A method for recognizing MHC class-I molecules neoantigens based on multiple-instance learning

Abstract

The present application belongs to the technical field of biological information calculation, and particularly relates to a MHC-I class molecule neoantigen recognition method based on multiple example learning. The present application comprises the following steps: designing an interactive convolution network structure, extracting high-dimensional features of an input peptide and each MHC-I molecule, learning and assigning weights to each high-dimensional representation by using an attention mechanism, obtaining fusion features of the input peptide and multiple MHC-I molecules by multiple example learning, and then performing a binary classification task for binding prediction; after training of multi-gene data, network knowledge is transferred to a downstream neoantigen recognition task by using a multi-task learning technology; paired neoantigen-wild type sequence data recorded in a database is used to construct an auxiliary data generation process, including using a ranking learning loss function and a binary classification cross-entropy loss function, so as to improve the absolute size of a model prediction score for a neoantigen. A large number of experiments on large-scale open source experimental data verify that the present application has superior performance.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics computing technology, specifically relating to a method for identifying neoantigens that bind to MHC-I class molecules. Background Technology

[0002] Neoantigens are important tumor-specific T-cell immune biomarkers because they are expressed only by mutated genes in human tumor cells and do not exist in normal tissue cells. Therefore, identifying neoantigens has attracted widespread attention in the field of T-cell immune bioinformatics computation. However, the highly random nature of mutations makes accurate neoantigen identification extremely challenging. In recent years, various deep learning techniques, especially Multiple Instance Learning (MIL), have been developed. [1] Introduced by this approach, models trained on large-scale in vivo multigene data have made progress in predicting whether a given short peptide sequence is an immunogenic neoantigen. MIL-based discrimination methods are all based on a biological background: multiple MHC-I molecules are co-expressed within human cells during the T-cell immune process and bind to antigen sequences. This allows for the prediction of neoantigens by learning the binding specificity of multiple MHC-I molecules through multigene data. However, these methods typically only consider the individual binding specificity of MHC-I molecules independently, neglecting their internal relationships.

[0003] Based on current literature, algorithms for identifying neoantigens can be categorized into three types according to their data sources: (1) based on in vitro binding strength data; (2) based on in vivo single-gene data; and (3) based on in vivo multi-gene data. In vitro binding strength data typically assesses the binding strength of a single MHC-I molecule to a short peptide using half-inhibitor concentrations. This data does not reflect relevant in vivo information and differs significantly from the distribution of neoantigen data. Therefore, methods based on this type of data are not suitable for this task. [2–4] In vivo single-gene data are generally obtained through specific experimental designs and do not conform to in vivo biological processes. They are usually used to assist models in learning the specificity of MHC-I molecule binding sequences in vivo. [6,5] In vivo multi-gene data best matches the in vivo biological process of neoantigen presentation by MHC-I molecules, and is currently the preferred approach for this task. [7] .

[0004] The rapid development of deep learning technology and the upgrading of liquid chromatography-mass spectrometry technology have driven tasks such as sequence-level binding prediction and antigen recognition. Deep learning-based models have achieved remarkable results in neoantigen discrimination. Designing and applying deep learning-based neoantigen discrimination models involves two key aspects: large-scale, high-quality data and network architecture; the former determines the model's learning ability for the relevant task. (Related literature...) [8]This indicates that, after applying collected clinical neoantigen data from patients, a simple artificial neural network achieved optimal performance on the specific task of predicting whether a mutated sequence within a patient's tumor is a neoantigen. However, the scale of this neoantigen data is relatively small, insufficient to support the model's generalization ability. Recent related literature shows that models trained on millions of somatic polygenic data obtained through mass spectrometry from within the human body exhibit better generalization ability. [7,9] They use a max pooling method based on multi-instance learning.

[10] A simplified strategy is employed: multiple MHC-I molecules are replaced with the MHC-I molecule with the highest predicted probability, simplifying multi-gene data into single-gene data for learning. However, these methods suffer from two main problems: First, they are often influenced by model bias, with different distributions of predicted scores for different molecules leading to erroneous preference distributions when comparing scores, thus reducing model accuracy. Furthermore, their simple network architecture limits their ability to uncover and learn the internal relationships within multi-gene data. Second, these methods do not effectively transfer learning to neoantigen prediction tasks, resulting in excessively high false positive rates that prevent them from effectively identifying neoantigens.

[0005] Unlike commonly used multi-gene data simplification strategies, this invention proposes a network architecture based on multi-instance learning. This architecture avoids model bias caused by direct score comparisons, and achieves more accurate multi-gene representations through fusion at each pMHC complex representation vector level. Simultaneously, it captures the inherent relationships within multi-gene data to learn the expected preference distribution under multi-gene conditions, thereby improving the prediction accuracy of multi-gene data. In addition to the network architecture design, this invention proposes a multi-task learning approach...

[11] The fine-tuning strategy incorporates knowledge of immunogenic neoantigens into the model, which can better distinguish between mutated neoantigen sequences and non-immunogenic wild-type sequences, thus enhancing the ability to distinguish neoantigens. Summary of the Invention

[0006] The purpose of this invention is to propose a method for identifying MHC-I class molecular neoantigens that can effectively determine the neoantigen sequence.

[0007] The method for recognizing MHC-I class molecular neoantigens provided by this invention is based on multi-instance learning technology. [1] Simultaneously, a novel interactive convolutional network structure for sequence fusion is designed. [12,13]The high-dimensional features of the input peptide and each MHC-I molecule are extracted. The attention mechanism is used to learn and assign weights to each high-dimensional representation. The fusion features of the input peptide and multiple MHC-I molecules are obtained through the permutation-invariant aggregation function in the multi-instance learning technique. Then, a binary classification task of binding prediction is performed. The advantages of this network architecture lie in its richer fusion representation information, avoiding self-bias caused by differences in the predicted score distributions of different MHC-I molecules, while not ignoring the internal relationships within multiple genes. Furthermore, it improves efficiency and accelerates training within a gradient-based update framework, allowing the model to converge to a better state. In addition to the attention-based multi-instance learning network architecture, after training on millions of multi-gene datasets, this invention transfers network knowledge to the downstream neoantigen recognition task using multi-task learning techniques: Given the background knowledge that neoantigens originate from wild-type sequence mutations with only a very small number of mutations possessing immunogenicity, an auxiliary data generation process is constructed using paired neoantigen-wild-type sequence data recorded in the database. First, a ranking learning loss function is used for Task 1, aiming to ensure the model's predicted score for neoantigens is higher than that for other random mutations. Then, a binary classification cross-entropy loss function is used for Task 2, classifying neoantigens as positive samples while maintaining the original predicted score for wild-type sequences, thus improving the absolute magnitude of the model's predicted score for neoantigens. This invention has demonstrated its superior performance through extensive experiments on large-scale open-source experimental data.

[0008] The method for predicting and identifying neoantigens based on MHC-I molecule binding, provided by this invention, comprises the following steps:

[0009] (1) First, to address the issue of thousands of MHC-I molecules in the human body, each with its own binding specificity, and to address the issue of the binding strength of short peptide sequences of different lengths being affected by the misalignment of binding residues with MHC-I molecules, a general neural network architecture, such as an artificial neural network (ANN), is used. [3] Convolutional Neural Network (CNN)

[14] Transformer [5] The ability to specifically bind to learning characteristics is relatively weak. This invention employs a novel network architecture to improve representation capabilities.

[0010] For an input pair of short peptide sequences P with L amino acids and an MHC-I molecule sequence Q, this invention uses a pseudo-sequence of the MHC-I molecule for substitution; this pseudo-sequence consists of 34 important sites on the MHC-I molecule when binding to the peptide and has been used in various methods. [7,6]This invention employs an embedding layer to map each amino acid in the sequence to a corresponding learnable vector, thereby obtaining the embedding matrix X∈R of the short peptide. L×d The embedding matrix Y∈R with MHC-I molecules 34×d d represents the dimension of the amino acid embedding vector.

[0011] For each amino acid representation vector in the short peptide embedding matrix, an independent gated linear unit (GLU) is used.

[15] Expression control is employed to differentiate binding residues in short peptide sequences, enabling adaptive identification of anchor sites:

[0012] GLU i (x i )=σ(W i x i +b i )⊙(V i x i +c i (1)

[0013] Among them, W i ,b i V i ,c i Here, σ(·) is a learnable parameter, ⊙ is the Sigmoid activation function, and ⊙ is the element-wise Hadmar product. The short peptide embedding matrix is ​​updated after passing through L GLUs. As input to subsequent convolutional layers. Unlike traditional convolutional layers, this invention utilizes a learnable parameter matrix U. i ∈R k×34 Using the embedding matrix Y of MHC-I molecules to generate convolution kernels Then, the short peptide embedding matrix was processed. Perform convolutional computations to construct the interaction information between the two:

[0014]

[0015] In the above formula, f(·) is the ReLU activation function. Here, represents a learnable bias term, and * denotes the convolution calculation symbol. To simplify the notation, the above calculation process is defined as follows: Where θ i These are the parameters in the network. This invention employs a two-layer BICL to implement a computational paradigm for residual connections:

[0016]

[0017] Next, this invention employs a multilayer perceptron network (MLP) to process the obtained interaction features. Further updates are performed, and finally, a max-pooling layer (MaxPool) is used to obtain the final representation vector:

[0018]

[0019] In summary, this invention first utilizes two layers of BICL, employing the original short peptide embedding matrix X and the controlled embedding matrix... The interaction features between short peptides and MHC-I molecules were constructed. The multilayer perceptron network is then used to further update it, and finally the max pooling layer is used to obtain the final representation vector h of a pair of short peptides and MHC-I molecules, which is used as the input for the subsequent MIL framework.

[0020] (2) In order to learn the intrinsic relationships of multi-gene data while avoiding bias caused by direct score comparison, this invention uses a multi-instance learning framework based on attention mechanism to replace the traditional simplified strategy of MIL-based max pooling; the computational framework is described below:

[0021] In multigene datasets, a sample consists of a short peptide and multiple, and an indeterminate number, MHC-I molecules. The data inherently possesses the following properties: a sample is considered positive when at least one MHC-I molecule binds to the short peptide, and negative when none of the MHC-I molecules bind to it. Traditionally, for the fuzzy uncertainty in positive samples, a hard processing approach is used, replacing the MHC-I molecule with the highest predicted score. This leads to direct information loss, and the training model's varying favoritism towards different molecules introduces bias. This invention adjusts the computational framework, employing a soft processing approach using attention weight allocation. This avoids losing information from the samples and simultaneously learns the binding preference patterns of different MHC-I molecule combinations in multigene datasets, effectively improving discrimination accuracy.

[0022] Specifically, this invention uses the network architecture described above to connect a short peptide P with multiple MHC-I molecules Q1, Q2, ... Q S The representation vectors h1, h2, ... h1, h2, ... h2, are calculated separately for each pair. S This invention uses the following formula to calculate the gated attention weights:

[0023]

[0024] Where w, V, and U are the learnable parameters, and exp(·) is the exponential function. Based on the above weights, the final comprehensive representation vector can be calculated:

[0025]

[0026] The present invention then employs a linear layer as a classifier for binary prediction:

[0027]

[0028] This part of the network is trained using a binary classification cross-entropy loss function:

[0029]

[0030] (3) The network architecture first proposed in this invention is applicable to multi-gene data to predict whether MHC presents peptides.

[0031] However, a large number of short peptides are presented by MHC-I molecules on the surface of cell membranes in the human body, but only a very small number are specifically recognized and trigger an immune response. Therefore, training this network framework solely with millions of multi-gene data from a database is not entirely suitable for the discrimination and prediction of neoantigens, and, consistent with traditional methods, easily leads to an excessively high false positive rate. Therefore, this invention uses paired neoantigen-wild-type sequence data, combined with multi-task learning for fine-tuning to enhance the ability to discriminate neoantigens.

[0032] Neoantigens are generated by tumor gene mutations. At the amino acid level, when these neoantigens are compared to the unmutated wild-type sequence, they generally show only a single amino acid substitution. Despite the extremely high sequence similarity, due to the influence of immune tolerance mechanisms, neoantigens exhibit a significantly increased affinity compared to wild-type sequences. Furthermore, although numerous mutations exist within cells, only a very small number will elicit a T-cell immune response.

[16] Therefore, an ideal model should predict neoantigens that elicit an immune response at a higher score than other random mutations. To enhance the ability to distinguish neoantigens, this method employs multi-task learning with different loss functions to fine-tune the model parameters, as detailed below:

[0033] This method employs a two-stage learning strategy of training and fine-tuning to determine the final model parameters.

[0034] During the training phase, this method performs supervised training on millions of washed data, including single-gene and multi-gene data, and adopts the standard binary cross-entropy loss function to obtain a model with presentation prediction capabilities.

[0035] The following describes the data construction and loss function calculation for Task 1. For the paired wild-type sequences and neoantigen sequences in the database, this method first iterates through the single amino acid mutation sequences of each wild-type sequence, and scores and sorts them using the model trained in the first step. For the single mutation set of each wild-type sequence, this method samples 10 single mutation sequences from the 10%, 30%, and 70% ranking positions to form three candidate pools: a strong binding pool and a strong binding pool. Weak binding pool Unbound pool For each neoantigen sequence S neo During training, a list x = (S) is formed by randomly selecting one mutation sequence from each of the three corresponding candidate pools. neo S S S W S N The list is expected to be ordered as y = (1, 2, 3, 4), meaning the neoantigen prediction score is greater than strongly bound mutation sequences, which are greater than weakly bound mutation sequences, which are greater than unbound mutation sequences. This method employs the previously proposed classic ranking learning loss function, ListMLE.

[11] The definition is as follows:

[0036] L1 ＝ -log(P(y|x;f w (9)

[0037] Among them, f w P(y|x;f) represents the model obtained from the first stage of training. w ) is defined as:

[0038]

[0039] Where, x y (j) represents the mutation sequence at position j in the true sequence.

[0040] Next, we will describe Task 2. Furthermore, this method also utilizes the labeling information of neoantigens for supervised classification. Specifically, for neoantigen S... neo The wild-type sequence S was labeled as a positive sample. wt The expected model should maintain its original prediction score, and the predicted value y obtained from the first stage should be used. wt As a self-distilled soft label. The specific loss function is as follows:

[0041] L2=H(1,f w (S neo ))+H(y wt ,fw(S wt (11)

[0042] Where H(·,·) represents the cross-entropy loss function. The loss function for the entire fine-tuning process is:

[0043] L=αL1+βL2, (12)

[0044] Where α and β are weighting coefficients. Typically, α = 0.5 and β = 1.0.

[0045] This invention can obtain high-quality multi-gene representations, more accurately capture the preference distribution of different pairs of molecules under multi-gene data, improve the accuracy of prediction under multi-gene data, and make targeted fine-tuning for the mutation characteristics of neoantigens, ultimately improving the accuracy of neoantigen discrimination.

[0046] This invention has demonstrated its superior performance both qualitatively and quantitatively through extensive experiments using real data. Attached Figure Description

[0047] Figure 1 This forms the overall framework of the model in this invention.

[0048] Figure 2 This describes the specific structure of the encoder of the present invention.

[0049] Figure 3 This describes the specific structure of the multi-example learning module of the present invention.

[0050] Figure 4 This is a visual comparison of the present invention and other methods for multi-gene binding motifs. Detailed Implementation

[0051] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0052] The real data used in this invention comes from open-source data in several classic works in the field, including data from the IEDB database.

[17] The cleaned and organized open-source dataset was used as a five-way cross-validation dataset (named EL2020). [7] This dataset is also used for parameter training, and the present invention also employs well-known methods in the field.

[18] The training set was used as an extension and merged with EL2020 to form the final training set. For testing and evaluation, this invention collected open-source test sets in the field, specifically including: a single-gene independent test set (named IS2020). [7]

[19] Compared with the multi-gene independent test set (named IM2020)

[18] This invention also utilizes the Neodb database.

[20] A fine-tuned dataset on neoantigens was cleaned and organized. Furthermore, to further verify the ability to discriminate neoantigens, this invention utilizes relevant articles...

[19] The compiled independent test set (named NE2023) was validated and evaluated, and clinical patient neoantigen data were collected as an independent test set (named PE2018). [8]The EL2020 dataset contains 4,032,815 single-gene data points covering 138 different MHC-I molecules and 8,841,551 multi-gene data points covering 114 different combinations. This data is used for method training and validation. The IS2020 dataset contains 946,001 single-gene data points covering 36 different human MHC-I molecules; this data is used for testing and evaluation. The IM2020 dataset covers 20 multi-gene combinations not present in the training set, totaling 2,700,700 data points, and is used for testing and evaluation. The NE2023 dataset covers 32 different MHC-I molecules, with a total of 6312 single-gene data. The positive sample data are neoantigens that have been verified to be immunogenic, while the negative sample data are sequences with high affinity but no immunogenicity. This high-quality independent test set can effectively evaluate the discriminative ability of the model. The PE2018 dataset contains 2023 mutation sequences caused by single nucleotide variants from 17 patients, of which only 26 are neoantigens that induce an immune response. This dataset is real clinical data and is used to test and evaluate the ability of different methods to select neoantigens from patient mutations.

[0053] In the experiments, classic classification metrics were used to measure the results: area under the ROC curve (AUC), area under the PR curve (AUPR), area under the ROC curve with a 10% error rate threshold (AUC<0.1), and topk. Comparison methods included well-known methods in the field, such as NetMHCpan-4.1. [7] MHCFlurry2.0

[18] MixMHCpred2.2 [9] BigMHC

[19] (BigMHC_EL and BigMHC_IM) and PRIME 2.0 [9] .

[0054] Experiment 1: Quantization Performance of the Algorithm

[0055] Table 1: Five-way crossover experiment comparison with the optimal method on EL2020.

[0056]

[0057] Table 2: Quantitative comparison of different methods on test sets IM2020 and IS2020.

[0058]

[0059] Table 3: Quantitative comparison of AUC metrics by length for different methods on the test set IM2020.

[0060] NetMHCpan-4.1 0.815 0.913 0.786 0.762 MHCFlurry2.0 0.855 0.933 0.830 0.823 MixMHCpred2.2 0.803 0.907 0.796 0.773 BigMHC_EL 0.859 0.926 0.831 0.819 NeoMHCI 0.897 0.952 0.896 0.897

[0061] Table 4: Quantitative comparison of AUC metrics by length for different methods on the IS2020 test set.

[0062] NetMHCpan-4.1 0.886 0.967 0.927 0.875 0.769 0.759 0.663 MHCFlurry2.0 0.918 0.979 0.952 0.910 0.839 0.832 0.782 MixMHCpred2.2 0.878 0.969 0.936 0.895 0.809 0.777 0.701 BigMHC_EL 0.953 0.985 0.966 0.953 0.880 0.886 0.854 NeoMHCI 0.949 0.982 0.970 0.957 0.939 0.931 0.938

[0063] Table 5: Quantitative comparison of different methods on the test set NE2023.

[0064] NetMHCpan-4.1 0.536 0.0451 MHCFlurry2.0 0.551 0.453 MixMHCpred2.2 0.550 0.417 BigMHC_IM 0.503 0.414 PRIME 2.0 0.590 0.441 NeoMHCI 0.579 0.451

[0065] Table 6: Quantitative comparison of different methods on the test set PE2018 with different gene expression levels (TPM) as thresholds.

[0066]

[0067] NeoMHCI is the method of this invention. The optimal results in Tables 1 to 6 are in bold. Tables 1 and 2 show that this method has significant superiority over other existing methods in antigen presentation prediction on large-scale multi-gene or single-gene data. Tables 3 and 4 show that this method significantly improves antigen presentation prediction at different lengths, maintaining high accuracy even with longer antigen sequences. Table 5 shows that this method outperforms other existing methods on high-quality neoantigen independent test sets, ranking second in AUC (2.3% improvement over the top-ranked PRIME 2.0) and second in AUPR (5.1% improvement over the top-ranked MHCFlurry 2.0). Table 6, based on clinical data, demonstrates that this method identified 6 and 4 immunogenic neoantigens within the Top-1 recommended mutation sequences, respectively, based on different gene expression thresholds, achieving optimal results compared to other methods. When the TPM>2 threshold was used for Top-3 and Top-5 recommendations, this method was only one less than the top-ranked PRIME2.0. However, when the TPM>0 threshold was used, it achieved first place in both methods, identifying 9 and 11 neoantigens respectively. Overall, this method also demonstrated superiority in clinical data, exhibiting more accurate neoantigen identification capabilities compared to other methods, and showcasing its potential value for clinical applications.

[0068] Experiment 2: Visual Comparison of Algorithms for Binding Motifs to Multi-Gene Data

[0069] Figure 4 shows the visualization results of binding motifs of the multi-gene combination CPH-08-TISSUE from different methods. This multi-gene combination was derived from the independent multi-gene test set IM2020, where our method significantly outperformed the comparative methods. By having each method predict the binding scores of 100,000 random 9-length and 9-short peptides from the natural environment with this multi-gene combination, the top 1% of samples were selected to construct a position-specific scoring matrix (PSSM) and visualized. The comparison revealed that the different MHC-I molecule preference ratios under multi-gene combinations directly affected the predictive ability of the methods. This invention first calculated the PSSM matrix constructed from 2106 positive sequences determined in the CPH-08-TISSUE experiment, and then calculated the Pearson correlation coefficient (PCC) with the PSSM matrices predicted by each method. The results showed that the proposed method NeoMHCI achieved the highest correlation PCC = 0.720. Furthermore, a difference was observed between the binding motifs predicted by existing methods and the actual motifs. The first position of the actual motif, as one of the anchoring points, exhibits a preference for amino acids E and D. This amino acid preference was difficult to observe in the comparative existing methods other than the proposed NeoMHCI method. By comparing the binding motifs of individual molecules, it can be found that this preference exhibited by multiple genes mainly originates from HLA-A26:01, which shows a clear preference for amino acids E and D at the first position. However, among the top 1% of short peptides, the proportion of preference for HLA-A26:01 varies significantly among different methods, with proportions of 16.7% for NeoMHCI, 11.3% for NetMHCpan-4.1, 3.5% for MHCflurry-2.0, 9.7% for MixMHCpred-2.2, and 3.8% for BigMHC_EL. This partly explains why the NeoMHCI method of this invention achieved the highest correlation in predicting the binding motifs of multiple genes, and also achieved the highest accuracy (AUC 0.1, 0.872) in the test. This indicates that the method based on the attention-based multi-instance learning module can better capture the preference distribution within multiple genes, thus providing more accurate prediction results.

[0070] In summary, the attention-based multi-instance learning-based method for predicting neoantigens by MHC-I molecule binding proposed in this invention achieves relatively optimal performance on large-scale and clinical data compared to other existing methods. The other algorithms compared are all well-known algorithms in this field.

[0071] References:

[0072] [1]Ilse M, Tomczak J, Welling M. Attention-based Deep Multiple Instance Learning[A]. Proceedings of the 35th International Conference on Machine Learning[C]. PMLR, 2018: 2127–2136.

[0073] [2]Feng P, Zeng J, Ma J. Predicting MHC-peptide binding affinity by differential boundary tree[J]. Bioinformatics, 2021, 37(Supplement_1): i254–i261.

[0074] [3]Nielsen M, Andreatta M. NetMHCpan-3.0; improved prediction of binding to MHC class I molecules integrating information from multiple receptor and peptide length datasets[J]. Genome medicine, BioMed Central, 2016, 8(1): 1–9.

[0075] [4]O’Donnell T J, Rubinsteyn A, Bonsack M, et al. MHCflurry: open-source class I MHC binding affinity prediction[J]. Cell systems, Elsevier, 2018, 7(1): 129–132.

[0076] [5]Chu Y, Zhang Y, Wang Q, et al. A transformer-based model to predict peptide–HLA class I binding and optimize mutated peptides for vaccine design[J]. Nature Machine Intelligence, Nature Publishing Group, 2022, 4(3): 300–311.

[0077] [6]Jurtz V, Paul S, Andreatta M, et al. NetMHCpan-4.0: improved peptide–MHC class I interaction predictions integrating eluted ligand and peptide binding affinity data[J]. The Journal of Immunology, Am Assoc Immnol, 2017, 199(9): 3360–3368.

[0078] [7]Reynisson B, Alvarez B, Paul S, et al. NetMHCpan-4.1 and NetMHCIIpan-4.0: improved predictions of MHC antigen presentation by concurrent motif deconvolution and integration of MS MHC eluted ligand data[J]. Nucleic Acids Research, Oxford University Press, 2020, 48(W1): W449–W454.

[0079] [8]Bulik-Sullivan B, Busby J, Palmer C D, et al. Deep learning using tumor HLA peptide mass spectrometry datasets improves neoantigen identification[J]. Nature Biotechnology, 2019, 37(1): 55–63.

[0080] [9]Gfeller D, Schmidt J, Croce G, et al. Improved predictions of antigen presentation and TCR recognition with MixMHCpred 2.2 and PRIME 2.0 reveal potent SARS-CoV-2 CD8+ T-cell epitopes[J]. Cell Systems, Elsevier, 2023, 14(1): 72-83.e5.

[0081]

[10] Feng J, Zhou Z-H. Deep MIML Network[A]. Proceedings of the AAAI Conference on Artificial Intelligence[C]. 2017, 31(1).

[0082]

[11] Xia F, Liu T-Y, Wang J, et al. Listwise Approach to Learning to Rank-Theory and Algorithm[J].

[0083]

[12] You R, Qu W, Mamitsuka H, et al. DeepMHCII: a novel binding core-aware deep interaction model for accurate MHC-II peptide binding affinity prediction[J]. Bioinformatics, 2022, 38(Supplement_1):i220–i228.

[0084]

[13] Qu W, You R, Mamitsuka H, et al. DeepMHCI: An anchor position-aware deep interaction model for accurate MHC-I peptide binding affinity prediction[J]. Bioinformatics, Oxford University Press, 2023:btad551.

[0085]

[14] Zeng H, Gifford D K. DeepLigand: accurate prediction of MHC class I ligands using peptide embedding[J]. Bioinformatics, Oxford University Press, 2019, 35(14):i278–i283.

[0086]

[15] Dauphin Y N, Fan A, Auli M, et al. Language Modeling with Gated Convolutional Networks[A]. Proceedings of the 34th International Conference on Machine Learning[C]. PMLR, 2017:933–941.

[0087]

[16] Bjerregaard A-M, Nielsen M, Jurtz V, et al. An Analysis of Natural T Cell Responses to Predicted Tumor Neoepitopes[J]. Frontiers in Immunology, 2017, 8.

[0088]

[17] Vita R, Mahajan S, Overton J A, et al. The immune epitope database (IEDB): 2018 update[J]. Nucleic acids research, Oxford University Press, 2019, 47(D1): D339–D343.

[0089]

[18] O’Donnell T J, Rubinsteyn A, Laserson U. MHCflurry 2.0: Improved pan-allele prediction of MHC class I-presented peptides by incorporating antigen processing[J]. Cell systems, Elsevier, 2020, 11(1): 42–48.

[0090]

[19] Albert B A, Yang Y, Shao X M, et al. Deep neural networks predict class I major histocompatibility complex epitope presentation and transfer learn neoepitope immunogenicity[J]. Nature Machine Intelligence, 2023, 5(8): 861–872.

[0091]

[20] Wu T, Chen J, Diao K, et al. Neodb: a comprehensive neoantigen database and discovery platform for cancer immunotherapy[J]. Database, 2023, 2023: baad041.

Claims

1. A method for recognizing MHC-I class molecular neoantigens based on multi-instance learning, characterized in that, Design an interactive convolutional network structure for sequence binding to extract high-dimensional features of the input peptide and each MHC-I molecule; Attention mechanisms are used to learn and assign weights to various high-dimensional representations. The fusion features of the input peptide binding to multiple MHC-I molecules are obtained through permutation-invariant aggregation functions in multi-instance learning techniques, followed by a binary classification task for binding prediction. After training on millions of multi-gene datasets, the network knowledge is transferred to the downstream neoantigen recognition task using multi-task learning techniques. For the background knowledge that neoantigens originate from wild-type sequence mutations with only a few mutations possessing immunogenicity, paired neoantigen-wild-type sequence data recorded in the database are used to construct an auxiliary data generation process. First, a ranking learning loss function is applied to Task 1, aiming to improve the model's prediction score for neoantigens compared to training with other random mutations. Then, a binary cross-entropy loss function is applied to Task 2, classifying neoantigens as positive samples while maintaining the original prediction score for wild-type sequences, thus increasing the absolute magnitude of the model's neoantigen prediction score. The specific steps are as follows: (1) First, to address the issue of thousands of MHC-I molecules in the human body, each with its own binding specificity, and to address the problem of the binding strength of short peptide sequences of different lengths being affected by the misalignment of binding residues with MHC-I molecules, a network architecture is adopted to improve characterization capabilities; specifically: For a pair of inputs with The short peptide sequence P of 1 amino acid and the MHC-I molecular sequence An embedding layer is used to map each amino acid in the sequence to a corresponding learnable vector, resulting in the embedding matrix of the short peptide. Embedding matrix with MHC-I molecules , The dimension of the amino acid embedding vector; For each amino acid representation vector in the short peptide embedding matrix, an independent gated linear unit (GLU) is used for expression control to distinguish binding residues in the short peptide sequence, thereby achieving adaptive identification of anchoring sites, as shown below: ， （1） in, For learnable parameters, It is the Sigmoid activation function. For the element-wise Hadema product; after The short peptide embedding matrix obtained after updating GLU is 1. As input to subsequent convolutional layers; utilizing the learnable parameter matrix Embedding matrix with MHC-I molecules To generate convolution kernels Then, the short peptide embedding matrix... Convolutional computations are performed to construct the interaction information between the two, represented as: ， （2） In the above formula, It is the ReLU activation function. For learnable bias terms, The symbol for convolution is ; to simplify the symbol, the calculation process of formula (2) is defined as follows: ,in The parameters in the network; a two-layer BICL is used to implement the computational paradigm for residual connections: ， （3） Next, a multilayer perceptron network (MLP) is used to analyze the obtained interaction features. Further updates are performed, and finally, a max pooling layer is used to obtain the final representation vector. This serves as input for subsequent multi-example learning of the MIL framework: ， （4） (2) In order to learn the intrinsic relationships of multi-gene data while avoiding bias caused by direct score comparison, a multi-instance learning framework based on attention mechanism is used to replace the traditional simplified strategy of max pooling based on multi-instance learning MIL; the multi-instance learning framework is as follows: Given that a sample in a multigene dataset consists of a short peptide and multiple, and an indeterminate number, MHC-I molecules, the data itself possesses the following properties: a sample is considered positive when at least one MHC-I molecule binds to the short peptide, and conversely, a sample is considered negative only when none of the MHC-I molecules bind to it. Soft processing is performed using attention weight allocation; specifically, the network architecture described above is used to process a short peptide... With multiple MHC-I molecules The multiple pairs are calculated separately and their representation vectors are obtained. The gated attention weights are calculated using the following formula: ， （5） in, These are the learnable parameters. It is an exponential function; based on the above weights, the final comprehensive representation vector is calculated: ， （6） Next, a linear layer is used as a classifier for binary prediction: ， （7） This multi-instance learning framework is trained using a binary classification cross-entropy loss function: ， （8） (3) Apply the above multi-example learning framework to multi-gene data to predict whether MHC presents peptides; Because there are a large number of short peptides presented by MHC-I molecules on the surface of cell membranes in the human body, only a very small number will be specifically recognized and trigger an immune response. Paired neoantigen-wildtype sequence data are used, combined with multi-task learning for fine-tuning to enhance the ability to distinguish neoantigens. The prediction model requires that the predicted score of neoantigens that elicit an immune response be higher than the predicted score of other random mutations. To enhance the ability to distinguish neoantigens, multi-task learning is used to fine-tune the model parameters by setting different loss functions, as follows: A two-stage learning strategy of training and fine-tuning is used to determine the final model parameters: During the training phase, supervised training was conducted on millions of washed data, including single-gene and multi-gene data, using the standard binary cross-entropy loss function to obtain a model with predictive capabilities. Task 1, Data Construction and Loss Function Calculation: For paired wild-type sequences and neoantigen sequences in the database, firstly, iterate through the single amino acid mutation sequences of each wild-type sequence, and score and rank them using the model trained in the first step; for each wild-type sequence's single mutation set, sample 10 single mutation sequences from the 10%, 30%, and 70% ranking positions respectively to form three candidate pools: a strong binding pool... Weak binding pool Unbound pool For each neoantigen sequence S neo During training, a list is formed by randomly selecting one mutation sequence from each of the three candidate pools. The list is expected to be sorted as y=(1,2,3,4), meaning the neoantigen prediction score is greater than strongly bound mutation sequences, which are greater than weakly bound mutation sequences, which are greater than unbound mutation sequences. The classic ranking learning loss function ListMLE is used, defined as follows: ， （9） in, This represents the model obtained from the first stage of training. Defined as: ， （10） in, This represents the mutation sequence at position j in the actual sorting; Task 2: Supervised classification using neoantigen labeling information; specifically, for neoantigen S... neo The wild-type sequence S was labeled as a positive sample. wt The expected model should maintain its original prediction score, and the predicted value y obtained from the first stage should be used. wt As a self-distilled soft label; the specific loss function is as follows: ， （11） Where H(·,·) represents the cross-entropy loss function; the loss function for the entire fine-tuning process is: L = αL1 + βL2, (12) Where α and β are weighting coefficients.

Images