Neoantigen identification with pan-allele models

Abstract

This invention provides a method for identifying neoantigens that can be presented by MHC pair genes on the surface of a subject's tumor cells. The peptide sequences of the tumor neoantigens and MHC pair genes are obtained by sequencing the subject's tumor cells. The tumor neoantigens and MHC pair gene peptide sequences are input into a machine learning presentation model to generate presentation probabilities of the tumor neoantigens, each presentation probability representing the likelihood that the neoantigen will be presented by at least one of the MHC pair genes on the surface of the subject's tumor cells. A subset of neoantigens is selected based on these presentation probabilities.

Description

[Previous Technology] Therapeutic vaccines and T-cell therapies based on tumor-specific neoantigens hold great promise as next-generation personalized cancer immunotherapies.1–3 Cancers with high mutation burden (such as non-small cell lung cancer (NSCLC) and melanoma) are particularly attractive targets for such therapies, given the relatively greater likelihood of neoantigen generation.4,5 Early evidence suggests that neoantigen-based vaccination can elicit T-cell responses,6 and neoantigen-targeting T-cell therapies can cause tumor regression in certain selected patients.7 Both MHC class I and MHC class II influence T-cell responses.70–71 However, the identification of neoantigens and neoantigen-recognizing T cells has become a central challenge in assessing tumor response 77,110, examining tumor evolution 111, and designing next-generation personalized therapies 112. Current neoantigen identification techniques are time-consuming and laborious 84,96 or not precise enough 87,91–93. Although neoantigen-recognizing T cells have recently been shown to be a major component of TILs 84,96,113,114 and circulating in the peripheral blood of cancer patients 107, current methods for identifying neoantigen-reactive T cells have some combination of the following three limitations: (1) they rely on difficult-to-obtain clinical samples, such as TILs 97,98 or leukaphereses 107 (2) they require impractical screening of large peptide libraries 95 or (3) they rely on MHC multimers that are practically available for only a small number of MHC pairs. Furthermore, initial approaches have been proposed that incorporate next-generation sequencing, RNA gene expression, and predictions of MHC binding affinity for candidate neoantigen peptides into mutation-based analyses.8 However, these proposed methods cannot model the complete antigenic determinant generation process, which involves many steps beyond gene expression and MHC binding (e.g., TAP transport, proteasome cleavage, MHC binding, transport of the peptide-MHC complex to the cell surface, and / or TCR recognition of MHC-I; endocytosis or autophagy, cleavage via extracellular or lysosomal proteases (e.g., cathepsins), HLA binding catalyzed by HLA-DM competing with CLIP peptides, transport of the peptide-MHC complex to the cell surface, and / or TCR recognition of MHC-II).9 Therefore, existing methods suffer from reduced positive predictive value (PPV). (Figure 1A) Indeed, analysis of multiple sets of peptides presented by tumor cells has shown that <5% of peptides predicted to be presented using gene expression and MHC binding affinity can be found on tumor surface MHC10,11 (Figure 1B). This low correlation between binding prediction and MHC presentation is further reinforced by the following recent observation: the lack of individually binding-restricting neoantigens improves the predictive accuracy of checkpoint inhibitor responses in multiple mutations12 The low positive predictive value (PPV) of existing methods for predicting presentation raises questions about neoantigen-based vaccine design and neoantigen-based T-cell therapy. If a predictive vaccine design with low PPV is used, most patients will be unable to tolerate the therapeutic neoantigen, and a small number will still be able to tolerate more than one (even assuming all presented peptides are immunogenic). Similarly, if therapeutic T-cells are designed based on predictive NPV with low PPV, most patients will be unable to tolerate T-cells responsive to tumor neoantigens, and the time and physical resource costs of predicting and identifying predictive neoantigens using downstream laboratory techniques may be prohibitively high. Therefore, neoantigen vaccination and T-cell therapy using current methods are unlikely to be successful in large numbers of subjects with cancer. (Figure 1C) Furthermore, previous methods only used cis-acting mutations to generate candidate neoantigens and largely ignored additional sources of neo-ORF, including mutations in splicing factors (which occur in multiple tumor types and lead to aberrant splicing of many genes) and mutations that produce or remove protease cleavage sites. Finally, standard methods for tumor genomic and transcriptomic analysis can miss somatic mutations that generate candidate neoantigens due to suboptimal conditions in library construction, exome and transcriptome capture, sequencing, or data analysis. Similarly, standard tumor analysis methods can unintentionally promote sequence artifacts or germline polymorphisms as neoantigens, leading to either ineffective use of vaccine capacity or risk of autoimmunity, respectively. [Summary of the Invention] This paper discloses an optimized approach for identifying and selecting neoantigens for personalized cancer vaccines, T-cell therapies, or both. First, it addresses the use of next-generation sequencing (NGS) to optimize tumor exome and transcriptome analyses for neoantigen candidate identification. These methods build upon standard NGS tumor analysis approaches to ensure the highest sensitivity and specificity for neoantigen candidates across all categories of genomic alterations. Second, it proposes a novel approach for high-PPV neoantigen selection to overcome specificity issues and ensure that neoantigens targeting vaccines and / or those intended as targets for T-cell therapies are more likely to elicit antitumor immunity. Depending on the embodiment, these methods include training a statistical regression or nonlinear deep learning model configured to predict the presentation of peptides of multiple lengths with shared statistical strength across peptides of varying lengths based on pan-pair genes. The model can predict the probability that a peptide will be presented via any MHC pair gene, including unknown MHC pairs not previously encountered by the model during training. Specifically, these nonlinear deep learning models can be designed and trained to treat different MHC pairs in the same cell as independent, thereby solving the problem of mutual interference when using linear models. Finally, additional considerations are addressed for the design and manufacture of neoantigen-based personalized vaccines and for the generation of personalized neoantigen-specific T cells for T-cell therapy. The model disclosed herein outperforms predictive factors trained on binding affinity at the time of patent application and earlier predictive factors based on MS peptide data by up to an order of magnitude. By more reliably predicting peptide presentation, this model enables more timely and cost-effective identification of neoantigen-specific or tumor antigen-specific T cells for personalized therapy using clinical practice methods that use a limited volume of patient peripheral blood, screen for a few peptides per patient, and unnecessarily rely on MHC multimers. However, in another embodiment, the model disclosed herein can be used to enable more timely and cost-effective identification of tumor antigen-specific T cells using MHC multimers by reducing the number of peptides binding to MHC multimers, which require screening for these MHC multimers to identify neoantigen-specific or tumor antigen-specific T cells. The model presented in this paper demonstrates predictive performance on TIL neoantigen determinant datasets and the task of identifying anticipated neoantigen-reactive T cells. Therapeuticly applicable neoantigen determinant predictions can now be obtained through modeling HLA processing and presentation. In summary, this work provides practical computer-simulated antigen identification for targeted immunotherapy, thereby accelerating the progression of disease in patients.

Implementation Method

[76] The specific paired genes used to train the four models for each of the four test samples depend on the HLA paired genes contained in the given test sample. Specifically, for the first test sample containing the HLA paired gene A*02:01, the training data used to train the four models to predict the presentation probability of the HLA paired gene A*02:01 includes the HLA paired gene A*02:01. For the second test sample containing the HLA paired gene A*02:01, the training data used to train the four models to predict the presentation probability of the HLA paired gene A*02:01 does not include the HLA paired gene A*02:01. For the third test sample containing the HLA paired gene B*44:02, the training data used to train the four models to predict the presentation probability of the HLA paired gene A*44:02 does not include the HLA paired gene B*44:02. For the fourth test sample containing the HLA dual gene C*08:02, the training data used to train the four models to predict the presentation probability of the HLA dual gene A*08:02 did not include the HLA dual gene C*08:02. During testing on each of the four samples, each model was tested on a single-pair gene dataset containing HLA pairs in a given sample and consisting of approximately 250,000 peptides (both presented and unpresented peptides were counted). Specifically, during testing on each of the four samples, the pan-pair gene presentation model, the random forest model, and the quadratic discriminant model each received the same input. Specifically, for each of the four samples, the pan-pair gene presentation model, the random forest model, and the quadratic discriminant model each received the 34-mer one-hot encoded HLA pair protein sequence of the HLA pair in the sample, and the 9-mer one-hot encoded (i.e., binarized) protein sequence of the peptide in question. On the other hand, for each of the four samples, the MHCFlurry model received the name of the HLA pair in the sample, and the 9-mer one-hot encoded (i.e., binarized) protein sequence of the peptide in question. As stated above, this difference in the inputs between these models is a result of constructing the MHCFlurry model to select which pair gene sub-model to use to produce the facts of the presented predictions. After these inputs are fed into four models, each model outputs the probability of predicting that the HLA-pair gene will present a peptide. These predicted probabilities are compared to known results of peptide presentation (i.e., marker yi) to produce the precision / recall curves shown in Figures 19 to 22. Specifically, Figure 19 corresponds to the data output by each of the four models using the first test sample, Figure 20 corresponds to the data output by each of the four models using the second test sample, Figure 21 corresponds to the data output by each of the four models using the third test sample, and Figure 22 corresponds to the data output by each of the four models using the fourth test sample. In each figure, the blue line represents the precision / recall curve of the pan-pair gene model, the orange line represents the precision / recall curve of the MHCFlurry model, the green line represents the precision / recall curve of the random forest model, and the red line represents the precision / recall curve of the quadratic discriminant model. Furthermore, each figure indicates the average prediction precision (i.e., positive predictive value) of each model. For example, as shown in Figure 19, the average prediction accuracy of the pan-dual gene model is 0.32. As shown in Figures 19 to 22, random forest and quadratic discriminant models using a generalized dual gene model architecture incorporating non-neural networks were run approximately twice, along with the MHCFlurry model. Furthermore, a generalized dual gene presentation model incorporating neural networks was run approximately twice, along with random forest and quadratic discriminant models using a generalized dual gene model architecture incorporating non-neural networks. In other words, the generalized dual gene presentation model incorporating neural networks achieved the highest accuracy compared to other models. However, the random forest and quadratic discriminant models using a generalized dual gene model architecture incorporating non-neural networks still outperformed the customized model based on dual gene binding affinity, MHCFlurry. Therefore, these results demonstrate that the generalized dual gene model architecture can be well generalized to other non-neural network machine learning models, such as random forests based on decision trees and variations of Bayesian methods (e.g., quadratic discriminant analysis), while still providing a high level of predictive accuracy. Furthermore, as shown in Figures 20 to 22, even though the pan-dual gene presentation model, the random forest model, and the secondary discriminant model never saw the HLA dual gene in the test, these models (including the random forest model and the secondary discriminant model, both of which use a pan-dual gene model architecture containing a non-neural network) achieved comparable performance to the model in Figure 19, which saw the HLA dual gene in the test during training. Therefore, these results demonstrate the ability of the pan-dual gene model architecture containing a non-neural network to accurately predict the presentation probability of HLA dual genes not used to train the model. IX. Example 5: Prediction Module Prediction module 320 uses a presentation model to receive sequence data and select candidate neoantigens from the sequence data. Specifically, the sequence data may be DNA, RNA, and / or protein sequences extracted from tumor tissue cells of a patient. Prediction module 320 processes the sequence data into multiple peptide sequences pk having 8 to 15 amino acids against MHC-I or 6 to 30 amino acids against MHC-II. For example, prediction module 320 may process a given sequence "IEFROEIFJEF" into three peptide sequences "IEFROEIFJ", "EFROEIFJE", and "FROEIFJEF" having 9 amino acids. In one embodiment, prediction module 320 may identify candidate neoantigens containing one or more mutations by comparing sequence data extracted from normal tissue cells of a patient with sequence data extracted from tumor tissue cells of a patient. Prediction module 320 applies one or more of the presentation models to the processed peptide sequence to assess the presentation probability of the peptide sequence. Specifically, prediction module 320 may select one or more candidate neoantigen peptide sequences that may be presented on tumor HLA molecules by applying the presentation model to candidate neoantigens. In one embodiment, prediction module 320 selects candidate neoantigen sequences with an estimated presentation probability above a predetermined threshold. In another embodiment, the presentation model selects v candidate neoantigen sequences with the highest estimated presentation probability (where v is generally the maximum number of antigenic determinants that can be delivered in a vaccine). A vaccine containing selected candidate neoantigens against a given patient can be injected into that patient to induce an immune response. X. Example 6: Patient Selection Module The patient selection module 324 selects a subset of patients for vaccine therapy and / or T-cell therapy based on whether patients meet the inclusion criteria. In one embodiment, the inclusion criteria are determined based on the presentation probability of patient neoantigen candidates, such as those generated by a presentation model. By adjusting the inclusion criteria, the patient selection module 324 can adjust the number of patients to receive vaccine and / or T-cell therapy based on the presentation probability of its neoantigen candidates. Specifically, strict inclusion criteria result in a smaller number of patients to be treated with vaccine and / or T-cell therapy, but may result in a higher percentage of patients treated with vaccine and / or T-cell therapy who receive effective treatment (e.g., one or more tumor-specific neoantigens (TSNAs) and / or one or more neoantigen-responsive T-cells). On the other hand, lenient inclusion criteria result in a larger number of patients to be treated with vaccine and / or T-cell therapy, but may result in a lower percentage of patients treated with vaccine and / or T-cell therapy who receive effective treatment. Patient selection module 324 incorporates criteria based on the required balance between the target ratio of patients receiving treatment and the ratio of patients receiving effective treatment. In some embodiments, the inclusion criteria for selecting patients to receive vaccine treatment are the same as those for selecting patients to receive T-cell therapy. However, in alternative embodiments, the inclusion criteria for selecting patients to receive vaccine treatment may differ from those for selecting patients to receive T-cell therapy. Sections XA and XB below each discuss the inclusion criteria for selecting patients to receive vaccine treatment and the inclusion criteria for selecting patients to receive T-cell therapy. XA: Selection of Patients for Vaccine Treatment In one embodiment, a patient is associated with a corresponding treatment subset of v neoantigen candidates, which may potentially be included in a custom vaccine for patients with a vaccine capacity of v. In one embodiment, the treatment subset for a patient is the neoantigen candidate with the highest presentation probability as determined by a presentation model. For example, if a vaccine may contain v=20 antigen determinants, then the vaccine may contain treatment subsets for each patient with the highest presentation probability as determined by a presentation model. However, it should be understood that in other embodiments, the treatment subset for a patient may be determined based on other methods. For example, the treatment subset for a patient may be randomly selected from the patient's neoantigen candidate set or may be determined based on a model of the binding affinity or stability of peptide sequences at the time of the current patent application, or a combination of factors including the presentation probability from the presentation model and information about the affinity or stability of those peptide sequences. In one embodiment, if a patient's tumor mutational burden is equal to or greater than the minimum mutational burden, then the patient selection module 324 determines that the patient meets the inclusion criteria. A patient's tumor mutational burden (TMB) indicates the total number of nonsynonymous mutations in the tumor exon. In one implementation, if the absolute number of a patient's TMB is equal to or greater than a predetermined threshold, then the patient selection module 324 may select the patient for vaccine therapy. In another implementation, if a patient's TMB is within the threshold percentile of TMB values ​​measured for a patient group, then the patient selection module 324 may select the patient for vaccine therapy. In another embodiment, if the utility score of a patient in a treatment subset is equal to or greater than the minimum utility score, then patient selection module 324 determines that the patient meets the inclusion criteria. In one implementation, the utility score is a measure of the estimated number of neoantigens presented from the treatment subset. The estimated number of neoantigens presented can be predicted by modeling neoantigen presentation as a random variable of one or more probability distributions. In one implementation, the utility score of patient i is the expected number of neoantigen candidates to be presented from a subset of treatments or some function thereof. As an example, the presentation of each neoantigen can be modeled as a Bernoulli random variable, where the probability of presentation (success) is given by the presentation likelihood of the neoantigen candidate. Specifically, for a subset Si of v neoantigen candidates pi1, pi2, ..., piv with the highest presentation likelihoods ui1, ui2, ..., uiv, the presentation of neoantigen candidate pij is given by a random variable Aij, where: [Figure] the expected number of neoantigens presented is given by summing the presentation likelihoods of each neoantigen candidate. In other words, the utility score of patient i can be represented as: [Figure] The patient selection module 324 selects a subset of patients with a utility score equal to or greater than the minimum utility for vaccine treatment. In another implementation, the utility score of patient i is the probability of presenting at least the minimum threshold number of neoantigens k. In one instance, the number of neoantigens presented in the therapeutic subset Si of neoantigen candidates is modeled as a Poisson binomial random variable, where the probability of presentation (success) is given by the presentation likelihood of each antigen determinant. Specifically, the number of neoantigens presented by patient i can be given by a random variable Ni, where: [Figure] where PBD(∙) denotes a Poisson binomial distribution. The probability of presenting at least the minimum threshold number of neoantigens k is given by the probability that the sum of the number of neoantigens presented Ni will be equal to or greater than k. In other words, the utility score of patient i can be represented as: [Figure] The patient selection module 324 selects a subset of patients with a utility score equal to or greater than the minimum utility score for vaccine treatment. In another implementation, the utility score for patient i is the number of neoantigens in a therapeutic subset Si of neoantigen candidates having a binding affinity or predicted affinity below a fixed threshold (e.g., 500 nM) for one or more of the patient's HLA pairs. In one example, this fixed threshold ranges from 1000 nM to 10 nM. Depending on the circumstances, the utility score may only include those neoantigens detected as expressed via RNA-seq. In another implementation, the utility score for patient i is the number of neoantigens in a therapeutic subset Si that are neoantigen candidates with a binding affinity for one or more of the patient's HLA pairs that is at or below the threshold percentile for the binding affinity of a random peptide for that HLA pair. In one example, the threshold percentile is the range from the 10th percentile to the 0.1st percentile. Depending on the circumstances, the utility score may only include neoantigens detected as expressed by RNA-seq. It should be understood that the examples of utility scores provided for equations (25) and (27) are illustrative only, and that other statistical or probability distributions may be used to generate utility scores for patient selection module 324. XB Patient Selection for T-Cell Therapy In another embodiment, instead of or in addition to vaccine treatment, the patient may receive T-cell therapy. Similar to vaccine treatment, in embodiments where a patient receives T-cell therapy, the patient may be associated with a corresponding therapeutic subset of the v neoantigen candidates as described above. This therapeutic subset of the v neoantigen candidates can be used to identify, in vitro, T cells from the patient that respond to one or more of the v neoantigen candidates. These identified T cells can then be expanded and fused into the patient for customized T-cell therapy. Patients can be selected to receive T-cell therapy at two different time points. The first point is after using these models to predict a subset of v neoantigen candidates for treatment, but before in vitro screening of T cells specific to the predicted treatment subset of v neoantigen candidates. The second point is after in vitro screening of T cells specific to the predicted treatment subset of v neoantigen candidates. First, patients can be selected for T-cell therapy after a treatment subset of v neoantigen candidates has been predicted for a patient, but before in vitro identification of T cells specific to the predicted subset of v neoantigen candidates. Specifically, because in vitro screening of neoantigen-specific T cells from a patient can be expensive, it is desirable to select only patients for neoantigen-specific T-cell screening if a patient is more likely to have neoantigen-specific T cells. To select patients prior to the in vitro T-cell screening step, the same criteria used for selecting patients for vaccine therapy can be used. Specifically, in some embodiments, patient selection module 324 may select a patient for T-cell therapy if the patient's tumor mutation burden is equal to or greater than the minimum mutation burden described above. In another embodiment, patient selection module 324 may select a patient for T-cell therapy if the utility score of a patient based on the treatment subset of v neoantigen candidates is equal to or greater than the minimum utility score described above. Secondly, in addition to or instead of selecting patients for T-cell therapy before in vitro identification of T cells specific to a predicted subset of v neoantigen candidates, patients can also be selected for T-cell therapy after in vitro identification of T cells specific to a predicted therapeutic subset of v neoantigen candidates. Specifically, if, during in vitro screening of T-cells for patients used for neoantigen identification, at least a critical limit number of neoantigen-specific TCRs are identified for the patient, then the patient can be selected for T-cell therapy. For example, the patient can be selected for T-cell therapy only if at least two neoantigen-specific TCRs are identified for the patient, or only if neoantigen-specific TCRs are identified for two different neoantigens. In another embodiment, a patient may be selected for T-cell therapy only if their TCR identifies at least a number of neoantigens within a therapeutic subset of the patient's v neoantigen candidates. For example, a patient may be selected for T-cell therapy only if their TCR identifies at least one neoantigen within a therapeutic subset of the patient's v neoantigen candidates. In other embodiments, a patient may be selected for T-cell therapy only if their TCR is identified as having neoantigen specificity to a specific HLA-restricted class of neoantigen peptides. For example, a patient may be selected for T-cell therapy only if at least one of their TCRs is identified as having a neoantigen-specific HLA class I-restricted neoantigen peptide. In even other embodiments, a patient may be selected for T-cell therapy only if their TCR identifies at least a number of specific HLA-restricted neoantigen peptides of a particular class that are at least at the threshold number. For example, a patient may be selected for T-cell therapy only if their TCR identifies at least one HLA class I restricted neoantigen peptide. As another example, a patient may be selected for T-cell therapy only if their TCR identifies at least two HLA class II restricted neoantigen peptides. Any combination of the above criteria may also be used to select a patient for T-cell therapy after in vitro identification of T-cells specific to a predicted therapeutic subset of v neoantigen candidates. XI. Example 7: Experimental results showing the patient selection performance in the example. The effectiveness of the patient selection method described in Section X was tested by selecting patients on a set of simulated patients associated with each of the test sets of simulated neoantigen candidates, wherein a subset of the simulated neoantigens is known to exist in the mass spectrometry data. Specifically, each simulated neoantigen candidate in the test set is associated with a marker indicating whether the neoantigen is presented in the mass spectrometry data sets of the multiply paired gene JY cell lines HLA-A*02:01 and HLA-B*07:02 from the Bassani-Sternberg dataset (dataset "D1") (data available at www.ebi.ac.uk / pride / archive / projects / PXD0000394). As described in more detail below with reference to Figure 23A, many of the simulated patient neoantigen candidates are sampled from the human proteome based on the known frequency distribution of mutation burden in non-small cell lung cancer (NSCLC) patients. The training set was used to train a pairwise gene presentation model for the same HLA pairwise gene. This training set was a subset of the mass spectrometry data of single-pairwise genes HLA-A*02:01 and HLA-B*07:02 from the IEDB dataset (dataset "D2") (data available at http: / / www.iedb.org / doc / mhc_ligand_full.zip). Specifically, the presentation model for each pairwise gene was the pairwise gene model shown in equation (8), which used network dependency functions gh(∙) and gw(∙) and an out-of-field function f(∙) to combine the N-terminal and C-terminal flanking sequences into pairwise gene non-interacting variables. The presentation model for the pairwise gene HLA-A*02:01 generated the presentation probability of a given peptide on the pairwise gene HLA-A*02:01, the given peptide sequence as a pairwise gene interaction variable, and the N-terminal and C-terminal flanking sequences as pairwise gene non-interacting variables. The presentation model of the HLA-B*07:02 pair gene generates the presentation probability of a given peptide on the HLA-B*07:02 pair gene, the given peptide sequence as a pair gene interaction variable, and the N-terminal and C-terminal side-joint sequences as pair gene non-interaction variables. As illustrated in the following examples and with reference to Figures 23A to 23E, various models (such as training and presentation models for peptide binding prediction and models of the current patent application level) are applied to a test set of neoantigen candidates for each simulated patient to identify different treatment subsets of patients based on these predictions. Patients meeting the inclusion criteria are selected for vaccine therapy and are associated with a customized vaccine that includes a patient-specific treatment subset based on antigen determination. The size of the treatment subset varies depending on the vaccine capacity. No overlap is introduced between the training set used to train the presentation model and the test set of simulated neoantigen candidates. In the following examples, the proportion of selected patients with at least a certain number of presented neoantigens contained in the antigenic determinants of the vaccine is analyzed. This statistic indicates the effectiveness of the potential neoantigen that the simulated vaccine delivery will elicit an immune response in the patient. Specifically, if a neoantigen is present in mass spectrometry dataset D2, a simulated neoantigen in the test set is presented. A high proportion of patients with presented neoantigens indicates the potential for successful treatment by inducing an immune response via a neoantigen vaccine. XI.A. Example 7A: Frequency distribution of mutation burden in NSCLC cancer patients Figure 23A illustrates the sample frequency distribution of mutation burden in NSCLC patients. Mutation burden and mutations in different tumor types (including NSCLC) can be identified, for example, using the Cancer Genome Atlas (TCGA) (https: / / cancergenome.nih.gov). The x-axis represents the number of nonsynonymous mutations in each patient, and the y-axis represents the percentage of sample patients with a given number of nonsynonymous mutations. The sample frequency distribution in Figure 23A shows a range of 3 to 1786 mutations, with 30% of patients having fewer than 100 mutations. Although not shown in Figure 23A, studies indicate that mutation burden is higher in smokers than in non-smokers, and that mutation burden can be a strong indicator of neoantigen load in patients. As introduced at the beginning of Section XI above, a number of simulated patients were associated with a test set of neoantigen candidates. The test set for each patient was generated by sampling mutation burden mi from the frequency distribution shown for each patient in Figure 23A. For each mutation, a 21-mer peptide sequence from the human proteome was randomly selected to represent the simulated mutation sequence. The test set for neoantigen candidate sequences for patient i was generated by identifying (8, 9, 10, 11)-mer peptide sequences across the 21-mer mutation. Each neoantigen candidate was associated with a marker indicating the presence of the neoantigen candidate sequence in the mass spectrometry D1 dataset. For example, a neoantigen candidate sequence present in dataset D1 could be associated with marker "1", while a sequence not present in dataset D1 could be associated with marker "0". Figures 23B to 23E illustrate the experimental results of patient selection based on presented neoantigens from patients in the test sets, as described in more detail below. XI.B. Example 7B: Proportion of selected patients with neoantigen presentation based on mutation burden inclusion criteria. Figure 23B illustrates the number of presented neoantigens in a mock vaccine for patients selected based on whether they met the inclusion criteria of minimum mutational burden. It identifies the proportion of selected patients who had at least a certain number of presented neoantigens in the corresponding test. In Figure 23B, the x-axis indicates the percentage of patients excluded from vaccine treatment based on minimum mutational burden, as indicated by the marker "Minimum Number of Mutations". For example, the data point at "Minimum Number of Mutations" of 200 indicates that patient selection module 324 selects only a subset of simulated patients with a mutational burden of at least 200 mutations. As another example, the data point at "Minimum Number of Mutations" of 300 indicates a lower percentage of simulated patients selected by patient selection module 324 with at least 300 mutations. The y-axis indicates the percentage of selected patients associated with at least a certain number of presented neoantigens in the test set without any vaccine capacity v. Specifically, the upper graph shows the percentage of selected patients presented with at least one neoantigen, the middle graph shows the percentage of selected patients presented with at least two neoantigens, and the lower graph shows the percentage of selected patients presented with at least three neoantigens. As shown in Figure 23B, the proportion of selected patients with presented neoantigens increases significantly with higher mutational burden. This mutational burden indicator, used as an inclusion criterion, is effective in selecting patients in whom neoantigen vaccines are more likely to successfully induce an immune response. XI.C. Example 7C: Comparison of neoantigen presentation of vaccines identified by a presentation model versus neoantigen presentation of vaccines identified by a model based on the technology at the time of patent application. Figure 23C compares the number of presented neoantigens in simulated vaccines between selected patients with a vaccine-associated treatment subset identified based on a presentation model and selected patients with a vaccine-associated treatment subset identified through a model based on the technology level at the time of the current patent application. The left figure assumes a limited vaccine capacity v=10, and the right figure assumes a limited vaccine capacity v=20. Patients are selected based on utility scores indicating the expected number of presented neoantigens. In Figure 23C, solid lines indicate patients associated with vaccines, including those identified using presentation models based on the HLA-A*02:01 and HLA-B*07:02 pairs of genes. Treatment subsets for each patient are identified by applying each presentation model to sequences in the test set and identifying v neoantigen candidates with the highest presentation probability. Dashed lines indicate patients associated with vaccines, including those identified using the NETMHCpan model, the then-current patent-pending technology model based on a single HLA-A*02:01 pair of genes. Detailed implementation information for NETMHCpan is available at http: / / www.cbs.dtu.dk / services / NetMHCpan. Treatment subsets for each patient are identified by applying the NETMHCpan model to sequences in the test set and identifying v neoantigen candidates with the highest predicted binding affinity. The x-axis of both graphs indicates the percentage of patients excluded from vaccine treatment based on expected utility scores, which are based on the expected number of presented neoantigens in the treatment subset identified by the presentation model. Expected utility scores are determined as described in Equation (25) in Section X. The y-axis indicates the percentage of selected patients presented with at least a certain number of neoantigens (1, 2, or 3 neoantigens) included in the vaccine. As shown in Figure 23C, patients associated with vaccines containing a subset of treatments based on a presentation model received vaccines containing presented neoantigens at a significantly higher rate than patients associated with vaccines containing a subset of treatments based on a model based on the technology level at the time of patent application. For example, as shown in the right figure, 80% of the selected patients associated with vaccines based on a presentation model received at least one presented neoantigen in the vaccines compared to only 40% of the selected patients associated with vaccines based on a model based on the technology level at the time of patent application. The results indicate that the presentation model described herein is effective for selecting neoantigen candidates for vaccines that may elicit an immune response to treat tumors. XI.D. Example 7D: The impact of HLA coverage on neoantigen presentation of vaccines identified by a presentation model. Figure 23D compares the number of neoantigens presented in simulated vaccines between selected patients in a treatment subset identified by a single-pair gene presentation model based on HLA-A*02:01 and selected patients in a treatment subset identified by both HLA-A*02:01 and HLA-B*07:02 gene presentation models. The vaccine capacity was set to v=20 antigen determinants. For each experiment, patients were selected based on the expected utility score determined for different treatment subsets. In Figure 23D, the solid line indicates patients associated with vaccines who received treatments comprising both presentation models based on the HLA-A*02:01 and HLA-B*07:02 pairs of genes. Treatment subsets for each patient were identified by applying each presentation model to sequences in the test set and identifying the v neoantigen candidates with the highest presentation probability. The dashed line indicates patients associated with vaccines who received treatments comprising a single presentation model based on the HLA-A*02:01 pair of genes. Treatment subsets for each patient were identified by applying a single HLA-A*02:01 presentation model to sequences in the test set and identifying the v neoantigen candidates with the highest presentation probability. For the solid line plot, the x-axis indicates the percentage of patients excluded from vaccine treatment based on the expected utility score of the treatment subset identified using both presentation models. For the dashed line plot, the x-axis indicates the percentage of patients excluded from vaccine treatment based on the expected utility score of the treatment subset identified using a single presentation model. The y-axis indicates the percentage of selected patients who presented with at least a certain number of neoantigens (1, 2, or 3 neoantigens). As shown in Figure 23D, patients associated with vaccines containing a subset of treatments identified using a presentation model of two HLA pairs presented at a significantly higher rate than patients associated with vaccines containing a subset of treatments identified using a single presentation model. These results highlight the importance of establishing presentation models with high HLA pair coverage. XI.E. Example 7E: Comparison of neoantigen presentation in patients selected by mutation burden versus patients selected by the expected number of presented neoantigens. Figure 23E compares the number of neoantigens presented in a simulated vaccine between patients selected based on mutational burden and those selected based on expected utility scores. Expected utility scores were determined based on a treatment subset identified using a presentation model with a size of v=20 antigenic determinants. In Figure 23E, the solid line indicates patients selected based on expected utility scores associated with vaccines that include a subset of treatments identified by the presentation model. The treatment subsets for each patient were identified by applying the presentation model to sequences in the test set and identifying v=20 neoantigen candidates with the highest presentation probability. The expected utility score was determined based on the presentation probability of the treatment subsets identified using Equation (25) in Section X. The dashed line indicates patients selected based on mutational burden associated with vaccines that also include a subset of treatments identified by the presentation model. The x-axis indicates the percentage of patients excluded from vaccine treatment based on expected utility scores (solid line) and the percentage of patients excluded based on mutational burden (dashed line). The y-axis indicates the percentage of selected patients who received vaccines containing at least a certain number of presented neoantigens (1, 2, or 3 neoantigens). As indicated in Figure 23E, patients selected based on expected utility scores received vaccines containing presented neoantigens at a higher rate than those selected based on mutational burden. However, patients selected based on mutational burden received vaccines containing presented neoantigens at a higher rate than unselected patients. Therefore, mutational burden is an effective patient selection criterion for successful neoantigen vaccine therapy, although the expected utility score is more effective. XII. Example 8: Evaluation of retained mass spectrometry data by a mass spectrometry training model. Because HLA peptide presentation by tumor cells is a key requirement for anti-tumor immunity 91, 96, 97, a large integrated dataset (N=74 patients) of human tumor and normal tissue samples with paired class I HLA peptide sequences, HLA types, and transcriptomic RNA-seq (method) was generated using these and publicly available data 92, 98, 99 to train a novel deep learning model 100 to predict antigen presentation in human cancers. Samples were selected from several tumor types of interest in immunotherapy development and based on tissue availability. Mass spectrometry identified an average of 3,704 peptides / sample at peptide levels FDR < 0.1 (range 344–11,301). These peptides followed a characteristic class I HLA length distribution: lengths of 8 to 15 aa, with a modal length of 9 (56% of peptides). Consistent with previous reports, most peptides (median 79%) were predicted to bind at least one patient HLA pair gene via MHCflurry90 at the standard 500 nM affinity threshold, but with substantial variability across samples (e.g., 33% of peptides in one sample had predicted affinity >500 nM). The commonly used 101 "strong binder" threshold of 50 nM captured only 42% of presented peptides at a median rate. Transcriptome sequencing yielded an average of 131 M unique reads / sample and 68% of genes were expressed at at least one transcript / million (TPM) level in at least one sample, emphasizing values ​​from large and diverse sample sets to observe the largest number of gene expressions. Peptide presentation via HLA was strongly correlated with mRNA expression. Significant and reproducible intergenetic differences in peptide presentation rates beyond those that could be explained by differences in RNA expression or sequence alone were observed. The observed HLA type matched expectations from samples from patients with a predominantly European ancestry group. For each patient, positive marker data points were peptides detected by mass spectrometry, and negative marker data points were peptides from the reference proteome (SwissProt) in the sample that were not detected by mass spectrometry. Data were divided into training, validation, and test sets (methods). The training set consisted of 142,844 HLA-presented peptides (FDR < ~0.02) from 101 samples (69 recently described and 32 previously disclosed in this study). The validation set (used for early cessation) consisted of 18,004 presented peptides from the same 101 samples. Two mass spectrometry datasets were used for testing: (1) a tumor sample test set consisting of 571 presented peptides from 5 additional tumor samples (2 lung, 2 colon, and 1 ovary) removed from the training data, and (2) a single-pair gene cell line test set consisting of 2,128 presented peptides from genomic localization windows (blocks) located in adjacent but different locations from the single-pair gene peptides included in the training data (see additional details on the training / testing method). Using these and publicly available HLA peptide data 92, 98, 99, neural network (NN) models were trained to predict HLA antigen presentation. Specifically, in Example 9, the pan-dual gene model discussed in Section VIII.D above was trained using the above data to predict HLA antigen presentation. On the other hand, in Example 11, the dual gene-specific model described in detail below was trained using the above data to predict HLA antigen presentation. In Example 10, both the pan-dual gene model discussed in Section VIII.D above and the dual gene-specific model described in detail below were trained using the above data to predict HLA antigen presentation. Specifically, in Examples 10 and 11, novel network architectures were developed for learning pair gene-specific models from tumor mass spectrometry data, where each peptide can be presented by any of the six HLA pair genes. These models can jointly learn pair gene peptide localization and pair gene-specific presentation motifs (see Section XVII.B below). Training data identified predictive models for 53 HLA pair genes. Compared to previous work, these models captured the dependence of HLA presentation on the sequence positions of peptides of multiple lengths. The models also correctly learned the critical dependence of mRNA abundance on gene RNA expression and gene-specific presentation tendency, and learned the per-gene tendency to independently combine to produce presentation rates with a maximum difference of approximately 60-fold between the least expressed and most expressed genes. Furthermore, the stability of the model's prediction of HLA / peptide complex measurements in IEDB88 was observed (for 10 paired genes, p < 1e-10), even after controlling for the predicted binding affinity (for 8 / 10 paired genes tested, p < 0.05). Together, these characteristics form the basis for improved prediction of immunogenic HLA class I peptides. XIII. Example 9: Experimental Results of Induced Presentation Hotspot Modeling To specifically evaluate the benefits of using presentation hotspot parameters in HLA modeling, the performance of pan-pair gene neural network presentation models incorporating presentation hotspot parameters was compared with that of pan-pair gene neural network presentation models without incorporating presentation hotspot parameters. The basic neural network architecture was identical for both pan-pair gene models and was the same as that described in Sections VII to VIII above. Briefly, these pan-pair gene models included peptide and flanking amino acid sequence parameters, RNA transcriptional sequencing (TPM) data, protein family data, per-sample identification, and HLA-A, B, and C serotypes. A set of 5 networks was used for each pan-pair gene model. The pan-pair gene model incorporating presentation hotspot parameters used Equation 12b described in Section VIII.B.3 above, with a per-gene proteome block size of 10 and peptide lengths of 8 to 12. The two pan-dual gene models were compared using the mass spectrometry dataset described in Section XII above. Specifically, for the purpose of fairly evaluating the competing models, five samples were retained from model training and validation. The remaining samples were randomly divided into 90% for model training and 10% for validation of the training. Figure 24 compares the positive predictive value (PPV) at 40% recall when these pan-pair gene models were tested on five types of retained test samples, using pan-pair gene presentation models with and without presentation hotspot parameters. As shown in Figure 24, the pan-pair gene presentation model incorporating presentation hotspot parameters consistently outperformed the pan-pair gene presentation model without them. XIV. Example 10: Model Evaluation of Retrospective Neoantigen T-cell Data We then evaluate whether the accurate prediction of HLA peptide presentation by the pan-pair gene model can be translated into the ability to identify human tumor CD8 T-cell antigenic determinants (i.e., immunotherapy targets). Defining a suitable test dataset for this evaluation is challenging because it requires peptides that are recognized by T cells and presented by HLA on the surface of tumor cells. Furthermore, formal performance assessment requires not only positively labeled (i.e., T-cell-recognized) peptides but also a sufficient number of negatively labeled (i.e., tested but unrecognized) peptides. Mass spectrometry datasets address tumor-presented but not T-cell-recognized peptides; conversely, initiation or post-vaccination T-cell assays address the presence of T-cell precursors and T-cell recognition but not tumor presentation (e.g., a strongly binding peptide whose source gene is expressed at too low a level in the tumor to support peptide presentation may produce a strong CD8 T-cell response after vaccination, but is not a therapeutically usable target because it is not presented by the tumor). To obtain a suitable data set, we collected publicly available data from four recent studies that met the required criteria regarding CD8 T-cell antigenic determinants: Study A140 examined TILs in nine patients with gastrointestinal tumors and reported T-cell recognition of 12 / 1,053 somatic SNV mutations in autologous dendritic cells using the longitudinal microgenetic assay (TMG) via IFN-γ ELISPOT. Study B84 also used TMG and reported T-cell recognition of 6 / 574 SNVs by CD8+PD-1+ circulating lymphocytes from four melanoma patients. Study C141 assessed TILs from three melanoma patients using pulsed peptide stimulation and found a response to 5 / 381 tested SNV mutations. Study D108 assessed TILs from a breast cancer patient using a combination of TMG assays and minimal antigenic determinant peptide pulses and reported the recognition of 2 / 62 SNVs. The combined dataset consists of 2,023 examined SNVs from 17 patients, including 26 TSNAs with pre-existing T-cell responses. Importantly, because this dataset primarily includes neoantigen recognition via tumor-infiltrating lymphocytes, successful prediction suggests not only the ability to identify neoantigens that elicit T-cell responses as described in references 81, 82, 141, but also, more rigorously, the ability to identify neoantigens presented to T-cells via tumors. We used standard HLA binding affinity prediction with a gene expression threshold >2 TPM to rank mutations in order of presentation probability, as determined by RNA-seq, the pair-specific neural network model described in Section VIII.B, and the pan-pair-specific neural network model described in Section VIII.D. Because the capacity for antigen-specific immunotherapy is limited by the number of target specificities (e.g., currently personalized vaccines encode approximately 10 to 20 mutations 6, 81, 82), we compared predictive methods by calculating the number of pre-existing T-cell responses among the top 5, 10, or 20 mutations in each patient. These results are illustrated in Figure 25A. Specifically, Figure 25A compares the ratio of somatic mutations identified by T-cells (e.g., pre-existing T-cell responses) for the top 5, 10, and 20 somatic mutations identified using standard HLA binding affinity prediction with gene expression >2 TPM thresholds, as determined by RNA-seq, pair-gene-specific neural network models, and pan-pair-gene neural network models for test sets containing 12 different test samples. As expected, affinity predictions included only a small number of pre-existing T-cell responses from preferred mutations, for example, 9 / 26 (35%) of the top 20. In contrast, the majority (19 / 26, 73%) of pre-existing T-cell responses ranked in the top 20 using both the pair-gene-specific model and the pan-pair-gene NN model (Figure 25A). These results confirm the ability of the pan-pair-gene model to identify human tumor CD8 T-cell antigenic determinants with comparable accuracy (not statistically significant) to the pair-gene-specific model. We then evaluate mutations at the minimal neoantigen determinant level (i.e., 8-11mers covering the mutation are identified) because they can be used to identify T-cell / TCRs for cell therapy. In other words, the minimal neoantigen determinants are ranked in order of presentation probability using standard HLA binding affinity predictions with gene expression >2 TPM thresholds, as determined by RNA-seq, the pair-specific neural network model described in Section VIII.B, and the pan-pair-specific neural network model described in Section VIII.D. As mentioned above, because antigen-specific immunotherapy is technically limited by the number of target specificities, the predictive methods are compared by calculating the number of pre-existing T-cell responses in the top 5, 10, or 20 minimal neoantigen determinants for each patient with at least one pre-existing T-cell response. Positive markers are those confirmed to be the minimal antigenic determinants of the peptide-based immunogen (instead of, or in addition to, TMG-based assays), and negative examples are all antigenic determinants not identified in peptide-based assays and all transmutation antigenic determinants contained in unidentified microgenes. The results are shown in Figure 25B. Specifically, Figure 25B compares the ratio of the smallest neoantigen determinants (NIDs) identified by T-cells (e.g., pre-existing T-cell responses) based on the top 5, 10, and 20 NIDs identified using standard HLA binding affinity prediction with gene expression >2 TPM thresholds, as determined by RNA-seq, pair gene-specific neural network models, and pan-pair gene neural network models for a test set containing 12 different test samples, each taken from patients with at least one pre-existing T-cell response. As shown in Figure 25B, when evaluating mutations at the level of minimal antigenic determinants, the pan-dual gene model continues to perform comparablely to the dual gene-specific model. (XIV.A. Data) We obtained mutant cell, HLA type, and T-cell identification data from supplementary information from Gros et al. 84, Tran et al. 140, Stronen et al. 141, and Zacharakis et al. We inferred that tumor RNA expression across different patients was associated with the same tumor type, replacing RNA-seq data from tumor type-matched patients from TCGA. This data was used for neural network prediction and for filtering RNA expression before binding affinity prediction. The addition of tumor type-matched RNA-seq data improved predictive performance. For mutation-level analysis (Figure 25A), positive marker data points for Gros et al., Tran et al., and Zacharakis et al. were mutations recognized by patient T-cells by either the TMG assay or the minimum antigenic determinant peptide-pulse assay. Negative marker data points were all other mutations tested in the TMG assay. For Stronen et al., positive marker mutations were mutations crossed by at least one identified peptide, and negative marker data points were all mutations tested in the tetramer assay but not identified. For Gros, Tran, and Zacharakis data, mutations were ranked by summing presentation probabilities or by taking the minimum binding affinity across all mutation-crossing peptides, since the 25-mer TMG assay for a mutation tests T-cell recognition across all peptides of that mutation. For Stronen data, mutations were ranked by summing presentation probabilities or by taking the minimum binding affinity across all mutation-crossing peptides tested in the tetramer assay. For antigenic determinant-level analysis, positive data points were all minimal antigenic determinants recognized by patient T cells via peptide-pulse or tetramer assays, and negative data points were all minimal antigenic determinants not recognized by T cells via peptide-pulse or tetramer assays and all mutant-crossing peptides from the tested TMG not recognized by patient T cells. In the cases of Gros et al., Tran et al., and Zacharakis et al., peptides crossing the minimal antigenic determinants not identified in TMG analyses tested by peptide-pulse assays were removed from the analysis because the T-cell recognition status of these peptides was not experimentally determined. XV. Example 11: Identification of Neoantigen-Reactive T Cells in Cancer Patients This example demonstrates that improved prediction can enable the identification of neoantigens from routine patient samples. To this end, archived FFPE tumor biopsies and 5–30 ml of peripheral blood were analyzed from 9 patients with metastatic NSCLC who had received anti-PD(L)1 therapy (Supplementary Table 1: Patient demographics and treatment information for the N=9 patients studied in Figures 26A–C. Key areas include tumor stage and subtype, anti-PD1 therapy received, and an overview of NGS results). Tumor whole-exome sequencing, tumor transcriptome sequencing, and matched normal exome sequencing yielded an average of 198 somatic mutations / patient (SNVs and short insertions or deletions), of which an average of 118 were expressed (Methods, Supplementary Table 1). A whole-MS model was used to prioritize 20 neoantigen determinants / patient for testing pre-existing anti-tumor T-cell responses. To focus the analysis on potential CD8 responses, a preferred peptide was synthesized as the 8-11mer minimal antigenic determinant (method), and peripheral blood mononuclear cells (PBMCs) were then cultured using the synthetic peptide in a short-lived in vitro stimulation (IVS) culture to expand neoantigen-reactive T cells (Supplementary Table 2). At two weeks later, the presence of antigen-specific T cells was assessed using IFN-γ ELISpot against the preferred neoantigen determinant. Separate experiments were also performed in 7 patients from whom adequate PBMCs were available to completely or partially deconvolve the identified specific antigen. The results are described in Figures 26A-C and 27A-30. Figure 26A illustrates the detection of T-cell responses in patient-specific neoantigen peptide pools from 9 patients. For each patient, predicted neoantigens were concatenated into two pools of 10 peptides, ordered according to the model and any sequence similarity (identical peptides were separated into different pools). Then, for each patient, in vitro expanded PBMCs were stimulated with two patient-specific neoantigen peptide pools in an IFN-γ ELISpot. The data in Figure 26A are represented as speckle-forming units (SFU) / 10⁵ plate-seeded cells, minus background (corresponding to the DMSO negative control). Background measurements (DMSO negative control) are shown in Figure 30. For patients 1-038-001, 1-050-001, 1-001-002, CU04, 1-024-001, 1-024-002, and CU05, single-well (patients 1-038-001, CU02, CU03, and 1-050-001) or replicates with mean and standard deviation (all other patients) were shown to respond to homologous peptide pools #1 and #2. For patients CU02 and CU03, cell number allowed for testing only of specific peptide pool #1. Samples with a value more than twice the background value were considered positive and designated with an asterisk (reactive donors included patients 1-038-001, CU04, 1-024-001, 1-024-002, and CU02). Non-reactive donors included patients 1-050-001, 1-001-002, CU05, and CU03. Figure 15C depicts an ELISpot well containing in vitro expanded PBMCs from IFN-γ ELISpots containing CU04 from patients stimulated with DMSO negative control, PHA positive control, CU04-specific neoantigen peptide pool #1, CU04-specific peptide 1, CU04-specific peptide 6, and CU04-specific peptide 8. Figures 27A and 27B illustrate the results of controlled experiments using patient neoantigens from HLA-matched healthy donors. These results demonstrate that in vitro culture conditions only amplify pre-existing in vivo-activated memory T cells, rather than enabling their reactivation in vitro. Figure 28 illustrates the detection of T-cell responses for each donor and for each in vitro expanded PHA-positive control described in Figure 26A. For each donor and for each in vitro expanded PHA-positive control in Figure 26A, in vitro expanded patient PBMCs were stimulated with PHA for maximum T-cell activation. Data in Figure 28 are expressed as speckle-forming units (SFU) / 10⁵ cells seeded on a plate, minus background (corresponding to DMSO negative controls). Single-well or bioreplicative responses are shown for patients 1-038-001, 1-050-001, 1-001-002, CU04, 1-024-001, 1-024-002, CU05, and CU03. No PHA-based testing was performed for patient CU02. Cells from patient CU02 were included in the analysis as a positive response to peptide pool #1, an indicator of active and functional T-cells (Figure 26A). As shown in Figure 26A, donors responsive to the peptide pool include patients 1-038-001, CU04, 1-024-001, and 1-024-002. Also as shown in Figure 26A, donors unresponsive to the peptide pool include patients 1-050-001, 1-001-002, CU05, and CU03. Figure 29A illustrates the detection of T-cell responses to individual patient-specific neoantigen peptides in pool #2 of patient CU04. Figure 29A also illustrates the detection of T-cell responses in a PHA-positive control of patient CU04. (This is positive control data, also shown in Figure 28.) For patient CU04, in vitro expanded PBMCs of this patient were stimulated in an IFN-γ ELISpot with individual patient-specific neoantigen peptides from pool #2 of patient CU04. In vitro expanded PBMCs of this patient were also stimulated in an IFN-γ ELISpot with PHA as a positive control. Data are expressed as spotted-forming units (SFU) / 10⁵ cells seeded on a plate, minus background (corresponding to the DMSO negative control). Figure 29B illustrates the detection of T-cell responses to individual patient-specific neoantigen peptides in three visits for patient CU04 and two visits for patient 1-024-002 (each visit occurred at different time points). For both patients, in vitro expanded PBMCs were stimulated with patient-specific individual neoantigen peptides in an IFN-γ ELISpot. For each patient, data from each visit are expressed as cumulative (additional) speckle-forming units (SFU) / 10⁵ plate-seeded cells, minus background (corresponding to the DMSO control). Data for patient CU04 are shown as cumulative SFUs from the three visits minus background. For patient CU04, SFUs minus background are shown for the first visit (T0) and subsequent visits at 2 months (T0 + 2 months) and 14 months (T0 + 14 months) after the first visit (T0). Data for patient 1-024-002 are shown as cumulative SFUs from the two visits minus background. For patient 1-024-002, SFU (surface unit count) minus background was observed at the initial visit (T0) and the subsequent visit 1 month after the initial visit (T0 + 1 month). Samples with a value more than twice the background level were considered positive and designated with an asterisk. Figure 29C illustrates the detection of T-cell responses to individual patient-specific neoantigen peptides and patient-specific neoantigen peptide pools in two visits to patient CU04 and two visits to patient 1-024-002 (each visit occurring at different time points). For both patients, in vitro expanded PBMCs were stimulated in an IFN-γ ELISpot with patient-specific individual neoantigen peptides and patient-specific neoantigen peptide pools. Specifically, for patient CU04, in vitro expanded PBMCs were stimulated in an IFN-γ ELISpot with CU04-specific individual neoantigen peptides 6 and 8, and with the CU04-specific neoantigen peptide pool; for patient 1-024-002, in vitro expanded PBMCs were stimulated in an IFN-γ ELISpot with 1-024-002-specific individual neoantigen peptide 16, and with the 1-024-002-specific neoantigen peptide pool. The data in Figure 29C are represented as speckle-forming units (SFU) / 10⁵ cells seeded on a plate, where background (corresponding to the DMSO control) is subtracted for replication of each technique with mean and range. Data for patient CU04 are shown as SFUs from two visits, subtracted for background. For patient CU04, SFUs from the first visit (T0; technique triplicate) and subsequent visits 2 months after the first visit (T0) (T0 + 2 months; technique triplicate) are shown. Data for patient 1-024-002 are shown as SFUs from two visits, subtracted for background. For patient 1-024-002, SFUs from the first visit (T0; technique triplicate) and subsequent visits 1 month after the first visit (T0) (T0 + 1 month; technique duplicate, except for samples stimulated with the patient 1-024-002-specific neoantigen peptide pool) are shown, subtracted for background. Figure 30 illustrates the detection of T-cell responses in two patient-specific neoantigen peptide pools and a DMSO negative control for the patient in Figure 26A. For each patient, in vitro expanded PBMCs were stimulated in an IFN-γ ELISpot with two patient-specific neoantigen peptide pools. For each donor and each in vitro expansion, in vitro expanded patient PBMCs were also stimulated in an IFN-γ ELISpot with DMSO as a negative control. The data in Figure 30 are represented as speckle-forming units (SFU) / 10⁵ cells seeded on a plate, including background (corresponding to the DMSO negative control) for the patient-specific neoantigen peptide pools and the corresponding DMSO control. For patients 1-038-001, 1-050-001, 1-001-002, CU04, 1-024-001, 1-024-002, and CU05, the response of a single well (1-038-001, CU02, CU03, and 1-050-001) or the average of the standard deviations of biological replication (all other samples) to homologous peptide pools #1 and #2 is shown. For patients CU02 and CU03, cell number allows for testing only of specific peptide pool #1. Samples with a value greater than 2 times the background level are considered positive and designated with an asterisk (reactive donors include patients 1-038-001, CU04, 1-024-001, 1-024-002, and CU02). Non-reactive donors include patients 1-050-001, 1-001-002, CU05, and CU03. As briefly discussed above regarding Figures 27A and 27B, to verify that in vitro culture conditions only amplify pre-existing in vivo-initiated memory T-cells, rather than enabling in vitro reactivation, a series of control experiments were conducted using neoantigens from HLA-matched healthy donors. The results of these experiments are described in Figures 27A and 27B and in Supplementary Table 4. Using IVS culture technology, the results of these experiments confirmed the absence of reactivation and detectable neoantigen-specific T-cell responses in healthy donors. In contrast, the use of IFN-γ ELISpot identified pre-existing neoantigen-responsive T cells in the majority (5 / 9, 56%) of patients tested using a patient-specific peptide pool (Figs. 26A and 29–30). Of the seven patients whose cell numbers permitted full or partial testing of individual neoantigen homologous peptides, four responded to at least one of the tested neoantigen peptides, and all of these patients had a corresponding pool response (Fig. 26B). The remaining three patients tested using individual neoantigens (patients 1-001-002, 1-050-001, and CU05) did not have a detectable response to a single peptide (data not shown), confirming a lack of visible response to the neoantigen pool in these patients (Fig. 26A). Of the four responding patients, samples from a single visit were available for two responding patients (patients 1-024-001 and 1-038-001), while samples from multiple visits were available for the other two responding patients (CU04 and 1-024-002). For the two patients with samples from multiple visits, accumulated (additional) speckled-forming units (SFUs) from three visits (patient CU04) or two visits (patient 1-024-002) are shown in Figure 26B and in Figure 29B by visit decomposition. Additional PBMC samples from the same visits were also available for patients 1-024-002 and CU04, and repeated IVS cultures and ELISpot tests confirmed a patient-specific neoantigen response (Figure 29C). In summary, in patients who identified at least one T-cell-recognized neoantigen determinant as shown by the response to the pool of 10 peptides in Figure 26A, the average number of identified neoantigen determinants was at least 2 per patient (at least 10 antigen determinants were identified in 5 patients, and the pool of identified peptides that could not be deconvolved as one identified peptide was counted). In addition to testing the IFN-γ response by ELISpot, culture supernatant was also tested for granzyme B by ELISA and for TNF-α, IL-2, and IL-5 by MSD cytokine multiplex assays. Cells from 4 / 5 patients with positive ELISpots secreted three or more analytes, including granzyme B (Supplementary Table 3), indicating the versatility of neoantigen-specific T-cells. Importantly, because combinatorial prediction and the IVS approach do not depend on a limited set of available MHC multimers, responses can be broadly tested across restricted HLA pairs. Furthermore, compared to longitudinal microgene screening for identifying identified mutations, this method directly identifies the smallest antigenic determinants and requires a separate deconvolution step for this purpose. In summary, the neoantigen identification yield is comparable to previous best methods 96, which utilize machine-sampled tests for all mutated TILs while screening only 20 synthetic peptides using conventional 5 to 30 mL whole blood. XV.A. peptide Custom-designed, recombinant lyophilized peptides were purchased from JPT Peptide Technologies (Berlin, Germany) or Genscript (Piscataway, NJ, USA) and reconstituted in sterile DMSO (VWR International, Pittsburgh, PA, USA) at 10 to 50 mM, aliquoted, and stored at -80°C. XV.B. Human peripheral blood mononuclear cells (PBMCs) HLA-typed PBMCs (confirmed negative for HIV, HCV, and HBV serological reactions) from healthy donors were purchased from Precision for Medicine (Gladstone, NJ, USA) or Cellular Technology, Ltd. (Cleveland, OH, USA) and stored in liquid nitrogen until use. Fresh blood samples were purchased from Research Blood Components (Boston, MA, USA). Leukopak and PBMCs from AllCell (Boston, MA, USA) were isolated using a Ficoll-Paque density gradient (GE Healthcare Bio, Marlborough, MA, USA) before refrigeration. Patient PBMCs were processed at the local clinical processing center according to local clinical standard operating procedures (SOPs) and approved IRBs. Approved IRBs were Quorum Review IRB, Comitato Etico Interaziendale AOU San Luigi Gonzaga di Orbassano, and Comité Ético de la Investigación del Grupo Hospitalario Quirón en Barcelona. In short, PBMCs were isolated by density gradient centrifugation, washed, counted, and cryopreserved at 5 x 10⁶ cells / ml in a CryoStor CS10 (STEMCELL Technologies, Vancouver, BC, V6A 1B6, Canada). Cryopreserved cells were shipped via cryoport and transferred to LN2 storage upon arrival. Patient demographics are listed in Supplementary Table 1. Cryopreserved cells were thawed and washed twice in OpTmizer T-cell expansion basal medium containing Benzonase (EMD Millipore, Billerica, MA, USA) (Gibco, Gaithersburg, MD, USA) and once in OpTmizer T-cell expansion basal medium without Benzonase. Cell counts and viability were assessed using the Guava ViaCount reagent and a module on a Guava easyCyte HT cell counter (EMD Millipore). The cells were then resuspended in a culture medium suitable for process assays at the specified concentration (see next section). XV.C. In Vitro Stimulation (IVS) Culture Pre-existing T-cells from healthy donors or patient samples were expanded in the presence of homologous peptides and IL-2 using a method similar to that applied by Ott et al. 81. Briefly, thawed PBMCs were allowed to stand overnight and stimulated for 14 days in 24-well tissue culture plates in ImmunoCult™-XF T-cell expansion medium (STEMCELL Technologies) containing 10 IU / ml rhIL-2 (R&D Systems Inc., Minneapolis, MN) in the presence of peptide pools (10 µM / peptide, 10 peptides / pool). Cells were seeded at 2 x 10⁶ cells / well and fed by replacing 2 / 3 of the medium every 2 to 3 days. One patient sample showed a deviation from the protocol and should be considered a potential false negative: after thawing, patient CU03 did not produce an adequate number of cells, and cells were seeded at 2 x 10⁵ cells / peptide pool (10 times less than in each protocol). XV.D.IFNg enzyme-linked immunospot (ELISpot) assay The detection of IFNg-producing T cells was performed using the ELISpot assay 142. In short, PBMCs (either in vitro or after in vivo expansion) were harvested, washed in serum-free RPMI (VWR International), and cultured in OpTmizer T-cell expansion basal medium (in vitro) or ImmunoCult™-XF T-cell expansion medium (expanded culture) coated with an anti-human IFNg capture antibody (Mabtech, Cincinatti, OH, USA) in the presence of a control or homologous peptide in an ELISpot Multiscreen plate (EMD Millipore). After 18 hours of incubation in a humidified incubator at 37°C and 5% CO2, the cells were removed from the plate and the membrane-bound IFNg was detected using an anti-human IFNg detection antibody (Mabtech), a vesartan-avidin-peroxidase complex (Vector Labs, Burlingame, CA, USA), and an AEC receptor (BD Biosciences, San Jose, CA, USA). ELISpot plates may be dried, stored protected from light, and transported to Zellnet Consulting, Inc., Fort Lee, NJ, USA for standardized evaluation.143 Data are expressed as spot-forming units (SFU) / number of cells seeded onto the plate. XV.E. Granulase B ELISA and MSD multiplex assays. The secreted IL-2, IL-5, and TNF-α in ELISpot supernatant were detected using a triple assay MSD U-PLEX Biomarker assay (catalog number K15067L-2). Assays were performed according to the manufacturer's instructions. Analyte concentrations (pg / ml) were calculated using serial dilutions of known standards for each cytokine. For graphical representation, values ​​below the minimum range of the standard curve were represented as zero. Granulase B in ELISpot supernatant was detected using a Granulase B DuoSet® ELISA (R & D Systems, Minneapolis, MN) according to the manufacturer's instructions. In short, the ELISpot supernatant was diluted 1:4 in sample diluent and run simultaneously with serially diluted standards of granase B to calculate concentrations (pg / ml). For graphical representation, values ​​below the minimum range of the standard curve were represented as zero. Negative control assay for XV.F. IVS – Neoantigen from tumor cell lines tested from healthy donors. Figure 27A illustrates a negative control assay for IVS testing of neoantigens from tumor cell lines tested from healthy donors. Healthy donor PBMCs were stimulated in IVS cultures using peptide pools containing positive control peptides (previously exposed to infectious diseases), HLA-matched neoantigens derived from tumor cell lines (unexposed), and peptides derived from pathogens to which the donors were serologically negative. Following stimulation with DMSO (negative control, black circle), PHA and common infectious disease peptides (positive control, red circle), neoantigens (unexposed, light blue circle), or HIV and HCV peptides (confirming negative serological reaction in the donor, navy blue, A and B), the expanded cells were subsequently analyzed by IFNγ ELISpot (10⁵ cells / well). Data are shown as speckled units (SFU) / 10⁵ seeded cells. Biological replication is shown with mean and SEM. No response to the neoantigens or to the pathogen-derived peptides was observed in the donors who had not been exposed to these pathogens (serologically negative). Negative control test for XV.G.IVS assay – neoantigen from patients tested on healthy donors. Figure 27A illustrates a negative control experiment of IVS assay for patient neoantigens tested for responsiveness in healthy donors. T-cell responses in healthy donors to a pool of HLA-matched neoantigen peptides were evaluated. Left panel: Healthy donor PBMCs were stimulated in vitro in an IFN-γ ELISpot with controls (DMSO, CEF, and PHA) or HLA-matched patient-derived neoantigen peptides. Data are expressed as speckle-forming units (SFU) / 2 x 10⁵ cells per ternary well plate. Right panel: Healthy donor PBMCs expanded by IVS culture in the presence of either a neoantigen pool or a CEF pool were stimulated in an IFN-γ ELISpot with controls (DMSO, CEF, and PHA) or HLA-matched patient-derived neoantigen peptides. Data are expressed as SFU / 1 x 10⁵ cells per ternary well plate. No response to the neoantigens in healthy donors was observed. XV.H. Supplementary Table 2: Peptides Tested for T-cell Recognition in NSCLC Patients Figures 26A to C detail the neoantigen peptides tested in a study of N=9 patients (identification of neoantigen-reactive T cells from NSCLC patients). Key regions include the source mutation, peptide sequence and pool, and observed individual peptide responses. The "most likely limitation" line indicates which paired genes, as predicted by the model, are most likely to present each peptide. It also includes the ordering of these peptides among all mutated peptides in each patient, as calculated using binding affinity prediction (method). There are four peptides that rank highly by the full MS model and are recognized by CD8 T cells, which have low predicted binding affinity or low ranking predicted by binding affinity. For three of these peptides, this discrepancy arises from the difference in HLA coverage between the model and MHCflurry 1.2.0. The predicted peptide YEHEDVKEA is presented via HLA-B*49:01, which is not covered by MHCflurry 1.2.0. Similarly, the predicted peptides SSAAAPFPL and FVSTSDIKSM are presented via HLA-C*03:04, which is also not covered by MHCflurry 1.2.0. The online NetMHCpan 4.0 (BA) predictor (a pan-specific binding affinity predictor that, in principle, covers all paired genes) listed SSAAAPFPL as a strong binder to HLA-C*03:04 (23.2 nM, ranked 2nd for patient 1-024-002), predicted FVSTSDIKSM as a weak binder to HLA-C*03:04 (943.4 nM, ranked 39th for patient 1-024-002) and YEHEDVKEA as a weak binder to HLA-B*49:01 (3387.8 nM), but predicted a stronger binder to HLA-B*41:01 (208.9 nM, ranked 11th for patient 1-038-001), which is also present in this patient but not covered by the model. Therefore, of these three peptides, FVSTSDIKSM will be predicted to be missed by binding affinity, SSAAAPFPL will be captured, and the HLA restriction of YEHEDVKEA is uncertain. The remaining five peptides, deconvolved from peptide-specific T-cell responses, were derived from patients, including those whose most likely presenting pairs, as determined by the model, were also covered by MHCflurry 1.2.0. Of these five peptides, 4 / 5 had predicted binding affinities stronger than the standard 500 nM threshold and ranked in the top 20, although they had slightly lower rankings than those from the model (peptides DENITTIQF, QDVSVQVER, EVADAATLTM, and DTVEYPYTSF ranked 0, 4, 5, and 7 respectively from the model compared to 2, 14, 7, and 9 from MHCflurry). Peptide GTKKDVDVLK was recognized by CD8 T-cells and ranked 1st from the model, but had a ranking of 70th from MHCflurry and a predicted binding affinity of 2169 nM. In summary, 6 / 8 of the individually identified peptides that rank highly using the full MS model also exhibit high ranking using binding affinity prediction and have predicted binding affinity <500 nM, while 2 / 8 of the individually identified peptides will be missed if binding affinity prediction is used instead of the full MS model. XV.I. Supplementary Table 3: MSD Cytokine Multiplex Assay and ELISA Assay for ELISA Supernatant of NSCLC Neoantigen Peptides For granzyme B (ELISA), TNFα, IL-2, and IL-5 (MSD), the analytes detected in the supernatant of a positive ELISA plate (IFNγ) are shown. Values ​​are presented as average pg / ml from technical replication. Positive values ​​are shown in italics. Granzyme B ELISA: A value ≥1.5 times the DMSO background is considered positive. U-Plex MSD test: A value ≥1.5 times the DMSO background is considered positive. XV.J. Supplementary Table 4: Neoantigens and Infectious Disease Antigen Determinants in IVS Control Experiments Details of the neoantigens and viral peptides of the tumor cell lines tested in the IVS control assay are shown in Figures 27A-B. Key areas include the source cell line or virus, peptide sequences, and predicted HLA-pairing genes. (XV.K. Data) The MS peptide dataset used for training and testing the predictive model (Figures 25A-B) is available in the Massive Archive (massive.ucsd.edu), accession number MSV000082648. Neoantigen peptides tested by ELISpot (Figures 26A-C and 27A-B) are included in the manuscript (Supplementary Tables 2 and 4). XVI. Methods of Examples 8-11 XVI.A. Mass Spectrometry XVI.A.1. Samples Archived frozen tissue samples used for mass spectrometry analysis were obtained from commercial sources, including BioServe (Beltsville, MD), ProteoGenex (Culver City, CA), iSpecimen (Lexington, MA), and Indivumed (Hamburg, Germany). A subset of samples was also prospectively collected from patients under a research agreement approved by the Comité de Protection des Personnes, Ile-de-France VII, at Hospital Marie Lannelongue (Le Plessis-Robinson, France). XVI.A.2. HLA Immunoprecipitation HLA-peptide molecules were isolated using an established immunoprecipitation (IP) method after lysis and lysis of tissue samples 87, 124-126. Fresh frozen tissue was pulverized (CryoPrep; Covaris, Woburn, MA), and lysis buffer (1% CHAPS, 20 mM Tris-HCl, 150 mM NaCl, protease and phosphatase inhibitors, pH=8) was added to lyse the tissue, and the resulting solution was centrifuged at 4°C for 2 hours to form spherical fragments. The clarified lysate was used for HLA-specific IP. Immunoprecipitation was performed using antibody W6 / 32 as previously described 127. The lysate was added to antibody beads and rotated overnight at 4°C for immunoprecipitation. After immunoprecipitation, the beads were removed from the lysate. The IP beads were washed to remove nonspecific binding and the HLA / peptide complex was dissociated from the beads with 2N acetic acid. Protein fractions were removed from the peptides using a molecular weight spin column. The resulting peptides were evaporated to dryness using SpeedVac and stored at -20°C before MS analysis. XVI.A.3 Peptide Sequencing The dried peptide was reconstituted in HPLC buffer A and loaded onto a C-18 microcapillary HPLC column for gradient dissociation in the mass spectrometer. The peptide was dissociated in a Fusion Lumos mass spectrometer (Thermo) over 180 minutes using a gradient of 0 to 40% B (solvent A – 0.1% formic acid, solvent B – 0.1% formic acid / 80% acetonitrile). MS1 spectra of the peptide mass / charge (m / z) were collected in an Orbitrap detector with a resolution of 120,000. Following HCD fragmentation of selected ions, 20 low-resolution MS2 scans were collected in an Orbitrap or ion trap detector. MS2 ion selection was performed using a data-dependent acquisition mode and dynamic exclusion 30 seconds after MS2 ion selection. Automatic gain control (AGC) for MS1 scans was set to 4 x 10⁵ and for MS2 scans to 1 x 10⁴. For sequencing HLA peptides, +1, +2, and +3 charge states can be selected for MS2 fragmentation. The protein database was searched using Comet 128 and 129 for MS2 spectra ...

Claims

1. A method for identifying at least one neoantigen from one or more tumor cells of a subject, the at least one neoantigen being presented by one or more MHC pair genes on the surface of the tumor cells, the method comprising the steps of: obtaining at least one of exome, transcriptome, or whole genome nucleotide sequence data, wherein the nucleotide sequence data is used to obtain peptide sequences representing each of the neoantigen set; obtaining peptide sequences representing each of the one or more MHC pair genes of the subject; applying a pan-allel machine learning presentation model to the peptide sequences representing each of the neoantigen set and to the peptide sequences representing each of the one or more MHC pair genes to determine a set of presentation probabilities for the neoantigen set, each presentation probability in the set representing a corresponding neoantigen. The probability of antigen presentation via one or more MHC pair genes on the surface of tumor cells of the subject, the pan-pair gene machine learning presentation model includes: at least a plurality of parameters identified based on a training dataset, the training dataset including: for each of the plurality of samples, a marker obtained by mass spectrometry, the mass spectrometry measuring the presence of a peptide bound to at least one MHC pair gene identified as present in the MHC pair gene set in the sample; for each of the samples, data on training peptide sequences representing neoantigens, the neoantigens comprising 6 to 30 amino acids; and for each of the samples, data on training peptide sequences representing at least one MHC pair gene; selecting a subset of the neoantigen set based on the presentation probability set to generate a selected neoantigen set; and returning the selected neoantigen set.

2. The method of claim 1, wherein applying the pan-dual gene machine learning presentation model comprises: generating a dependency score for each of the one or more MHC dual genes, the score indicating whether the MHC dual gene will present the neoantigen based on specific amino acids at specific positions of the peptide sequences.

3. The method of claim 2, wherein applying the pan-dual gene machine learning presentation model further comprises: a) transforming the dependency scores to generate corresponding per-dual gene probabilities for each MHC dual gene, the per-dual gene probabilities indicating the probability that the corresponding MHC dual gene will present the corresponding neoantigen; and combining the per-dual gene probabilities to generate the presentation probability of the neoantigen; or b) transforming the combination of the dependency scores to generate the presentation probability, wherein the combination of transforming the dependency scores models the presentation of the neoantigen as interference between the one or more MHC dual genes.

4. The method of claim 2 or 3, wherein the presentation probability set is further identified by at least one or more dual gene non-interaction features, and further comprises: applying the pan-dual gene machine learning presentation model to the dual gene non-interaction features to generate a dependency score of the dual gene non-interaction features, the score indicating whether the peptide sequence of the corresponding neoantigen will be presented based on the dual gene non-interaction features.

5. The method of claim 4, further comprising: a) combining the dependency score of each MHC pair of the one or more MHC pairs with the dependency score of the non-interacting features of the pairs; converting the dependency scores of the combinations of the MHC pairs to generate a pairwise probability of each MHC pair, the pairwise probability indicating the probability that the corresponding MHC pair will present the corresponding neoantigen; and combining the pairwise probabilities to generate the presentation probability; or b) combining the dependency scores of each MHC pair with the dependency score of the non-interacting features of the pairs; and converting the dependency scores of the combinations to generate the presentation probability.

6. The method of any one of requests 1 to 3, wherein: a) The one or more MHC pairs comprise two or more distinct MHC pairs; b) The peptide sequences comprise a peptide sequence having a length other than 9 amino acids; c) The encoding peptide sequence comprises a peptide sequence encoded using a one-hot encoding scheme; d) The plurality of samples comprise at least one of the following: (i) cell lines engineered to express one or more single MHC pairs; (ii) cell lines engineered to express one or more of the plurality of MHC pairs; (iii) obtained from or derived from one or more human cell lines from a plurality of patients; (iv) obtained from recombinant human cells. (v) Fresh or frozen tumor samples from several patients; and (e) fresh or frozen tissue samples from several patients; the training dataset further includes at least one of the following: (i) data associated with peptide-MHC binding affinity measurements of at least one of the peptides; and (ii) data associated with peptide-MHC binding stability measurements of at least one of the peptides; and / or (f) the presentation probability set is further identified by at least the expression level of the one or more MHC pair genes in the subject, such expression level as measured by RNA-seq or mass spectrometry.

7. The method of any one of requests 1 to 3, wherein: a) The presentation probability set is further identified by features including at least one of the following: (i) the predicted affinity between the neoantigen in the neoantigen set and the one or more MHC pair genes; and (ii) the predicted stability of the peptide-MHC complex encoded by the neoantigen; and / or b) The numerical probability set is further identified by features including at least one of the following: (i) its source protein sequence is internally attached to the C-terminal sequence of the peptide sequence encoded by the neoantigen; and (ii) its source protein sequence is internally attached to the N-terminal sequence of the peptide sequence encoded by the neoantigen.

8. The method of any one of claims 1 to 3, wherein selecting the selected neoantigen set comprises: a) selecting neoantigens with an increased likelihood of presentation on the surface of tumor cells relative to unselected neoantigens based on the pan-dual gene machine learning presentation model; b) selecting neoantigens with an increased likelihood of inducing a tumor-specific immune response in the subject relative to unselected neoantigens based on the pan-dual gene machine learning presentation model; c) selecting neoantigens with an increased likelihood of presentation to naive T cells via specialized antigen-presenting cells (APCs) relative to unselected neoantigens based on the presentation model; d) selecting neoantigens with a reduced likelihood of being suppressed by central or peripheral tolerance relative to unselected neoantigens based on the pan-dual gene machine learning presentation model; and / or e) selecting neoantigens with a reduced likelihood of inducing an autoimmune response in the subject's normal tissues based on the pan-dual gene machine learning presentation model relative to unselected neoantigens.

9. The method of any one of claims 1 to 3, wherein the one or more tumor cell lines 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.

10. The method of any one of claims 1 to 3, further comprising generating an output for constructing a personalized cancer vaccine from the selected neoantigen set.

11. The method of any one of claims 1 to 3, wherein the pan-dual gene machine learning presentation model is a neural network model.

12. The method of claim 11, wherein the neural network model comprises a single neural network model containing a series of nodes arranged in one or more layers, the single neural network model being configured to receive numerical vectors encoding peptide sequences of multiple different MHC pair genes.

13. The method of request item 11, wherein the generalized dual gene machine learning presentation model is a deep learning model containing one or more node layers.

14. The method of any one of requests 1 to 3, wherein: a) The training peptide sequences derived from at least one MHC pair gene or the data do not contain peptide sequences of the subject's MHC pair gene, and the peptide sequences are input into the pan-pair gene machine learning presentation model to generate the presentation probability set of the neoantigen set; b) The at least one MHC pair gene of the peptides in each of the plurality of samples bound to the training data set belongs to the gene family to which the subject's one or more MHC pair genes belong; c) The at least one MHC pair gene of the peptides in each of the plurality of samples bound to the training data set contains one or more MHC pair genes; and / or d) The one or more MHC pair genes are class I MHC pair genes.

15. A computer system comprising: a computer processor; and memory storing computer program instructions that, when executed by the computer processor, cause the computer processor to perform the following: obtaining at least one of exome, transcriptome, or whole genome nucleotide sequence data, wherein the nucleotide sequence data is used to obtain peptide sequences representing each of a neoantigen set; obtaining peptide sequences representing each of one or more MHC pair genes of the subject; applying a pan-pair gene machine learning presentation model to the peptide sequences representing each of the neoantigen set and to the peptide sequences representing each of the one or more MHC pair genes to determine a set of presentation probabilities for the neoantigen set, each presentation probability in the set representing the expression of a corresponding neoantigen by the tumor cells of the subject. The pan-dual gene machine learning presentation model, which considers the likelihood of presentation of one or more MHC dual genes on a surface, comprises: at least a plurality of parameters identified based on a training dataset, the training dataset comprising: for each of a plurality of samples, a marker obtained by mass spectrometry, the mass spectrometry measuring the presence of a peptide bound to at least one MHC dual gene in a set of MHC dual genes identified as being presented in the sample; for each of the samples, data on training peptide sequences representing neoantigens, the neoantigens comprising 6 to 30 amino acids; and for each of the samples, data on training peptide sequences representing at least one MHC dual gene; selecting a subset of the neoantigen set based on the presentation likelihood set to generate a selected neoantigen set; and returning the selected neoantigen set.

Images