TW202228153A - System and method for predicting and identifying the immunogenicity of a peptide based on machine learning - Google Patents

Abstract

The present disclosure relates to a system and a method to train a machine-learning HLA-peptide presentation prediction and identification model. The system comprises: an encoding module, a neural network training module, an ensemble learning module and a immunogenic prediction module. The system outperforms existing predictors trained on binding affinity.

Description

Machine learning-based peptide immunogenicity prediction and identification system and method

This application claims the priority of the patent application filed on December 9, 2020 (application number CN202011450578.9) and the patent application filed on July 5, 2021 (application number CN202110756286.6).

The present disclosure belongs to the field of biomedicine, and provides a method for predicting and/or identifying the possibility of peptide presentation by human leukocyte antigen (HLA) molecules, for preparing a vaccine for immunotherapy of tumors.

Immunotherapy is a new type of tumor treatment that has emerged in recent years. Compared with surgical resection, traditional radiotherapy and chemotherapy, targeted therapy and other methods, immunotherapy has more significant effect, less side effects, and longer time for patients to benefit. The classic immunotherapy scheme uses the injection of immune checkpoint inhibitors to activate the immune cells of patients, but the activated immune cells often cannot specifically attack tumor cells, but may attack normal cells, resulting in immunotherapy only applicable to the difference between tumors and normal tissues Larger patients greatly limit the scope, safety and efficacy of immunotherapy. On this basis, methods such as tumor neoantigen vaccine and tumor neoantigen in vitro T cell initiation culture have been developed to make the activated immune cells more specific to tumor cells, thereby improving the universality and safety of immunotherapy.

The core of methods such as tumor neoantigen vaccines and tumor neoantigen in vitro T cell priming culture is the immunogenicity prediction of antigens. Only vaccines/priming T cells made from tumor neoantigens with higher immunogenicity are effective. One step is currently a big challenge. The immunogenicity of peptides can be accurately obtained by using the ELISPOT experimental method, but this experiment can only detect dozens of peptides at a time, which cannot meet the clinical needs. The use of mass spectrometry experiments can detect the immunogenicity of a large number of peptides at one time, but the period of mass spectrometry experiments is often several months, and the experimental conditions are still unstable, making clinical application difficult.

With the rapid development of machine learning and the continuous integration of medicine and artificial intelligence, the use of computer methods to assist in the study of related problems in biology and medicine has become a powerful tool. The first computational method to predict HLA binding was SYFPEITHI, but SYFPEITHI can only predict 9 amino acids and 10 amino acid peptides for most HLA types. At present, the immunogenicity determination of tumor neoantigens mostly relies on public software, such as netMHCpan, MHCflurry, etc. for prediction. NetMHCpan 4.0 integrates both binding affinity and MS eluted ligand data for training, and obtains better prediction results than training with a single data. When trained on affinity measures, MHCflurry 1.2 outperforms standard predictors NetMHC 4.0 and NetMHCpan 3.0 overall (see O'Donnell, Timothy J., et al. "MHCflurry: open-source class I MHC affinity binding prediction." Cell systems 7.1 (2018): 129-132.). However, the accuracy of these software is less than 10%, and the study by Bonsack et al. shows that NetMHCpan 4.0 and MHCflurry 1.2 do not significantly outperform PickPocket 1.1, IEDB SMM, IEDB SMMPMBEC, and SYFPEITHI methods (see Bonsack, Maria, et al. "Performance evaluation of MHC class-I binding prediction tools based on an experimentally validated MHC-peptide binding data set.” Cancer immunology research 7.5(2019): 719-736.). It is therefore necessary to develop new predictive antigens presented by human leukocyte antigen molecules possible approaches to facilitate the development of tumor immunotherapy.

The ensemble learning model is a machine learning method that has developed rapidly in recent years. Its main principle is to integrate multiple models with weak effects to obtain a model with strong effects. The ensemble learning method first trains a model with weaker effect, and then trains the next model with weaker effect according to the result obtained by the previous model, and repeats the operation for many times, and finally integrates a series of models with weaker effect to obtain the final result. a more powerful model. The advantage of ensemble learning is that it can adjust the parameters by self-adjustment, and the integration is repeated many times, so the accuracy is often high.

Existing ensemble learning models include AdaBoost, XGBoost, and LightGBM. AdaBoost is the original ensemble learning model, which can optimize parameters and integrate itself during training, but it does not have multi-threaded function, and will include all data for training at one time. When the data set is too large, it often takes too long; XGBoost is based on this. Multi-threaded optimization is adopted, and some data are randomly used for training, thereby improving the training speed and model performance; LighGBM uses different parameter self-optimization methods, and its training speed is faster than XGBoost, but the accuracy of the model is often slightly lower. . At present, there is no report that the integrated learning model is applied to the immunogenicity prediction of antigens, and there is still a lack of accurate and effective prediction models for antigen immunogenicity.

The present disclosure provides a system and method for the prediction and/or identification of peptide immunogenicity based on machine learning, using high-quality mass spectrometry data for parameter training, and integrating neural network The road model and integrated learning model can improve the accuracy of existing prediction and/or identification methods, solve the problem of low immunogenicity detection accuracy, and improve the effectiveness and safety of downstream immunotherapy therapies.

The present disclosure provides a method for predicting and/or identifying the likelihood of peptide presentation by HLA molecules based on a fusion neural network model and an ensemble learning model. Each presentation likelihood represents the likelihood that the corresponding peptide will be presented on the surface of tumor cells in the subject by one or more HLA alleles.

In some embodiments, the method includes a model building step and a prediction and/or identification application step.

In some embodiments, the model building step includes building a dataset; training a neural network model and an ensemble learning model, and performing model fusion. In some embodiments, parameter training is included on data provided by mass spectrometry experiments.

In some embodiments, the predicting and/or identifying applying step includes predicting and/or identifying the likelihood that the test peptide will be presented by an HLA allele.

The present disclosure obtains the construction data set data obtained by mass spectrometry experiments from public databases. Suitable repositories include, but are not limited to, sequences accessible in NCBI, EMBL, GenBank, RefSeq, UCSC repositories, etc., GenPept, EBI expression profiles (https://www.ebi.ac.uk/gxa/home), UniProtKB/Swiss-Prot (http://www.uniprot.org/), Protein Information Resource (PIR) (http://pir.georgetown.edu/pirwww/index.shtml), COSMIC, MOWSE, DDBJ, PDB, EST, STS, GSS, HTGS, IEDB, SYFPEITHI, and MassIVE, among others, have computerized databases known to those of ordinary skill in the art.

In some embodiments, the data set data includes data for positive peptides that are frequently presented by HLA typing in the population.

In some specific embodiments, peptides, immunogenicity data and expression data obtained from mass spectrometry experiments are obtained from public databases IEDB, SYFPEITHI and MassIVE.

The lengths of peptide sequences used as training and test sets can be of any length. In some embodiments, the length is less than or equal to 50, in some embodiments, less than or equal to 40, in some embodiments, less than or equal to 30, and in some embodiments, less than or equal to 20 or less than or equal to 10.

In some embodiments, the peptide sequences used as training set and test set are 5-15 amino acids, 7-12 amino acids, 8-11 amino acids in length, such as 5, 6, 7, 8, 9, 10, 11, 12, 13, 14 or 15 amino acids.

In some embodiments, the training set includes positive peptides that are frequently presented by HLA typing in the population.

In some embodiments, the dataset includes a set of positive peptides and a set of negative peptides.

The positive peptide set includes entries for peptides identified or inferred from surface-bound or secreted HLA/peptide complexes encoded by one or more distinct HLA alleles. In some embodiments, the set of positive peptides includes positive peptides that are presented by HLA typing with a high frequency in the population.

This negative peptide set includes entries for peptides not identified or inferred from surface-bound or secreted HLA-peptide complexes.

In some embodiments, the dataset is divided into one or more training datasets (training set) and one or more testing datasets (testing set).

In some embodiments, the positive peptide set and the negative peptide set in the dataset are mixed 1:1 to 1:20000, in some embodiments, 1:1 to 1:10000, 1:1 to 1:5000, 1:1 To 1:4000 mix, 1:1 to 1:3000 mix, 1:1 to 1:2000 mix, 1:1 to 1:1000 mix, 1:1 to 1:500 mix. For example 1: 20000, 1: 10000, 1:5000, 1:2000, 1:1000, 1:500, 1:300, 1:150, 1:100, 1:50, 1:10, 1:1.

In some embodiments, the datasets include monoallelic datasets and/or multiallelic datasets.

In some embodiments, the training set trains the model based on training data generated by expressing a single HLA allele.

In some embodiments, the training set trains the model based on training data generated by expressing a single HLA allele, expressing multiple HLA alleles, or a combination thereof.

The model building step includes data set preprocessing. In some embodiments, data in the data set is data encoded before model training. In some embodiments, it includes digitally encoding peptides, expression levels, and HLA typing data in the data set. For example, HLA typing and peptide sequence amino acid codes, along with expression data and protein family IDs to which the peptides belong, are input into the model for training.

In some embodiments, one-hot encoding (one-hot), BLOMAP, PSSM, word2vec, or BLOSUM62 is used to encode peptide, HLA typing data into data lines consisting of multiple digits.

In some embodiments, one-hot encoding is used. N states are encoded using an N-bit state register. The amino acids of the peptide sequences in the disclosed embodiments have a total of 21-bit states (20 amino acids plus 1 vacancy, the vacancy is represented by Z, and the state sequence is A-Z in the alphabetical order of amino acid abbreviations). For the first state amino acid A (Alanine), it is one-hot encoded.

In some embodiments, the code is [1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ]; for the last state vacancy, its one-hot coding. One-hot encoded as [0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1]; other states And so on.

In some embodiments, one-hot encoding is implemented using Scikit-learn 0.20.3.

In some embodiments, the neural network model includes a fully connected neural network, a convolutional neural network, a long short-term memory neural network, and the like. In some specific embodiments, the neural network model is a fully connected neural network model.

In some embodiments, parameter training is performed on mass spectrometry experimental data using a fully connected neural network model, specifically:

Suppose there are m training peptides, each corresponding to n HLA types:

為各訓練肽經資料編碼後的資料矩陣(不包含HLA分型)，

為各訓練肽在質譜實驗中的陽性/陰性標籤，β _target 1為神經網路係數矩陣，經交叉驗證訓練得到準確度最大時的係數。 in

is the data matrix encoded by the data for each training peptide (excluding HLA typing),

is the positive/negative label of each training peptide in the mass spectrometry experiment, β _target 1 is the neural network coefficient matrix, and the coefficient when the accuracy is the largest is obtained through cross-validation training.

In some embodiments, training is performed with 3-fold or 5-fold cross-validation.

In some embodiments, the fully connected neural network model architecture is 3-5 layers, eg, 3 layers, 4 layers, 5 layers.

In some specific embodiments, the fully-connected neural network model has 4 layers, and the number of neurons in each layer is 256, 32, 16, and 1 in sequence.

In some embodiments, the activation function is sigmoid, tanh, relu, softmax.

In some embodiments, the loss function is Negative Bernoulli's log loss or Cross entropy. In some embodiments the cross-entropy is binary crossentropy.

In some embodiments, the neural network initialization is Glorot initialization, Kaiming initialization, lecun initialization, or Batch Normalization.

In some embodiments, the ensemble learning model includes Bagging and Boosting. For example Random Forest, Adaboost, Gradient Boosting Decision Tree (GBDT), XGboost, LightGBM.

In some specific embodiments, the ensemble learning model is XGBoost.

In some specific embodiments, the XGBoost integrated learning model is used to perform parameter training on mass spectrometry experimental data, specifically:

Suppose there are m training peptides, each corresponding to n HLA types:

為各訓練肽經資料編碼後的資料矩陣(包含HLA分型)，

為各訓練肽在質譜實驗中的陽性/陰性標籤，β _target2為集成學習係數矩陣，經交叉驗證訓練得到準確度最大時的係數。 in

is the data matrix (including HLA typing) encoded by the data for each training peptide,

is the positive/negative label of each training peptide in the mass spectrometry experiment, β _{target 2} is the ensemble learning coefficient matrix, and the coefficient when the accuracy is maximum is obtained after cross-validation training.

In some embodiments, training is performed with 3-fold or 5-fold cross-validation.

In some embodiments, the model parameters and architecture are optimized with the test set positive predictive values (PPV) as the preferred objective.

The parameters for optimizing the ensemble learning model include the maximum depth of the classification tree (max depth), the minimum sum of instance weights required for a child node (min child weight), the subsampling ratio of columns when constructing each tree (colsample bytree), leaf nodes Minimum required for upper division Loss reduction (gamma), maximum number of weak learners (n estimators), learning rate (learning rate), and subsample ratio of training instances.

In some specific embodiments, the max depth is selected from 3-10, 4-8, eg, 3, 4, 5, 6, 7, 8, 9, 10.

In some specific embodiments, the min child weight is selected from 2-10, 3-9, 4-8, eg, 2, 3, 4, 5, 6.

In some specific embodiments, the colsample bytree is selected from 0.40-1.1, 0.5.-0.90, 0.50-0.80, 0.50-0.60, 0.45-0.55, 0.50-0.54.

In some specific embodiments, gamma is selected from 0.01-1.0, 0.05-1.0, 0.1-1.0, 0.2-0.9, 0.3-0.8, 0.4-0.7, 0.5-0.6.

In some specific embodiments, n estimators are selected from 100-2500, 500-2300, 1000-1800, 1500-1700, 1550-1650.

In some specific embodiments, the learning rate is selected from 0.01-0.5, eg, 0.02-0.5, 0.03-0.45, 0.04-0.40, 0.05-0.35, 0.06-0.30, 0.07-0.25.

In some specific embodiments, the subsample is selected from 0.5-1, 0.6-0.9, 0.7-0.8.

In some embodiments, predicting and/or identifying the likelihood of peptide presentation by HLA molecules includes a fusion neural network model coefficient matrix and an ensemble learning model coefficient matrix.

In some embodiments, a fusion of neural network models and ensemble learning models is included.

In some embodiments, strategies for fusing neural network models and ensemble learning models include averaging, voting, and learning, wherein averaging includes simple averaging or weighted averaging.

In some embodiments, fusing includes multiplying the scores of the outputs of the ensemble learning model in the neural network model by a weighting factor to form a final model.

In some embodiments, the weight coefficients are optimized with positive predictive values (PPV) as the preferred objective.

In some embodiments, the weighting factor is selected from the group consisting of 0.1-0.00001, 0.005-0.0001, 0.05-0.0001, 0.02-0.001, 0.01-0.005, eg, 0.1, 0.05, 0.02, 0.01, 0.005, 0.002, 0.001, 0.0005, 0.0002, 0.000.

In some embodiments, the method for predicting and/or identifying the likelihood of peptide presentation by HLA molecules includes predicting and/or identifying the likelihood of each peptide being presented by HLA molecules in the test sample based on the test sample.

In some embodiments, it includes predicting and/or identifying the possibility that each peptide of the test sample is presented by HLA molecules according to the peptide, expression level and HLA typing data of the test sample.

In some embodiments, the sample under test is subjected to the same preprocessing as the model building stage. The processed data to be predicted and/or identified is logged into the model for prediction and/or identification, and the threshold is used for classification, and the immunogenicity classification result of the data to be predicted and/or identified is output.

In some embodiments, the sample to be tested is encoded with data.

In some embodiments, one-hot encoding (one-hot), BLOMAP, PSSM, word2vec, or BLOSUM62 is used to encode the test sample peptides, HLA typing data into data lines composed of multiple digits.

In some embodiments, one-hot encoding is used. N states are encoded using an N-bit state register. The amino acids of the peptide sequences in the disclosed embodiments have a total of 21-bit states (20 amino acids plus 1 vacancy, the vacancy is represented by Z, and the state sequence is A-Z in the alphabetical order of amino acid abbreviations). For the first state amino acid A (Alanine), it is one-hot encoded.

In some embodiments, the code is [1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 ]; for the last state slot, its one-hot encoding. One-hot encoded as [0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1]; other states And so on.

In some embodiments, one-hot encoding is implemented using Scikit-learn 0.20.3.

In some embodiments, the peptides of the samples to be tested and/or identified are subjected to the same preprocessing as in the model building phase, and the processed data to be predicted and/or identified are logged into the model for prediction and/or identification, and delineated thresholds are used. A classification is performed to output a likelihood of presentation by HLA molecules for the data to be predicted and/or identified. Peptides in the test sample above this threshold are predicted and/or identified as being presented by HLA molecules.

In some embodiments, the data set to be predicted and/or identified further includes at least one of: data related to peptide-HLA binding affinity measurements of at least one of the peptides; and stabilization of peptide-HLA binding to at least one of the peptides Information about sex measurements.

，具體為： In some embodiments, for predicting and/or identifying the immunogenicity of test peptides

,Specifically:

Wherein X is the data row obtained after each peptide is encoded by the data encoding module.

In some embodiments, the top 0.1% of the scores are selected as positive (top 0.1%) or the positive predictive value (recall 0.4) at 40% recall as the threshold above which peptides in the test sample are predicted and/or or identified as being presented by HLA molecules.

In some embodiments, recall 0.4 is used as the threshold.

The present disclosure is to predict and/or identify peptides of a sample, including obtaining at least one of exome, transcriptome, or genome-wide tumor nucleotide sequencing data from tumor cells of a subject, wherein the tumor nucleotide sequencing data was used to obtain peptide data.

In some embodiments, the sample to be predicted and/or identified includes at least one of: a cell line engineered to express a single HLA allele; a cell line engineered to express multiple HLA alleles; obtained or obtained human cell lines; fresh or frozen tumor samples obtained from multiple patients; and fresh or frozen tissue samples obtained from multiple patients.

In some embodiments, presentation potential is optionally further identified by measuring the expression level of one or more HLA alleles in the subject, such as by RNA-seq or mass spectrometry. In some specific embodiments, the preprocessing of the sample data set to be tested is the same as the model building. In some specific embodiments, the test sample data set uses one-hot encoding (one-hot), BLOMAP, PSSM, word2vec or BLOSUM62 to encode the test sample peptide and HLA typing data into data lines composed of multiple digits.

In some embodiments, the prediction and/or identification is selected from the group consisting of single HLA allele prediction and/or identification, multi-allele prediction and/or identification.

In some embodiments, the neural network model and the ensemble learning model fused in the present disclosure are selected from a monoallelic model and a multiallelic model.

In some embodiments, a monoallelic model builds a model based on a single HLA allele and predicts and/or identifies the likelihood that a peptide will be presented by the associated HLA allele.

In some embodiments, the multi-allelic gene model builds a model in a multi-allelic environment where two or more HLA alleles are present, and predicts and/or identifies whether a peptide will be presented by multiple HLA alleles possibility.

In some embodiments, the two or more HLA alleles comprise two or more different HLA alleles.

In some embodiments, the HLA alleles comprise class I HLA alleles.

In some embodiments, the HLA alleles include class II HLA alleles.

In some embodiments, HLA alleles include HLA-A, HLA-B, HLA-C, HLA-E, HLA-F, HLA-G, HLA-DQ, HLA-DR, HLA-DP.

In some embodiments, the HLA allele is selected from the group consisting of A*01:01, A*02:01, A*02:03, A*02:04, A*02:07, A*03:01, A* 24:02,A*29:02,A*31:01,A*68:02,B*35:01,B*44:02,B*44:03,B*51:01,B*54: 01,B57:01,C*03:02,C*03:04,C*04:01,C*05:01,C*06:02,C*08:01,C*08:02,C* 12:02, C*14:02, C*14:03, C*15:02 and C*16:01. It is known to those of ordinary skill in the art that there are allelic variants of the above-mentioned HLA types, all of which are encompassed by the present invention. A complete list of HLA class alleles can be found at http://hla.alleles.org/alleles/. For example, a complete list of HLA class I alleles can be found at http://hla.alleles.org/alleles/class1.html.

In some embodiments, the present disclosure provides methods for predicting and/or identifying the likelihood of peptides or combinations of peptides being presented by HLA molecules based on fused neural network models and ensemble learning models, comprising:

Step 1: Data coding, digital coding of peptides, expression levels, HLA typing data, etc.;

Step 2: Neural network model training, using the fully connected neural network model to perform parameter training on the data provided by the mass spectrometry experiment;

Step 3: Integrated learning model training, using the integrated learning model to perform parameter training on the immunogenicity data provided by the mass spectrometry experiment;

Step 4: The peptides are predicted and/or identified by the possibility of presentation of HLA molecules, and the peptides, expression levels and HLA typing data of the samples to be predicted and/or identified are predicted and/or identified according to the neural network coefficient matrix and the integrated learning coefficient matrix. Or identify the likelihood that each peptide in the sample is presented by HLA molecules.

The present disclosure provides a method for predicting and/or identifying the possibility of peptide presentation by HLA molecules based on a fusion neural network model and an ensemble learning model. The method includes: constructing a data set and data encoding; training, fusing the neural network model and integrating Learning model; predicting and/or identifying the likelihood of peptide presentation by HLA molecules, wherein the dataset, data encoding, training, fusion, neural network model, ensemble learning model, prediction and/or identification are as described in this disclosure.

The present disclosure provides the application of machine learning models in the preparation of mRNA, polypeptide vaccines, antitumor drugs or tumor vaccines. In some specific embodiments, the machine learning model is the fused neural network model and the ensemble learning model provided by the present disclosure.

The present disclosure provides methods for identifying at least one peptide from one or more tumor cells of a subject that is likely to be presented on the surface of the tumor cell by one or more HLA alleles, comprising predicting by the present disclosure and/or methods of identifying the likelihood of peptide presentation by HLA molecules.

The present disclosure provides the application of the method for predicting and/or identifying the possibility of peptide presentation by HLA molecules described in the present disclosure in the preparation of mRNA, polypeptide vaccines, anti-tumor drugs or tumor vaccines.

In some embodiments, the tumor is selected from the group consisting of: lung cancer, melanoma, breast cancer, ovarian cancer, prostate cancer, kidney cancer, stomach cancer, colon cancer, testicular cancer, head and neck cancer, pancreatic cancer, brain cancer, B cell lymphoma tumor, acute myeloid leukemia, chronic myeloid leukemia, chronic lymphocytic leukemia and T-cell lymphocytic leukemia, non-small cell lung cancer and small cell lung cancer.

In some embodiments, the method further comprises generating an output for constructing a personalized cancer vaccine from a peptide or combination of peptides predicted and/or identified to be presented by HLA molecules. In such embodiments, the output of the personalized cancer vaccine may include at least one peptide or at least one polynucleotide encoding the selected peptide or combination of peptides.

In another aspect, methods are provided for identifying variant or allelic mutations in tumors based on the fusion-based neural network models and ensemble learning models of the present disclosure.

In some embodiments, identification of peptides with mutations in tumor cells or mutant polypeptides resulting from, for example, splice site mutations, frameshift mutations, readthrough mutations, or gene fusion mutations is included.

The present disclosure provides a peptide or combination of peptides comprising one or more peptides predicted and/or identified by the methods of the present disclosure.

The present disclosure provides a peptide or combination of peptides comprising one or more peptides predicted and/or identified as presented by HLA molecules by the methods of the present disclosure.

In some embodiments, the peptide is presented on the HLA protein with a higher affinity than the wild-type peptide.

In some embodiments, the peptide can have an IC50 value of at least below 5000 nM, at least below 1000 nM, at least below 500 nM, at least below 250 nM, at least below 200 nM, at least below 150 nM, at least below 100 nM, at least below 100 nM 50nM or less.

In some embodiments, the length of the peptide includes, but is not limited to, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 60, 70, 80, 90, 100, 110, 120 or more amine molecular residues, and any range derivable therefrom. In certain embodiments, the peptide is equal to or less than 50 amino acids in length.

In some embodiments, the combination of peptides includes at least two or more peptides.

In some embodiments, the combination contains at least two different peptides.

In some embodiments, a peptide with a desired activity or property can be modified to provide certain desired properties, such as improved pharmacological properties, while increasing or at least maintaining substantially all of the biological activity of the unmodified peptide to Binds the desired HLA molecule and activates the appropriate T cells.

The present disclosure provides methods for identifying from one or more tumor cells of a subject at least one peptide selected from the group consisting of the one or more HLA alleles likely to be presented on the surface of the tumor cell by one or more HLA alleles peptides predicted and/or identified by the methods described above.

The present disclosure provides an isolated T cell specific for a peptide or combination of peptides described in the present disclosure.

The present disclosure provides a polynucleotide encoding a peptide or combination of peptides described in the present disclosure.

In some embodiments, the polynucleotide may be, for example, single- and/or double-stranded DNA, cDNA, PNA, CAN, RNA (eg, mRNA), or a natural or chemically modified form of the polynucleotide, or a combination thereof, and the polynucleotide The nucleotides may or may not contain introns.

In some specific embodiments, the polynucleotide is selected from mRNA.

In some embodiments, the polynucleotide is delivered directly.

In some embodiments, the polynucleotide is delivered by a drug delivery system. Various delivery systems are known and can be used with the polynucleotides of the present disclosure, eg, encapsulation in viral vectors, mRNA vectors, DNA vectors, liposomes.

In some specific embodiments, the polynucleotides are delivered in complexes with cationic compounds, such as cationic lipids.

The present disclosure provides a vector containing the polynucleotide described in the present disclosure, and the vector is a eukaryotic expression vector, a prokaryotic expression vector or a viral vector.

In another aspect, the present disclosure provides a host cell comprising the vector described in the present disclosure; in some embodiments, the host cell is bacteria, yeast, mammalian cells, and in some embodiments, the host cell is Escherichia coli, Pichia pastoris, Chinese hamster ovary cells or human embryonic kidney 293 cells.

The present disclosure provides a method of making a peptide predicted and/or identified by the method described in the present disclosure.

In some embodiments, the peptide is predicted and/or identified by the methods described herein to be presented by an HLA molecule.

In some embodiments, the peptide is a peptide or combination of peptides described in this disclosure.

The present disclosure provides a method for preparing a peptide comprising the steps of the method described in the present disclosure for predicting and/or identifying the likelihood of peptide presentation by HLA molecules.

The present disclosure provides a method of making a polynucleotide.

In some embodiments, the polynucleotide encodes a peptide presented by an HLA molecule predicted and/or identified by the methods described herein.

In some embodiments, the polynucleotide encodes a peptide or combination of peptides described herein.

In some embodiments, the polynucleotide is the polynucleotide of the disclosure.

The present disclosure provides antigens comprising the polynucleotides or peptides described in the present disclosure.

In some embodiments, the antigen includes a polynucleotide encoding a peptide or portion thereof. The polynucleotide may be, for example, single- and/or double-stranded DNA, cDNA, PNA, CAN, RNA (eg, mRNA), or a natural or chemically modified form of the polynucleotide, or a combination thereof, and the polynucleotide may contain Or may be free of introns.

In some embodiments, antigens include isolated peptides with tumor-specific mutations predicted and/or identified by the methods disclosed herein, peptides comprising known tumor-specific mutations, and predicted and/or identified by the methods disclosed herein peptides or fragments thereof. In some embodiments, one or more antigens can be presented on the tumor surface.

In some embodiments, one or more antigens may be immunogenic in a subject with a tumor, eg, capable of eliciting a T cell response or a B cell response in the subject.

In some embodiments, where a vaccine is produced for a subject with a tumor, it may be contemplated to exclude one or more antigens that induce an autoimmune response in the subject.

In some embodiments, an expression vector capable of expressing a peptide or portion thereof is provided.

The antigens of the present disclosure can be produced using methods known in the art, including culturing a host cell under conditions suitable for expression of the antigen or peptide or vector, wherein the host cell comprises at least one polynucleotide encoding the antigen or peptide or vector ; and purifying the antigen or peptide or carrier. Standard purification methods include chromatographic techniques, electrophoresis techniques, immunological techniques, precipitation, dialysis, filtration, concentration and isoelectric focusing techniques.

In another aspect, the present disclosure provides a composition capable of eliciting a specific immune response (eg, a tumor-specific immune response) comprising a plurality of peptides or encoded peptides predicted and/or identified using the methods described in the present disclosure, or part of the polynucleotide.

In some embodiments, the composition is an mRNA, a polypeptide vaccine, an anti-tumor drug, or a tumor vaccine.

In some embodiments, the composition is capable of eliciting a T cell response or a B cell response in the subject.

In some embodiments, the composition is an immunogenic composition such as a vaccine composition, or a pharmaceutical composition.

In some embodiments, the vaccine composition or pharmaceutical composition comprises between 1 and 30 peptides, ie 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13 , 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29 or 30 different peptides.

In some embodiments, longer peptides can be designed in several ways. In one case, when the likelihood of peptide presentation on HLA alleles is predicted and/or identified or known, longer peptides may consist of either: (1) towards the N-terminus of each corresponding gene product and individually presented peptides extending 2-5 amino acids from the C-terminus; (2) concatenation of some or all of the presented peptides with the respective extension sequences. In another case, when sequencing revealed the presence of longer (>10 residues) neo-epitopes in the tumor (eg, caused by frameshifts, read-throughs, or intron inclusions to generate novel peptides), longer peptides Will consist of: (3) an entire stretch consisting of novel tumor-specific amino acids, thereby bypassing the need to select shorter peptides with the strongest HLA presentation based on computational or in vitro testing. In both cases, the use of longer chains enables endogenous processing in patient cells and can result in more efficient antigen presentation and induction of T cell responses.

In some embodiments, the peptides may include post-translational modifications.

Peptides can be prepared by any technique known to those of ordinary skill in the art, including expression of polypeptides or peptides by standard molecular biology techniques, isolation of peptides from natural sources, or chemical synthesis of peptides. Polynucleotides and proteins, polypeptides and peptides corresponding to various genes have been previously disclosed and can be found in computerized databases known to those of ordinary skill in the art. One such repository is the Genbank and GenPept repositories of the National Center for Biotechnology Information located on the National Institutes of Health website. The coding regions of known genes can be amplified and/or expressed using techniques disclosed herein or known to those of ordinary skill in the art. Alternatively, there are various commercially available formulations of peptides and peptide sequences known to those of ordinary skill in the art.

In some embodiments, the vaccine composition or pharmaceutical composition comprises between 1 and 100 or more polynucleotide sequences, ie 2, 3, 4, 5, 6, 7, 8, 9, 10 , 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35 , 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60 , 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85 , 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100 or more different polynucleotide sequences.

In some embodiments, the selection of different peptides or polynucleotides encoding the same enables these peptides to associate with different HLA molecules, such as different class I HLA molecules and/or different class II HLA molecules.

In some embodiments, the vaccine composition or pharmaceutical composition comprises a peptide capable of associating with the most commonly occurring class I HLA molecules and/or class II HAL molecules. For example, the vaccine composition may comprise different fragments capable of associating with at least 2, at least 3, or at least 4 HLA class I molecules and/or class II HLA molecules.

In some embodiments, the vaccine composition or pharmaceutical composition is capable of eliciting a specific cytotoxic T cell response and/or a specific helper T cell response.

In some embodiments, the vaccine composition further includes an adjuvant and/or carrier.

In some embodiments, the pharmaceutical composition further includes a pharmaceutically acceptable carrier.

In some embodiments, the composition may also be included in a viral vector-based vaccine or drug platform, such as vaccinia, fowlpox, self-replicating alphavirus, Maraba virus, adenovirus, or lentivirus.

In some embodiments, the composition can be included in liposomes.

In some embodiments, the composition may be administered parenterally, eg, intravenously, subcutaneously, intradermally, or intramuscularly. For example, compositions for intravenous injection, subcutaneous injection, intradermal injection, intraperitoneal injection, intraperitoneal injection, intramuscular injection can be prepared.

In some embodiments, a composition of the present disclosure comprises a solution of a peptide or polynucleotide encoding a peptide or portion thereof and is dissolved or suspended in an acceptable carrier, eg, an aqueous carrier. These compositions can be sterilized by well-known conventional sterilization techniques, or can be subjected to sterile filtration. The resulting formulation can be packaged for use as is, or lyophilized; the lyophilized formulation is administered after reconstitution.

In another aspect, the present disclosure provides a method for inducing tumor specificity in a subject by administering to the subject one or more antigens or peptides, such as multiple antigens or peptides predicted and/or identified using the methods disclosed herein Methods of immune response, vaccination against tumors, treatment, and or alleviation of symptoms of cancer in a subject.

In some embodiments, the tumor can be any solid tumor, such as breast tumor, ovarian tumor, prostate tumor, lung tumor, kidney tumor, stomach tumor, colon tumor, testicular tumor, head and neck tumor, pancreatic tumor, brain tumor, melanoma and other tissue and organ tumors; and hematological tumors, such as lymphomas and leukemias, including acute myeloid leukemia, chronic myeloid leukemia, chronic lymphocytic leukemia, T-cell lymphocytic leukemia, and B-cell lymphoma.

In some embodiments, the antigen or peptide can be administered alone or in combination with other therapeutic agents.

In some embodiments, the therapeutic agent is, for example, a chemotherapeutic agent, radiation or immunotherapy. Any suitable therapeutic treatment for a particular cancer can be administered.

In another aspect, the present disclosure provides a method of manufacturing a tumor vaccine, the method comprising performing the various steps of the method of the present disclosure; and producing a tumor vaccine comprising a plurality of antigens or peptides or a subset of the plurality of antigens or peptides. In some embodiments, the method of making a tumor vaccine comprises using By the step of identifying one or more antigens or peptides from one or more tumor cells of the subject that may be presented on the surface of the tumor cells.

In some embodiments, the method of making a tumor vaccine comprises the steps of: obtaining at least one of exome, transcriptome, or genome-wide tumor nucleotide sequencing data from tumor cells of the subject, wherein the tumor nucleotide sequenced The data is used to obtain profiles of peptides or combinations of peptides; the profiles of each peptide are logged into one or more machine learning systems to generate tumor cells of each of the peptides or combinations of peptides in the subject's tumor cells A numerical likelihood set ostensibly presented by one or more HLA alleles, the numerical likelihood set optionally identified based on the received mass spectrometry data; and selecting the peptide or peptides based on the numerical likelihood set A subset of combinations to produce a selected peptide or combination of peptides; and a tumor vaccine comprising the selected peptide or combination of peptides is produced or has been produced. The dosage of a compound or composition used in the methods of treatment and/or prevention of the present disclosure will generally vary with the severity of the disease, the weight of the patient, and the relative efficacy of the compound. However, as a general guide, a suitable unit dose may range from 0.1 to 1000 mg.

As is well known to those of ordinary skill in the art, the dosage of a drug to be administered depends on a variety of factors, including but not limited to the following factors: the activity of the particular compound used, the age of the patient, the weight of the patient, the medical condition of the patient, the behavior, patient's diet, administration time, administration mode, excretion rate, combination of drugs, etc.; in addition, the optimal treatment mode such as the mode of treatment, the daily dosage of the compound (I) or the pharmaceutically acceptable The type of salt can be verified according to traditional treatment protocols.

The present disclosure also provides a computer system that includes a computer processor and a memory storing computer program instructions that, when executed by the computer processor, cause the computer processor to execute embodiments of the above-described methods of the present disclosure.

In some embodiments, the computer system of the present disclosure is capable of predicting and/or identifying the likelihood of peptide presentation by HLA molecules based on a fusion neural network model and an ensemble learning model.

In some specific embodiments, the disclosed system includes:

Data encoding module: used to digitize data such as peptides, expression levels, and HLA typing;

Neural network training module: connected to the data coding module, using the fully connected neural network model to perform parameter training on mass spectrometry experimental data;

Integrated learning and training module: connected to the data coding module, using the integrated learning model to perform parameter training on mass spectrometry experimental data;

Immunogenicity Prediction and/or Identification Module: Linked to Neural Network Training Module and Ensemble Learning Module. Based on the peptide, expression level and HLA typing data of the test sample, according to the neural network coefficient matrix and the ensemble learning coefficient matrix, it is used to predict and/or identify the possibility of HLA molecule presentation.

The present disclosure also provides computer-readable media having computer-executable instructions stored thereon for implementing the methods described in the present disclosure.

In some embodiments, a computer-readable medium is provided having computer-executable instructions stored thereon for implementing the methods of the present disclosure for predicting and/or identifying the likelihood of peptide presentation by HLA molecules.

The present disclosure also provides an apparatus including a memory for storing a program and a processor for executing the program, so as to implement the method described in the present disclosure.

In some embodiments, an apparatus for predicting and/or identifying the likelihood of peptide presentation by an HLA molecule is provided, comprising a memory for storing a program and a processor for executing the program to achieve the prediction and/or the present disclosure. or identification method.

The present disclosure also provides an HLA allele-specific binding information database, which includes information on peptides predicted and/or identified by the methods described in the present disclosure and/or encodes peptides predicted and/or identified by the methods described in the present disclosure. peptide polynucleotide information.

The present disclosure also provides an HLA allele-specific binding peptide sequence database, including peptide sequence information predicted and/or identified according to the method of the present disclosure.

1 is a flowchart of a system and method for predicting the likelihood of peptide presentation by HLA molecules based on a fusion neural network model and an ensemble learning model.

Figure 2 shows the positive predictive value of the neural network model when the recall rate is 40%.

FIG. 3 is a comparison of the positive peptide characteristics of the present disclosure and mass spectrometry experimental data in Example 4. FIG.

FIG. 4 shows the positive predictive value when the recall rate is 40% after the positive peptide set of the machine learning model system is expanded in Example 6.

Figure 5 is a Western blot experiment of mass spectrometry samples.

Figure 6 is the silver staining experiment of mass spectrometry samples.

the term:

For easier understanding of this application, certain technical and scientific terms are specifically defined below. Unless otherwise clearly defined elsewhere in this document, all other technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

The three-letter and one-letter codes for amino acids used in this application are as in J. Biol. Chem, 243, p3558 (1968). An "antigen" is a substance that induces an immune response, including peptides or proteins that specifically react with antibodies or T lymphocytes (T cells). Include at least one antigen with a change that makes it different from the corresponding wild-type parent antigen, eg, the change is a tumor cell mutation or a tumor cell-specific post-translational modification. Antigens can include polypeptides or polynucleotides. Mutations may include frameshift or non-frameshift indels, missense or nonsense substitutions, splice site changes, genomic rearrangements or gene fusions, or mutations that produce tumor-specific open reading frames resulting from mutations or other abnormalities such as splicing. Any genomic or expression changes. Mutations can also include splice variants. Tumor cell-specific post-translational modifications can include aberrant phosphorylation. Tumor cell-specific post-translational modifications can also include proteasome-generated splicing antigens.

A "tumor antigen" is an antigen that is present in tumor cells or tissues in a subject but not in corresponding normal cells or tissues in the subject.

"Inducing an immune response" and "enhancing an immune response" are used interchangeably and refer to the stimulation (ie, passive or adaptive) of an immune response to a particular antigen. After inducing an immune response in an object, the object is protected from developing a disease (eg, cancer disease) or ameliorated by inducing an immune response. For example, an immune response to an antigen expressed by a tumor can be induced in a patient suffering from a cancerous disease or in an object at risk of developing a cancerous disease. In this case, induction of an immune response may mean that the disease status of the object has improved, that the object has not metastasized, or that the object at risk of developing cancer disease has not developed cancer disease.

"Immunogenicity" is the ability to elicit an immune response, eg, by T cells, B cells, or both.

"Peptide", "peptide sequence", "peptide fragment" and "polypeptide" are used interchangeably and refer to a chain of amino acid residues, usually of a defined sequence, including variants or fragments of peptides, which may be single bulk or aggregate. A "peptide" comprises any peptide comprising at least two amino acids joined by modified or unmodified peptide bonds, both L-amino acids and D-amino acids can be used. Peptides may contain modified amino acids, for example, may be modified by natural processes such as post-transcriptional modification or by chemical processes. Some examples of these modifications are: acetylation, acylation, ADP-ribosylation, acylation, covalent bonding to flavin, Covalent bonding of erythrocytes, covalent bonding to nucleotides or nucleotide derivatives, covalent bonding to modified or unmodified carbohydrate moieties, bonding to lipids or lipid derivatives, to phospholipids Inositol covalent bonding, cross-linking, cyclization, disulfide bond formation, demethylation, cysteine molecule formation, pyroglutamate formation, methylation, gamma-carboxylation, hydroxylation, iodine Methylation, methylation, oxidation, phosphorylation, racemization, hydroxylation, etc.

"Polynucleotide" or "nucleic acid" refers to a chain of nucleotides of any length, including DNA and RNA. Nucleotides can be deoxyribonucleotides, ribonucleotides, modified nucleotides or bases and/or their analogs, or any substrate that can be incorporated into the chain by DNA or RNA polymerase . Polynucleotides may comprise modified nucleotides, such as methylated nucleotides and analogs thereof. Modifications to the nucleotide structure, if present, can be imparted before strand assembly or after strand assembly. Polynucleotides may also contain analogous forms of ribose or deoxyribose sugars generally known in the art, including, for example, 2 followed by para-methyl-, 2-based para-allyl, 2-propyl fluoro- or 2-propane Nitro-ribose, carbocyclic sugar analogs, alpha- or beta-anomeric sugars, epimeric sugars (such as arabinose, xylose or lyxose, pyranose, furanose, sedum heptulose) , acyclic analogs and abasic nucleoside analogs such as methyl riboside.

"Antigen processing" or "processing" refers to the degradation (eg, polypeptide-to-peptide degradation) of a polypeptide or antigen into a processed product that is a fragment of the polypeptide or antigen and the interaction of one or more of these fragments with an MHC molecule Association (eg, by binding) is presented to specific T cells by cells, preferably antigen-presenting cells.

"Antigen-presenting cells" (APCs) are cells that display peptide fragments of protein antigens associated with MHC molecules on their cell surface. Some APCs activate antigen-specific T cells.

"HLA affinity" is the binding affinity between a particular antigen and a particular HLA allele.

By "vaccine" is meant an immune response, particularly a cellular immune response, which, upon administration, induces an immune response that recognizes and attacks pathogens or diseased cells such as cancer cells. Vaccines can be used to prevent or treat disease.

"Personalized cancer vaccine" refers to a particular cancer patient and means that the cancer vaccine is tailored to the needs or particular circumstances of the individual cancer patient.

A "neural network" is a machine learning model for classification or regression that consists of multiple layers of linear transformations followed by element-wise nonlinearity, usually trained by stochastic gradient descent and backpropagation.

"Ensemble learning" refers to combining multiple learning models to obtain better prediction results, so that the combined model has stronger generalization ability, or more universality.

"XGBoost" uses the ensemble learning idea for result/label prediction. XGBoost can generally be used to solve two kinds of problems, including classification problems and regression problems.

"Training set" refers to the set of samples used for training mainly used to train the parameters in the learning model.

"Validation set" refers to the sample set used to verify the performance of the model. After different learning models are trained on the training set, the performance of each model is compared and judged by the validation set.

"Program" or "computer program" generally refers to a syntactic unit conforming to the rules of a particular programming language, consisting of declarations and statements or instructions, which can be divided into "code fragments" required to solve or perform a certain function, task or problem.

"System" or "computer system" generally refers to one or more computers, peripherals, and software that process data. A "user" or "system operator" generally includes a person who accesses and uses a computer network via a "user device" (eg, computer, wireless device, etc.) for the purpose of data processing and information exchange.

A "computer" is usually a functional unit capable of performing substantial computation, including a large number of arithmetic and logical operations without human intervention.

"Application software" or "application" generally refers to software or programs designed to solve application problems.

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

"Nucleic acid" is deoxyribonucleic acid (DNA) or ribonucleic acid (RNA). According to the present disclosure, nucleic acid includes genomic DNA, cDNA, mRNA, recombinantly produced and chemically synthesized molecules. According to the present disclosure, nucleic acids can exist as single- or double-stranded and linear or covalently circularly closed molecules. Nucleic acids according to the present disclosure may be isolated. According to the present disclosure, the term "isolated nucleic acid" means that the nucleic acid is (i) amplified in vitro, eg, by polymerase chain reaction (PCR), (ii) produced by clonal recombination, (iii) purified , eg by cleavage and separation by gel electrophoresis, or (iv) synthesized eg by chemical synthesis. Nucleic acids can be introduced into (ie, transfected) cells, in particular, RNA forms that can be prepared from DNA templates by in vitro transcription. RNA can also be modified by stabilizing sequences, capping and polyadenylation prior to application.

"Preparation" includes preparation in large-scale production processes and laboratory preparations.

"Prediction" means a prospective determination or identification.

"Administration" and "treatment" when applied to animals, humans, experimental subjects, cells, tissues, organs, or biological fluids refer to the interaction of exogenous drugs, therapeutic agents, diagnostic agents, or compositions with the animal, human, subject, or biological fluid. contact with subjects, cells, tissues, organs or biological fluids. "Administering" and "treatment" can refer to, for example, therapeutic, pharmacokinetic, diagnostic, research, and experimental methods. Treatment of cells includes contact of reagents with cells, and contact of reagents with fluids, wherein the fluids are in contact with cells. "Administering" and "treating" also mean by an agent, a diagnosis, a binding composition, or by another cell body In vitro and ex vivo treatment of eg cells. "Treatment" when applied to human, veterinary or research subjects refers to therapeutic treatment, prophylactic or preventive measures, research and diagnostic applications.

"Treatment" means administering an internal or external therapeutic agent, such as a composition comprising any of the binding compounds herein, to a patient having one or more disease symptoms for which the therapeutic agent is known to have a therapeutic effect. Typically, the therapeutic agent is administered in an amount effective to alleviate one or more symptoms of a disease in a patient or population being treated, either by inducing regression of such symptoms or inhibiting the progression of such symptoms to any clinically measured extent. The amount of a therapeutic agent effective to relieve symptoms of any particular disease (also referred to as a "therapeutically effective amount") can vary depending on factors such as the patient's disease state, age and weight, and the ability of the drug to produce the desired effect in the patient. Whether symptoms of a disease have been alleviated can be assessed by any clinical test commonly used by physicians or other health care professionals to assess the severity or progression of the symptoms. Although embodiments of the present invention (eg, methods of treatment or articles of manufacture) may be ineffective in alleviating symptoms of a target disease in a single patient, the method of The U test, Kruskal-Wallis test (H test), Jonckheere-Terpstra checked sum Wilcoxon test determined that it should reduce symptoms of the target disease in a statistically significant number of patients.

"Conservative modification" or "conservative substitution or substitution" refers to the replacement of amino acids in proteins by other amino acids with similar characteristics (eg, charge, side chain size, hydrophobicity/hydrophilicity, backbone conformation and rigidity, etc.), This allows frequent changes without altering the biological activity of the protein. It is known to those of ordinary skill in the art that, in general, single amino acid substitutions in non-essential regions of polypeptides do not substantially alter biological activity (see, eg, Watson et al. (1987) Molecular Biology of the Gene, The Benjamin/Cummings Pub. Co., p. 224, (4th ed.). In addition, substitution of structurally or functionally similar amino acids is unlikely to disrupt biological activity.

The term "naturally occurring" as applied to an item refers to the fact that the item can be found in nature. For example, peptides or polynucleotides that are present in organisms (including viruses) that can be isolated from natural sources and have not been intentionally modified artificially in the laboratory are naturally occurring.

An "effective amount" includes an amount sufficient to ameliorate or prevent the symptoms or conditions of the medical condition. An effective amount also means an amount sufficient to allow or facilitate diagnosis. The effective amount for a particular patient or veterinary subject may vary depending on factors such as the condition being treated, the general health of the patient, the method, route and dosage of administration, and the severity of side effects. An effective amount can be the maximum dose or dosing regimen that avoids significant side effects or toxic effects.

"Exogenous" refers to a substance produced outside an organism, cell, or human body depending on the context. "Endogenous" refers to a substance produced in a cell, organism or human body depending on the context.

"Homology" or "identity" refers to the sequence similarity between two polynucleotide sequences or between two polypeptides. When a position in the two compared sequences is occupied by the same base or amino acid monomer subunit, for example if every position in two DNA molecules is occupied by an adenine, then the molecules are homologous at that position of. The percent homology between the two sequences is a function of the number of matches or homologous positions shared by the two sequences divided by the number of positions compared x 100%. For example, when sequences are optimally aligned, two sequences are 60% homologous if 6 of 10 positions in the sequences are matched or homologous. In general, comparisons are made when the two sequences are aligned for the greatest percent homology. As used herein, "at least 85% sequence identity" means that the variant has at least 85% homology between the two sequences compared to the parental sequence, in some embodiments, it has at least 86%, 87%, 88%, 89%, 90%, 91%, 92%, 93%, 94%, 95%, 96%, 97%, 98%, or 99% sequence homology; in some specific schemes, it has 90%, 95% or 99% % or more; in other specific schemes, it has at least 95% sequence homology. the amine group having at least 85% sequence identity Acid sequences include those obtained by mutagenesis of one or more amino acid deletions, insertions or substitutions of the parental sequence.

As used herein, the expressions "cell", "cell line" and "cell culture" are used interchangeably and all such designations include progeny thereof. Thus, the words "transformants" and "transformed cells" include primary test cells and cultures derived therefrom, regardless of the number of transfers. It should also be understood that, due to deliberate or unintentional mutations, all progeny may not be exactly the same in terms of DNA content. Mutant progeny that have the same function or biological activity as screened in the original transformed cell are included.

"Optional" or "optionally" means that the subsequently described event or circumstance can but need not occur, and that the description includes instances where the event or circumstance occurs or does not occur. For example, "optionally comprising 1-3 antibody heavy chain variable regions" means that an antibody heavy chain variable region of a particular sequence may, but need not, be present.

"Pharmaceutical composition" means a mixture containing one or more of the compounds described herein, or a physiologically/pharmaceutically acceptable salt or prodrug thereof, with other chemical components, and other components such as a physiological/pharmaceutically acceptable carrier and excipients. The purpose of the pharmaceutical composition is to facilitate the administration to the organism, facilitate the absorption of the active ingredient and then exert the biological activity.

The term "vector" refers to a nucleic acid molecule capable of transporting another nucleic acid to which it has been linked. In one embodiment, the vector is a "plasmid," which refers to a circular double-stranded DNA loop into which additional DNA segments can be ligated. In another embodiment, the vector is a viral vector in which additional DNA segments can be ligated into the viral genome. The vectors disclosed herein are capable of autonomous replication in the host cells into which they have been introduced (eg, bacterial vectors and episomal mammalian vectors with a bacterial origin of replication) or may integrate into the host cell's genome after introduction into the host cell, thereby following The host genome replicates together (eg, a non-episomal mammalian vector).

The present disclosure is further described below with reference to the embodiments, but these embodiments do not limit the scope of the present disclosure. The experimental methods that do not specify specific conditions in the embodiments of the present disclosure generally follow conventional conditions, such as Cold Spring Harbor Antibody Technology Experiment Manual, Molecular Cloning Manual; or conditions suggested by raw material or commodity manufacturers. Reagents with no specific source indicated are conventional reagents purchased in the market.

Example 1. Machine learning model system construction

Dataset construction process:

First, the immunogenicity data and expression data of peptides obtained by mass spectrometry experiments were obtained from public databases such as IEDB, SYFPEITHI and MassIVE, and only the peptides with a length of 8-11 amino acids and expressed were retained, and a total of 154,103 mass spectra were obtained. Peptide positive data, as a positive peptide set.

Then, find all the expressed protein data of the human reference genome hg38 from the UCSC database, and exhaustively cut it into 8-11 amino acid lengths as a negative peptide set. At the same time, the UCSC database was used to obtain the annotation information of each peptide, and the BioMart website was used to convert the annotation information into the protein family ID (ie the PANTHER family ID column) to which each peptide belongs.

Finally, the positive peptide set and the negative peptide set are mixed, and then divided into two categories, one is the test data set (test set), which includes a multi-allele test set and a monoallele test set. The two test sets are calculated separately. The second category is the training data set (training set), which is used for model training. The multiallelic test set contained 5,000,500 peptides, the monoallelic test set contained 32,003,200 peptides, and the training set contained 105,724,241 peptides.

The dataset construction process is consistent in neural network model training and ensemble learning model training.

Model training optimization process:

After a lot of experiments, in this embodiment, the neural network model and the ensemble learning model are selected to be fused to form a final optimal model, and a threshold is set to make the regression model have the ability to classify. The neural network model is constructed in Example 2, with relu as the startup function and a fully connected neural network model initialized by Lecun.

Before model training, the data in the data set needs to be preprocessed, the HLA typing and peptide sequence amino acids are converted into one-hot encoding, and the expression data and the protein family ID to which the peptide belongs are input into the model for training.

The CentOS 7.0 operating system was used during training, and python 2.17.15 and Perl 5.10 were used as programming languages. A fully connected network built with Keras 2.2.4 and TensorFlow 1.15.0 was used as the neural network architecture, and XGBoost 0.82 was used as the ensemble learner for training. The fully connected network architecture is 4 layers, and the number of neurons in each layer is 256, 32, 16, and 1 in turn. Binary crossentropy is used as the loss function, and Adam is used as the optimizer for parameter optimization. XGBoost uses the Bayesian method to determine the maximum depth of the classification tree (max depth), the minimum sum of instance weights required for a child node (min child weight), and the subsampling ratio of columns when constructing each tree (colsample bytree) , the minimum loss reduction (gamma) required for division on leaf nodes, the maximum number of weak learners (n estimators), the learning rate (learning rate) and the subsample ratio of training instances (subsample) and other parameters to optimize the search.

During the training process, 10% of the training set was taken out to verify the performance of the current training model, and repeated ten times to make the verification data cover all the training sets. After ten trainings are completed, ten models are used to predict the multi-allele test set, and the test set recall rate (recall) = 0.4 is used to define a threshold, and the PPV on the test set is used as a better target to optimize the model parameters and architecture. .

After training and optimization, each example uses 1:1000 as the mixing ratio of positive and negative data in the training set, uses relu as the neural network startup function, and uses Lecun to initialize As a neural network initializer, it multiplies the score of the output result of the ensemble learning model in the neural network model by a weight coefficient of 0.01 to form the final optimal model-based fusion model.

For the prediction application stage, first preprocess the data set to be predicted in the same model construction stage, log the transformed data to be predicted into the model for prediction, use the threshold for classification, and output the immunogenicity classification result of the data to be predicted. Peptides above the threshold were predicted and/or identified as being presented by HLA molecules.

Example 2. Neural network model construction

Five groups of positive peptides with HLA type A*02:07 from public databases IEDB, SYFPEITHI and MassIVE were used, with an average of 586.8 in each group, and the negative control peptide was added at a ratio of 1:10,000.

Use the fully connected neural network model with relu as the startup function and Lecun initialization in the present invention, and the fully connected neural network that does not use relu as the startup function (uses tanh as the startup function) and does not use Lecun initialization (does not initialize) The model makes predictions of the likelihood of HLA molecule presentation on the test dataset. The sample detection results are shown in Table 1 below. The results show that the average detection accuracy of the five groups of test data using relu as the startup function and the fully connected neural network initialized by Lecun has increased by more than 1 times (as shown in Figure 2). Among them, true positives represent peptides predicted to be positive and positive in the database; false negatives represent peptides predicted to be negative and positive in the database; false positives represent peptides predicted to be positive and negative in the database; true negatives represent Peptides predicted to be negative and negative in the database.

Example 3. Prediction of the likelihood of HLA molecule presentation by fusion model single HLA allele genotyping

Five groups of positive peptides with HLA type A*02:07 from public databases IEDB, SYFPEITHI and MassIVE were used, with an average of 586.8 in each group, and the negative control peptide was added at a ratio of 1:10,000.

The public software netMHCpan and MHCflurry and the fusion model obtained in Example 1 were used to predict the probability of HLA molecule presentation on the test data set. The sample detection results are shown in Table 2 below. The results show that, compared with the published software results, the average detection accuracy of the five sets of test data for the fusion model obtained in Example 1 is improved by more than 10 times. Among them, true positives represent peptides predicted to be positive and positive in the database; false negatives represent peptides predicted to be negative and positive in the database; false positives represent peptides predicted to be positive and negative in the database; true negatives represent Peptides predicted to be negative and negative in the database.

Some examples of test results are shown in Table 3. The second column is the immunogenicity of the corresponding peptides predicted by the fusion model obtained in Example 1 under A*02:07 typing. The model was positive for the peptide immunogenicity.

Example 4. Prediction of the likelihood of HLA molecule presentation by 16-group single HLA allele typing of the fusion model

The negative control peptides were added at a ratio of 1:10,000 using 200 mass spectrometry positive peptide data in each group (3200 in total) of 16 groups of different HLA types from the MassIVE database. Use the fusion model obtained in Example 1 to predict the possibility of HLA molecule presentation on the test data set. The sample detection results are shown in Table 4 below. Using the positive predictive value when the recall rate is 40% as the evaluation standard, the accuracy of this disclosure is increased to 52.52%, which is significantly improved compared to the accuracy of the public software.

Using the mass spectrometry positive peptide data of 2 groups of different HLA types (A*24:02, A*02:01) and the positive peptide data predicted by the present invention, respectively draw the positive peptide characteristic map on each HLA type, and the results are as follows As shown in FIG. 3 , the characteristics of the positive peptides learned in this example are similar to the mass spectrometry data, indicating that the learning results are good.

Example 5. Multi-HLA allele genotyping mass spectrometry data for HLA molecule presentation likelihood prediction

This example identifies peptides from positive and negative control samples, with two replicates each, by the following methods:

Using 2.5× ^{10 8} A375 cells, lyse with 5 ml of cell lysis buffer (20 mM Tris 8.0, 1 mM EDTA, 100 mM NaCl, 1% Triton X-100) for 1 hour at 4°C in a cold storage shaker. The protein lysate was then pre-cooled in a centrifuge at 4°C at 14,000 rpm for 30 minutes. After the protein lysate was quantified for BCA protein, 30μL protein A/G magnetic beads (Protein A/G beads), 6μg antibody (positive sample) or 6μg antibody (negative control) were used for co-immunoprecipitation sample mixture according to the total protein amount of 4mg. (co-IP samples, HLA-related complexes) were prepared, and the samples from co-IP were placed in a 4°C freezer with rotation and mixed evenly overnight. Subsequently, magnetic beads (beads) deposited by HLA-associated complexes were washed 4 times with co-IP wash solution (20 mM Tris 8.0, 1 mM EDTA, 100 mM NaCl). Finally, acetic acid at a concentration of 1-2 mol/L was incubated with the beads deposited by HLA-related complexes at room temperature for 10 min, the HLA complexes were washed off the beads, and centrifuged at 1000 rpm for 1 min. The collected supernatant is the sample for MS analysis.

Mass spectrometry samples will be used to identify isolated peptides for HLA-presenting antigenic peptides using LC/MS. Data obtained by LC/MS were analyzed using Mascot (V 2.3.0) software. The retrieval parameters are set as follows: the database is set to SwissProt Homo_sapiens respectively; the enzyme digestion method is set to None; the missed cleavage site is set to 0; The main instruments used include Thermo Scientific EASY-nLC 1200 UPLC Thermo nanoliter ultra-high performance liquid chromatograph, Thermo Scientific Q ExactiveTM HF-X quadrupole-orbital hydrazine mass spectrometer and Thermo Scientific C18 chromatographic column (1.9 μm, 250 mm × 100 μm i.d.).

Using the multi-HLA typing positive peptide data obtained by the above mass spectrometry experiment and the negative control peptide data obtained by the negative control, the fusion model obtained in Example 1 was used to predict the immunogenicity of the antigen, and the built-in threshold was used to judge the immunogenicity. The built-in threshold indicates that the recall rate of the training set is 40%, so the prediction result of the positive data is also based on the recall rate as the evaluation standard. The test results are shown in Table 5 below. The recall rate in the test sample is greater than 40%, and the recall rate in the negative control sample is much lower than 40%, indicating that the present invention has good application results on multi-HLA typing data.

The test results and statistics show that the antigen immunogenicity prediction method based on the machine learning method of the present disclosure can accurately predict the immunogenicity of the antigen, and the positive predictive value is significantly higher than the current public software.

Example 6. Mass spectrometry analysis for high-frequency HLA typing in the population to expand the positive peptide set of the machine learning model system

In this example, the following methods were used to identify positive polypeptides that were not reported in public databases and that frequently appeared in the human population for HLA typing binding. This example mainly includes the following three steps: construction of a single HLA typing cell line, HLA mass spectrometry sample preparation and quantification, and mass spectrometry analysis.

Lentivirus-infected K562 cells were used to construct stable transfected cell lines with single HLA typing. According to the HLA expression levels of different typing cell lines, the amount of cells prepared for mass spectrometry samples was adjusted. Tris 8.0, 1mM EDTA, 100mM NaCl, 1% Triton X-100) 5-20ml, lysed at 4°C for 1 hour, pre-cooled at 4°C and centrifuged at 14,000 rpm for 30 minutes; after the protein lysate was quantified by BCA protein, Prepare co-immunoprecipitation sample mixture according to the ratio of 4mg total protein corresponding to 30μL Protein A/G beads, 6μg MHC-I antibody (positive sample) or 6μg normal mouse IgG antibody (negative control). -IP samples, HLA-associated complexes). The co-IP samples were mixed homogeneously at 4°C overnight, then washed with wash buffer (20 mM Tris 8.0, 1 mM EDTA, 100 mM NaCl) for 4 washes of co-IP beads. Finally, incubate with co-IP beads with a concentration of 1-2mol/L acetic acid for 10min, wash out the HLA complex, and centrifuge at 1000rpm for 1min. The supernatant was collected as the sample for mass spectrometry analysis (MS sample).

5% of the samples analyzed by mass spectrometry were taken out for western blotting and silver staining to qualitatively evaluate the specificity of mass spectrometry samples and quantitatively analyze the total amount of HLA-related protein complexes, respectively. First, the samples were premixed with SDS-PAGE protein loading buffer (5X) and boiled at 100°C for 5 minutes; then the samples were divided into two parts for SDS-PAGE electrophoresis. One of them was transferred to PVDF membrane, incubated with rabbit anti-human MHC-I antibody, and HRP-conjugated anti-rabbit IgG secondary antibody for western blotting. As shown in Figure 5, by comparing the normal mouse IgG group (negative control) and the MHC-I antibody group (positive sample), MHC-I signal appeared in the positive sample group, but no signal in the negative control group (Figure 5, lane #6, # 7), indicating that the sample preparation process is specific; by comparing the HLA protein content in the cell lysate before and after co-IP (Figure 5, lane #1, #4), the co-IP enrichment efficiency can be calculated. In Figure 5, lane #1 is the protein lysate sample before Co-IP, lane #2 is Marker, lane #3 is the protein lysate after IgG Co-IP, and lane #4 is the protein lysis after HLA antibody Co-IP solution, lane #5 is an empty lane, lane #6 is normal mouse IgG group (negative control), and lane #7 is MHC-I antibody group (positive sample). The results show that the current experimental system can effectively enrich more than 90% of HLA proteins in the protein lysate, indicating that the sample preparation process is efficient.

Another SDS-PAGE was directly stained with silver. By comparing the signal values of 1ng, 5ng, and 10ng protein standard BSA, the total protein amount of the mass spectrometry sample could be calculated to assist mass spectrometry analysis. As shown in Figure 6, lane #1 is Marker, lane #2 is the negative control of mass spec sample 1, lane #3 is the positive control of mass spec sample 1, lane #4 is the negative control of mass spec sample 2, and lane #5 is mass spec sample 2 The positive controls for , lane #6 is Marker, lane #7 is 1 ng protein standard (BSA), lane #8 is 5 ng protein standard (BSA), and lane #9 is 10 ng protein standard (BSA). Arrows indicate MHC-I antibody group (positive samples) specific enriched proteins, in which the bar near 45KDa The band was identified as HLA protein (arrow HLA), and a total protein amount of 1.0-12.4ug was obtained in this experiment.

Mass spectrometry samples will be used to identify isolated peptides for HLA-presenting antigenic peptides using LC/MS. Data obtained by LC/MS were analyzed using Mascot (V 2.3.0) software. The retrieval parameters are set as follows: the database is set to SwissProt Homo_sapiens respectively; the enzyme digestion method is set to None; the missed cleavage site is set to 0; The main instruments used include Thermo Scientific EASY-nLC 1200 UPLC Thermo nanoliter ultra-high performance liquid chromatograph, Thermo Scientific Q ExactiveTM HF-X quadrupole-orbital hydrazine mass spectrometer and Thermo Scientific C18 chromatographic column (1.9 μm, 250 mm × 100 μm i.d.).

As shown in Figure 4, using these positive peptides to retrain the fusion model, the resulting model performs better than the model without addition, with an accuracy improvement of about 15%, and the positive predictive value is significantly higher than the current public software.

The above are only specific application examples of the present invention, and do not constitute any limitation to the protection scope of the present invention; for those with ordinary knowledge in the technical field, other different forms of changes or modifications can also be made on the basis of the above description. It is not necessary and impossible to enumerate and describe all the embodiments here. Any similar technical solutions formed by equivalent transformation or equivalent replacement fall within the scope of the patent application of the present invention.

Claims (23)

A method of predicting and/or identifying the likelihood of peptide presentation by HLA molecules, comprising: Step 1: Construct data sets and data coding, including digital coding of peptides, expression levels, and HLA typing data in the data set; Step 2: Train the neural network model and the integrated learning model, including parameter training on the data provided by the mass spectrometry experiment; Step 3: Predict and/or identify the possibility of peptide presentation by HLA molecules, including fusion neural network model coefficient matrix and ensemble learning model coefficient matrix, predict and/or identify according to the peptide, expression level and HLA typing data of the sample to be tested The possibility that each peptide of the sample to be tested is presented by HLA molecules. The method of claim 1, wherein the neural network model is selected from a fully connected neural network model, a convolutional neural network model, and a long short-term memory neural network model, preferably a fully connected neural network model. The method according to claim 1 or 2, wherein the neural network model architecture is 3-5 layers; preferably, the neural network model architecture is 4 layers, and the number of neurons in each layer is 256, 32, 16 and 1. The method of any one of claims 1 to 3, wherein the ensemble learning model is selected from random forest, Adaboost, XGboost, LightGBM, preferably XGboost. The method of any one of claims 1 to 4, wherein the data set includes a positive peptide set and a negative peptide set, divided into one or more training data sets and one or more test data sets. The method of any one of claims 1 to 5, wherein the length of the peptides in the dataset is 5-15 amino acids, preferably 7-12 amino acids, most preferably 8-11 amino acids . The method of any one of claims 1 to 6, wherein the data encoding is selected from one-hot encoding, BLOMAP, PSSM, word2vec and BLOSUM62, preferably one-hot encoding. The method of any one of claims 1 to 7, wherein the neural network model performs parameter training on mass spectrometry experimental data: Suppose there are m training peptides, each corresponding to n HLA types:

in

is the data matrix coded by the data coding module for each training peptide without HLA typing, and each row in the matrix is a data row coded by the data coding module for one peptide;

is the positive/negative label of each training peptide in the mass spectrometry experiment; β _{target 1} is the neural network coefficient matrix, which is obtained through cross-validation training when the accuracy is the largest. The method of any one of claims 1 to 8, wherein the activation function of the neural network model is relu. The method of any one of claims 1 to 9, wherein the initializer of the neural network model is a Lecun initializer. The method of any one of claims 1 to 10, wherein the ensemble learning model performs parameter training on mass spectrometry experimental data: Suppose there are m training peptides, each corresponding to n HLA types:

in

is a data matrix containing HLA typing for each training peptide encoded by the data, and each row in the matrix is a data row of a peptide encoded by the data;

is the positive/negative label of each training peptide in the mass spectrometry experiment; β _{target 2} is the ensemble learning coefficient matrix, and the coefficient when the accuracy is the maximum is obtained through cross-validation training. The method according to any one of claims 1 to 11, wherein the strategy of fusing the neural network model and the ensemble learning model is selected from averaging method, voting method and learning method; preferably, the fusion comprises combining the ensemble learning model The score of the output result in the neural network model is multiplied by the weight factor. The method according to claim 12, wherein the weight coefficient is 0.1-0.00001, preferably 0.05-0.0001, and optimally 0.01. The method of any one of claims 1 to 13, wherein the prediction and/or identification is selected from single HLA allele prediction and/or identification, multi-allele prediction and/or identification. The method of any one of claims 1 to 14, wherein the predicting and/or identifying an expected value of the probability of presentation of the peptide by HLA molecules

,Specifically:

Wherein X is the data line obtained after each peptide is encoded by the data. The method according to any one of claims 1 to 15, wherein the 0.1% with the highest score is selected as the positive value or the positive predictive value under 40% recall rate is used as the threshold value, and the peptides in the sample to be tested that are higher than the threshold value are predicted and/or Identified as being presented by HLA molecules. A method of preparing a peptide predicted and/or identified by a method as in any one of claims 1 to 16. A method of preparing a polynucleotide encoding a peptide prepared by the method of claim 17. An application of the method as described in any one of claims 1 to 18 in the preparation of mRNA, polypeptide vaccine, antitumor drug or tumor vaccine, preferably, the tumor is selected from the group consisting of: lung cancer, melanoma, Breast cancer, ovarian cancer, prostate cancer, kidney cancer, stomach cancer, colon cancer, testicular cancer, head and neck cancer, pancreatic cancer, brain cancer, B-cell lymphoma, acute myeloid leukemia, chronic myeloid leukemia, chronic lymphocytic leukemia and T Cell lymphocytic leukemia, non-small cell lung cancer, and small cell lung cancer. A device for predicting and/or identifying the likelihood of peptide presentation by HLA molecules, characterized in that the device comprises a memory for storing a program and a processor for executing the program to achieve any of claims 1 to 16 one of the methods described. A computer-readable storage medium is characterized by comprising a program, the program being executed by a processor to perform the method according to any one of claims 1 to 16. A method for predicting and/or identifying the possibility of one or more peptides being presented by HLA molecules, comprising the steps of: constructing a data set and data encoding; training, fusing a neural network model and an ensemble learning model; predicting and/or identifying peptides by HLA molecule presentation possibilities. A database of HLA allele-specific binding information comprising information or coding of peptides predicted and/or identified by the method as described in any one of claims 1 to 16 or encoded as in any one of claims 1 to 16 The method predicts and/or identifies the polynucleotide profile of the peptide.

Images