamster Ovary (CHO) cells, NS0 cells, yeast, or HEK293 cells. The host cell can be transformed with one or more polynucleotides comprising at least one nucleic acid sequence encoding a neoantigen or vector disclosed herein, optionally the isolated polynucleotide comprising: It further comprises a promoter sequence operably linked to at least one nucleic acid sequence encoding the neoantigen or vector. In certain embodiments, the isolated polynucleotide can be cDNA.

[0217] VI. Identification of Neoantigens VI.A. Identification of Neoantigen Candidates Describes and applies research methods for NGS analysis of tumor and normal exomes and transcriptomes to the specific space of neoantigens 6,14,15 . The examples below consider certain optimizations for greater sensitivity and specificity for identifying neoantigens in the clinical setting. These optimizations can be grouped into two areas, those related to laboratory processes and those related to NGS data analysis.

[0218] VI.A.1. Laboratory process optimization The process improvements presented here are based on concepts developed for the evaluation of reliable cancer driver genes in targeted cancer panels. 16 to the whole-exome and whole-transcriptome settings required for identification of neoantigens, addressing the challenges in high-fidelity neoantigen discovery from low tumor content and low volume clinical specimens. Specifically, these improvements include: 1. Targeting deep (greater than 500x) unique average coverage across the tumor exome to detect mutations present at low mutant allele frequencies, either due to low tumor content or subclonal status. 2. Less than 5% of the bases covered at less than 100x so as to miss the least possible neoantigens, e.g. a. Use of DNA-based capture probes with individual probe QCs 17 b. Inclusion of additional bait for areas not sufficiently covered 3. 5 bases covered less than 20x so that potential neoantigens are least likely to remain unclassified for somatic / germline status (thus not usable as TSNA) Homogeneous coverage targeting across normal exomes that is less than %. 4. To minimize the total amount of sequencing required, sequence capture probes are designed only for the coding regions of genes, since non-coding RNA cannot give rise to neoantigens. Additional optimizations include: a. Complementary probes for HLA genes that are GC-rich and not well captured by standard exome sequencing 18 . b. Elimination of genes predicted to produce few or no candidate neoantigens due to factors such as poor expression, suboptimal digestion by the proteasome, or unusual sequence characteristics. 5. Tumor RNA is also sequenced at high depth (greater than 100M reads) to allow mutation detection, quantification of gene and splice variant (“isoform”) expression, and fusion detection. be. RNA from FFPE samples was probe-based enriched with the same or similar probes used to capture exomes in DNA 19 is extracted using

[0219] VI.A.2. Optimization of NGS data analysis Improvements in analytical methods address the sub-optimal sensitivity and specificity of common research variant calling approaches, and specifically allow for relevant customizations for identification of neoantigens in the clinical setting. These include: 1. Use of the HG38 reference human genome or a later version for the alignment, as it contains multiple MHC region assemblies that better reflect population polymorphisms, in contrast to previous genome releases. . 2. Various programs 5 Overcoming the limitation of single mutation caller 20 by merging the results from . a. Single nucleotide mutations and indels are detected in tumor DNA, tumor RNA and normal DNA with a range of tools including: Strelka 21 and Mutec t22 programs based on comparison of tumor and normal DNA, such as 23 , UNCeqR, and other programs that incorporate tumor DNA, tumor RNA, and normal DNA. b. Indels are Strelka and ABRA 24 determined by a program that performs local reassembly, such as c. Structural reorganization is Pindel 25 or Breakseq 26 determined using a dedicated tool such as 3. Mutation calls from samples for the same patient are compared at a selected number of polymorphic sites to detect and prevent sample swapping. 4. Extensive filtering of artificial calls is performed, for example by: a. Elimination of mutations found in normal DNA, potentially with loose detection parameters in the case of low coverage and permissive proximity criteria in the case of insertion 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 not observed in their normal counterparts 27 . Examples include mutations detected primarily on one strand. d. Elimination of Mutations Detected in a Set of Unrelated Controls 27 . 5.seq2HLA 28 , ATHLATES 29 , or using one of the Optitypes, and also combining exome and RNA sequencing data 28 , correct HLA calling from normal exomes. Additional potential optimization is the adoption of dedicated assays for HLA typing, such as long-read DNA sequencing 30 , or adaptation of methods to ligate RNA fragments and preserve continuity 31 including. 6. Robust detection of nascent ORFs arising from tumor-specific splice variants is a CLASS 32 , Bayesembler 33 , String Tie 34 , or from RNA-seq data using a similar program in its reference-guided mode (i.e., using known transcript structures rather than attempting to recreate their entire transcript from each experiment). It is done by assembling transcripts. Cufflinks 35 Although commonly used for this purpose, it frequently produces an implausibly large number of splice variants, many of which are much shorter than the full-length gene and recover simple positive controls. may not be possible. The coding sequence and potential nonsense mutation-dependent degradation machinery were reintroduced mutant sequences, SpliceR 36 and MAMBA 37 determined by tools such as Gene Expression Cufflinks 35 Or determined by tools such as Express (Roberts and Pachter, 2013). Wild-type and mutant-specific expression counts and / or relative levels of ASE 38 or HTSeq 39 determined by tools developed for these purposes, such as Potential filtering steps include: a. Elimination of candidate nascent ORFs believed to be underexpressed. b. Elimination of candidate nascent ORFs predicted to trigger a nonsense mutation-dependent degradation mechanism (NMD). 7. Candidate neoantigens (e.g., nascent ORFs) observed only in RNA that cannot be directly validated as tumor-specific can be made tumor-specific according to additional parameters, for example by considering the following: Categorized as probable: Presence of supporting cis-acting frameshifts or splice site mutations in tumor DNA only. b. Presence of confirmatory tumor DNA-only trans-acting mutations in splicing factors. As an example, in three independently published experiments with the R625 mutant SF3B1, the most differentially spliced ​​gene was found to be 40 , a second experiment examined the uveal melanoma cell line 41 , and a third study examined breast cancer patients 42 nevertheless, they were in agreement. c. For novel splicing isoforms, presence of confirmation of "novel" splice-junction reads in RNASeq data. d. For de novo rearrangements, the presence of confirmation of near-exonic reads in tumor DNA that are not present in normal DNA. e.GTEx 43 Absence from gene expression compendium, such as (ie, making germline origin less likely). 8. Complementing reference genome alignment-based analysis by comparing tumor and normal reads (or kmers derived from such reads) of assembled DNA to directly avoid alignment and annotation-based errors and artifacts. (eg, for germline mutations or somatic mutations that occur near repeat-context insertion-deletions).

[0220] In samples with polyadenylated RNA, the presence of viral and microbial RNA in the RNA-seq data is assessed using RNA CoMPASS44 or similar methods toward identifying additional factors that may predict patient response. .

[0221] VI.B. Isolation and detection of HLA peptides Isolation of HLA peptide molecules was performed using classical immunoprecipitation (IP) methods after lysis and solubilization of tissue samples. 55~58 . Clarified lysates were used for HLA-specific IP.

[0222] Immunoprecipitation was performed with bead-coupled antibodies, where the antibodies are specific for HLA molecules. For pan-class I HLA immunoprecipitation use pan-class I CR antibody and for class II HLA-DR use HLA-DR antibody. Antibodies are covalently attached to NHS-Sepharose beads during an overnight incubation. After covalent attachment, beads were washed and aliquoted for IP 59、60 . Immunoprecipitation can also be performed with 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. Some antibodies that can be used to selectively enrich for MHC / peptide complexes are shown below. TIFF2022524328000002.tif42149

[0223] Clarified tissue lysates are added to the antibody beads for immunoprecipitation. After immunoprecipitation, the beads are removed from the lysate and the lysate is saved for further experiments including additional IP. The IP beads are washed to remove non-specific binding and the HLA / peptide complexes are eluted from the beads using standard techniques. Protein components are removed from peptides using molecular weight spin columns or C18 fractionation. The resulting peptides are dried by SpeedVac evaporation and in some cases stored at -20°C prior to MS analysis.

[0224] Dried peptides are reconstituted in an HPLC buffer suitable for reverse-phase chromatography and loaded onto a C-18 microcapillary HPLC column for gradient elution on a Fusion Lumos mass spectrometer (Thermo). MS1 spectra of peptide mass / charge (m / z) were collected at high resolution on the Orbitrap detector, followed by MS2 low resolution scans on the ion trap detector after HCD fragmentation of selected ions. Additionally, MS2 spectra can be acquired using either the CID or ETD fragmentation method, or any combination of the three techniques to obtain greater amino acid coverage of the peptide. MS2 spectra can also be measured with high resolution mass accuracy on an Orbitrap detector.

[0225] The MS2 spectra from each analysis were collected by Comet 61、62 Search against a protein database using the Percolator 63~65 Score using Further sequencing using PEAKS studio (Bioinformatics Solutions Inc.) and other search engines, or spectral matching and de novo sequencing 75 can be used.

[0226] VI.B.1. MS Detection Limit Studies for Comprehensive HLA Peptide Sequencing With peptide YVYVADVAAK, what was the limit of detection was determined with varying amounts of peptide loaded onto the LC column. The amount of peptide tested was 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 lowest limit of detection (LoD) is in the attomole range (10 -18 ), dynamic range spanning five orders of magnitude, and signal-to-noise in the low femtomole range (10 -15 ) appears to be sufficient for sequencing. TIFF2022524328000003.tif56128

[0227] VII. Presentation model VII.A. System Overview FIG. 2A is an overview of an environment 100 for identifying likelihoods of peptide presentation in a patient, according to one embodiment. Environment 100 itself provides a context for introducing presentation specific system 160 including presentation information storage 165 .

[0228] The presentation identification system 160 is one or a computer model, embodied in a computer computing system as discussed below with respect to FIG. 30, that receives peptide sequences associated with a set of MHC alleles, the peptide sequences determine the likelihood that one or more of the set of MHC alleles will be presented. The presentation specific system 160 can be applied to both class I and class II MHC alleles. This is useful in various contexts. One specific example of use of the presentation identification system 160 is to receive the nucleotide sequences of candidate neoantigens associated with a set of MHC alleles from tumor cells of a patient 110, wherein the candidate neoantigens are associated with the tumor's associated MHC alleles. It is possible to determine the likelihood presented by one or more and / or that it will induce an immunogenic response in the patient's 110 immune system. Those candidate neoantigens that have a high likelihood as determined by the system 160 can be selected for inclusion in the vaccine 118, and such anti-tumor immune responses provide tumor cells of the patient 110. It can be induced from the immune system. In addition, T cells with TCRs having reactivity against candidate neoantigens with high likelihood of presentation can be generated for use in T cell therapy, thereby enhancing anti-tumor immunity from the patient's 110 immune system. A response is also elicited.

[0229] Presentation identification system 160 determines presentation likelihood through one or more presentation models. Specifically, the presentation model generates likelihoods of whether a given peptide sequence is presented for a set of related MHC alleles, the likelihoods being based on presentation information stored in storage 165. generated. For example, the presented model is that the peptide sequence "YVYVADVAAK" has the set of alleles HLA-A*02:01, HLA-A*03:01, HLA-B*07:02, HLA-B* on the cell surface of the sample. 08:03, we can generate the likelihood of being presented for HLA-C*01:04. The presentation information 165 contains information about whether these peptides bind to various types of MHC alleles such that the peptides are presented by the MHC allele, which in the model corresponds to the amino acid position in the peptide sequence. determined accordingly. A presentation model can predict whether an unrecognized peptide sequence will be presented in association with a related set of MHC alleles based on presentation information 165 . As noted above, the presentation model can be applied to both class I and class II MHC alleles.

[0230] VII.B. Presentation information FIG. 2 illustrates a method of 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 affecting presentation of peptide sequences that is dependent on the type of MHC allele. Allele non-interaction information includes information that affects presentation of peptide sequences that is independent of MHC allele type.

[0231] VII.B.1. Allelic interaction information Allelic interaction information primarily includes identified peptide sequences known to be presented by one or more identified MHC molecules from humans, mice, and the like. Of note, this may or may not include data obtained from tumor samples. The displayed peptide sequences may be identified from cells expressing a single MHC allele. In this example, the displayed peptide sequences are generally collected from single-allelic cell lines that have been engineered to express predetermined MHC alleles and subsequently exposed to synthetic proteins. Peptides displayed on MHC alleles are isolated by techniques such as acid elution and identified by mass spectrometry. FIG. 2B shows an example of this where an exemplary peptide YEMFNDKSQRAPDDKMF presented on a predetermined MHC allele HLA-DRB1*12:01 was isolated and identified by mass spectrometry. In this context, the peptide is specified through cells engineered to express a single, predetermined MHC protein, thus direct association between the presented peptide and the MHC protein to which it is bound. is known definitively.

[0232] Displayed peptide sequences may also be collected from cells expressing multiple MHC alleles. Typically in humans, 6 different types of MHCI molecules and up to 12 different types of MHCII molecules are expressed in cells. Such displayed peptide sequences may be identified from a polyallelic cell line that has been engineered to express multiple predetermined MHC alleles. Such displayed peptide sequences may also be identified from tissue samples, either normal or tumor tissue samples. Specifically in this instance, 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. FIG. 2C shows six exemplary peptides YEMFNDKSF, HROEIFSHDFJ, FJIEJFOESS, NEIOREIREI, JFKSIFEMMSJDSSUIFLKSJFIEIFJ, and KNFLENFIESOFI 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 presented above, isolated and characterized by mass spectrometry , an example of this is shown. In contrast to single-allelic cell lines, the direct association between the presented peptide and the MHC protein to which it binds is unknown, since the bound peptide is isolated from the MHC molecule before it has been identified. could be.

[0233] Allelic interaction information can also include mass spectrometric ion currents, which depend on both the concentration of peptide-MHC molecule complexes and the ionization efficiency of the peptide. Ionization efficiency varies from peptide to peptide in a sequence-dependent manner. In general, ionic efficiencies vary from peptide to peptide over approximately two orders of magnitude, while concentrations of peptide-MHC complexes vary over a larger range.

[0234] Allelic interaction information can also include measured or predicted binding affinities between a given MHC allele and a given peptide. One or more affinity models can generate such predictive value (72,73,74). For example, returning to the example shown in FIG. * A binding affinity prediction value of 1000 nM between 01:01 can be included. Few peptides with IC50>1000 nm are presented by MHC, and lower IC50 values ​​increase the probability of presentation. Presentation information 165 may include binding affinity predictions between peptide KNFLENFIESOFI and class II allele HLA-DRB1:11:01.

[0235] Allelic interaction information can also include measurements or predictions of MHC complex stability. One or more stability models can generate such predictors. A more stable peptide-MHC complex (ie, a complex with a longer half-life) is more likely to be presented in high copy number on tumor cells and on antigen-presenting cells that encounter the vaccine antigen. For example, returning to the example shown in FIG. 2C, presentation information 165 may include a half-life stability prediction value of 1 hour for class I molecule HLA-A*01:01. The presentation information 165 may also include half-life stability predictions for the class II molecule HLA-DRB1:11:01.

[0236] Allelic interaction information can also include measured or predicted rates of peptide-MHC complex formation reactions. Complexes that form at a faster rate are more likely to be presented on the cell surface at high concentrations.

[0237] Allelic interaction information can also include peptide sequence and length. MHC class I molecules typically prefer to present peptides with a length of 8-15 peptides. 60-80% of the presented peptides have a length of 9. MHC class II molecules generally tend to present peptides with a length of 6-30 peptides.

[0238] Allelic interaction information can also include the presence of kinase sequence motifs on the neoantigen-encoding peptide and the presence or absence of specific post-translational modifications on the neoantigen-encoding peptide. The presence of kinase motifs influences the probability of post-translational modifications that can enhance or interfere with MHC binding.

[0239] Allelic interaction information should also include expression or activity levels of proteins involved in post-translational modification processes, such as kinases (as measured or predicted by RNA seq, mass spectrometry, or other methods). can be done.

[0240] Allele interaction information can also include the probability of presentation of peptides with similar sequences in cells from other individuals expressing a particular MHC allele as assessed by mass spectrometry proteomics or other means. can.

[0241] Allelic interaction information can also include expression levels of particular MHC alleles in the individual in question (eg, as determined by RNA-seq or mass spectrometry). Peptides that bind most strongly to MHC alleles that are expressed at high levels are more likely to be presented than peptides that bind most strongly to MHC alleles that are expressed at low levels.

[0242] Allelic interaction information can also include global neoantigen-encoding peptide sequence-independent probabilities of presentation by a particular MHC allele in other individuals who express the particular MHC allele.

[0243] Allele interaction information can also be used to assess the overall presentation by MHC alleles of the same family of molecules (e.g., HLA-A, HLA-B, HLA-C, HLA-DQ, HLA-DR, HLA-DP) in other individuals. It can also include probabilities that are independent of specific peptide sequences. For example, HLA-C molecules are typically expressed at lower levels than HLA-A or HLA-B molecules, and thus presentation of peptides by HLA-C is less than that of HLA-A or HLA-B II. is a priori less probable than As another example, HLA-DP is generally expressed at lower levels than HLA-DR or HLA-DQ, so presentation of peptides by HLA-DP is more It is assumed that the probability is low.

[0244] Allele interaction information can also include the protein sequence of a particular MHC allele.

[0245] Any MHC allele non-interacting information listed in the section below can also be modeled as MHC allele interacting information.

[0246] VII.B.2. Allele non-interacting information Allelic non-interacting information can include C-terminal sequences adjacent to neoantigen-encoding peptides within the source protein sequence. In MHC-I, C-terminal flanking sequences (adjacent sequences) can influence proteasomal processing of peptides. However, the C-terminal flanking sequences are cleaved from the peptide by the proteasome before the peptide is transported to the endoplasmic reticulum and encounters the MHC allele on the surface of the cell. As a result, the MHC molecule does not receive any information about the C-terminal flanking sequences and therefore the effect of the C-terminal flanking sequences cannot vary depending on the MHC allele type. For example, returning to the example shown in FIG. 2C, the presentation information 165 may include the C-terminal flanking sequence FOEIFNDKSLDKFJI of the presented peptide FJIEJFOESS identified from the peptide's source protein.

[0247] Allele non-interaction information can also include mRNA quantification measurements. For example, mRNA quantification data can be obtained for the same sample that provided the mass spectrometry training data. As described below with respect to Figure 13H, RNA expression was identified as a strong predictor of peptide presentation. In one embodiment, mRNA quantification measurements are determined from the software tool RSEM. A detailed implementation 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 of transcript per million mapped reads (FPKM).

[0248] Allelic non-interaction information can also include N-terminal sequences adjacent to the peptide within the source protein sequence.

[0249] Allelic non-interacting information can also include the source gene of the peptide sequence. A source gene can be defined as an Ensembl protein family of peptide sequences. In other examples, the source gene can be defined as the source DNA or source RNA of the peptide sequence. A source gene can be represented, for example, as a string of nucleotides encoding a protein, or alternatively based on a named set of known DNA or RNA sequences known to encode a particular protein. It can be expressed in a more categorical form. In another example, allelic non-interaction information can also include a source transcript or isoform or set of potential source transcripts or isoforms of peptide sequences extracted from databases such as Ensembl or RefSeq.

[0250] Allelic non-interaction information can also include the tissue type, cell type, or tumor type of the cell from which the peptide sequence was derived.

[0251] Allele non-interaction information can also include the presence of protease cleavage motifs in peptides, optionally weighted according to the expression of the corresponding protease in tumor cells (as determined by RNA-seq or mass spectrometry). can. Peptides containing protease cleavage motifs are less likely to be presented because they are more readily degraded by proteases and therefore are less stable in cells.

[0252] Allele non-interaction information can also include the turnover rate of the source protein as measured in appropriate cell types. A faster turnover rate (ie, lower half-life) increases the probability of presentation, but the predictive power of this property is low when measured in dissimilar cell types.

[0253] Allelic non-interaction information may also be determined by RNA-seq or proteome mass spectrometry, or predicted from annotation of germline or somatic splicing variants detected in DNA or RNA sequence data. Optionally, the length of the source protein can also be included, taking into account the specific splice variant (“isoform”) that is most highly expressed in tumor cells.

[0254] Allele non-interaction information can also include levels of expression 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. Greater weight is given to the cleavage preference of each type of proteasome in proportion to its expression level.

[0255] Allelic non-interaction information can also include expression of the source gene of the peptide (eg, as measured 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 within the tumor sample. Peptides from higher expressed genes are more likely to be presented. Peptides derived from genes with undetectable levels of expression can be eliminated from consideration.

[0256] Allele non-interaction information also indicates that the source mRNAs of neoantigen-encoding peptides are subject to nonsense-dependent degradation mechanisms, such as those predicted by models from Rivas et al, Science 2015. It can also include the probability that it will be done.

[0257] Allelic non-interaction information can also include typical tissue-specific expression of the peptide's source gene during various stages of the cell cycle. Expressed at low levels overall (as measured by RNA-seq or sample analysis proteomics) but known to be expressed at high levels during specific stages of the cell cycle A gene is more likely to produce a more displayed peptide than a gene that is stably expressed at very low levels.

[0258] Allele non-interacting information can also include a comprehensive catalog of properties of the source protein, for example as provided in uniProt or PDB http: / / www.rcsb.org / pdb / home / home.do . These properties may include protein secondary and tertiary structure, subcellular localization11, gene ontology (GO) terminology, among others. Specifically, this information may include annotations that operate at the protein level, such as the 5'UTR length, and annotations that operate at the level of specific residues, such as the helical motif at residues 300-310. . These properties can also include turn motifs, sheet motifs, and disordered residues.

[0259] Allelic non-interaction information can also include properties that describe the nature of the domains of the source protein containing the peptide, such as secondary or tertiary structure (eg, α-helix vs. β-sheet); alternative splicing.

[0260] Allelic 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.

[0261] Allele non-interacting information also adjusted for the probability of presentation of peptides derived from the source protein of the peptide in question in other individuals (the level of expression of the source protein in those individuals and the influence of different HLA types in those individuals). after) can also be included.

[0262] Allelic non-interaction information can also include the probability that a peptide will be undetected or overrepresented by mass spectrometry due to technical bias.

[0263] Measured by gene expression assays such as RNASeq, microarrays, targeted panels such as Nanostrings, or assays such as RT-PCR that are informative about the status of tumor cells, stroma, or tumor infiltrating lymphocytes (TILs) Expression of different gene modules / pathways as measured by single / multiple genes representing the gene modules (not necessarily containing the source protein of the peptide).

[0264] Allelic non-interaction information can also include the copy number of the peptide's source gene in the tumor cell. For example, peptides derived from genes that are subject to homozygous deletion in tumor cells can be assigned a presentation probability of zero.

[0265] Allelic non-interaction information can also include the probability that the peptide binds to TAP, or the measured or predicted binding affinity of the peptide for TAP. Peptides that are more likely to bind TAP, or that bind TAP with higher affinity, are more likely to be presented by MHC-I.

[0266] Allele non-interaction information can also include expression levels of TAP in tumor cells (which can be measured by RNA-seq, proteome mass spectrometry, immunohistochemistry). In MHC-I, higher TAP expression levels increase the probability of presentation of all peptides.

[0267] Allelic non-interacting information can also include the presence or absence of tumor mutations, including but not limited to: Driver mutations in known cancer driver genes such as i.EGFR, KRAS, ALK, RET, ROS1, TP53, CDKN2A, CDKN2B, NTRK1, NTRK2, NTRK3. ii. 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 of the genes encoding the proteasome or immunoproteasome components). Peptides whose presentation relies on components of the antigen presentation machinery that are under the influence of loss-of-function mutations in tumors have a reduced probability of presentation.

[0268] Presence or absence of functional germline polymorphisms, including but not limited to: i. 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 of the genes encoding the proteasome or immunoproteasome components).

[0269] Allele non-interacting information can also include tumor type (eg, NSCLC, melanoma).

[0270] Allele non-interaction information can also include known functionality of HLA alleles, as reflected by HLA allele suffixes, for example. For example, the N suffix in the allele name HLA-A*24:09N indicates a null allele that is not expressed and therefore unlikely to present an epitope; full HLA allele suffix nomenclature is available at https: / / See www.ebi.ac.uk / ipd / imgt / hla / nomenclature / suffixes.html.

[0271] Allele non-interaction information can also include clinical tumor subtype (eg, squamous lung cancer vs. nonsquamous).

[0272] Allele non-interaction information can also include smoking history.

[0273] Allele non-interaction information can also include a history of sunburn, sun exposure, or exposure to other mutagens.

[0274] Allelic non-interaction information can also include local expression of the peptide's source gene in relevant tumor types or clinical subtypes, optionally stratified by driver mutations. Genes that are typically expressed at high levels in relevant tumor types are more likely to be represented.

[0275] Allele non-interacting information is also available in all tumors, or in tumors of the same type, or in tumors from individuals with at least one shared MHC allele, or in individuals with at least one shared MHC allele. The frequency of mutations in tumors of the same type can also be included.

[0276] In the example of a mutated tumor-specific peptide, the list of properties used to predict the probability of presentation can also be an annotation of the mutation (e.g., missense, readthrough, frameshift, fusion, etc.), or whether the mutation is nonsense. It may also include whether it is predicted to result in a mutation dependent degradation mechanism (NMD). For example, peptides from protein segments that are not translated in tumor cells due to homozygous premature termination mutations can be assigned a presentation probability of zero. NMD results in decreased mRNA translation, which reduces the probability of presentation.

[0277] VII.C. Presentation specific system FIG. 3 is a high-level block diagram illustrating the computer logic components of presentation identification system 160, according to one embodiment. In this exemplary embodiment, presentation identification system 160 includes data management module 312 , encoding module 314 , training module 316 , and prediction module 320 . The presentation identification system 160 also includes a training data store 170 and a presentation model store 175 . Some embodiments of model management system 160 have different modules than those described herein. Likewise, functionality may be distributed among modules in ways different from those described herein.

[0278] VII.C.1. Data management module Data management module 312 generates a set of training data 170 from presentation information 165 . Each training data set contains a number of data examples, each data example i at least one peptide sequence p i and the peptide sequence p i One or more relevant MHC alleles a combined with i and the dependent variable y i and the independent variable z i contains a set of

[0279] In one particular implementation referred to throughout the remainder of this specification, the dependent variable y i is the peptide p i but one or more relevant MHC alleles a i is a binary label indicating whether it was presented by However, in other implementations, the dependent variable y i , the presented identification system 160 determines that the independent variable z i It is recognized that it can represent any other kind of information that we are interested in predicting depending on . For example, in another implementation, the dependent variable y i can also be a numerical value indicative of the mass spectrometric ion current specified for the example data.

[0280] Peptide sequence p for data example i i is k i is a sequence of amino acids, and k i can vary within a range between data instances i. For example, the range can 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 data set i may have the same length, eg 9. The number of amino acids in a peptide sequence can vary depending on the type of MHC allele (eg, MHC alleles in humans, etc.). MHC allele a for example data i i is the peptide sequence p i indicates whether it existed in combination with

[0281] Data management module 312 also manages peptide sequences p contained in training data 170 i and combined MHC allele a i together with the binding affinity b i and stability s i Additional allele interaction variables may also be included, such as the predicted value of For example, the training data 170 is the peptide p i and a i Binding affinity prediction value b between each of the bound MHC molecules shown in i can contain As another example, the training data 170 is a i The stability predictive value s for each of the MHC alleles shown in i can contain

[0282] The data management module 312 also stores the peptide sequence p i along with allele non-interacting variables such as C-terminal flanking sequences and mRNA quantification measurements w i can also include

[0283] Data management module 312 also identifies peptide sequences not presented by MHC alleles to generate training data 170 . Generally, this involves identifying a "longer" sequence of the source protein that contains the displayed peptide sequence prior to presentation. If the presentation information contains an engineered cell line, the data management module 312 identifies a set of peptide sequences in the cell-exposed synthetic protein that were not presented on the cell's MHC allele. If the presentation information contains a tissue sample, the data management module 312 identifies the source protein from which the presented peptide sequence originates and extracts a series of sequences in the source protein that were not presented on the MHC alleles of the tissue sample cell. Identify the peptide sequence.

[0284] 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 MHC alleles. This can be accomplished by randomly generating peptide sequences, allowing data management module 312 to easily generate large amounts of synthetic data for peptides not presented on MHC alleles. In fact, because a small percentage of peptide sequences are presented by MHC alleles, synthetically produced peptide sequences are not presented by MHC alleles, even if they are included in proteins processed by the cell. Very likely.

[0285] FIG. 4 illustrates an exemplary set of training data 170A, according to one embodiment. Specifically, the first three data examples in training data 170A are peptides from a single allele cell line containing allele HLA-C*01:03 and three peptide sequences QCEIOWAREFLKEIGJ, FIEUHFWI, and FEWRHRJTRUJR. Indicates presentation information. The fourth data example in training data 170A is from a multi-allelic cell line containing alleles HLA-B*07:02, HLA-C*01:03, HLA-A*01:01 and peptide sequence QIEJOEIJE. Peptide information is shown. The first data example shows that the peptide sequence QCEIOWARE was not presented by the allele HLA-DRB3:01:01. As discussed in the previous two paragraphs, the negatively labeled peptide sequences may be randomly generated by the data management module 312 or identified from the source protein of the presented peptides. Training data 170A also includes a binding affinity prediction value of 1000 nM and a half-life stability prediction value of 1 hour for peptide sequence-allele pairs. Training data 170A also includes the C-terminal flanking sequence of peptide FJELFISBOSJFIE, and 10 2 Also include allelic non-interacting variables, such as TPM mRNA quantification measurements. The fourth data example shows that peptide sequence QIEJOEIJE was presented by one of 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 of the alleles, as well as peptide C-terminal flanking sequences and mRNA quantification measurements for the peptide.

[0286] VII.C.2. Encoding module Encoding module 314 encodes information contained in training data 170 into numerical representations that can be used to generate one or more presentation models. In one implementation, the encoding module 314 one-hot encodes sequences (eg, peptide sequences or C-terminal flanking sequences) for a predetermined 20-letter amino acid alphabet. Specifically, k i a peptide sequence p with amino acids i is 20 k i p, represented as a row vector of elements and corresponding to the amino acid alphabet at the j-th position of the peptide sequence i 20·(j-1)+1 , p i 20·(j-1)+2 , …, p i 20·j A single element in has a value of 1. Otherwise, the 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 data The 3-amino acid peptide sequence EAF of example i is represented by the 60-element row vector p i =[0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]. C-terminal flanking sequence c i , as well as the protein sequence d for the MHC allele h , and other sequence data in the presentation information can similarly be encoded as described above.

[0287] If the training data 170 contains sequences of amino acids of different lengths, the encoding module 314 further divides the peptides into equivalent lengths by adding PAD letters to extend the predetermined alphabet. can be encoded into a vector. For example, this can be done by left-padding the peptide sequence with PAD characters until the length of the peptide sequence reaches the peptide sequence with the largest length in the training data 170 . Therefore, the peptide sequence with maximum length is k 最大 amino acids, encoding module 314 converts each sequence to (20+1) k最大 Represent numerically as a row vector of elements. As an 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 represented by the 105 element row vector p i =[1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]. C-terminal flanking sequence c i Or other sequence data can be similarly encoded as above. Therefore, the peptide sequence p i or c i Each independent variable or column in represents the presence of a particular amino acid at a particular position in the sequence.

[0288] Although the above methods of encoding sequence data have been described in terms of sequences having amino acid sequences, the methods can be extended to other types of sequence data as well, such as, for example, DNA or RNA sequence data. .

[0289] Encoding module 314 also generates one or more MHC alleles a for data instance i. i into an m-element row vector, where each element h=1, 2, . . . , m corresponds to a unique identified MHC allele. The element corresponding to the MHC allele identified for data instance i has a value of one. Otherwise, the remaining elements have a value of 0. As an example, among m = 4 unique identified MHC allele types {HLA-A*01:01, HLA-C*01:08, HLA-B*07:02, HLA-DRB1*10:01} , alleles HLA-B*07:02 and HLA-DRB1*10:01 for data example i corresponding to a multi-allelic cell line are represented by a four-element row vector a i =[0 0 1 1] and can be represented by a 3 i =1 and a 4 i =1. Examples with four identified MHC allele types are described herein, although the number of MHC allele types can actually be in the hundreds or thousands. As noted above, each data instance i typically contains a peptide sequence p i up to 6 different MHC allele class I types and / or peptide sequences p i Up to 4 different MHC class IIDR allele types and / or peptide sequences p i contains up to 12 different MHC class II alleletypes related to

[0290] Encoding module 314 also creates a label y for each data instance i i is encoded as a binary variable with values ​​from the set of {0,1}, where a value of 1 corresponds to the peptide x i but the associated MHC allele a i A value of 0 indicates that it was presented by one of the peptides x i but the associated MHC allele a i indicates that it was not presented by any of the dependent variable y i represents the mass spectrometric ion current, the encoding module 314 uses various functions such as the log function with a range of [-∞, ∞] for ion current values ​​between [0, ∞] to convert the value to It can be scaled additionally.

[0291] Coding module 314 contains peptide p i and the allelic interaction variable x for the associated MHC allele h h i can be represented as row vectors in which the numerical representations of the allelic interaction variables are concatenated one after the other. For example, the encoding module 314 may use x h i , [p i ], [p i b h i ], [p i s h i ], or [p i b h i s h i ] and can be represented as a row vector equivalent to b h i is the peptide p i and the binding affinity predictive value for the related MHC allele h, as well as s h i is about stability. Alternatively, one or more combinations of allele interaction variables may be stored individually (eg, as individual vectors or matrices).

[0292] In one example, encoding module 314 converts the measured or predicted value for binding affinity into the allele interaction variable x h i represents the binding affinity information by incorporating

[0293] In one example, encoding module 314 converts the measured or predicted value for binding stability to the allele interaction variable x h i represents the binding stability information by incorporating

[0294] In one example, the encoding module 314 converts the measured or predicted value for the binding on-rate into the allele interaction variable x h i represents the binding on-rate information by incorporating

[0295] In one example, for peptides presented by Class I MHC molecules, encoding module 314 stores the peptide length in vector TIFF2022524328000004.tif4128 (however, TIFF2022524328000005.tif3128 is the indicator function and L k is peptide p k (meaning the length of ). Vector T k , the allele interaction variable x h i can be included in In another example, for peptides presented by class II MHC molecules, the encoding module 314 vectorizes the peptide length TIFF2022524328000006.tif18146 (however, TIFF2022524328000007.tif3128 is the indicator function and L k is peptide p k (meaning the length of ). Vector T k , the allele interaction variable x h i can be included in

[0296] In one example, the encoding module 314 converts the RNA-seq-based expression levels of MHC alleles into the allele interaction variable x h i represents the RNA expression information of the MHC alleles by incorporating .

[0297] Similarly, the encoding module 314 uses the allelic non-interacting variable w i can be represented as a row vector in which numerical representations of allelic non-interacting variables are concatenated one after the other. For example, w i is [c i ] or [c i m i w i ], and w i is the peptide p i mRNA quantification measurements associated with the C-terminal flanking sequences and peptides of m i is a row vector representing any other allelic non-interacting variables in addition to . Alternatively, one or more combinations of allelic non-interacting variables may be stored individually (eg, as individual vectors or matrices).

[0298] In one example, the encoding module 314 assigns the turnover rate or half-life to the allele non-interacting variable w i represents the turnover rate of the source protein for the peptide sequence.

[0299] In one example, encoding module 314 converts protein length to allele non-interacting variable w i represents the length of the source protein or isoform by incorporating .

[0300] In one example, encoding module 314 uses β1 i , β2 i , β5 i The average expression of immunoproteasome-specific proteasome subunit containing subunits was defined by the allele non-interacting variable w i represents activation of the immunoproteasome by incorporating .

[0301] In one example, the encoding module 314 compares the RNA-seq abundance of the source protein of the peptide (quantified in units of FPKM, TPM by a technique such as RSEM), or the gene or transcript of the peptide, to that of the source protein. Abundances of allele non-interacting variables w i by incorporating into

[0302] In one example, the coding module 314 is such that the peptide originating transcript will undergo a nonsense mutation-dependent degradation mechanism (NMD), e.g., as predicted by the model in Rivas et.al.Science, 2015. This probability is the allele non-interacting variable w i by incorporating into

[0303] In one example, the encoding module 314 provides the activation status of gene modules or pathways assessed via RNA-seq, e.g. quantifying the expression of the genes in the pathway, and then representing by computing a summary statistic, such as the mean, over the genes in the pathway. Let the mean be the allelic non-interacting variable w i can be incorporated into

[0304] In one example, the encoding module 314 converts the copy number of the source gene to the allele non-interacting variable w i by incorporating into

[0305] In one example, encoding module 314 converts the measured or predicted TAP binding affinities (eg, in nanomolar units) to allele non-interaction variables w i represents the TAP binding affinity by including in

[0306] In one example, the encoding module 314 converts TAP expression levels measured by RNA-seq (and quantified in units of TPM, eg, by RSEM) into allele non-interacting variables w i represents TAP expression levels by including in

[0307] In one example, encoding module 314 maps tumor mutations to allele non-interacting variables w i A vector of indicator variables (i.e., peptides p k is from a sample with the KRAS G12D mutation, d k = 1, otherwise 0).

[0308] In one example, encoding module 314 maps germline polymorphisms in antigen-presenting genes into a vector of indicator variables (i.e., peptide p k is from a sample with a specific germline polymorphism in TAP, d k =1). Let these indicator variables be the allelic non-interacting variables w i can be included in

[0309] In one example, the encoding module 314 represents the tumor type as a one-hot encoded vector of length 1 for the alphabet of tumor types (eg, NSCLC, melanoma, colon cancer, etc.). These one-hot coded variables are allele non-interacting variables w i can be included in

[0310] In one example, the encoding module 314 represents MHC allele suffixes by treating the 4-digit HLA alleles with various suffixes. For example, HLA-A*24:09N is considered a different allele than HLA-A*24:09 for modeling purposes. Alternatively, the probability of presentation by N-suffixed MHC alleles can be set to zero for all peptides, since HLA alleles ending in the N-suffix are not expressed.

[0311] In one example, the encoding module 314 represents the tumor subtype as a one-hot encoded vector of length 1 for the alphabet of tumor subtypes (eg, lung adenocarcinoma, lung squamous cell carcinoma, etc.). These one-hot coded variables are allele non-interacting variables w i can be included in

[0312] In one example, encoding module 314 combines smoking history with allele non-interacting variable w i can be included in the binary index variable (d k = 1, otherwise 0). Alternatively, smoking history can be coded as a length 1 one-hot coded variable for the smoking severity alphabet. For example, smoking status can be assessed on a scale of 1-5, where 1 indicates a nonsmoker and 5 indicates a current heavy smoker. Smoking history is primarily associated with lung tumors, so when training models for multiple tumor types, this variable should be set to 1 if the patient has a history of smoking and the tumor type is lung tumor. It can also be defined as equal and zero otherwise.

[0313] In one example, the encoding module 314 combines the history of sunburn with the allelic non-interacting variable w i a binary index variable (if the patient has a history of severe sunburn, d k = 1, otherwise 0). Severe sunburn is primarily associated with melanoma, so when training a model for multiple tumor types, this variable is useful if the patient has a history of severe sunburn and the tumor type is melanoma. It can also be defined to be equal to 1 if it is, and zero otherwise.

[0314] In one example, the encoding module 314 calculates the distribution of expression levels of a particular gene or transcript for each gene or transcript in the human genome by using a reference database such as TCGA to summarize expression level distribution statistics. Express as a quantity (eg mean, median). Specifically, peptide p in samples with tumor type melanoma k for peptide p k The measured gene or transcript expression level of the originating gene or transcript of the allele non-interacting variable w i as well as peptide p in melanoma, as measured by TCGA k Mean and / or median gene or transcript expression of the genes or transcripts of origin can also be included.

[0315] In one example, the encoding module 314 represents the mutation type as a one-hot encoded variable of length 1 for the alphabet of mutation types (eg, missense, frameshift, NMD-inducible, etc.). These one-hot coded variables are allele non-interacting variables w i can be included in

[0316] In one example, the encoding module 314 expresses the protein-level properties of a protein as the value of the source protein's annotation (e.g., 5'UTR length) as the allele non-interacting variable w i Represented in In another example, the encoding module 314 includes the peptide p i Residue-level annotation of the source protein for peptide p i is equivalent to 1 if overlaps the helical motif, otherwise 0, or the peptide p i Let the allelic non-interacting variable w i represented by including in In another example, the peptide p contained within the helix motif annotation i Let the allelic non-interacting variable w i can be included in

[0317] In one example, the encoding module 314 identifies the types of proteins or isoforms in the human proteome as an index vector o having a length equal to the number of proteins or isoforms in the human proteome. k and the corresponding element o k i is the peptide p k is 1 if is from protein i, 0 otherwise.

[0318] In one example, encoding module 314 encodes peptide p i source gene G=gene(p i ) as a categorical variable with L possible categories, where L denotes the upper bound 1, 2, . . . , L of the number of subscripted source genes.

[0319] In one example, encoding module 314 encodes peptide p i tissue type, cell type, tumor type, or tumor histology type T = tissue (p i ) as a categorical variable with M possible categories, where M indicates an upper bound on the number of subscripted types 1, 2, ..., M). Tissue types can include, for example, lung tissue, heart tissue, intestinal tissue, nerve tissue, and the like. Cell types can include dendritic cells, macrophages, CD4 T cells, and the like. Examples include lung adenocarcinoma, lung squamous cell carcinoma, melanoma, non-Hodgkin's lymphoma, and the like.

[0320] Encoding module 314 also includes peptide p i and the variable z for the associated MHC allele h i the overall set of allele interaction variables x i and the allelic non-interacting variable w i can also be represented as a row vector in which the numerical representations of are concatenated one after the other. For example, the encoding module 314 may use z h i , [x h i w i ] or [w i x h i ] can be represented as a row vector equivalent to

[0321] VIII. TRAINING MODULE Training module 316 builds one or more presentation models that generate likelihoods of whether a peptide sequence will be presented by an MHC allele associated with the peptide sequence. Specifically, the peptide sequence p k and peptide sequence p k MHC allele a related to k , and each presented model has the peptide sequence p k but the associated MHC allele a k an estimate u, which indicates the likelihood that it would be offered by one or more of k to generate

[0322] VIII.A. Overview Training module 316 builds one or more presentation models based on training data sets stored in storage 170 generated from presentation information stored in 165 . In general, 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 such that the loss function is minimized. Specifically, the loss function (y i∈S ,u i∈S ;θ) is the dependent variable y for one or more data examples S in the training data 170 i∈S and the estimated likelihood u for the data example S generated by the presented model i∈S represents a contradiction between In one particular implementation that will be referred to throughout the rest of this specification, the loss function (y i∈S ,u i∈S ; θ) is the negative log-likelihood function given by equation (1a) as follows. TIFF2022524328000008.tif10128 However, in practice a different loss function may be used. For example, if predictions are made for the mass analysis ion current, the loss function is the mean squared loss given by Equation 1b below. TIFF2022524328000009.tif10128

[0323] The presentation model can be a parametric model in which one or more parameters θ mathematically specify the dependencies between the independent and dependent variables. Typically, the loss function (y i∈S ,u i∈S Various parameters of the parametric type presentation model that minimizes ; θ) are determined through gradient-based numerical optimization algorithms, such as batch gradient algorithms, stochastic gradient algorithms, and the like. Alternatively, the presented model can be a non-parametric model in which the model structure is determined from the training data 170 and not strictly based on a fixed set of parameters.

[0324] VIII.B. Models per allele The training module 316 may build a presentation model to predict the presentation likelihood of peptides on an allele-by-allele basis. In this example, training module 316 may train a presentation model based on data examples S in training data 170 generated from cells expressing a single MHC allele.

[0325] In one implementation, training module 316 includes: Peptide p for a particular allele h by TIFF2022524328000010.tif7128 k estimated presentation likelihood u k where the peptide sequence x h k is the peptide p k and the encoded allele interaction variable for the corresponding MHC allele h, where f(•) is an arbitrary function, referred to throughout this specification as a transformation function for convenience of description. Furthermore, g h ( ) is an arbitrary function, for convenience of description referred to throughout this specification as the dependency function, determined for the MHC allele h, the parameter θ h based on the set of allelic interaction variables x h k Generate a dependency score for Parameter θ for each MHC allele h h The set of values ​​for θ h where i is each instance in a subset S of training data 170 generated from cells expressing a single MHC allele h.

[0326] dependency function g h (x h k ;θ h ) that the MHC allele h has at least the allele interaction characteristic x h k and in particular the peptide p k Dependency score for MHC allele h indicating whether it presents the corresponding neoantigen based on the amino acid position of the peptide sequence of . For example, the dependence score for MHC allele h is that MHC allele h is k It may have a high value if it is likely to present , and may have a low value if it is not likely to present. The transformation function f( ) transforms the input, more specifically g h (x h k ;θ h ), the dependence score generated by the peptide p k to an appropriate value indicating the likelihood that it would be presented by the MHC allele.

[0327] In one particular implementation referred to throughout the remainder of this specification, f(•) is a function with a range in [0,1] for the appropriate domain range. In one example, f( ) is The expit function given by TIFF2022524328000011.tif10128. As another example, f( ) can also be It can also be the hyperbolic tangent function given by TIFF2022524328000012.tif6128. Alternatively, if predictions are made for mass spectrometry ion currents with values ​​outside the range [0,1], f( ) is any function, e.g., identity function, exponential function, log function, etc. be able to.

[0328] Therefore, the peptide sequence p k is presented by MHC allele h is the dependence function g for MHC allele h h ( ) is the peptide sequence p k , to generate the corresponding dependency score. The dependence score is the peptide sequence p k may be transformed by a transformation function f(·) to produce the likelihood for each allele that would be presented by MHC allele h.

[0329] VIII.B.1 Dependence functions for allele interaction variables In one particular implementation referred to throughout this specification, the dependency function g h ( ) is x h k Each allele interaction variable in the parameter θ determined for the associated MHC allele h h a linear combination with the corresponding parameters in the set of is the affine function given by TIFF2022524328000013.tif5128.

[0330] In another particular implementation referred to throughout this specification, the dependency function g h ( ) is a network model NN with a series of nodes arranged in one or more layers h represented by ( ), The network function given by TIFF2022524328000014.tif5128. The node uses the parameter θ h can be connected to other nodes through connections each having an associated parameter in the set of . The value at one particular node can be expressed as the sum of the values ​​of the nodes connected to the particular node weighted by the relevant parameters mapped by the activation function associated with the particular node. In contrast to affine functions, network models are advantageous because the presentation model can incorporate non-linearity 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 a peptide sequence and how this interaction affects peptide presentation.

[0331] In general, the network model NNh ( ) denotes feedforward networks such as artificial neural networks (ANN), convolutional neural networks (CNN), deep neural networks (DNN), and / or long-short-term memory networks (LSTM), bidirectional LSTM networks, both It can be structured as a recurrent neural network (RNN) such as a directed recurrent network, a deep bidirectional recurrent network, a multi-layered perceptual network (NLP), and the like.

[0332] In one example referred to throughout the rest of this specification, each MHC allele at h=1, 2, . h (·) means the output from the network model associated with MHC allele h.

[0333] Figure 5 shows an exemplary network model NN associated with an arbitrary MHC allele h=3 3 ( ) is explained. As shown in Figure 5, the network model NN for MHC allele h=3 3 ( ) has 3 types of input nodes at layer l=1, 4 types of nodes at layer l=2, 2 types of nodes at layer l=3, and 1 type of output at layer l=4 Contains nodes. network model NN 3 ( ) are 10 parameters θ 3 (1), θ 3 (2), …, θ 3 Related to the set of (10). network model NN 3 ( ) are three allele interaction variables x for MHC allele h=3 3 k (1), x 3 k (2), and x 3 k Receive input values ​​for (3) (individual data instances, including encoded polypeptide sequence data and any other training data used), and value NN 3 (x 3 k ). A network function may include one or more network models each taking a different allelic interaction variable as input.

[0334] In another example, the identified MHC alleles h=1, 2, . H ( ) and NN h (·) means one or more outputs of a single network model associated with MHC allele h. In such an example, the parameter θ h may correspond to the set of parameters for a single network model, so the parameters θ h set can be shared by all MHC alleles.

[0335] FIG. 6A shows an exemplary network model NN shared by MHC alleles h=1, 2, . H ( ) is explained. As shown in Figure 6A, the network model NN H (·) contains m output nodes, each corresponding to an MHC allele. network model NN 3 ( ) is allelic interaction variable x for MHC allele h=3 3 k and the value NN corresponding to the MHC allele h=3 3 (x 3 k ), output m values.

[0336] In yet another example, a single network model NN H ( ) is the allelic interaction variable x for MHC allele h h k and the encoded protein sequence d h It can be a network model that outputs a dependency score given In such an example, the parameter θ h can again correspond to the set of parameters for a single network model, so the parameters θ h set can be shared by all MHC alleles. Therefore, in such an example, NNh( ) is the input [x h k d h ], a single network model NN H means the output of ( ). Such network models are advantageous because they can correctly predict peptide presentation probabilities for MHC alleles that were unknown in the training data simply by specifying their protein sequences.

[0337] Figure 6B shows an exemplary network model NN shared by MHC alleles H ( ) is explained. As shown in Figure 6B, the network model NN H ( ) receives the allele interaction variable and protein sequence of MHC allele h=3 as input, and the dependency score NN corresponding to MHC allele h=3 3 (x 3 k ).

[0338] In yet another example, the dependency function g h (·)teeth, TIFF2022524328000015.tif5128, where g' h (x h k ;θ’ h ) is an affine function, network function, etc., with a set of parameters θ'h, the bias parameter θ h 0 Accompanied by

[0339] In another implementation, the bias parameter θ h 0 may be shared according to the gene family of MHC allele h. That is, the bias parameter θ for MHC allele h h 0 is θ 遺伝子(h) 0 and gene (h) is the gene family of MHC allele h. For example, the class I MHC alleles HLA-A*02:01, HLA-A*02:02, and HLA-A*02:03 may be assigned to the "HLA-A" gene family, and these MHC bias parameter θ for each of the alleles h 0 may be shared. As another example, class II MHC alleles HLA-DRB1:10:01, HLA-DRB1:11:01, and HLA-DRB3:01:01 are assigned to the "HLA-DRB" gene family, and these MHC alleles Each bias parameter θ h 0 can be shared.

[0340] As an example, returning to equation (2), the affine dependence function g h Peptide p by MHC allele h=3 among m=4 different identified MHC alleles with ( ) k The likelihood that TIFF2022524328000016.tif5128, where x3k is the allele interaction variable identified for MHC allele h=3 and θ3 is the parameter determined for MHC allele h=3 through loss function minimization. is a set of

[0341] As another example, peptide p by MHC allele h = 3 among m = 4 different identified MHC alleles using separate network transformation functions gh( ) k The likelihood that TIFF2022524328000017.tif5128, where x 3 k is the allelic interaction variable identified for MHC allele h=3, and θ 3 is the network model NN associated with the MHC allele h=3 3 is the set of parameters determined for (·).

[0342] Figure 7 shows an exemplary network model NN 3 Peptide p associated with MHC allele h=3 with ( ) k We describe the generation of the presentation likelihood of . As shown in Figure 7, the network model NN 3 ( ) is allelic interaction variable x for MHC allele h=3 3 k and output NN 3 (x 3 k ). The output is mapped by the function f( ) to give the estimated presentation likelihood u k to generate

[0343] VIII.B.2. Per allele with allelic non-interacting variables In one implementation, the training module 316 incorporates allele non-interacting variables to Peptide p k model the estimated presentation likelihood uk of , where w k is the peptide p k Mean coded allelic non-interacting variables for g w ( ) is the parameter θ determined for allele non-interacting variablesw Allele non-interacting variable w k is a function about Specifically, the parameter θ for each MHC allele h h and the parameter θ for allelic non-interacting variables w , the value of the set of θ h and θ w where i is each instance in a subset S of training data 170 generated from cells expressing a single MHC allele.

[0344] dependency function g w (w k ;θ w ) outputs peptide p k Represents the dependence score for allele non-interacting variables, indicating whether or not is presented. For example, the dependence score for allele non-interacting variables is peptide p k C-terminal flanking sequences and peptide p k may have high values ​​if the peptide p k C-terminal flanking sequences and peptide p k is bound, it may have a low value.

[0345] According to equation (8), the peptide sequence p k is presented by MHC allele h is the function g h ( ) is the peptide sequence p k to generate the corresponding dependency scores for the allelic interaction variables. function g for allelic non-interacting variables w (·) is also applied to the coded version of the allele non-interacting variables to generate a dependency score for the allele non-interacting variables. Both scores are combined and the combined score is the peptide sequence p k is transformed by the transformation function f(⋅) to generate the likelihood for each allele that would be presented.

[0346] Alternatively, the training module 316 uses the allele non-interacting variable w k the allele interaction variable x h k may include allele non-interacting variables wk in the prediction by appending to . Therefore, the suggested likelihood is TIFF2022524328000019.tif7128.

[0347] VIII.B.3 Dependence functions for allele non-interacting variables dependency function g for allele interaction variables h Similar to ( ), the dependence function g for allele non-interacting variables w ( ) is the affine function, or separate network model is the allelic non-interacting variable w k can be a network function associated with

[0348] Specifically, the dependency function g w ( ) is w k Allele non-interacting variables at the parameter θ w a linear combination with the corresponding parameters in the set of is the affine function given by TIFF2022524328000020.tif6128.

[0349] dependency function g w ( ) is also the parameter θ w A network model NN with associated parameters in the set of w represented by ( ), The network function given by TIFF2022524328000021.tif6128. A network function may include one or more network models each taking a different allelic non-interacting variable as input.

[0350] In another example, the dependence function g for allele non-interacting variables w (·)teeth, TIFF2022524328000022.tif5128, where g' w (w k ;θ’ w ) is the allelic non-interaction parameter θ’ w are affine functions, network functions, etc., with the set of m k is the peptide p k is the mRNA quantification measurement for, h(·) is the function that transforms the quantification measurement, and θ w m is the parameter in the set of parameters for the allele non-interacting variables that is combined with the mRNA quantification measurement 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, although in practice h( ) can be any one of a variety of different functions. can be one.

[0351] In yet another example, the dependence function g w (·)teeth, Given by TIFF2022524328000023.tif5128, where g’ w (w k ;θ’ w ) is the allelic non-interaction parameter θ’ w are affine functions, network functions, etc. with a set of o k is the peptide p k is an index vector, as described in Section VII.C.2, representing proteins and isoforms in the human proteome for , and θ w o is the set of parameters in the set of parameters for the allelic non-interacting variables combined with the index vector. In one variation, o k and parameter θ w o If the dimensionality of the set of is significantly higher, λ||θ w o A parameter regularization term such as || (where || || stands for L1 norm, L2 norm, combination, etc.) can be added to the loss function when determining the values ​​of the parameters. An optimal value for the hyperparameter λ can be determined through any suitable method.

[0352] In yet another example, the dependence function g w ( ) is given by the following equation. i.e. TIFF2022524328000024.tif13128 but g’ w (w k ;θ’ w ) is the allelic non-interaction parameter θ’ w are affine functions, network functions, etc., with the set of TIFF2022524328000025.tif5128 is peptide p k is an indicator function equal to 1 if is from the source gene l mentioned above for allele non-interacting variables, and θ w l is a parameter indicating the "antigenicity" of the source gene l. In one variation, L is large enough so that the number of parameters θ w l=1, 2,...,L is large enough, λ||θ w l A parameter regularization term such as || (where ||·|| is the L1 norm, L2 norm, combination, etc.) can be added to the loss function when determining the values ​​of the parameters. An optimal value for the hyperparameter λ can be determined by any suitable method.

[0353] In yet another example, the dependence function g w ( ) is given by the following equation. i.e. TIFF2022524328000026.tif13147 but g’ w (w k ;θ’ w ) is the allelic non-interaction parameter θ’ w are affine functions, network functions, etc. with the set of TIFF2022524328000027.tif5128 is peptide p k is derived from source gene l, and peptide p k is an indicator function equal to 1 if is from tissue type m, and θ w lm is a parameter that indicates the antigenicity of the combination of source gene l and tissue type m. Specifically, the antigenicity of gene l in tissue type m may indicate a residual propensity for tissue type m cells to present peptides from gene l after regulation for RNA expression and peptide sequence context.

[0354] In one variation, L or M is large enough so that the number of parameters θ w lm=1, 2,...,LM is large enough, λ||θ w lmA parameter regularization term such as || (where ||·|| is the L1 norm, L2 norm, combination, etc.) can be added to the loss function when determining the values ​​of the parameters. An optimal value for the hyperparameter λ can be determined by any suitable 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, a penalty term such as: TIFF2022524328000028.tif18128 (however, TIFF2022524328000029.tif5128 is the average antigenicity across tissue types for source gene l) can be penalized for the standard deviation of antigenicity across different tissue types in the loss function.

[0355] In practice, the dependence function g w ( ) can be generated. For example, by adding together the term h(·) representing the mRNA quantification measurement in equation (10) and the term representing the antigenicity of the source gene in equation (12) together with any other affine or network function, Dependency functions can be generated for allele non-interacting variables.

[0356] As an example, returning to equation (8), the affine transformation function g h ( ), g w Peptide p by MHC allele h=3 among m=4 different identified MHC alleles with ( ) k The likelihood that TIFF2022524328000030.tif5128, where w k is the peptide p k are allelic non-interacting variables identified for θ w is the set of parameters determined for allelic non-interacting variables.

[0357] As another example, the network transformation function g h ( ), g w Peptide p by MHC allele h=3 among m=4 different identified MHC alleles with ( ) k The likelihood that TIFF2022524328000031.tif5128, where w k is the peptide p k is the allelic interaction variable identified for θ w is the set of parameters determined for allelic non-interacting variables.

[0358] Figure 8 shows an exemplary network model NN 3 ( ) and NN w Peptide p associated with MHC allele h=3 with ( ) k We describe the generation of the presentation likelihood of . As shown in Figure 8, the network model NN 3 ( ) is allelic interaction variable x for MHC allele h=3 3 k and output NN 3 (x 3 k ). network model NN w ( ) is the peptide p k allelic non-interacting variable w k and output NN w (w k ). The outputs are combined and mapped by the function f( ) to obtain the estimated presentation likelihood u k to generate

[0359] VIII.C. Multiple allele model The training module 316 may also build a presentation model for predicting the presentation likelihood of peptides in a multi-allele setting where two or more MHC alleles are present. In this example, training module 316 trains the presentation model based on data examples S in training data 170 generated from cells expressing a single MHC allele, cells expressing multiple MHC alleles, or a combination thereof. can be trained.

[0360] VIII.C.1. Example 1: Maximum model values ​​per allele In one implementation, the training module 316 trains the peptide p associated with a set of multiple MHC allele H k estimated presentation likelihood u k be the presentation likelihood u k h∈H modeled as a function of Specifically, the presentation likelihood u k is u k h∈H can be any function of In one implementation, the function is the maximum function and the presentation likelihood u k can be determined as the maximum presentation likelihood for each MHC allele h in set H. TIFF2022524328000032.tif5128

[0361] VIII.C.2. Example 2.1: Sum functional model In one implementation, the training module 316 trains the peptide p k estimated presentation likelihood u k of, Modeled by TIFF2022524328000033.tif13128, where the element a h k is the peptide sequence p k 1 for multiple MHC alleles H associated with x h k is the peptide p k and the coded allele interaction variables for the corresponding MHC alleles. Parameter θ for each MHC allele h h The set of values ​​for θ h , where i is each in a subset S of training data 170 generated from cells expressing a single MHC allele and / or cells expressing multiple MHC alleles For example. dependency function g h is the dependency function g h can be in the form of either

[0362] According to equation (13), the peptide sequence p k will be presented by more than one MHC allele h is the dependency function g h ( ) for each MHC allele H peptide sequence p k to generate the corresponding scores for the allelic interaction variables. The scores for each MHC allele h are combined to give the peptide sequence p k is transformed by the transformation function f(⋅) to produce the presentation likelihood that the set of MHC alleles H would be presented.

[0363] The model presented in equation (13) is that each peptide p k It differs from the allele-by-allele model of equation (2) in that the number of relevant alleles for can be greater than one. In other words, a h k more than one element in the peptide sequence p k can have a value of 1 for multiple MHC allele H associated with

[0364] As an example, the affine transformation function g h Peptide p by MHC allele h=2, h=3 among m=4 different identified MHC alleles with ( ) k The likelihood that TIFF2022524328000034.tif5128, where x 2 k , x 3 k is the allelic interaction variable identified for MHC alleles h=2, h=3, and θ 2 , θ 3 is the set of parameters determined for MHC alleles h=2, h=3.

[0365] As another example, the network transformation function g h ( ), g w Peptide p by MHC allele h=2, h=3 among m=4 different identified MHC alleles with ( ) k The likelihood that TIFF2022524328000035.tif5128 can be generated by NN 2 ( ), NN 3 ( ) is the network model identified for MHC alleles h=2, h=3, θ 2 , θ 3 is the set of parameters determined for MHC alleles h=2, h=3.

[0366] Figure 9 shows an exemplary network model NN 2 ( ) and NN 3 Peptide p associated with MHC alleles h=2, h=3 with ( ) k We describe the generation of the presentation likelihood of . As shown in Figure 9, the network model NN2 ( ) is allele interaction variable x for MHC allele h=2 2 k and output NN 2 (x 2 k ) and generate the network model NN 3 ( ) is allelic interaction variable x for MHC allele h=3 3 k and output NN 3 (x 3 k ). The outputs are combined and mapped by the function f( ) to obtain the estimated presentation likelihood u k to generate

[0367] VIII.C.3. Example 2.2: Sum Functional Model with Allelic Non-Interacting Variables In one implementation, the training module 316 incorporates allele non-interacting variables to Peptide p k estimated presentation likelihood u k is modeled, where w k is the peptide p k means coded allelic non-interacting variables for Specifically, the parameter θ for each MHC allele h h and the parameter θ for allelic non-interacting variables w , the value of the set of θ h and θ w , where i is each in a subset S of training data 170 generated from cells expressing a single MHC allele and / or cells expressing multiple MHC alleles For example. dependency function g w is the dependency function g w can be in the form of either

[0368] Therefore, according to equation (14), the peptide sequence p k is presented is the function g h ( ) for each MHC allele H peptide sequence p k to generate a corresponding dependency score for the allelic interaction variables for each MHC allele h. function g for allelic non-interacting variables w (·) is also applied to the coded version of the allele non-interacting variables to generate a dependency score for the allele non-interacting variables. The scores are combined and the combined score is the peptide sequence p k is transformed by the transformation function f(·) to generate the presentation likelihood that will be presented.

[0369] In the presentation model of equation (14), each peptide p k The number of related alleles for can be greater than one. In other words, a h k more than one element in the peptide sequence p k can have a value of 1 for multiple MHC allele H associated with

[0370] As an example, the affine transformation function g h ( ), g w Peptide p by MHC allele h=2, h=3 among m=4 different identified MHC alleles with ( ) k The likelihood that TIFF2022524328000037.tif5128, where w k is the peptide p k are allelic non-interacting variables identified for θ w is the set of parameters determined for allelic non-interacting variables.

[0371] As another example, the network transformation function g h ( ), g w Peptide p by MHC allele h=2, h=3 among m=4 different identified MHC alleles with ( ) k The likelihood that TIFF2022524328000038.tif5128, where w k is the peptide p k is the allelic interaction variable identified for θ w is the set of parameters determined for allelic non-interacting variables.

[0372] Figure 10 shows an exemplary network model NN 2 ( ), NN 3 ( ), and NN w As shown in Fig. 10, using ( ), the network model NN 2 ( ) is allele interaction variable x for MHC allele h=2 2 k and output NN 2 (x 2 k ). network model NN 3 ( ) is allelic interaction variable x for MHC allele h=3 3 k and output NN 3 (x 3 k ). network model NN w ( ) is the peptide p k allelic non-interacting variable w k and output NN w (w k ). The outputs are combined and mapped by the function f( ) to obtain the estimated presentation likelihood u k to generate

[0373] Alternatively, the training module 316 uses the allele non-interacting variable w k the allele interaction variable x h k allele non-interacting variable w k may include Therefore, the suggested likelihood is TIFF2022524328000039.tif13128.

[0374] VIII.C.4. Example 3.1: A Model Using Potential Allele-Specific Likelihoods In another implementation, the training module 316 trains the peptide p k estimated presentation likelihood u k of, Modeled by TIFF2022524328000040.tif7128, where the element a h k is the peptide sequence p k 1 for multiple MHC alleles h∈H associated with u’ k h is the potential per-allele presentation likelihood for MHC allele h, and vector v is the element v h but a h k u’ k h , s(·) is a function mapping the elements of v, and r(·) is a clipping function that clips the input values ​​into a given range. As described in more detail below, s( ) may be a summation function or a quadratic function, but in other embodiments s( ) is any function, such as a maximum function It is recognized that it is possible. A set of values ​​for the parameter θ for the likelihood of each potential allele can be determined by minimizing a loss function for θ, where i is the number of cells expressing a single MHC allele and / or multiple are examples in subset S of training data 170 generated from cells expressing MHC alleles of .

[0375] The presentation likelihood in the presentation model of equation (17) is that each of the peptides p k The presentation likelihood u' for each potential allele, corresponding to the likelihood that k h is modeled as a function of Potential per-allele likelihood parameters for the potential per-allele likelihood are used in single-allele settings as well as for unknown direct associations between presented peptides and corresponding MHC alleles. It differs from the per-allele presentation likelihood of section VIII.B in that it can be learned from some multi-allele setting. Therefore, in a multi-allele setting, the presentation model is that peptide p k is presented by the set of MHC alleles H as a whole, as well as which MHC allele h is represented by peptide p k individual likelihood u’ k h∈H can also be provided. The advantage of this is that the presentation model can generate potential likelihoods without training data for cells expressing a single MHC allele.

[0376] In one particular implementation referred to throughout the remainder of this specification, r(·) is a function with range [0,1]. For example, r( ) is the clip function: r(z)=min(max(z,0),1) , where the minimum value between z and 1 is the proposed likelihood u k selected as In another implementation, r( ) is r(z)=tanh(z) is the hyperbolic tangent function given as , where the value of domain z is greater than or equal to 0.

[0377] VIII.C.5. Example 3.2: Sum Model of Functions In one particular implementation, s(·) is a summation function and the presentation likelihood is given by summing the presentation likelihood for each potential allele. TIFF2022524328000041.tif15128

[0378] In one implementation, the presentation likelihood for each potential allele for MHC allele h is Generated by TIFF2022524328000042.tif7128, the presentation likelihood is Let it be estimated by TIFF2022524328000043.tif13128.

[0379] According to equation (19), the peptide sequence p k is presented is the function g h ( ) for each MHC allele H peptide sequence p k to generate the corresponding dependency scores for the allelic interaction variables. Each dependence score is first evaluated by the potential allele-wise presentation likelihood u’ k h is transformed by the function f( ) to produce Likelihood for each allele u’ k h are combined and a clipping function is applied to the combined likelihoods to clip the values ​​into the range [0,1], yielding the peptide sequence p k will be presented by the set of MHC allele H can be generated. dependency function g h is the dependency function g h can be in the form of either

[0380] As an example, the affine transformation function g h Peptide p by MHC allele h=2, h=3 among m=4 different identified MHC alleles with ( ) k The likelihood that TIFF2022524328000044.tif7128, where x 2 k , x 3 k is the allelic interaction variable identified for MHC alleles h=2, h=3, and θ 2 , θ 3 is the set of parameters determined for MHC alleles h=2, h=3.

[0381] As another example, the network transformation function g h ( ), g w Peptide p by MHC allele h=2, h=3 among m=4 different identified MHC alleles with ( ) k The likelihood that TIFF2022524328000045.tif7128 can be generated by NN 2 ( ), NN 3 ( ) is the network model identified for MHC alleles h=2, h=3, θ 2 , θ 3 is the set of parameters determined for MHC alleles h=2, h=3.

[0382] Figure 11 shows an exemplary network model NN 2 ( ) and NN 3 Peptide p associated with MHC alleles h=2, h=3 with ( ) k We describe the generation of the presentation likelihood of . As shown in Figure 9, the network model NN 2 ( ) is allele interaction variable x for MHC allele h=2 2 k and output NN 2 (x 2 k ) and generate the network model NN 3 ( ) is allelic interaction variable x for MHC allele h=3 3 k and output NN 3 (x 3 k ). Each output is mapped by a function f( ) and combined to obtain an estimated presentation likelihood u k to generate

[0383] In another implementation, if predictions are made for the log of the mass spectrometry ion current, r(·) is a log function and f(·) is an exponential function.

[0384] VIII.C.6. Example 3.3: Sum of Functions Model with Allelic Non-Interacting Variables In one implementation, the presentation likelihood for each potential allele for MHC allele h is Generated by TIFF2022524328000046.tif7128, the presentation likelihood is Incorporating the influence of allelic non-interacting variables into peptide presentation as generated by TIFF2022524328000047.tif13128.

[0385] According to equation (21), the peptide sequence p k is presented is the function g h ( ) for each MHC allele H peptide sequence p k to generate a corresponding dependency score for the allelic interaction variables for each MHC allele h. function g for allelic non-interacting variables w (·) is also applied to the coded version of the allele non-interacting variables to generate a dependency score for the allele non-interacting variables. The allele non-interacting variable score is combined with each of the allele interacting variable dependency scores. Each of the combined scores is transformed by the function f(•) to generate a presentation likelihood for each potential allele. The latent likelihoods are combined and a clipping function is applied to the combined output to clip the values ​​into the range [0,1], resulting in peptide sequence p k A presentation likelihood that would be presented may be generated. dependency function g w is the dependency function g w can be in the form of either

[0386] As an example, the affine transformation function g h ( ), g w Peptide p by MHC allele h=2, h=3 among m=4 different identified MHC alleles with ( ) k The likelihood that TIFF2022524328000048.tif7128, where w k is the peptide p k is the allele non-interacting variable identified for and θ w is the set of parameters determined for the allele non-interacting variable.

[0387] As another example, the network transformation function g h ( ), g w Peptide p by MHC allele h=2, h=3 among m=4 different identified MHC alleles with ( ) k The likelihood that TIFF2022524328000049.tif7128, where w k is the peptide p k is the allelic interaction variable identified for θ w is the set of parameters determined for allelic non-interacting variables.

[0388] Figure 12 shows an exemplary network model NN 2 ( ), NN 3 ( ), and NN w Peptide p associated with MHC alleles h=2, h=3 with ( ) k We describe the generation of the presentation likelihood of . As shown in Figure 12, the network model NN 2 ( ) is allele interaction variable x for MHC allele h=2 2 k and output NN 2 (x 2 k ). network model NN w ( ) is the peptide p k allelic non-interacting variable w k and output NN w (w k ). The outputs are combined and mapped by the function f(·). network model NN 3 ( ) is allelic interaction variable x for MHC allele h=3 3 k and output NN 3 (x 3 k ), which also generates the same network model NN w ( ) output NN w (w k ) and mapped by the function f( ). Both outputs are combined to produce the estimated presentation likelihood uk.

[0389] In another implementation, the presentation likelihood for each potential allele for MHC allele h is Generated by TIFF2022524328000050.tif7128, the presentation likelihood is Make it generated by TIFF2022524328000051.tif13128.

[0390] VIII.C.7. Example 4: Quadratic Model In one implementation, s(·) is a quadratic function and the peptide p k estimated presentation likelihood u k teeth, given by TIFF2022524328000052.tif13128, where the element u' k h is the potential allele-wise presentation likelihood for MHC allele h. A set of values ​​for the parameter θ for the likelihood of each potential allele can be determined by minimizing a loss function for θ, where i is the number of cells expressing a single MHC allele and / or multiple are examples in subset S of training data 170 generated from cells expressing MHC alleles of . The presentation likelihood for each potential allele can be in any of the forms shown in equations (18), (20), and (22) above.

[0391] In one embodiment, the model of equation (23) is the peptide sequence p k However, there exists the possibility that the presentation by the two MHC alleles may be presented simultaneously, implying that the presentation by the two HLA alleles is statistically independent.

[0392] According to equation (23), the peptide sequence p k would be presented by combining the presentation likelihoods of each potential allele and the peptide sequence p k Each pair of MHC alleles is associated with peptide p k can be generated by subtracting from the summation the likelihood of simultaneously presenting

[0393] As an example, the affine transformation function g h Peptide p by HLA allele h=2, h=3 among m=4 different identified HLA alleles using ( ) k The likelihood that TIFF2022524328000053.tif5128, where x 2 k , x 3 k is the allele interaction variable identified for HLA alleles h=2, h=3, and θ 2 , θ 3 is the set of parameters determined for HLA alleles h=2, h=3.

[0394] As another example, the network transformation function g h ( ), g w Peptide p by HLA allele h=2, h=3 among m=4 different identified HLA alleles using ( ) k The likelihood that TIFF2022524328000054.tif5128 can be generated by NN 2 ( ), NN 3 ( ) is the network model identified for HLA alleles h=2, h=3, θ 2 , θ 3 is the set of parameters determined for HLA alleles h=2, h=3.

[0395] IX Example 5: Prediction Module A prediction module 320 receives sequence data and uses a presentation model to select candidate neoantigens in the sequence data. Specifically, the sequence data may be DNA, RNA and / or protein sequences extracted from the patient's tumor tissue cells. The prediction module 320 converts the sequence data into multiple peptide sequences p having 8-15 amino acids for MHC-I or 6-30 amino acids for MHC-II. k to process. For example, the prediction module 320 can process the given sequence “IEFROEIFJEF” into three peptide sequences “IEFROEIFJ”, “EFROEIFJE” and “FROEIFJEF” having nine amino acids. In one embodiment, the prediction module 320 compares sequence data extracted from normal tissue cells of the patient with sequence data extracted from tumor tissue cells of the patient to identify portions with one or more mutations. can identify candidate neoantigens that are mutated peptide sequences.

[0396] The prediction module 320 applies one or more of the presentation models to the processed peptide sequences to estimate the presentation likelihood of the peptide sequences. Specifically, the prediction module 320 can select one or more candidate neoantigen peptide sequences that are likely to be presented on tumor HLA molecules by applying a presentation model to the candidate neoantigens. In one implementation, the prediction module 320 selects candidate neoantigen sequences that have an estimated likelihood of presentation above a predetermined threshold. In another implementation, the presentation model selects v candidate neoantigen sequences with the highest estimated likelihood of presentation (where v is generally the maximum number of epitopes that can be delivered in a vaccine). ). A vaccine containing candidate neoantigens selected for a given patient can be injected into the patient to induce an immune response.

[0397] X. Example 6: Patient Selection Module A patient selection module 324 selects a subset of patients for vaccine therapy and / or T cell therapy based on whether the patients meet the inclusion criteria. In one embodiment, the selection criteria are determined based on the presentation likelihood of the patient's neoantigen candidates generated by the presentation model. By adjusting the selection criteria, the patient selection module 324 can adjust the number of patients to receive vaccination and / or T-cell therapy based on the likelihood of patient presentation of candidate neoantigens. Specifically, stringent selection criteria result in fewer patients being treated with vaccines and / or T-cell therapy, but with effective treatments (e.g., one or more tumor-specific neoantigen (TSNA) and The proportion of patients treated with vaccines and / or T cell therapy who receive one or more neoantigen-responsive T cells) may be high. In contrast, loose selection criteria lead to a higher number of patients treated with vaccines and / or T-cell therapy, but a lower proportion of patients treated with vaccines and / or T-cell therapy receive effective treatment. obtain. The patient selection module 324 modifies the selection criteria based on the desired balance between the target proportion of patients receiving treatment and the proportion of patients receiving effective treatment.

[0398] In some embodiments, the selection criteria for selecting patients to receive vaccine therapy are the same as the selection criteria for selecting patients to receive T cell therapy. However, in alternative embodiments, the selection criteria for selecting patients to receive vaccine therapy may differ from the selection criteria for selecting patients to receive T cell therapy. Sections X.A and X.B below discuss selection criteria for selecting patients to receive vaccine therapy and selection criteria for selecting patients to receive T cell therapy, respectively.

[0399] X.A. Patient Selection for Vaccine Treatment In one embodiment, a patient is associated with a corresponding therapeutic subset of v neoantigen candidates that can potentially be included in a personalized vaccine for that patient with vaccine volume v. In one embodiment, the therapeutic subset for a patient is the neoantigen candidates with the highest likelihood of presentation as determined by the presentation model. For example, if a vaccine may contain v=20 epitopes, the vaccine may contain the therapeutic subset of each patient with the highest likelihood of presentation as determined by the presentation model. However, it is recognized that in other embodiments, the treatment subset for a patient can be determined based on other methods. For example, therapeutic subsets for a given patient can be randomly selected from a set of neoantigen candidates for that patient, or from prior art models that model binding affinity or stability of peptide sequences, or presentation models. Decisions can be made based, in part, on a combination of certain factors, including the resulting presentation likelihood and affinity or stability information about these peptide sequences.

[0400] In one embodiment, the patient selection module 324 determines that a patient meets the inclusion criteria if the patient's tumor mutational burden is equal to or greater than the minimum mutational burden. A patient's tumor mutational burden (TMB) represents the total number of nonsynonymous mutations in the tumor exome. In one embodiment, the patient selection module 324 selects a patient for vaccination if the patient's absolute count of TMB is equal to or higher than a predetermined threshold. In another implementation, the patient selection module 324 selects patients for vaccination if their TMB is within a threshold percentile between TMBs determined for the set of patients.

[0401] 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 greater than the minimum utility score. In one embodiment, the utility value score is a measure of the estimated number of presented antigens from the therapeutic subset.

[0402] The estimated number of presented antigens can be predicted by modeling neoantigen presentation as a random variable with one or more probability distributions. In one implementation, the utility value score for patient i is the expected number of presented neoantigen candidates from the treatment subset, or a specific function thereof. As an example, the presentation of each neoantigen can be modeled as a Bernoulli random variable whose probability of presentation (success) is given by the presentation likelihood of the neoantigen candidate. Specifically, each has the highest presentation likelihood u i1 , u i2 , …, u iv v neoantigen candidates p i1 , p i2 , …, p iv The therapeutic subset of S i for the neoantigen candidate p ij is a random variable A ij , where TIFF2022524328000055.tif6128. The expected number of neoantigens presented is given by the sum of the presentation likelihoods of each neoantigen candidate. In other words, patient i's utility score is expressed as: TIFF2022524328000056.tif15128 The patient selection module 324 selects a subset of patients with utility score equal to or higher than the minimum utility value for vaccine treatment.

[0403] In another implementation, the utility score for patient i is the probability that at least a threshold number of neoantigens k will be presented. In one example, the therapeutic subset S of neoantigen candidates i The number of presented antigens in is modeled as a Poisson-binomial random variable whose probability of presentation (success) is given by the respective likelihood of presentation of the epitope. Specifically, the number of antigens presented by patient i is determined by the random variable N i can be given by In the TIFF2022524328000057.tif13128 formula, PBD(•) indicates a Poisson binomial distribution. The probability that at least a threshold number of neoantigens k are presented is the number of presented antigens N i is given by the operation of probabilities equal to or greater than k. In other words, patient i's utility score is expressed as: TIFF2022524328000058.tif13128 The patient selection module 324 selects a subset of patients with utility score equal to or higher than the minimum utility value for vaccine treatment.

[0404] In another implementation, the utility value score for patient i is evaluated for neoantigen candidates with binding affinities below a fixed threshold (e.g., 500 nM) or predicted binding affinities for one or more of the patient's HLA alleles. therapeutic subset S i is the number of neoantigens in In one example, the fixed threshold is in the range of 1000nM-10nM. In some cases, the utility value score may count only neoantigens detected as expressed by RNA-seq.

[0405] In another implementation, a utility value score for patient i is a candidate neoantigen whose binding affinity for one or more HLA alleles for that patient is equal to or less than the threshold percentile of the binding affinity for a random peptide for that HLA allele. therapeutic subset S i is the number of neoantigens in In one example, the threshold percentile ranges from the 10th percentile to the 0.1 percentile. In some cases, the utility score may count only neoantigens detected as expressed by RNA-seq.

[0406] The examples of utility scores described with respect to equations (25) and (27) are merely illustrative, and patient selection module 324 may use other statistics or probability distributions to generate utility scores. Points will be recognized.

[0407] X.B. Patient selection for T cell therapy In another embodiment, the patient may receive T cell therapy instead of or in addition to receiving vaccine therapy. Similar to vaccine therapy, in embodiments in which a patient receives T cell therapy, the patient can be associated with a corresponding therapeutic subset of the v neoantigen candidates as described above. This therapeutic subset of v neoantigen candidates can be used for the in vitro identification of patient-derived T cells with reactivity to one or more of the v neoantigen candidates. These identified T cells can then be expanded and infused into patients in personalized T cell therapy.

[0408] Patients can be selected to receive T cell therapy at two different time points. The first time point is after a therapeutic subset of v neoantigen candidates for the patient has been predicted using the model, but is specific for the predicted therapeutic subset of v neoantigen candidates. Prior to in vitro screening of cells. The second time point is after in vitro screening of T cells that are specific for therapeutic subsets of the predicted v neoantigen candidates.

[0409] First, after a therapeutic subset of v neoantigen candidates for that patient has been predicted, and in vitro T cells from the patient that are specific for the predicted therapeutic subset of v neoantigen candidates Patients can be selected for T-cell therapy before the identification of Specifically, because in vitro screening of patient-derived neoantigen-specific T cells can be costly, neoantigen-specific T cells should only be tested if the patient is likely to have neoantigen-specific T cells. It may be desirable to select patients for screening. The same criteria used to select patients for vaccination can be used to select patients prior to the in vitro T cell screening step. Specifically, in some embodiments, the patient selection module 324 selects patients to receive T cell therapy if the patient's tumor mutational burden is equal to or greater than the minimum mutational burden as described above. can be selected. In another embodiment, the patient selection module 324 determines that the patient's utility score based on the therapeutic subset of the v neoantigen candidates for that patient is equal to or greater than the minimum utility score as described above. Patients can be selected to receive T cell therapy if they are higher than

[0410] Second, in addition to selecting patients to undergo T cell therapy prior to in vitro identification of patient-derived T cells that are specific for a therapeutic subset of the predicted v neoantigen candidates, or Alternatively, patients can be selected to receive T cell therapy after in vitro identification of T cells that are specific for therapeutic subsets of the predicted v neoantigen candidates. Specifically, a patient is selected to receive T cell therapy if an in vitro screening of the patient's T cells for neoantigen recognition identifies at least a threshold amount of neoantigen-specific TCRs for that patient. can be done. For example, a patient will only have a TCR if at least two neoantigen-specific TCRs have been identified for that patient, or if neoantigen-specific TCRs have been identified for two different neoantigens. You can choose to receive cell therapy.

[0411] In another embodiment, a patient is selected to receive T cell therapy only if a threshold amount of neoantigens of a therapeutic subset of v neoantigen candidates for that patient is recognized by the patient's TCR. can be done. For example, a patient can be selected to receive T cell therapy only if at least one neoantigen out of a therapeutic subset of v neoantigen candidates for that patient is recognized by the patient's TCR. . In a further embodiment, a patient receives T cell therapy only if at least a threshold amount of TCR for that patient is identified as neoantigen-specific for a particular HLA-restricted class of neoantigen peptides. can be selected as For example, a patient can be selected to receive T cell therapy only if at least one TCR for that patient is identified as a neoantigen-specific HLA class I restricted neoantigen peptide.

[0412] In still further embodiments, a patient can be selected to receive T cell therapy only if at least a threshold amount of neoantigen peptides of a particular HLA-restricted class are recognized by the patient's TCR. For example, a patient can be selected to receive T cell therapy only if at least one HLA class I restricted neoantigen peptide is recognized by the patient's TCR. For example, a patient can be selected to receive T cell therapy only if at least two HLA class II restricted neoantigen peptides are recognized by the patient's TCR. Any combination of the above criteria to select a patient to receive T cell therapy after in vitro identification of T cells that are specific for a therapeutic subset of the v neoantigen candidates predicted for that patient can also be used for

[0413] XI. Example 7: Experimental Results Demonstrating Exemplary Patient Selection Performance Validity of patient selection as described in Section X is based on a simulated test set associated with each of the simulated neoantigen candidates that are known to represent a subset of the simulated neoantigens in the mass spectrometry data. Validate by making patient selections on a set of selected patients. Specifically, for each simulated neoantigen candidate in the test set, the neoantigen is associated with the Bassani-Sternberg dataset (dataset "D1") (data available at www.ebi.ac.uk / pride / archive / projects / PXD0000394) are presented in the mass spectrometry data sets of the multi-allelic JY cell lines HLA-A*02:01 and HLA-B*07:02. Associate. As detailed below in conjunction with FIG. 13A, a number of neoantigen candidates for simulated patients are sampled from the human proteome based on the known frequency distribution of mutation burden in patients with non-small cell lung cancer (NSCLC).

[0414] Allele-by-allele presentation models for the same HLA alleles were presented in a single model from the IEDB dataset (dataset 'D2') (data can be found at http: / / www.iedb.org / doc / mhc_ligand_full.zip). Train with a training set that is a subset of monoallelic HLA-A*02:01 and HLA-B*07:02 mass spectrometry data. In detail, the presentation model of each allele is represented by the network-dependent function g h ( ) and g w (·) and the model for each allele shown in the formula (8) incorporated with the expit function f(·). The presentation model of the allele HLA-A*02:01 assumes that the peptide sequence is given as an allele-interacting variable, and the N-terminal and C-terminal flanking sequences are given as allele non-interacting variables. Generate presentation likelihoods presented on HLA-A*02:01. The presentation model of the allele HLA-B*07:02 assumes that the peptide sequence is given as an allele-interacting variable, and the N-terminal and C-terminal flanking sequences are given as allele non-interacting variables. Generate presentation likelihoods to be presented on HLA-B*07:02.

[0415] As disclosed in the examples below with reference to Figures 13A-13E, different models, such as presentation models trained for peptide binding prediction and prior art models, were tested on each simulated patient with neoantigen candidates. Applying to the set identifies different treatment subsets for the patient based on prediction. Patients meeting the inclusion criteria for vaccine therapy are selected and associated with personalized vaccines containing epitopes for the therapeutic subset of patients. The size of the therapeutic subset will vary for different vaccine doses. No overlap is introduced between the training set used to train the presentation model and the test set of simulated neoantigen candidates.

[0416] The following example analyzes the proportion of selected patients with at least a specified number of presented neoantigens among the epitopes included in the vaccine. This statistic demonstrates the efficacy of simulated vaccines in delivering potential neoantigens to induce an immune response in patients. Specifically, a simulated neoantigen within a test set is presented if that neoantigen is presented in mass spectrometric data set D2. The high proportion of patients with presented neoantigens indicates the potential success of treatment with neoantigen vaccines by inducing an immune response.

[0417] XI.A. Example 7A: Frequency Distribution of Mutational Burden in NSCLC Cancer Patients FIG. 13A shows the sample frequency distribution of mutation 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 the specified number of non-synonymous mutations. The sample frequency distribution in Figure 13A shows a range of 3-1786 mutations, with 30% of patients having less than 100 mutations. Although not shown in Figure 13A, studies have shown that mutational burden is higher in smokers compared to nonsmokers, and that mutational burden can be a strong indicator of neoantigen burden in patients.

[0418] As introduced at the beginning of Section XI above, each number of simulated patients is associated with a test set of neoantigen candidates. The test set for each patient is the mutation burden m i is generated by sampling For each mutation, a 21-mer peptide sequence from the human proteome is randomly selected to represent the simulated mutant sequence. A test set of neoantigen candidate sequences is generated for patient i by identifying the peptide sequences of each (8, 9, 10, 11) mer spanning the mutations within the 21 mer. Each neoantigen candidate is associated with a label that indicates whether the neoantigen candidate sequence is present in the mass spectrometric D1 data set. For example, neoantigen candidate sequences that are present in data set D1 can be associated with the label "1" and sequences that are not in data set D1 can be associated with the label "0". As described in more detail below, FIGS. 13B-13E show experimental results of patient selection based on patient-presented neoantigens within the test set.

[0419] XI.B. Example 7B: Proportion of Selected Patients with Neoantigen Presentation Based on Mutational Burden Selection Criteria FIG. 13B shows the number of presented neoantigens in the simulated vaccine for patients selected based on selection criteria whether the patient meets the minimum mutational burden. Identify the proportion of selected patients who have at least a specified number of presented neoantigens in the corresponding studies.

[0420] In FIG. 13B, the x-axis shows the proportion of patients excluded from vaccine treatment based on tumor mutational burden, labeled "Minimum number of mutations." For example, a data point at "minimum number of mutations" 200 indicates that the patient selection module 324 selected only a subset of simulated patients with a mutation load of at least 200 mutations. As another example, the data points in "minimum number of mutations" 300 indicate that the patient selection module 324 selected a lower proportion of simulated patients with at least 300 mutations. The y-axis shows the proportion of selected patients who were associated with at least the specified number of presented neoantigens in the test set without vaccine dose v. In detail, the top plot shows the proportion of selected patients presenting at least one neoantigen, 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 two antigens. plot shows the proportion of selected patients presenting at least 3 antigens.

[0421] As shown in FIG. 13B, the proportion of patients with presented neoantigens increases significantly with higher mutational burden. This indicates that mutational burden as a selection criterion can be effective in selecting patients in whom neoantigen vaccines are likely to induce effective immune responses.

[0422] XI.C. Example 7C: Comparison of Neoantigen Presentation in Vaccines Identified by Presentation Models and Prior Art Models FIG. 13C shows selected patients associated with vaccines including therapeutic subsets identified based on the presented model and selected patients associated with vaccines including therapeutic subsets identified by prior art models. Comparison of the number of neoantigens presented in simulated vaccines between The left plot assumes v=10 as the limiting vaccine volume and the right plot assumes v=20 as the limiting vaccine volume. Patients are selected based on utility score, which indicates the expected number of neoantigens presented.

[0423] In FIG. 13C, solid lines indicate vaccine-associated patients with therapeutic subsets identified based on the presentation model for alleles HLA-A*02:01 and HLA-B*07:02. A therapeutic subset for each patient is identified by applying each of the presentation models to the sequences in the test set and identifying the v neoantigen candidates with the highest likelihood of presentation. Dotted lines indicate patients associated with vaccines including therapeutic subsets identified based on the prior art model NETMHCpan against single allele HLA-A*02:01. Implementation details for NETMHCpan are provided at http: / / www.cbs.dtu.dk / services / NetMHCpan. A therapeutic subset for each patient is identified by applying the NETMHCpan model to the sequences in the test set and identifying the v neoantigen candidates with the highest putative binding affinities. The x-axis of both graphs shows the proportion of patients excluded from vaccine treatment based on the expected utility value score, which indicates the expected number of presented neoantigens within the treatment 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 shows the proportion of selected patients presenting at least the specified number of neoantigens included in the vaccine (1, 2, or 3 neoantigens).

[0424] As shown in FIG. 13C, a significantly higher proportion of patients associated with the vaccine including the therapeutic subset based on the presented model had a significantly higher proportion of patients associated with the vaccine including the therapeutic subset based on the prior art model. A vaccine containing the presented neoantigen is administered. For example, as shown in the graph on the right, selected patients associated with vaccines based on the proposed model compared to only 40% of selected patients associated with vaccines based on prior art models 80% of the population receive at least one presented neoantigen in the vaccine. These results demonstrate that the presentation model described herein is effective in selecting neoantigen candidates for vaccines that are likely to elicit an immune response to treat tumors.

[0425] XI.D. Example 7D: Effect of HLA Coverage on Neoantigen Presentation of Vaccines Identified by Presentation Models Figure 13D shows selected patients associated with vaccines including therapeutic subsets identified based on a single allele-by-allele presentation model for HLA-A*02:01 and HLA-A*02:01. Simulated vaccines between selected patients associated with vaccines containing therapeutic subsets identified based on both the allele-by-allele presentation model for HLA-B*07:02 and the allele-by-allele presentation model for HLA-B*07:02 Figure 2 compares the number of neoantigens presented in the medium. The vaccine volume is set at v=20 epitopes. For each experiment, patients are selected based on expected utility scores determined based on different treatment subsets.

[0426] In FIG. 13D, solid lines indicate vaccine-associated patients with treatment subsets based on both presentation models for HLA alleles HLA-A*02:01 and HLA-B*07:02. A therapeutic subset for each patient is identified by applying each of the presentation models to the sequences in the test set and identifying the v neoantigen candidates with the highest likelihood of presentation. Dotted lines indicate patients associated with vaccines including therapeutic subsets based on a single presentation model for the HLA allele HLA-A*02:01. A therapeutic subset for each patient is identified by applying a presentation model for only single HLA alleles to the sequences in the test set and identifying the v neoantigen candidates with the highest likelihood of presentation. In solid line plots, the x-axis shows the proportion of patients excluded from vaccine treatment based on expected utility value scores for the treatment subsets identified by both presentation models. In dotted plots, the x-axis shows the proportion of patients excluded from vaccine treatment based on expected utility value scores for treatment subsets identified by a single presentation model. The y-axis shows the proportion of selected patients presenting at least a specified number of neoantigens (1, 2, or 3 neoantigens).

[0427] As shown in FIG. 13D, vaccine-associated patients with therapeutic subsets identified by the presentation model for both HLA alleles are more likely than patients associated with vaccines with therapeutic subsets identified by a single presentation model. presents neoantigens at a significantly higher rate than These results demonstrate the importance of establishing presentation models with high HLA coverage.

[0428] XI.E. Example 7E: Comparison of Neoantigen Presentation in Patients Selected by Mutational Burden and Expected Number of Presented Neoantigens FIG. 13E compares the number of presented neoantigens in the simulated vaccine between patients selected based on mutational load and patients selected by expected utility score. Expected utility scores are determined based on the therapeutic subset identified by a presentation model with a size of v=20 epitopes.

[0429] In FIG. 13E, solid lines indicate patients selected based on expected utility scores associated with vaccines containing therapeutic subsets identified by the presented model. A therapeutic subset for each patient is identified by applying each of the presentation models to the sequences in the test set and identifying the v=20 neoantigen candidates with the highest likelihood of presentation. A therapeutic utility score is determined based on the likelihood of presentation of the therapeutic subsets identified in section X based on equation (25). Dotted lines indicate patients selected based on mutational burden associated with vaccines that include the therapeutic subset identified by the presented model. The x-axis shows the proportion of patients excluded from vaccine treatment based on expected utility score in the solid plot and the proportion of patients excluded based on mutational burden in the dotted plot. The y-axis shows the proportion of selected patients receiving vaccines containing at least a specified number of presented neoantigens (1, 2, or 3 neoantigens). As shown in FIG. 13E, patients selected based on expected utility score receive a higher percentage of vaccines containing presented neoantigens than patients selected based on mutational burden. However, patients selected on the basis of mutational burden receive vaccines containing presented neoantigens at a higher rate than unselected patients. Therefore, mutational burden is an effective patient selection criterion in effective neoantigen vaccine delayed leak, whereas expected utility score is more effective.

[0430] XII. Example 8: Evaluation of Mass Spectrometry Trained MHC Class II Presentation Models on Excluded MHC Class II Mass Spec Data The effectiveness of the various presentation models described above can be evaluated from training data 170 that are subsets of training data 170 that were not used to train the presentation model, or that have similar variables and data structures as training data 170. Another data set, test data T, was tested.

[0431] As a relevant indicator of the performance of the presented model, There is TIFF2022524328000059.tif10128, which shows the ratio of the number of peptide instances correctly predicted to be presented on the relevant HLA allele to the number of peptide instances predicted to be presented on that HLA allele. . In one implementation, peptide p in test data T i is the corresponding likelihood estimate u i was predicted to be presented on one or more relevant HLA alleles if was greater than or equal to a predetermined threshold t. Another relevant indicator of the performance of the presented model is There is TIFF2022524328000060.tif10128, which is the ratio of the number of peptide instances correctly predicted to be presented on the relevant HLA allele to the number of peptide instances known to be presented on that HLA allele. indicates Another relevant indicator of the performance of the presentation model is the area under the curve (AUC) of the receiver operating characteristic (ROC). ROC plots recall against false positive rate (FPR) given by the formula: TIFF2022524328000061.tif10128

[0432] XII.A. Performance of the proposed model on mass spectrometry data XII.A.1. Example 1 FIG. 14A is a length histogram of peptides eluted from class II MHC alleles on human tumor cells and tumor infiltrating lymphocytes (TIL) using mass spectrometry. Specifically, HLA-DRB1*12:01 homozygous allele (“Dataset 1”) and HLA-DRB1*12:01, HLA-DRB1*10:01 multi-allele samples (“Dataset 2”) was subjected to mass spectrometry peptidomics. The results show that the length of peptides eluted from class II MHC alleles ranged from 6 to 30 amino acids. The frequency distribution shown in Figure 14A is similar to the length of peptides eluted from class II MHC using state-of-the-art mass spectrometry, shown in Figure 1C of reference 69.

[0433] FIG. 14B shows the dependence of mRNA quantification and presented peptides per residue for Dataset 1 and Dataset 2. FIG. The results indicate a strong dependence between mRNA expression and peptide presentation of class II MHC alleles.

[0434] Specifically, the horizontal axis of FIG. 14B is log 10 Expression of mRNA expressed in (transcripts per million bins (TPM)) is shown. The vertical axis of FIG. 14B represents the amount of peptide presentation per residue, 10 -2 <log 10 TPM<10 -1 are shown as folds of the lowest bin corresponding to mRNA expression levels between . One solid line is the plot associated with mRNA quantification and peptide presentation for dataset 1, the other solid line is for dataset 2. As shown in Figure 14B, there is a strong correlation between mRNA expression level and peptide presentation level per residue within the corresponding gene. Specifically, when the RNA expression level is 10 1 <log 10 TPM<10 2 Peptides from genes that range from are more than five times more likely to be presented compared to the lowest bin.

[0435] These results show that the performance of presentation models can be greatly improved by incorporating these measurements, as measurements of mRNA quantification are strongly predictive of peptide presentation.

[0436] FIG. 14C compares the performance results of exemplary presentation models trained and tested using Dataset 1 and Dataset 2. FIG. For each set of model features of the exemplary presentation model, FIG. The PPV values ​​at a recall rate of 10% when classified as interaction characteristic variables are shown. As seen in FIG. 14C, for each set of model traits of the exemplary presentation model, the 10% recall identified if the traits in that set of model traits were classified as allelic interaction traits. The PPV values ​​are shown on the left and the PPV values ​​at 10% recall determined when the trait in that set of model traits is classified as an allelic non-interacting trait are shown on the right. Note that peptide sequence properties were always classified as allelic interaction properties for the purposes of Figure 14C. The results show that the presented model achieves PPV values ​​at 10% recall ranging from 14% to 29%, which are significantly higher (approximately 500-fold) than the PPV in random prediction.

[0437] Peptide sequences of 9-20 lengths were considered in this experiment. The data were divided into training, validation and test sets. Blocks of 50-residue peptides from both dataset 1 and dataset 2 were assigned to the training and test sets. Duplicate peptides anywhere in the proteome were removed so that no peptide sequences appeared in both the training and test sets. The incidence of peptide presentation in training and test sets was increased 50-fold by removing non-presenting peptides. This indicates that Dataset 1 and Dataset 2 were from human tumor samples in which only a fraction of the cells were class II HLA alleles, resulting in approximately 10-fold lower peptide yields than pure samples of class II HLA alleles. (which is still an underestimation due to imperfect mass spectrometry sensitivity). The training set contained 1064 presented peptides and 3,810,070 non-presented peptides. The test set contained 314 presented peptides and 807,400 non-presented peptides.

[0438] Example model 1 uses the network dependence function g h ( ), the expit function f( ), and the identity function r( ) were used as the summation model of the function of Equation (22). network dependency function g h ( ) was structured as a multi-layer perceptron (MLP) with 256 hidden nodes and normalized linear unit (ReLU) activations. Besides the peptide sequence, the allelic interaction variable w is one-hot encoded C-terminal and N-terminal flanking sequences, peptide p i The index of the source gene of G=gene(p i ), and variables representing mRNA quantification measurements. Exemplary Model 2 was the same as Exemplary Model 1, except that the C-terminal and N-terminal flanking sequences were omitted from the allelic interaction variables. Exemplary Model 3 was the same as Exemplary Model 1, except that the source gene subscript was omitted from the allelic interaction variables. Exemplary Model 4 was the same as Exemplary Model 1, except that mRNA measurements were omitted from the allelic interaction variables.

[0439] Example model 5 uses the network dependence function g h ( ), the expit function f( ), the identity function r( ), and the dependency function g of Eq. (12) w ( ) was used as the sum model of the function of Eq. (20). dependency function g w ( ), with mRNA quantification measurements as input, network model structured as MLP with 16 hidden nodes and normalized linear unit (ReLU) activations, and C-terminal flanking sequences, with 32 We also included a network model structured as an MLP with hidden nodes and rectified linear unit (ReLU) activations. network dependency function g h ( ) was structured as a multi-layer perceptron with 256 hidden nodes and rectified linear unit (ReLU) activations. Exemplary model 6 was the same as exemplary model 5, except that the network model for the C-terminal and N-terminal flanking sequences was omitted. Exemplary Model 7 was the same as Exemplary Model 5, except that the source gene subscript was omitted from allele non-interacting variables. Exemplary model 8 was the same as exemplary model 5, except that the network model for mRNA quantification measurements was omitted.

[0440] The incidence of presented peptides in the test set is approximately 1 / 2400, so the PPV of random predictions would also be approximately 1 / 2400=0.00042. As shown in FIG. 14C, the best presentation model achieves a PPV value of approximately 29%, which is approximately 500 times better than the random prediction PPV value.

[0441] XII.A.2. Example 2 FIG. 14D shows the amount of peptides sequenced using mass spectrometry for each sample of a total of 73 human tumors (NSCLC, lymphoma, and ovarian cancer) and cell lines (EBV) containing HLA class II molecules. is a histogram showing An average of 900 peptides were sequenced for each sample, as shown in Figure 14D. Furthermore, for each of the multiple samples, the histogram shown in FIG. 14D shows the amount of peptides sequenced using mass spectrometry at different q-value thresholds. Specifically, for each of the multiple samples, Figure 14D shows the amount of peptides sequenced using mass spectrometry with a q value of less than 0.01, a q value of less than 0.05, and a q value of less than 0.2. there is

[0442] As noted above, each of the 73 samples in Figure 14D contained HLA class II molecules. More specifically, each of the 73 samples in Figure 14D contained HLA-DR molecules. HLA-DR molecules are one type of HLA class II molecules. Even more specifically, each of the 73 samples in Figure 14D contained HLA-DRB1, HLA-DRB3, HLA-DRB4, and / or HLA-DRB5 molecules. HLA-DRB1 molecules, HLA-DRB3 molecules, HLA-DRB4 molecules, and HLA-DRB5 molecules are types of HLA-DR molecules.

[0443] Although this particular experiment was performed using a sample containing HLA-DR molecules, in particular HLA-DRB1, HLA-DRB3, HLA-DRB4 and HLA-DRB5 molecules, in alternative embodiments this Experiments can be performed with samples containing one or more of any type(s) of HLA class II molecules. For example, in alternative embodiments, the same experiment can be performed using samples containing HLA-DP and / or HLA-DQ molecules. It is well known to those skilled in the art that the same method can be used to model any type(s) of MHC class II molecules and still give reliable results. For example, Jensen, Kamilla Kjaergaard et al. 76 is an example of a recent scientific paper using the same method to model binding affinities to HLA-DR molecules and to HLA-DP and HLA-DQ molecules. Thus, one skilled in the art would be able to use the experiments and models described herein to model not only HLA-DR molecules, but any other MHC class II molecule separately or simultaneously, and still be confident. It will be appreciated that high results can be obtained with

[0444] Mass spectrometry was performed on each sample to sequence the peptides of each sample for a total of 73 samples. The mass spectra obtained for the samples were searched with Comet, scored with Percolator and peptides were sequenced. The amount of peptides sequenced in the samples was then determined for several different Percolator q-value thresholds. Specifically, for that sample, the amount of peptides sequenced with a Percolator q value of less than 0.01, a Percolator q value of less than 0.05, and a Percolator q value of less than 0.2 was determined.

[0445] The amount of peptides sequenced at each of the different Percolator q-value thresholds for each of the 73 samples is shown in Figure 14D. For example, as seen in FIG. 14D, in the first sample approximately 4700 peptides were sequenced using mass spectrometry with a q value of less than 0.2 and Approximately 3600 peptides were sequenced and approximately 3200 peptides were sequenced using mass spectrometry with a q value of less than 0.01.

[0446] Overall, FIG. 14D demonstrates that mass spectrometry can be used to sequence large amounts of peptides with low q-values ​​from samples containing MHC class II molecules. In other words, the data shown in Figure 14D demonstrate that mass spectrometry can be used to reliably sequence peptides that can be presented by MHC class II molecules.

[0447] FIG. 14E is a histogram showing the amount of samples in which specific MHC class II molecule alleles were identified. More specifically, Figure 14E shows the amount of samples in which a particular MHC class II molecule was identified for a total of 73 samples containing HLA class II molecules.

[0448] As described above with respect to Figure 14D, each of the 73 samples in Figure 14D contained HLA-DRB1, HLA-DRB3, HLA-DRB4, and / or HLA-DRB5 molecules. Accordingly, FIG. 14E shows the amount of samples in which certain alleles were identified for HLA-DRB1, HLA-DRB3, HLA-DRB4, and HLA-DRB5 molecules. To identify the HLA alleles present in a sample, the sample is typed for HLA class II DR. Then, to determine the amount of samples in which a particular HLA allele was identified, simply sum the number of samples in which the HLA allele was identified using HLA class II DR typing. For example, as shown in Figure 14E, 17 out of a total of 73 samples contained the HLA class II molecule allele HLA-DRB3*01:01. In other words, 17 samples out of a total of 73 samples contained the allele HLA-DRB3*01:01 for the HLA-DRB3 molecule. Overall, FIG. 14E shows that a wide range of HLA class II molecule alleles can be identified from 73 samples containing HLA class II molecules.

[0449] FIG. 14F is a histogram showing the percentage of peptides presented by MHC class II molecules in a total of 73 samples for each peptide length over a range of peptide lengths. To determine the length of each peptide in each of the 73 samples in total, each peptide was sequenced using mass spectrometry as described above with respect to FIG. The number of groups was simply quantified.

[0450] As mentioned above, MHC class II molecules present peptides that are generally 9-20 amino acids in length. Thus, Figure 14F shows the percentage of peptides presented by MHC class II molecules in 73 samples for each peptide length of 9-20 amino acids. For example, as shown in Figure 14F, approximately 23% of peptides presented by MHC class II molecules in 73 samples have a length of 14 amino acids.

[0451] Based on the data shown in Figure 14F, the most frequent lengths of peptides presented by MHC class II molecules in 73 samples were identified as 14 and 15 amino acids in length. These modal lengths identified for peptides presented by MHC class II molecules in 73 samples are consistent with previous reports on the modal lengths of peptides presented by MHC class II molecules. Furthermore, also consistent with previous reports, the data in Figure 14F show that more than 60% of the peptides presented by MHC class II molecules from 73 samples had lengths other than 14 and 15 amino acids. indicates that it has In other words, Figure 14F shows that peptides presented by MHC class II molecules are most frequently 14 or 15 amino acids in length, whereas a large proportion of peptides presented by MHC class II molecules are 14 amino acids long. This indicates that the length is neither 1 nor 15. Therefore, it is reasonable to assume that peptides of all lengths have an equal probability of being presented by MHC class II molecules, or that only peptides with a length of 14 or 15 amino acids are presented by MHC II class molecules. This is an incorrect assumption. As detailed below with respect to FIG. 14L, these erroneous assumptions are currently used in many current models for predicting presentation by MHC class II molecules and are therefore predicted by these models. The suggested likelihood is often unreliable.

[0452] FIG. 14G is a line graph showing the relationship between gene expression and incidence of presentation of gene expression products by MHC class II molecules for genes present in 73 samples. More specifically, Figure 14G shows the relationship between gene expression and the proportion of residues resulting from that gene expression that form the N-terminus of peptides presented by MHC class II molecules. RNA sequencing is performed on the RNA contained in each sample to quantify gene expression in each of the 73 samples in total. In Figure 14G, gene expression is measured by RNA sequencing in terms of transcripts per million (TPM). To determine the incidence of gene expression presentation for each of the 73 samples, identification of HLA class II DR peptideome data was performed for each sample.

[0453] As shown in Figure 14G, there is a strong correlation between gene expression levels and the presentation of residues of expressed gene products by MHC class II molecules for 73 samples. Specifically, as shown in FIG. 14G, peptides resulting from expression of the lowest-expressed genes were more than 100-fold more likely to express MHC class II molecules than peptides resulting from expression of the highest-expressed genes. difficult to present by Simply put, the products of more highly expressed genes are more frequently presented by MHC class II molecules.

[0454] Figures 14H-I and Figures 14K-L show different presentations in predicting the likelihood that a peptide in a test data set of peptides will be presented by at least one of the MHC class II molecules present in the test data set. 4 is a line graph comparing the performance of models; As shown in Figures 14H-I and 14K-L, the model's performance in predicting the likelihood that a peptide will be presented by at least one of the MHC class II molecules present in the test dataset is: It is determined by specifying the ratio of the true positive rate to the false positive rate for each prediction generated by the model. These ratios specified for a given model are plotted as a receiver operator characteristic (ROC) curve, a line graph with the x-axis quantifying the false positive rate and the y-axis quantifying the true positive rate. can be visualized in Area under the curve (AUC) is used to quantify model performance. Specifically, models with higher AUC have better performance (ie, higher accuracy) compared to models with lower AUC. In Figures 14H, 14I, and 14L, the dashed black line with a slope of 1 (ratio of true positive rate to false positive rate of 1) represents the prediction curve of the likelihood of randomly estimated peptide presentation. The dashed AUC is 0.5. ROC curves and AUC measurements are discussed in detail with respect to the first half of Section XII above.

[0455] FIG. 14H shows five different approaches for predicting the likelihood that a peptide in the test dataset of peptides will be presented by an MHC class II molecule given different sets of allelic interaction and allele non-interaction variables. FIG. 4 is a line graph comparing the performance of example models; FIG. In other words, Figure 14H quantifies the relative importance of different allelic interaction and allele non-interaction variables in predicting the likelihood that a peptide will be presented by an MHC class II molecule.

[0456] The model architecture of each of the five exemplary models used to generate the ROC curve of the line graph of Figure 14H included a set of five sigmoidal summation models. Each sigmoidal summation model within the set was constructed to model peptide presentation for up to four unique HLA-DR alleles per sample. In addition, each sigmoidal summation model in the set estimates peptide presentation likelihood based on the following allelic interaction and allele non-interaction variables: peptide sequence, flanking sequence, RNA expression per TPM, gene identifier, and sample identifier. It was configured to make a degree prediction. The allelic interaction component of each sigmoidal summation model in the ensemble was one hidden layer of MLPs with ReLU activation as 256 hidden units.

[0457] An example model was trained and validated prior to using it to predict the likelihood that a peptide in a test dataset of peptides would be presented by an MHC class II molecule. To train, validate, and finally test the exemplary model, the data described above for the 73 samples were divided into training, validation, and test datasets.

[0458] To prevent peptides from appearing in more than one of the training, validation, and test datasets, the following steps were taken. First, all peptides from a total of 73 samples that appeared at multiple positions within the proteome were removed. Peptides from a total of 73 samples were then divided into blocks of 10 contiguous peptides. Each block of peptides from a total of 73 samples was individually assigned to a training, validation, or test dataset. This resulted in no peptides appearing in multiple datasets for training, validation and test datasets.

[0459] Of the 38,035,453 peptides in a total of 73 samples, the training dataset contained 33,570 peptides presented by MHC class II molecules from 69 of the 73 total samples. The 33,570 peptides included in the training dataset were 9-20 amino acids in length. The example model used to generate the ROC curve of Figure 14H was trained on the training data set using the ADAM optimization algorithm and early stopping.

[0460] The validation dataset consisted of 3,925 peptides presented by MHC class II molecules derived from the same 69 samples used in the training dataset. The validation set was used for early stopping only.

[0461] The test dataset included peptides presented by MHC class II molecules identified from tumor samples using mass spectrometry. Specifically, the test dataset included 232 peptides identified from 4 tumor samples. Peptides included in the test dataset were excluded from the training dataset described above.

[0462] As noted above, FIG. 14H quantifies the relative importance of different allele-interacting and non-allelic-interacting variables in predicting the likelihood that a peptide will be presented by an MHC class II molecule. It is. As also noted above, the exemplary model used to generate the ROC curve for the line graph of FIG. 14H includes the following allelic interaction and allele non-interaction variables: It was configured to predict the likelihood of peptide presentation based on the RNA expression, gene identifier, and sample identifier. We calculated the relative importance of four of these five variables (peptide sequence, flanking sequence, RNA expression, and gene identifier) ​​to predict the likelihood that a peptide is presented by an MHC class II molecule. For quantification, each example model of the five example models described above was tested with data from the test data set using different combinations of the four variables. Specifically, for each peptide in the test dataset, Exemplary Model 1 generated a prediction of peptide presentation likelihood based on peptide sequence, flanking sequence, gene identifier, and sample identifier, excluding RNA expression. Similarly, for each peptide in the test data set, Exemplary Model 2 generated a prediction of peptide presentation likelihood based on peptide sequence, RNA expression, gene identifier, and sample identifier, excluding flanking sequences. Similarly, for each peptide in the test data set, Exemplary Model 3 generated predictions of peptide presentation likelihood based on flanking sequence, RNA expression, gene identifier, and sample identifier, excluding peptide sequence. Similarly, for each peptide in the test data set, Exemplary Model 4 generated a prediction of peptide presentation likelihood based on flanking sequence, RNA expression, peptide sequence, and sample identifier, excluding gene identifiers. Finally, for each peptide in the test dataset, exemplary model 5 generated predictions of peptide presentation likelihood based on all five variables: flanking sequence, RNA expression, peptide sequence, gene identifier, and sample identifier. .

[0463] The performance of each of these five example models is shown in the line graph of Figure 14H. Specifically, each of the five example models is associated with an ROC curve showing the ratio of true positive rate to false positive rate for each prediction generated by the model. For example, FIG. 14H shows curves for exemplary model 1 that generated predictions of peptide presentation likelihood based on peptide sequence, flanking sequence, gene identifier, and sample identifier, excluding RNA expression. FIG. 14H shows curves for exemplary model 2 that generated predictions of peptide presentation likelihood based on peptide sequence, RNA expression, gene identifier, and sample identifier, excluding flanking sequences. FIG. 14H also shows curves for exemplary model 3 that generated predictions of peptide presentation likelihood based on flanking sequences, RNA expression, gene identifiers, and sample identifiers, excluding peptide sequences. FIG. 14H also shows curves for exemplary model 4 that generated predictions of peptide presentation likelihood based on flanking sequence, RNA expression, peptide sequence, and sample identifier, excluding gene identifiers. And finally, FIG. 14H shows curves for exemplary model 5 that generated predictions of peptide presentation likelihood based on all five variables: flanking sequence, RNA expression, peptide sequence, sample identifier, and gene identifier. there is

[0464] As noted above, the performance of a model in predicting the likelihood that a given peptide is presented by an MHC class II molecule is the ratio of the true positive rate to the false positive rate for each prediction generated by the model. Quantified by determining the AUC of the ROC curve showing the ratio. A model with a higher AUC has better performance (ie, higher accuracy) compared to a model with a lower AUC. As shown in Figure 14H, the curve for exemplary model 5, which generated predictions of peptide presentation likelihood based on all five variables: flanking sequence, RNA expression, peptide sequence, sample identifier, and gene identifier, was 0.98. Achieved the highest AUC. Thus, example model 5, which used all five variables to generate predictions of peptide presentation, achieved the best performance. The example model 2 curve, which generated predictions of peptide presentation likelihood based on peptide sequence, RNA expression, gene identifier, and sample identifier, excluding flanking sequences, achieved the second highest AUC of 0.97. Flanking sequences can therefore be identified as the least important variable in predicting the likelihood that a peptide will be presented by an MHC class II molecule. The example model 4 curve, which generated predictions of peptide presentation likelihood based on flanking sequence, RNA expression, peptide sequence, and sample identifier, excluding gene identifiers, achieved the third highest AUC of 0.96. Gene identifiers can therefore be identified as the second least important variable in predicting the likelihood that a peptide will be presented by an MHC class II molecule. The example model 3 curve, which generated predictions of peptide presentation likelihood based on flanking sequences, RNA expression, gene identifiers, and sample identifiers, excluding peptide sequences, achieved the lowest AUC of 0.88. Peptide sequence can therefore be identified as the most important variable in predicting the likelihood that a peptide will be presented by an MHC class II molecule. The Example Model 1 curve, which generated predictions of peptide presentation likelihood based on peptide sequence, flanking sequence, gene identifier, and sample identifier, excluding RNA expression, achieved the second lowest AUC at 0.95. RNA expression can therefore be identified as the second most important variable in predicting the likelihood that a peptide will be presented by an MHC class II molecule.

[0465] FIG. 14I is a line graph comparing the performance of four different presentation models in predicting the likelihood that a peptide within a test dataset of peptides will be presented by an MHC class II molecule.

[0466] The first model tested in Figure 14I is referred to herein as the "binding affinity model." The binding affinity model in Figure 14I is the best-in-class model, the NetMHCII2.3 model, using the lowest NetMHCII2.3 predicted binding affinity as the basis for generating predictions. Specifically, the NetMHCII2.3 model generates predictions of peptide presentation likelihood based on MHC class II molecule type and peptide sequence. The NetMHCII2.3 model can be found on the NetMHCII2.3 website (www.cbs.dtu.dk / services / NetMHCII / , PMID 29315598) 76 was tested using

[0467] The second model tested in Figure 14I is referred to herein as "MLP". The MLP (multilayer perceptron) model uses allele non-interacting variables w k and the allelic interaction variable x h k is input to separate dependency functions, such as neural networks, and then the outputs of these separate dependency functions are added together. Specifically, the complete non-interacting model is based on the allelic non-interacting variable w k is the dependency function g w and the allele interaction variable x h k is another dependency function g h and the dependency function g w and the dependency function g h are added together, one embodiment of the presentation model described above. Therefore, in some embodiments, the perfect non-interaction model determines the likelihood of peptide presentation using Equation 8 shown above. In addition, allelic non-interacting variable w k is the dependency function g w and the allele interaction variable x h k is another dependency function g h and the dependency function g w and the dependency function g h An embodiment of a fully non-interacting model to which the outputs of are summed is the first half of Section VIII.B.2, the second half of Section VIII.B.3, the first half of Section VIII.C.3, and the first half of Section VIII.B.3. The first half of C.6 is detailed above.

[0468] The third model tested in Figure 14I is referred to herein as "RNN." RNN models include recurrent neural networks and are similar to the fully non-interacting models described above. However, the layers of the recurrent neural network of the RNN model are different from the layers of the neural network of the MLP model. Specifically, the input layer of the recurrent neural network of the RNN model accepts peptide strings of variable length, modeling one peptide at a time. This peptide is fed, one amino acid at a time, to a neural network node whose output is piped into the input of the node along with the next amino acid in the sequence until the entire sequence is modeled. Regression layers are useful for modeling MHC class II peptides for two reasons: (1) data continuity is captured in the model, and (2) peptide lengths can vary without artificial padding. It is particularly applicable to The next layers of the recurrent neural network are a dropout layer with p=0.2 and finally a fully connected 64-node layer with ReLu activations.

[0469] The fourth model tested in Figure 14I is referred to herein as "Bi-LSTM." The Bi-LSTM model consists of a bidirectional long-term short-term memory neural network. The Bi-LSTM model is the same as the non-interacting model except for the peptide input layer. The input layer of the Bi-LSTM model accepts a 20-mer peptide string and then embeds this 20-mer peptide string as a (n,20,21) tensor. The following layers of bi-directional long-short term memory neural network of the Bi-LSTM model consisted of a recurrent long-term memory layer with 128 nodes, a dropout layer with p=0.2, and finally a fully connected 64 with ReLu activations. Consists of node layers. In conventional LSTM models, the order of continuous data is assumed to be directional (eg, read left-to-right or right-to-left). In a bidirectional LSTM, continuous data is processed in both directions, going left to right and right to left. Since peptide binding is essentially a non-directional task, modeling a sequence in both directions ensures that the information from both ends of the sequence has equal weight in the model's predictions.

[0470] Briefly referring to FIG. 14J, FIG. 14J shows an exemplary embodiment of the Bi-LSTM model of FIG. 14I configured to predict peptide presentation by HLA-DRB (MHC class II gene). . As shown in Figure 14J, the Bi-LSTM model consists of a shared neural network that accepts allele non-interacting properties (e.g., RNA sequence, sample number, protein number, and flanking sequence) and HLA-DRB alleles that each differ. and a set of distinct neural networks configured to accept encoded peptide sequences (allelic interaction properties) associated with . Each different neural network of the set of neural networks comprises a Bi-LSTM neural network. In the exemplary embodiment of the Bi-LSTM model of FIG. 14J, separate sets of neural networks associated with different alleles are associated with up to 4 different alleles to the HLA-DRB gene per patient sample. , it contains four separate neural networks. However, in another embodiment, in which the Bi-LSTM model is configured to predict peptide presentation by different HLA genes, separate sets of neural networks are used to generate the maximum It contains as many distinct neural networks as possible. Each separate neural network in the set of neural networks determines the likelihood that a peptide input to the model will be presented by the HLA-DRB alleles associated with a given neural network. Each of these likelihoods is then combined with the output from the shared neural network. Finally, the combined likelihoods are summed to generate the overall likelihood that the peptide is presented by the HLA-DRB gene.

[0471] Referring again to FIG. 14I, prior to using each of the four models in FIG. 14I to predict the likelihood that peptides in the test dataset of peptides will be presented by MHC class II verified. The binding affinity model was trained and validated using its own training and validation dataset based on HLA peptide binding affinity assays deposited in the Immune Epitope Database (IEDB, www.iedb.org). The other three models were trained using the training dataset of 69 samples mentioned above and validated using the validation dataset mentioned above. Following this training and validation of each model, each of the four models was tested with four excluded tumor samples from the test dataset described above. Specifically, for each of the four models, each peptide of the four excluded tumor samples from the test dataset was input to the model, and the model subsequently output the likelihood of presentation for that peptide.

[0472] The performance of each of these four models is shown in the line graph of Figure 14I. Specifically, each of the four models is associated with an ROC curve showing the ratio of true positive rate to false positive rate for each prediction generated by the model. For example, FIG. 14I shows the ROC curve of the binding affinity model, the ROC curve of the RNN model, the ROC curve of the MLP model, and the ROC curve of the Bi-LSTM model.

[0473] As noted above, the performance of a model in predicting the likelihood that a peptide is presented by an MHC class II molecule is the ratio of true positivity to false positivity for each prediction generated by the model. Quantified by determining the AUC of the ROC curve showing the ratio. A model with a higher AUC has better performance (ie, higher accuracy) compared to a model with a lower AUC. As shown in Figure 14I, the Bi-LSTM model curve achieved the highest AUC of 0.98. Therefore, the Bi-LSTM model achieved the best performance. Such peak performance of the Bi-LSTM model is partly due to the Bi-LSTM's greatest ability to accurately predict peptides of variable length, relatively long lengths, and peptides with repeated amino acids. It depends. The MLP and RNN model curves achieved the second highest AUC of 0.97. Therefore, MLP and RNN models achieved the second best performance. The binding affinity model curve achieved the lowest AUC of 0.79. Therefore, the binding affinity model performed the worst. Note that each of the Bi-LSTM, MLP, and RNN models tested in Figure 14I has an AUC greater than 0.9. Thus, despite the differences in peptide input layer architecture between these models, these models can yield relatively accurate predictions of peptide presentation, unlike binding affinity models that have significantly lower AUCs. .

[0474] Figure 14K is a line graph showing the perfect fit-recall curves for the "Bi-LSTM", "MLP", "RNN", and "Binding Affinity" models described above with respect to Figure 14I. As shown in Figure 14K and expected based on Figure 14I, the 'Bi-LSTM' model achieved the best performance with AUC = 0.23 and the 'RNN' model with AUC = 0.16 second. The 'MLP' model achieved the third best performance with AUC=0.11 and the 'Binding Affinity' model the worst with AUC=0.01. Specifically, the Bi-LSTM model trained on mass spectrometry data greatly outperformed the binding affinity model, with more than a 20-fold increase in AUC.

[0475] FIG. 14L shows two exemplary best-in-class conventional models given two different criteria, and two examples given two different sets of allele-interacting and non-allele-interacting variables. FIG. 4 is a line graph comparing the performance of presentation models in predicting the likelihood that a peptide in a test dataset of peptides will be presented by an MHC class II molecule. Specifically, FIG. 14L shows an example best-in-class conventional model (exemplary model 1) that uses the predicted binding affinity at minimum NetMHCII2.3 as the basis for generating predictions, generating predictions. Example best-in-class conventional model (exemplary model 2) using the binding rank predicted at minimum NetMHCII 2.3 as a criterion, predicting the likelihood of peptide presentation based on MHC class II molecule type and peptide sequence (Exemplary model 4), and an exemplary model (Exemplary model 3) is a line graph comparing performance.

[0476] The best-in-class conventional model used as example model 1 and example model 2 in FIG. 14L is the NetMHCII2.3 model. The NetMHCII2.3 model generates predictions of peptide presentation likelihood based on MHC class II molecule type and peptide sequence. The NetMHCII2.3 model can be found on the NetMHCII2.3 website (www.cbs.dtu.dk / services / NetMHCII / ,PMID 29315598) 76 was tested using

[0477] As mentioned above, the NetMHCII2.3 model was tested according to two criteria. Specifically, example model 1 generates predictions of peptide presentation likelihoods according to the predicted binding affinities at minimum NetMHCII2.3, and example model 2 according to the predicted binding ranks at minimum NetMHCII2.3. A prediction of peptide presentation likelihood was generated.

[0478] The presentation models used as example model 3 and example model 4 are embodiments of the presentation models disclosed herein that are trained using data obtained by mass spectrometry. As noted above, the presentation model generated predictions of peptide presentation likelihoods based on two different sets of allelic interaction and allele non-interaction variables. Specifically, exemplary model 4 generates predictions of peptide presentation likelihood based on MHC class II molecule type and peptide sequence (the same variables used in the NetMHCII2.3 model), and exemplary model 3: Predictions of peptide presentation likelihood were generated based on MHC class II molecule type, peptide sequence, RNA expression, gene identifier, and flanking sequence.

[0479] Each model was trained and validated prior to predicting the likelihood that a peptide in the test dataset of peptides would be presented by an MHC class II molecule using the exemplary model of FIG. 14L. The NetMHCII2.3 models (Exemplary Model 1 and Exemplary Model 2) used their own training and validation datasets based on HLA peptide binding affinity assays deposited in the Immune Epitope Database (IEDB, www.iedb.org). trained and verified. The training dataset used to train the NetMHCII2.3 model is known to contain almost exclusively 15-mer peptides. In contrast, example models 3 and 4 were trained using the training data set described above with respect to Figure 14H and validated using the validation data set described above with respect to Figure 14H.

[0480] Following training and validation of each model, each of the models was tested using the test dataset. As mentioned above, the NetMHCII2.3 model is trained on datasets containing mostly only 15-mer peptides, which means that NetMHCII2.3 cannot give different priorities to peptides of different weights. However, this reduces the expected performance of NetMHCII2.3 for mass spectrometry data of HLA class II presentations containing peptides of all lengths. Therefore, the test data set contained only 15-mer peptides to give a fair comparison between models unaffected by varying peptide lengths. Specifically, the test dataset contained 933 15-mer peptides. 40 of the 933 peptides in the test data set are mediated by MHC class II molecules, specifically HLA-DRB1*07:01, HLA-DRB1*15:01, HLA-DRB4*01:03 , and presented by HLA-DRB5*01:01 molecules. Peptides included in the test dataset were excluded from the training dataset described above.

[0481] To test each example model with the test data set, for each example model, the model generated predictions of the likelihood of peptide presentation for each peptide of the 933 peptides in the test data set. Specifically, for each peptide in the test data set, the model in Exemplary 1 uses the MHC class II molecule type and peptide sequence to determine the presentation score of that peptide by the MHC class II molecule, and the 4 generated by ranking peptides by their binding affinities predicted by the minimum NetMHCII2.3 across the three HLA class II DR alleles. Similarly, for each peptide in the test dataset, the model in Exemplary 2 uses the presentation score of that peptide by MHC class II molecules, the type of MHC class II molecule and the peptide sequence, and the four HLA Generated by ranking peptides by binding rank (ie, quantile-normalized binding affinities) predicted by the smallest NetMHCII2.3 across class II DR alleles. For each peptide in the test data set, the model of Example 4 generated the likelihood of presentation of that peptide by MHC class II molecules based on the type of MHC class II molecule and the peptide sequence. Similarly, for each peptide in the test data set, the model of Exemplary 3 determines the presentation of that peptide by an MHC class II molecule based on the MHC class II molecule type, peptide sequence, RNA expression, gene identifier, and flanking sequence. generated the likelihood.

[0482] The performance of each of these four example models is shown in the line graph of Figure 14L. Specifically, each of the four example models is associated with an ROC curve showing the ratio of true positive rate to false positive rate for each prediction generated by the model. For example, FIG. 14L shows the ROC curve of the model of Example 1 using binding affinities predicted by minimum NetMHCII2.3 to generate predictions, binding ranks predicted by minimum NetMHCII2.3 to generate predictions , the ROC curve of the model of Exemplification 4 that generated the likelihood of peptide presentation based on MHC class II molecule type and peptide sequence, and the MHC class II molecule type, peptide sequence, FIG. 13 shows the ROC curve for the model of Example 3 that generated peptide presentation likelihoods based on RNA expression, gene identifiers, and flanking sequences.

[0483] As noted above, the performance of a model in predicting the likelihood that a given peptide is presented by an MHC class II molecule is the ratio of the true positive rate to the false positive rate for each prediction generated by the model. Quantified by determining the AUC of the ROC curve showing the ratio. A model with a larger AUC has better performance (ie, higher accuracy) compared to a model with a smaller AUC. As shown in FIG. 14L, the curve for the model in Exemplary 3, which generated peptide presentation likelihoods based on MHC class II molecule type, peptide sequence, RNA expression, gene identifier, and flanking sequence, is the highest at 0.95. Achieved AUC. Thus, the model of Example 3, which generated peptide presentation likelihoods based on MHC class II molecule type, peptide sequence, RNA expression, gene identifier, and flanking sequence, achieved the best performance. The curve for the model in Exemplification 4, which generated the likelihood of peptide presentation based on MHC class II molecule type and peptide sequence, achieved the second highest AUC of 0.91. Thus, the model of Example 4, which generated peptide presentation likelihoods based on MHC class II molecule type and peptide sequence, achieved the second best performance. The curve for the model in Exemplification 1, which used binding affinities predicted by the minimum NetMHCII2.3 to generate predictions, resulted in the lowest AUC of 0.75. Therefore, the curve for the model in Exemplification 1, which used binding affinities predicted by minimum NetMHCII2.3 to generate predictions, performed the worst. The curve for the model in Example 2, using the binding rank predicted by minimum NetMHCII2.3 to generate predictions, resulted in the second lowest AUC of 0.76. Therefore, the curve for the model in Example 2, which used the binding rank predicted by minimum NetMHCII2.3 to generate the prediction, performed the second worst.

[0484] As shown in FIG. 14L, the performance gap between example models 1 and 2 and example models 3 and 4 is large. Specifically, the performance of the NetMHCII2.3 model (using either criteria of minimum NetMHCII2.3 predicted binding affinity or minimum NetMHCII2.3 predicted binding rank) is disclosed herein. performance of a presentation model (that generates peptide presentation likelihoods based on MHC class II molecule type and peptide sequence, or MHC class II molecule type, peptide sequence, RNA expression, gene identifier, and flanking sequence) is also almost 25% lower. Therefore, Figure 14L shows that the presentation model disclosed herein can achieve significantly more accurate presentation prediction than the current best-in-class conventional model, the NetMHCII2.3 model. is shown.

[0485] Furthermore, as mentioned above, the NetMHCII2.3 model is trained on a training dataset containing almost exclusively 15-mer peptides. As a result, the NetMHCII2.3 model is not trained to learn which peptide lengths are more likely to be presented by MHC class II molecules. Therefore, the NetMHCII2.3 model does not weight its prediction of likelihood of peptide presentation by MHC class II molecules according to peptide length. In other words, the NetMHCII2.3 model does not alter its prediction of the likelihood of peptide presentation by MHC class II molecules for peptides with lengths outside the 15 amino acid modal peptide length. As a result, the NetMHCII2.3 model over-predicts the likelihood of presentation of peptides with lengths longer or shorter than 15 amino acids.

[0486] In contrast, since the presentation model disclosed herein is trained using peptide data obtained by mass spectrometry, it can be trained on a training dataset containing peptides of all different lengths. can. As a result, the disclosed model disclosed herein can learn which peptide lengths are more likely to be presented by MHC class II molecules. Thus, the presentation model disclosed herein can weight its prediction of the likelihood of peptide presentation by MHC class II molecules according to the length of the peptide. In other words, the presentation models disclosed herein alter their predictions of the likelihood of peptide presentation by MHC class II molecules for peptides with lengths outside the 15 amino acid modal peptide length. can be done. As a result, the presentation model disclosed herein provides a higher or less , it is possible to achieve significantly more accurate presentation predictions. This is one of the advantages of using the presentation model disclosed herein to predict the likelihood of peptide presentation by MHC class II molecules.

[0487] XII.A.3. Example 3 FIG. 14M shows sequencing using mass spectrometry with a q value of less than 0.1 for each of a total of 230 samples, including human tumors (NSCLC, lymphoma, and ovarian cancer) and cell lines (EBV) containing HLA class II molecules. Fig. 4 is a histogram showing the amount of peptides singed. An average of 1300 peptides were sequenced for each sample with a q value of less than 0.1, as shown in Figure 14M.

[0488] As noted above with respect to Figure 14D, each of the 230 samples in Figure 14M contained HLA class II molecules. More specifically, each of the 230 samples in Figure 14M contained HLA-DR molecules. HLA-DR molecules are one type of HLA class II molecules. Even more specifically, each of the 230 samples in Figure 14M contained HLA-DRB1, HLA-DRB3, HLA-DRB4, and / or HLA-DRB5 molecules. HLA-DRB1 molecules, HLA-DRB3 molecules, HLA-DRB4 molecules, and HLA-DRB5 molecules are types of HLA-DR molecules.

[0489] Although this particular experiment was performed using a sample containing HLA-DR molecules, in particular HLA-DRB1, HLA-DRB3, HLA-DRB4 and HLA-DRB5 molecules, in alternative embodiments this Experiments can be performed with samples containing one or more of any type(s) of HLA class II molecules. For example, in alternative embodiments, the same experiment can be performed using samples containing HLA-DP and / or HLA-DQ molecules. It is well known to those skilled in the art that the same method can be used to model any type(s) of MHC class II molecules and still give reliable results. For example, Jensen, Kamilla Kjaergaard et al. 76 is an example of a recent scientific paper using the same method to model binding affinities for HLA-DR molecules and for HLA-DP and HLA-DQ molecules. Thus, one skilled in the art would be able to use the experiments and models described herein to model not only HLA-DR molecules, but any other MHC class II molecule separately or simultaneously, and still be confident. It will be appreciated that high results can be obtained with

[0490] Mass spectrometry was performed on each sample to sequence the peptides of each sample for a total of 230 samples. The mass spectra obtained for the samples were searched with Comet, scored with Percolator and peptides were sequenced. The amount of sequenced peptides in the samples was then determined for several different Percolator q-value thresholds. Specifically, for that sample, the amount of peptides sequenced with a Percolator q value of less than 0.01, a Percolator q value of less than 0.05, and a Percolator q value of less than 0.2 was determined.

[0491] The amount of peptides sequenced at each of the different Percolator q-value thresholds for each of the 203 samples is shown in Figure 14M. For example, as seen in Figure 14M, for the first sample approximately 8000 peptides with q-values ​​less than 0.1 were sequenced using mass spectrometry.

[0492] Overall, FIG. 14M demonstrates that mass spectrometry can be used to sequence large amounts of peptides with low q-values ​​from samples containing MHC class II molecules. In other words, the data shown in Figure 14M demonstrate that mass spectrometry can be used to reliably sequence peptides that can be presented by MHC class II molecules.

[0493] FIG. 14N is a histogram showing the amount of samples in which specific MHC class II molecule alleles were identified. More specifically, Figure 14N shows the amount of samples in which a particular MHC class II molecule was identified for a total of 230 samples containing HLA class II molecules.

[0494] As described above with respect to Figure 14M, each of the 230 samples in Figure 14M contained HLA-DRB1, HLA-DRB3, HLA-DRB4, and / or HLA-DRB5 molecules. Accordingly, FIG. 14N shows the amount of samples in which certain alleles were identified for HLA-DRB1, HLA-DRB3, HLA-DRB4, and HLA-DRB5 molecules.

[0495] To identify which HLA-DRB1, HLA-DRB3, HLA-DRB4, and HLA-DRB5 alleles were present in the sample, the sample was HLA class II DR typed. The number of samples in which HLA alleles were identified using HLA class II DR typing was then simply summed to determine the amount of samples in which a particular HLA allele was identified. For example, as shown in Figure 14N, 28 out of a total of 230 samples contained the HLA class II molecule allele HLA-DRB3*03:01. In other words, 28 samples out of a total of 230 samples contained the allele HLA-DRB3*03:01 for the HLA-DRB3 molecule. Overall, FIG. 14N demonstrates that a wide range of HLA class II molecule alleles can be identified from 230 samples containing HLA class II molecules. For the human population, allele frequencies for the HLA-DRB1 allele in the Caucasian population of US patent application number can be found in Maiers, M, et al. 161 .

[0496] FIG. 14O shows peptides bound to MHC class I molecules and peptides bound to MHC class II molecules 162 . As shown in Figure 14O, each peptide comprises a peptide backbone and multiple amino acids. Each MHC molecule has a binding groove. However, as discussed below, the manner in which peptides bind within the binding grooves of MHC class I and MHC class II molecules differs.

[0497] As noted throughout this disclosure, the length of peptides presented by MHC molecules can vary. Specifically, peptides presented by MHC molecules can be 9-20 amino acids in length. When a peptide binds to and is presented by an MHC molecule, the "binding core" of the peptide is located within the binding groove of the MHC molecule. Specifically, the binding core of a peptide is the sequence of amino acids of the peptide that are located within the binding groove of the MHC molecule when the peptide binds to and is presented by the MHC molecule. Furthermore, when a peptide binds to and is presented by an MHC molecule, the "binding anchor" of the binding core of the peptide physically binds within the binding groove of the MHC molecule. Specifically, a binding anchor of a binding core of a peptide is a specific amino acid of the binding core that binds to the binding groove of the MHC molecule when the peptide binds to and is presented by the MHC molecule.

[0498] As shown in FIG. 14O, the binding core of peptides presented by MHC class I molecules contains the entire length of the peptide. Specifically, as shown in Figure 14O, the entire peptide presented by the MHC class I molecule is located within the binding groove of the MHC class I molecule. In contrast, in molecules presented by MHC class II molecules, only a subsequence of the peptide's amino acids can be included in the binding core of the peptide. Specifically, as shown in Figure 14O, the ends of the peptide presented by the MHC class II molecule are not located within the binding groove of the MHC class II molecule. The partial sequence of amino acids that make up the binding core of peptides presented by MHC class II may be unknown. However, as stated in the literature, the most common length of the binding core of MHC class II-presenting peptides is 9 amino acids.

[0499] Furthermore, in addition to the unknown binding core of MHC class II-presenting peptides, the amounts and positions of the amino acids that make up the binding anchors of the binding core of the peptide may also be known. However, as stated in the literature, the binding core of MHC class II-presenting peptides usually contains 3-4 binding anchors, which usually contain amino acids located at both ends of the binding core.

[0500] Due to differences in peptide binding to MHC class I and MHC class II molecules, to ensure optimal peptide presentation prediction performance, the peptide presentation prediction model should be adapted to predict peptide presentation by MHC class II molecules in detail. must be configured to Specifically, the MHC class II peptide presentation prediction model takes into account this uncertainty, since the subsequence of amino acids comprising the binding core and the binding anchor of the binding core of peptides presented by MHC class II molecules may be unknown. must be configured to model Specifically, an inception model was developed to model the uncertainty of the binding core and binding anchor of peptides presented by MHC class II molecules.

[0501] FIG. 14P shows an exemplary embodiment of the inception neural network of the inception model of FIG. 14Q configured to predict peptide presentation by MHC class II molecules. Inception models are configured to identify binding cores and binding anchors for peptides presented by MHC class II molecules and to predict presentation of peptides by MHC class II molecules using these identified binding cores and binding anchors. This is the presented model. The inception model consists of a shared neural network that accepts allelic non-interacting properties (e.g., RNA sequence, sample number, protein number, and flanking sequences) and a separate inception neural network that accepts allelic-interacting properties (e.g., peptide sequence). and a set of Specifically, each distinct inception neural network in the set of distinct inception neural networks is configured to accept peptide sequences encoded associated with different MHC class II alleles (e.g., HLA-DRB alleles). be done. As mentioned above, FIG. 14P shows an exemplary embodiment of the Inception Neural Network of the Inception model.

[0502] First, since the length of peptides presented by MHC class II molecules is variable (e.g., 9-20 amino acids), peptides shorter than the maximum length of 20 amino acids were selected to have a length of 20 amino acids. padding to . Specifically, if the peptide is less than 20 amino acids in length, a particular amino acid Z is added to the left side of the peptide and then to the right side of the peptide. This pattern of padding the peptide is repeated until the peptide has a length of 20 amino acids. By padding the sides of the peptide, the binding core of the peptide remains unchanged while the peptide length is kept constant across all peptides.

[0503] The input layer of the Inception neural network accepts padded peptide sequences. The padded peptide is then one-hot encoded. As shown in FIG. 14P, each Inception neural network contains three one-dimensional CNN layers. One of the three CNN layers has 16 filters of size 8. One of the three CNN layers has 16 filters of size 10. One of the three CNN layers has 16 filters of size 12. These filter sizes, as mentioned above, are suitable for inception neural binding cores of about 9 amino acids, which has been shown in the literature to be the most common binding core length for peptides presented by MHC class II. It was deliberately chosen to be network specific.

[0504] The outputs of each of the three CNN layers are input to two one-dimensional CNN layers. One of the two CNN layers has 32 filters of size 1. One of the two CNN layers has 32 filters of size 2. These filter sizes were deliberately chosen to locate binding anchors within the binding core of peptides presented by MHC class II.

[0505] The outputs of these two CNN layers are concatenated. Each concatenated output is then fed to a bi-LSTM layer. The output of each bi-LSTM layer is concatenated and the concatenated result is sent to the multi-layer perceptron. The output of the multi-layer perceptron includes the output of separate inception neural networks. In other words, the output of the multi-layer perceptron includes the likelihood that a peptide input to a separate inception neural network will be presented by the MHC class II alleles associated with the separate inception neural network. The presentation likelihoods from each separate inception neural network are combined with the output from the shared neural network. Finally, the combined likelihoods are summed to generate an overall likelihood that the peptide is presented by one or more of the MHC class II alleles.

[0506] Figure 14Q shows "Bi-LSTM" and "Inception" in predicting the likelihood that a peptide in a test dataset of peptides will be presented by at least one MHC class II molecule present in the test dataset. 4 is a line graph comparing the performance of the presented models; Specifically, FIG. 14Q is a line graph showing the perfect precision-recall curves for the 'Bi-LSTM' and 'Inception' models. AUC is used to quantify the performance of each model.

[0507] The first model tested in Figure 14Q is the "Bi-LSTM" model. The Bi-LSTM model is the model detailed above with respect to Figures 14I and 14J.

[0508] The second model tested in Figure 14Q is the "Inception" model. The Inception model is the model described in detail above with respect to Figure 14P.

[0509] Each exemplary model was trained and validated prior to using these models to predict the likelihood that a peptide within a test dataset of peptides would be presented by an MHC class II molecule. To train, validate, and finally test the exemplary model, the data described above for the 230 samples were divided into training, validation, and test datasets.

[0510] To prevent peptides from appearing in more than one of the training, validation, and test datasets, the following steps were taken. First, all peptides from 230 samples that appeared at multiple locations within the proteome were removed. The remaining peptides from the 230 samples were then divided into blocks of 10 contiguous peptides. Each block of adjacent peptides was uniquely assigned to a training, validation, or test dataset. This resulted in no peptides appearing in multiple datasets for training, validation and test datasets.

[0511] The training dataset contained 188,210 peptides presented by MHC class II molecules derived from 226 of a total of 230 samples. The 188,210 peptides included in the training dataset were 9-20 amino acids in length. A Bi-LSTM model and an Inception model were trained on the training dataset using the ADAM optimization algorithm and early stopping, respectively.

[0512] The validation dataset contained 21,764 peptides presented by MHC class II molecules derived from the same 226 samples used in the training dataset. Validation datasets were used for early stopping only.

[0513] The test dataset included peptides presented by MHC class II molecules identified from tumor samples using mass spectrometry. Specifically, it consisted of 232 peptides identified from four tumor samples. Peptides included in the test dataset were excluded from the training dataset described above.

[0514] After training and validating the Bi-LSTM and Inception models using the training and validation datasets, respectively, each model was tested using the test dataset. The performance of the Bi-LSTM and Inception models on the test data set is shown in Figure 14Q as perfect precision-recall curves and AUC scores. As shown in Figure 14Q, the Bi-LSTM model outperformed the Inception model, achieving AUC=0.347. Bi-LSTM achieved AUC=0.238.

[0515] XII.A.4. Example 4 To further evaluate the applicability of the predictive model disclosed herein to class II HLA peptides, published class II mass spectrometry data were analyzed using peptides that each express a single HLA class I allele. obtained for two cell lines. One cell line expressed HLA-DRB1*15:01 and the other HLA-DRB5*01:01 150 . These two cell lines were used for training data. For test data, class II mass spectrometry data were obtained from separate cell lines expressing both HLA-DRB1*15:01 and HLA-DRB5*01:01. 151 . RNA-sequencing data from a different B-cell line B721.221 was used because RNA-sequencing data were not available for either training or test cell lines. 92 was substituted with

[0516] The peptide set was split into training, validation, and test sets using the same procedure as for the HLA class I data, except that the class II data included peptides with lengths of 9-20. The training data included 330 peptides presented by HLA-DRB1*15:01 and 103 peptides presented by HLA-DRB5*01:01. The test data set contained 223 peptides presented by HLA-DRB1*15:01 or HLA-DRB5*01:01 and 4708 non-presented peptides.

[0517] We trained an ensemble of 10 models on the training dataset to predict HLA class II peptide presentation. The architecture and training procedure of these models are similar to those of class I, except that the class II model adopted sequences that were one-hot coded to length 20 instead of length 11 and padded to 0 as input peptides. It was assumed to be the same as that used to predict the presentation of

[0518] Figure 15 takes the lowest NetMHCIIpan percentile rank across the "MS model", HLA-DRB1*15:01 and HLA-DRB5*01:01, "NetMHCIIpan rank": NetMHCIIpan 3.1 152 , and "NetMHCIIpan nM", which adopts the strongest affinity in nM across HLA-DRB1*15:01 and HLA-DRB5*01:01: NetMHCIIpan 3.1, HLA-DRB1*15:01 / HLA-DRB5*01 :01 compares predictive performance in ranking peptides within the test dataset. A "MS model" is the MHC class II presentation prediction model disclosed herein.

[0519] Specifically, FIG. 15 shows the Receiver Operating Characteristic (ROC) curve and the area under the ROC curve AUC (Panel A) and AUC 0.1 (Panel B) Statistics are shown for these ranking methods. AUC 0.1 is the AUC between 0 and 0.1 FPR*10, commonly considered in the epitope prediction field 19 . The performance of the NetMHCIIpan nM method and the NetMHCIIpan rank method were similar. The MS model performed best, significantly outperforming the compared methods, especially in the critical high-specificity region of the ROC curve (AUC 0.1 0.41 compared to 0.27).

[0520] XII.B. Examples of presented model parameters determined for MHC class II alleles The following describes a variation of the multi-allele presentation model (equation (16)) that generates presentation likelihoods for each potential allele of the class II MHC alleles HLA-DRB1*12:01 and HLA-DRB1*10:01. Figure 3 shows the determined set of parameters; u=expit(relu(X W 1 +b 1 ) W 2 +b 2 ) where relu( ) is the rectified linear unit (ReLU) function, W 1 , b 1 , W 2 , and b 2 is the set of parameters θ determined for the model. Allelic interaction variables X are contained in a 1×399 matrix consisting of one-hot-coded, middle-padded peptide sequences, one row per input peptide. W. 1 The size of b is (399×256) and b 1 is (1×256), W 2 The size of is (256×2), b 2 The size of is (1×2). The first row of output shows the probabilities of potential per allelic presentation of that peptide sequence by allele HLA-DRB1*12:01, the second row of output by allele HLA-DRB1*10:01. The probabilities of presentation for each potential allele of the peptide sequence are shown. For demonstration purposes, b 1 , b 2 , W 1 , and W 2 are shown in Appendix A.

[0521] XIII. Example 9: Evaluation of the MHC Class II Presentation Model of T Cell Data A published CD4+ T-cell multimer / tetramer assay to assess whether accurate prediction of peptide presentation by MHC class II alleles will lead to the ability to identify human tumor CD4 epitopes (i.e., immunotherapeutic targets) Immune Epitope Database (IEDB) 88 downloaded from. These data have 18 distinct HLA-DRB alleles, including 14 HLA-DRB1 alleles, 2 HLA-DRB3 alleles, 1 HLA-DRB4 allele, and 1 HLA-DRB5 allele. It consisted of 3,470 peptides of 9-20 residues in length from human samples. On average, each allele had 33 samples containing that allele. A full MHC class II MS model (the same model as described above in Section XII.A.2) was compared against the binding affinity prediction tool NetMHCII2.3. Across 18 alleles, the mean area under the ROC curve (ROC AUC) for the full MHC class II MS model was 0.81 with a standard deviation of 0.08, compared to only 0.65 for NetMHCII2.3. , with a standard deviation greater than 0.13. These results demonstrate the superior ability of the full MHC class II MS model to predict CD4 T cell epitopes. On an allele-by-allele basis, ROC AUCs were significantly similar between the two models for some of the more common alleles such as HLA-DRB1*01:01. For example, for HLA-DRB1*01:01, the full MHC class II MS model had a ROC AUC of 0.83 and the NetMHCII2.3 model had a ROC AUC of 0.81. However, for most alleles there was a wider difference in performance between the two models. Of the 18 allele-specific tests, the full MHC class II MS model outperformed the NetMHCII2.3 model in 17 alleles. NetMHCII2.3 outperformed the full MHC class II MS model in only one allele of HLA-DRB1*15:02. However, this allele was poorly represented in the training data for a full MHC class II MS model in which only one sample containing the allele was included.

[0522] XIV. Example 10: Evaluation of retrospective neoantigen T cell data by MHC class II presentation model This example further evaluates whether accurate prediction of peptide presentation by MHC class II molecules leads to the ability to identify human tumor CD4 T cell epitopes. To make this assessment, we ranked the CD4+ immunogenicity of the peptides predicted by the MHC class II presentation model.

[0523] A test data set suitable for this evaluation includes peptides presented by MHC class II molecules on the surface of tumor cells and recognized by T cells. In addition, formal performance evaluation included not only peptides shown as positive (i.e., recognized by T cells), but also a sufficient number of negatively labeled peptides (i.e., tested but not recognized by T cells). ) peptide was also required. Mass spectrometry data sets correspond to tumor presentation but not T cell recognition. Conversely, post-vaccination priming or T cell assays correspond to the presence of T cell precursors but not tumor presentation. For example, potent HLA-binding peptides with low levels of expression in the tumor of the source gene can generate strong post-immune CD4 T cell responses of no therapeutic utility because such peptides are not presented by the tumor.

[0524] To obtain a trial dataset suitable for this evaluation, we collected publicly available data from recent studies. 163 . The collected test data set included 69 positively labeled single nucleotide variant (SNV) mutations CD4+ responsive to TILs across 45 patients. As noted above, negatively labeled SNV mutations were also included in the collected test dataset. Specifically, there were an average of 104 and a median of 106 negatively labeled SNV mutations per patient.

[0525] Each SNV mutation in the test dataset was represented as a 25 amino acid sequence with the SNV mutation at amino acid position 13 in the middle of the sequence. Next, for each sequence of 25 amino acids, all possible peptides of 9-20 amino acids in length containing SNV mutations were generated. Each sequence of 25 amino acids yielded 118 possible peptides. For each possible peptide, flanking sequences of 5 amino acids were added to the left and right sides of the peptide.

[0526] To simulate the selection of antigens for personalized immunotherapy, the SNV mutations for each patient in the study dataset were matched to the Inception model and NetMHCIIPan3.2 binding affinity model disclosed herein with the TPM=1 gene. Using the expression threshold, the likelihood of presentation by the patient's MHC class II allele was ranked in order. The inception model used was trained to predict peptide presentation by 32 different MHC class II alleles, covering 25 of the 30 MHC class II alleles present in patients in the study dataset.

[0527] In the Inception model, we calculated the presentation score of each of the 118 possible peptides of the SNV mutation for each of the patient's identified MHC class II alleles in order to calculate the likelihood of presentation of each SNV mutation for each patient. determined using a model. We then identified the highest presentation score determined by the Inception model for each of the patient's MHC class II alleles. Finally, these highest presentation scores for each of the patient's MHC class II alleles were summed to determine the overall likelihood of presentation of the patient's SNV mutations.

[0528] In the NetMHCIIPan3.2 model, to calculate the likelihood of presentation of each SNV mutation in each patient, we calculated the binding affinity of each of the 118 possible peptides of the SNV mutation for each of the patient's identified MHC class II alleles. Gender was determined using the NetMHCIIPan3.2 model. We then identified the highest reverse binding affinities determined by the NetMHCIIPan3.2 model for each of the patient's MHC class II alleles. Note that the highest reverse binding affinities were identified, as lower binding affinities indicate a higher likelihood of presentation. Finally, these highest reverse binding affinities for each of the patient's MHC class II alleles were summed to determine the overall likelihood of presentation of the patient's SNV mutations.

[0529] Each patient's SNV mutations were then ranked in order of likelihood of presentation by the patient's MHC class II as determined by both the Inception and NetMHCIIPan3.2 models. Antigen-specific immunotherapy is technically limited in the number of MHC class II specificities it targets (e.g., current personalized vaccines encode approximately 10–20 somatic mutations). 80~82 , of which approximately 10 may be MHC class II-specific), the top 1, 2, 3, 4, 5, and 10 SNV mutations were ranked for each patient.

[0530] Additionally, as a control, for each patient, each of the patient's SNV mutations from genes with a TPM of 1 or greater was randomly ranked. Specifically, for each patient, each of the patient's SNV mutations from genes with a TPM of 1 or greater was randomly ranked in 100 trials to obtain an overall ranking of each SNV mutation for each patient. Decided.

[0531] After performing the ranking of SNV mutations, the predictive model was applied to the pre-existing Comparisons were made by counting the number of T cell responses. T cells (e.g., pre-existing We compared the proportion of SNV mutations recognized by the T-cell responses of the Specifically, Table 2 below shows the positive labeling SNV mutations predicted as the top 1, 2, 3, 4, 5, and 10 predictions by a particular model out of a total of 69 positive labeling SNV mutations. Indicates percentage (%). As shown in Table 2, the Inception model is more likely to accurately predict CD4+ immunogenic MHC class II-presenting peptides compared to the NetMHCIIPan3.2 model and random prediction. (Table 2) TIFF2022524328000062.tif19149

[0532] Therefore, this evaluation is in line with previous literature. 81,82,97 This establishes the superior performance of the Inception model, whi...

Claims

1. 1. A method for identifying one or more T cells derived from one or more tumor cells in a subject, the T cells being antigen-specific for at least one neoantigen likely to be presented on the surface of said tumor cells by one or more class II MHC alleles, the method comprising: obtaining at least one of exome, transcriptome, or whole genome nucleotide sequencing data from the tumor cells and normal cells of the subject, wherein the nucleotide sequencing data is used to obtain data representing a peptide sequence for each neoantigen of a set of neoantigens identified by comparing the nucleotide sequencing data from the tumor cells with the nucleotide sequencing data from the normal cells, wherein the peptide sequence for each neoantigen includes at least one alteration that makes the peptide sequence different from a corresponding wild-type peptide sequence identified from the normal cells of the subject; encoding the peptide sequence of each of the neoantigens into a corresponding numerical vector, each numerical vector containing information about a set of amino acids constituting the peptide sequence and the positions of the amino acids in the peptide sequence; using a computer processor to input the numerical vectors into a machine-learned presentation model to generate a set of presentation likelihoods for the set of neoantigens, wherein each presentation likelihood in the set represents the likelihood that a corresponding neoantigen will be presented on the surface of the tumor cells of the subject by the one or more class II MHC alleles; a label obtained by mass spectrometry to measure, for each sample of the plurality of samples, the presence of a peptide bound to at least one class II MHC allele from the set of class II MHC alleles identified as being present in said sample; and a training peptide sequence, encoded as a numerical vector containing information about a set of amino acids that make up the peptide and the positions of the amino acids in the peptide, for each of the samples; a plurality of parameters identified based at least on a training dataset including: a function representing a relationship between the numerical vector received as an input and the proposed likelihood generated as an output based on the numerical vector and the parameters; the steps of: selecting a subset of the set of neoantigens based on the set of presentation likelihoods to generate a set of selected neoantigens; identifying one or more T cells in said subset that are antigen-specific for at least one of said neoantigens; and returning the one or more identified T cells The method comprising:

2. The step of inputting the numerical vector into the machine-learned presentation model includes: applying the machine-learned presentation model to the peptide sequence of the neoantigen to generate a dependency score for each of the one or more class II MHC alleles indicating whether the class II MHC allele presents the neoantigen based on the particular amino acid at the particular position of the peptide sequence. The method of claim 1 , comprising:

3. The step of inputting the numerical vector into the machine-learned presentation model includes: transforming the dependency scores to generate a corresponding allele-specific likelihood for each Class II MHC allele, the likelihood indicating the likelihood that the corresponding Class II MHC allele will present the corresponding neoantigen; and combining the likelihoods for each allele to generate the presentation likelihood of the neoantigen. The method of claim 2 further comprising:

4. 4. The method of claim 3, wherein transforming the dependency score models presentation of the neoantigen as mutually exclusive across the one or more class II MHC alleles.

5. The step of inputting the numerical vector into the machine-learned presentation model includes: transforming the combination of dependency scores to generate the presentation likelihood, wherein transforming the combination of dependency scores models presentation of the neoantigen as interference between the one or more class II MHC alleles. The method of claim 2 further comprising:

6. The set of representation likelihoods is further specified by at least one or more allele non-interaction characteristics, and the method comprises: applying the machine-learned presentation model to the allele-non-interacting trait to generate a dependency score for the allele-non-interacting trait, the dependency score indicating whether the corresponding neoantigen peptide sequence will be presented based on the allele-non-interacting trait. The method of any one of claims 2 to 5, further comprising:

7. combining the dependency score for each class II MHC allele of the one or more class II MHC alleles with the dependency score for the non-allele interacting trait; transforming the combined dependency scores for each class II MHC allele to generate a per-allele likelihood for each class II MHC allele, the likelihood being that the corresponding class II MHC allele will present the corresponding neoantigen; and combining the likelihoods for each allele to generate the representation likelihood. The method of claim 6 further comprising:

8. combining the dependency scores for each of the class II MHC alleles with the dependency scores for the allele-non-interacting trait; and transforming the combined dependency scores to generate the presentation likelihood. The method of claim 6 further comprising:

9. The method of any one of claims 1 to 8, wherein the one or more class II MHC alleles comprises two or more different class II MHC alleles.

10. The method of any one of claims 1 to 9, wherein the at least one class II MHC allele comprises two or more different types of class II MHC alleles.

11. The method of any one of claims 1 to 10, wherein the peptide sequence comprises a peptide sequence having a length other than 9 amino acids.

12. The method of any one of claims 1 to 11, wherein the step of encoding the peptide sequence comprises encoding the peptide sequence using a one-hot coding scheme.

13. The plurality of samples (a) one or more cell lines engineered to express a single class II MHC allele; (b) one or more cell lines engineered to express multiple class II MHC alleles; (c) one or more human cell lines obtained 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 12, comprising at least one of:

14. the training data set is (a) data relating to measurements of peptide-MHC binding affinity for at least one of said peptides; and (b) data relating to a measure of peptide-MHC binding stability for at least one of said peptides; The method of any one of claims 1 to 13, further comprising at least one of:

15. 15. The method of any one of claims 1-14, wherein the set of representation likelihoods is further identified by at least the expression level of the one or more class II MHC alleles in the subject, wherein the expression level is measured by RNA-seq or mass spectrometry.

16. The set of presentation likelihoods is: (a) the predicted affinity between a neoantigen in the set of neoantigens and the one or more class II MHC alleles; and (b) the predicted stability of the neoantigen-encoded peptide-MHC complex. The method of any one of claims 1 to 15, further characterized by characteristics including at least one of:

17. The set of numerical likelihoods is: (a) a C-terminal sequence adjacent to the neoantigen-encoding peptide sequence within the source protein sequence; and (b) an N-terminal sequence adjacent to the neoantigen-encoding peptide sequence within the source protein sequence; The method of any one of claims 1 to 16, further characterized by characteristics including at least one of:

18. 18. The method of any one of claims 1 to 17, wherein the step of selecting the set of selected neoantigens comprises selecting neoantigens that are more likely to be presented on the tumor cell surface than neoantigens that are not selected based on the machine-learned presentation model.

19. 19. The method of any one of claims 1 to 18, wherein the step of selecting the set of selected neoantigens comprises selecting neoantigens that are more likely to be able to induce a tumor-specific immune response in the subject compared to neoantigens that are not selected based on the machine-learned presentation model.

20. 20. The method of any one of claims 1 to 19, wherein the step of selecting the set of selected neoantigens comprises selecting neoantigens that have a higher likelihood of being presented to naive T cells by professional antigen-presenting cells (APCs) compared to unselected neoantigens based on the presentation model, and optionally the APCs are dendritic cells (DCs).

21. 21. The method of any one of claims 1 to 20, wherein the step of selecting the set of selected neoantigens comprises selecting neoantigens that are less likely to be inhibited by central or peripheral tolerance compared to neoantigens that are not selected based on the machine-learned presentation model.

22. 22. The method of any one of claims 1 to 21, wherein the step of selecting the set of selected neoantigens comprises selecting neoantigens that are less likely to be able to induce an autoimmune response against normal tissue in the subject compared to neoantigens that are not selected based on the machine-learned presentation model.

23. 23. The method of any one of claims 1 to 22, 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.

24. 24. The method of any one of claims 1 to 23, further comprising generating an output for constructing a personalized cancer vaccine from the set of selected neoantigens.

25. 25. The method of claim 24, 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 neoantigens.

26. The method according to any one of claims 1 to 25, wherein the machine-learned presentation model is a neural network model.

27. 27. The method of claim 26, wherein the neural network model comprises a plurality of network models for the class II MHC alleles, each network model comprising a set of nodes assigned to a corresponding one of the class II MHC alleles and arranged in one or more layers.

28. 28. The method of claim 27, wherein each network model further comprises one or more convolutional neural networks, each of the one or more convolutional neural networks comprising a series of nodes arranged in one or more layers and having filters of different sizes, the filters of each of the one or more convolutional neural networks being sized to identify the position of an amino acid in the peptide sequence of each neoantigen that comprises a binding core or a binding anchor of the peptide sequence.

29. 29. The method of claim 27 or 28, wherein the neural network models are trained by updating parameters of the neural network models, and the parameters of at least two network models are updated together in at least one training iteration.

30. 30. The method of any one of claims 26 to 29, wherein the machine-learned presentation model is a deep learning model comprising one or more layers of nodes.

31. 31. The method of any one of claims 1-30, wherein the step of identifying the one or more T cells comprises co-culturing the one or more T cells with one or more of the neoantigens in the subset under conditions that expand the one or more T cells.

32. 32. The method of any one of claims 1 to 31, wherein the step of identifying the one or more T cells comprises contacting the one or more T cells with an MHC multimer comprising one or more of the neoantigens in the subset under conditions that allow binding of the T cells to the MHC multimer.

33. 33. The method of any one of claims 1 to 32, further comprising identifying one or more T cell receptors (TCR) of the one or more identified T cells.

34. 34. The method of claim 33, wherein the step of identifying the one or more T cell receptors comprises sequencing the T cell receptor sequences of the one or more identified T cells.

35. 35. An isolated T cell antigen-specific for at least one selected neoantigen in the subset according to any one of claims 1 to 34.

36. genetically engineering a plurality of T cells to express at least one of the one or more identified T cell receptors; culturing the plurality of T cells under conditions that expand the plurality of T cells; and injecting the expanded T cells into the subject.

35. The method of claim 34, further comprising:

37. said step of genetically engineering said plurality of T cells to express at least one of said one or more identified T cell receptors comprises: cloning the T cell receptor sequences of the one or more identified T cells into an expression vector; and transfecting each of said plurality of T cells with said expression vector.

37. The method of claim 36, comprising:

38. culturing the one or more identified T cells under conditions that expand the one or more identified T cells; and injecting the expanded T cells into the subject. The method of any one of claims 1 to 37, further comprising:

Images