Training method of antigen presentation prediction model, prediction method, device and medium thereof

Abstract

The present disclosure provides a training method of an antigen presentation prediction model and a prediction method, device and medium thereof, the training method comprising: obtaining antigen sequence training data and presentation probability training data for characterizing the presentation probability of the antigen sequence; inputting the antigen sequence training data into the antigen presentation prediction model to be trained; performing presentation prediction processing on the antigen sequence training data by the antigen presentation prediction model to predict the presentation probability corresponding to the antigen sequence training data, wherein the antigen sequence training data is subjected to amino acid encoding and vector mapping processing by a feature encoding module of the antigen presentation prediction model; and training the antigen presentation prediction model according to the predicted presentation probability and the presentation probability training data. The present disclosure effectively establishes a neural network algorithm model based on antigen sequence by using deep learning technology, effectively solves the problem of poor prediction of new antigen presentation ability, and improves the accuracy and efficiency of predicting antigen presentation ability.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of artificial intelligence, and in particular, to a method for training an antigen presentation prediction model, a method for predicting the antigen presentation prediction model, a device and a medium. BACKGROUND

[0002] Recent successes in immunotherapy have validated a long-standing hypothesis that tumors are immunogenic and can elicit an adaptive immune response. Major histocompatibility complex (MHC) or human leukocyte antigen (HLA) in tumor cells can present tumor-specific antigens, which can be recognized by CD8+ (a white blood cell differentiation antigen) killer T cells. Researchers have been exploring the nature of antigens that can stimulate T cells to produce effective responses to tumors for many years.

[0003] Tumor antigens can be divided into two categories: self and non-self antigens. Tumor-associated self antigens include cancer-testis (germline) antigens, differentiation antigens and antigens overexpressed by cancer cells. The advantage of self antigens is that they are shared among different tumor patients and tumors. However, these self antigens are limited by immune tolerance, so they can only elicit weak or no T cell response. The accumulation of nonsynonymous somatic mutations in tumors has long been hypothesized to be non-self antigens (also known as tumor neoantigens) recognized by the adaptive immune system, so as to elicit a more intense immune response. The first report of tumor neoantigens has been twenty years, but the comprehensive identification of neoantigens has always been a very difficult thing, which has been improved with the advent of next-generation sequencing technology. With the widespread use of next-generation sequencing, several studies have shown that nonsynonymous mutations are highly related to immunotherapy. The mutation load in the tumor can predict the response of patients to immune checkpoint inhibitor therapy, and T cells specific to neoantigens are detected in an increasing number of tumor patients. Direct evidence of the role of neoantigens in tumors comes from the fact that tumor-infiltrating lymphocyte (TIL) therapy targeting neoantigens caused the regression of metastatic tumors in patients. In addition, in preclinical studies, therapeutic vaccines targeting neoantigens caused the regression of tumors. In addition to nonsynonymous mutations, there are other sources of neoantigens, including post-translational modifications of polypeptides due to tumor-specific splicing or changes in MHC I presentation mechanisms in tumor cells, endogenous retroviruses and other viral infections found in cancer, etc.

[0004] Tumor neoantigens are important drivers of anti-tumor immunity, and their identification has important clinical applications, including as predictive and pharmacodynamic biomarkers, design of novel individualized tumor immunotherapies, including TIL therapy and vaccines. However, a major challenge is that only a small fraction of mutation-introduced antigens are MHC-presented, and of the presented antigens, only a small fraction are immunogenic. Because antigen immunogenicity validation requires significant manpower and resources, the throughput is small, resulting in less data accumulated, which further leads to difficulty in establishing an immunogenicity prediction model. Whether an antigen can bind to HLA and be presented to the cell surface can be verified in high throughput and a large amount of data has been accumulated, so more laboratories hope to select immunogenic antigens by predicting antigen presentation or binding to HLA. The study of Tumor Neoantigen Selection Alliance (TESLA) shows that antigen presentation mainly includes the characteristics of antigen and HLA affinity and expression amount. Although there are many HLA affinity software and the accuracy has reached a certain level, there are few antigen presentation prediction software integrating multiple characteristics, and the accuracy needs to be evaluated and determined by more data.

[0005] However, the current antigen presentation prediction algorithm is not very good for antigen presentation prediction, and is poor in external mass spectrometry data independent evaluation. SUMMARY

[0006] The main purpose of the present disclosure is to provide a training method and prediction method, device and medium of an antigen presentation prediction model to improve the above-mentioned defects in the prior art.

[0007] The present disclosure solves the above technical problems by the following technical solutions:

[0008] As an aspect of the present disclosure, a training method of an antigen presentation prediction model is provided, comprising:

[0009] obtaining antigen sequence training data and presentation probability training data for representing the presentation probability of the antigen sequence;

[0010] inputting the antigen sequence training data into an antigen presentation prediction model to be trained;

[0011] performing presentation prediction processing on the antigen sequence training data by the antigen presentation prediction model to predict the presentation probability corresponding to the antigen sequence training data, wherein the antigen sequence training data is subjected to amino acid coding and vector mapping processing by a feature coding module of the antigen presentation prediction model; and

[0012] training the antigen presentation prediction model according to the predicted presentation probability and the presentation probability training data.

[0013] As an optional implementation, the antigen sequence training data includes polypeptide sequences mainly composed of amino acids, short peptide sequences, and pseudo sequences corresponding to the short peptides (when encoding MHC sequences, only those residues that bind tightly with the peptides are considered, which form a short sequence called MHC pseudo sequence (MHC Pseudo sequence)).

[0014] The presentation probability training data includes label values (0 or 1, 0 represents poor presentation, and 1 represents strong presentation).

[0015] As an optional implementation, the feature encoding module includes an embedding module (a data processing module).

[0016] As an optional implementation, the step of performing presentation prediction processing on the antigen sequence training data by the antigen presentation prediction model to predict the presentation probability corresponding to the antigen sequence training data further includes:

[0017] The first feature extraction module of the antigen presentation prediction model performs polypeptide sequence one-dimensional convolution feature extraction processing on the antigen sequence training data after the amino acid encoding and vector mapping processing.

[0018] As an optional implementation, the first feature extraction module includes a one-dimensional CNN (Convolutional Neural Network) module.

[0019] As an optional implementation, the one-dimensional CNN module includes three parallel convolution structures.

[0020] As an optional implementation, the three parallel convolution structures include a convolution layer with a convolution kernel size of 1*1, 1*5, and 1*9.

[0021] As an optional implementation, the step of performing presentation prediction processing on the antigen sequence training data by the antigen presentation prediction model to predict the presentation probability corresponding to the antigen sequence training data further includes:

[0022] The second feature extraction module of the antigen presentation prediction model performs polypeptide sequence ordered feature extraction processing on the antigen sequence training data after the polypeptide sequence one-dimensional convolution feature extraction processing.

[0023] As an optional implementation, the second feature extraction module includes an RNN (Recurrent Neural Network) module.

[0024] As an optional implementation, the RNN module includes at least two layers of BiGRU (a neural network structure).

[0025] As an optional implementation manner, the step of performing presentation prediction processing on the antigen sequence training data by the antigen presentation prediction model to predict a presentation probability corresponding to the antigen sequence training data further includes:

[0026] The antigen sequence training data subjected to the ordered feature extraction processing on the polypeptide sequence is subjected to feature fusion processing by a feature fusion module of the antigen presentation prediction model.

[0027] As an optional implementation manner, the feature fusion module includes a full connection module.

[0028] As an optional implementation manner, before the step of inputting the antigen sequence training data into the antigen presentation prediction model to be trained, the training method further includes:

[0029] The obtained antigen sequence training data is preprocessed.

[0030] As an optional implementation manner, the preprocessing includes at least one of invalid data filtering processing, string sequence digitization processing, and normalization processing.

[0031] As an optional implementation manner, the method further includes:

[0032] Obtaining antigen sequence test data;

[0033] Inputting the antigen sequence test data into the trained antigen presentation prediction model to test the trained antigen presentation prediction model to obtain the trained and tested antigen presentation prediction model.

[0034] As another aspect of the present disclosure, an antigen presentation prediction method is provided, including:

[0035] Obtaining antigen sequence data;

[0036] Inputting the antigen sequence data into an antigen presentation prediction model trained by the antigen presentation prediction model training method as described above;

[0037] Performing presentation prediction processing on the antigen sequence data by the trained antigen presentation prediction model; and,

[0038] Outputting a predicted presentation probability corresponding to the antigen sequence data.

[0039] As another aspect of the present disclosure, an antigen presentation prediction method is provided, including:

[0040] Obtaining antigen sequence data;

[0041] Inputting the antigen sequence data into an antigen presentation prediction model;

[0042] performing antigen sequence data on the antigen presentation prediction model to predict the presentation probability corresponding to the antigen sequence data, wherein the antigen sequence data is subjected to amino acid encoding and vector mapping processing by a feature encoding module of the antigen presentation prediction model; and

[0043] outputting the predicted presentation probability corresponding to the antigen sequence data.

[0044] As an optional implementation, the antigen sequence data includes polypeptide sequences, short peptide sequences, and pseudo sequences corresponding to short peptides, which are mainly composed of amino acids.

[0045] As an optional implementation, the step of performing antigen sequence data on the antigen presentation prediction model to predict the presentation probability corresponding to the antigen sequence data further includes:

[0046] performing polypeptide sequence one-dimensional convolution feature extraction processing on the antigen sequence training data subjected to amino acid encoding and vector mapping processing by a first feature extraction module of the antigen presentation prediction model;

[0047] performing polypeptide sequence ordered feature extraction processing on the antigen sequence training data subjected to polypeptide sequence one-dimensional convolution feature extraction processing by a second feature extraction module of the antigen presentation prediction model.

[0048] As an optional implementation, the step of performing antigen sequence data on the antigen presentation prediction model to predict the presentation probability corresponding to the antigen sequence data further includes:

[0049] performing feature fusion processing on the antigen sequence training data subjected to polypeptide sequence ordered feature extraction processing by a feature fusion module of the antigen presentation prediction model.

[0050] As another aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the training method of the antigen presentation prediction model or the antigen presentation prediction method as described above when executing the computer program.

[0051] As another aspect of the present disclosure, a computer readable medium is provided, which stores computer instructions, wherein the computer instructions implement the training method of the antigen presentation prediction model or the antigen presentation prediction method as described above when executed by a processor.

[0052] According to the present disclosure, other aspects of the present disclosure can be understood by those skilled in the art.

[0053] The positive progress effect of the present disclosure is that:

[0054] The disclosure effectively establishes a neural network algorithm model based on antigen sequences by using deep learning technology. After a large amount of antigen sequence training data and corresponding presentation probability data (label data) representing the presentation ability are trained, the model can mine the deep features of the input antigen sequence and learn the one-to-one correspondence between the deep features and the label. After the model completes automatic optimization of parameters, the model has the prediction and decision ability for the input antigen sequence data, thereby effectively solving the problem of poor prediction of new antigen presentation ability and improving the accuracy and efficiency of predicting antigen presentation ability.

[0055] The disclosure proposes to use the wordembedding method commonly used in NLP (natural language processing) to initialize the vector of amino acid sites, which is suitable for the corresponding scene of the data, so that the initialization vector of the amino acid can express more meaningful.

[0056] The disclosure considers using a one-dimensional CNN neural network structure with stronger feature extraction ability to extract the features of the sequence.

[0057] The disclosure uses an RNN module, a neural network technology for sequence modeling with memory function, to consider the correlation of the sites in the sequence and the integrity of the sequence. BRIEF DESCRIPTION OF DRAWINGS

[0058] The features and advantages of the disclosure can be better understood after reading the detailed description of embodiments of the disclosure in conjunction with the following drawings. In the drawings, the components are not necessarily drawn to scale, and components with similar related properties or features can have the same or similar reference numerals.

[0059] Figure 1 A flowchart of a training method of an antigen presentation prediction model according to an embodiment of the disclosure.

[0060] Figure 2 A flowchart of a training model scenario according to an optional embodiment of the disclosure.

[0061] Figure 3 A structure diagram of an antigen presentation prediction model according to an optional embodiment of the disclosure.

[0062] Figure 4 A flowchart of an antigen presentation prediction method according to another embodiment of the disclosure.

[0063] Figure 5A structural diagram of an electronic device for implementing a training method of an antigen presentation prediction model or an antigen presentation prediction method according to another embodiment of the disclosure. DETAILED DESCRIPTION

[0064] The present disclosure is further illustrated by the following examples without thereby limiting the present disclosure to the described examples.

[0065] It should be noted that the reference to "one embodiment", "an optional embodiment", "another embodiment", and the like, in the description indicates that the described embodiment can include a particular feature, structure, or characteristic, but every embodiment can not necessarily include the particular feature, structure, or characteristic. Moreover, such phrases are not necessarily referring to the same embodiment. Furthermore, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is submitted that implementation of such feature, structure, or characteristic is within the purview of one of ordinary skill in the art, whether or not explicitly described in connection with another embodiment.

[0066] In the description of the present disclosure, it should be understood that the terms "center", "transverse", "upper", "lower", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the purpose of facilitating the description of the present disclosure and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present disclosure. In addition, the terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present disclosure, unless otherwise specified, the meaning of "a plurality of" is two or more. In addition, the term "comprising" and any variation thereof is intended to cover non-exclusive inclusion.

[0067] In the description of the present disclosure, it should be noted that, unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connecting" should be understood in a broad sense, for example, can be fixedly connected, can be detachably connected, or integrally connected; can be mechanically connected, can be electrically connected; can be directly connected, can be indirectly connected through an intermediate medium, and can be the communication between the two elements inside. For those skilled in the art, the specific meaning of the above terms in the present disclosure can be understood according to the specific circumstances.

[0068] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising", when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0069] Currently, the antigen presentation prediction algorithm commonly used mainly has the following defects:

[0070] 1. In data processing, the current antigen presentation prediction algorithm uses a general traditional encoding method, which lacks certain adaptability and is not necessarily suitable for presentation data scenarios.

[0071] 2. In algorithm modeling, a simple linear network structure is used to establish the model, ignoring the mining of deep features of input data and the mutual relationship of sequence sites. Such algorithm model obviously lacks expression ability.

[0072] In summary, the current antigen presentation prediction algorithm is not very good for antigen presentation prediction, and is poor in external mass spectrometry data independent evaluation.

[0073] In order to overcome the above defects existing at present, the embodiment provides a training method of an antigen presentation prediction model, comprising: obtaining antigen sequence training data and presentation probability training data for representing antigen sequence presentation probability; inputting the antigen sequence training data into the antigen presentation prediction model to be trained; performing presentation prediction processing on the antigen sequence training data by the antigen presentation prediction model to predict the corresponding presentation probability of the antigen sequence training data, wherein the antigen sequence training data is subjected to amino acid encoding and vector mapping processing by a feature encoding module of the antigen presentation prediction model; and training the antigen presentation prediction model according to the predicted presentation probability and the presentation probability training data.

[0074] In the embodiment, the training method of the antigen presentation prediction model is preferably applicable to the training method of the HLA antigen presentation prediction model, and of course the embodiment can be applicable to the affinity of all polypeptide sequences, antigen presentation and immunogenicity model, including but not limited to the prediction of class I and class II human leukocyte antigens, so the embodiment is not limited to application scenarios, and can be selected and adjusted according to actual needs or possible needs.

[0075] In the embodiment, the feature encoding module is preferably an embedding module, but is not limited thereto, and can be selected and adjusted according to actual needs or possible needs.

[0076] In the embodiment, the neural network algorithm model based on antigen sequence is effectively established by using deep learning technology. After a large amount of antigen sequence training data and corresponding presentation probability data (label data) representing the presentation ability are trained, the model can mine the deep features of the input antigen sequence and learn the one-to-one correspondence between the deep features and the label. After the model completes automatic optimization of parameters, the model has the prediction and decision ability for the input antigen sequence data, thereby effectively solving the problem of poor prediction of new antigen presentation ability and improving the accuracy and efficiency of predicting antigen presentation ability.

[0077] In the embodiment, for the input sequence coding problem, the wordembedding method commonly used in NLP (natural language processing) can be used to initialize the vector of the amino acid site. This coding method is suitable for the corresponding scene of the data, so that the amino acid initialization vector expression is more meaningful.

[0078] In order to overcome the above-mentioned defects existing at present, first, the antigen presentation prediction model of the embodiment considers the coding mode of the input sequence digitization. The universal coding mode is difficult to make the coded data suitable for the antigen presentation scene, and a coding mode with strong adaptability is needed for antigen sequence coding. Therefore, what type of coding model is used for the input sequence to make the amino acid initialization vector expression more meaningful is one of the technical problems solved by the embodiment.

[0079] Secondly, the antigen presentation prediction model of the embodiment also considers mining the deep features of the input sequence coding vector, that is, the feature expression of the nonlinear mapping is more meaningful, and then various nonlinear combinations can be performed, thereby showing stronger expression ability. In addition, the mutual relationship of the sequence sites needs to be combined, the locality and integrity of the sequence are considered, so that the sequence feature expression can be more comprehensive. Therefore, what type of algorithm structure is used for modeling is one of the technical problems solved by the embodiment.

[0080] Based on this, as shown in Figure 2 In the embodiment, first, the training data set is digitally preprocessed and input into the initialized algorithm model for training to obtain the trained antigen presentation prediction model with prediction ability. Then, the test input data is also digitally preprocessed and input into the trained antigen presentation prediction model. Finally, the presentation score of each input data is obtained, which is located in 0-1, representing the presentation probability, and the closer to 1, the stronger the presentation ability.

[0081] Specifically, as an embodiment, as shown in Figure 1As shown, the training method of the antigen presentation prediction model provided in the embodiment mainly includes the following steps:

[0082] Step 101, obtaining antigen sequence related training data and test data.

[0083] In this step, antigen sequence training data, presentation probability training data for representing the presentation probability of the antigen sequence, and antigen sequence test data are obtained.

[0084] Specifically, all the antigen sequence related data are divided into training data and test data, and the only difference between the two is that the training data contains output data corresponding to the input data (antigen sequence training data) in addition to the input data, that is, the label value (0 or 1, 0 represents poor presentation and 1 represents strong presentation) representing the presentation of the input antigen sequence.

[0085] Step 102, data preprocessing.

[0086] In this step, the obtained antigen sequence training data and antigen sequence test data are preprocessed.

[0087] As an optional implementation, in this step, the preprocessing includes invalid data filtering processing, string sequence digitization processing and normalization processing.

[0088] Specifically, no processing is performed on the output data, and for the input data, whether it is the input data in the training data or the input data in the test data, the same preprocessing is required, including invalid data filtering processing, string sequence digitization processing.

[0089] Among them, invalid data filtering is mainly to filter out some data that does not meet the requirements, such as string data with bit position character error, string length, and short peptide length <8 (the length of the polypeptide input in this embodiment is not fixed), etc. These data are considered as extremely small probability sample data or error data in the data set of this embodiment, and do not participate in model training and need to be filtered in the training set. For the test set, such data is not directly filtered, but also cannot be input into the model, and the score value representing the presentation is not obtained by model prediction, but is directly assigned a value of 0, indicating low presentation.

[0090] The string sequence digitization process is mainly for the short peptide sequence in the input and the pseudo sequence corresponding to the short peptide (when encoding the MHC sequence, only consider those residues that bind closely with the peptide, which form a short sequence, called MHC pseudo sequence). These sequences are mainly composed of amino acids. There are 20 kinds of amino acids, which can be encoded as 0-19 respectively. Considering that there may be some errors or unknown sites in the sequence, these sites can be set as the character "UNK", corresponding to the encoding 20. In addition, a line feed "EOS" is set, encoded as 21, to facilitate the training of the embedding module in the model structure. The detailed amino acids and their corresponding encoding methods are as follows:

[0091] alphabet_dict = {'A': 0, 'C': 1, 'D': 2, 'E': 3, 'F': 4, 'G': 5, 'H': 6, 'I': 7, 'K': 8, 'L': 9, 'M': 10, 'N': 11, 'P': 12, 'Q': 13, 'R': 14, 'S': 15, 'T': 16, 'V': 17, 'W': 18, 'Y': 19, 'UNK': 20, 'EOS': 21} <unk> ':20,' <eos>':21}.

[0092] The alphabet_dict is an encoding dictionary, according to which the encoding number corresponding to each amino acid can be queried, for example, the amino acids 'E' and 'H' are encoded as 3 and 6 respectively.

[0093] In addition, the preprocessing operation that needs to be performed is the normalization operation, which is mainly aimed at the expression amount characteristics in the input data. In the present embodiment, the numerical range is between 0-170 (the expression amount greater than 170 is directly set to 170), in order to facilitate the convergence of model training, it is necessary to normalize the expression amount characteristic numerical range in the input data to 0-1.

[0094] Step 103, input the antigen sequence training data into the model to be trained.

[0095] In the present step, the antigen sequence training data is input into the initialized antigen presentation prediction model to be trained.

[0096] As an optional implementation, the antigen presentation prediction model mainly includes a feature encoding module, a first feature extraction module, a second feature extraction module and a feature fusion module, of course, the main structure of the antigen presentation prediction model is not specifically limited in the present embodiment, and can be selected and adjusted according to actual needs or possible needs.

[0097] As a preferred implementation, the first feature extraction module mainly includes a one-dimensional CNN module, preferably, the one-dimensional CNN module includes three parallel convolution structures, more preferably, referring to Figure 3 As shown, the three parallel convolution structures include a convolution layer with convolution kernel size of 1*1 (Conv 1d 1), 1*5 (Conv 1d 5) and 1*9 (Conv 1d 9) respectively, of course, the main structure of the first feature extraction module is not specifically limited in the present embodiment, and can be selected and adjusted according to actual needs or possible needs.

[0098] As a preferred implementation, the second feature extraction module mainly includes an RNN module, preferably, referring to Figure 3 As shown, the RNN module includes at least two layers of BiGRU (BiGRU_1, BiGRU_2), of course, the main structure of the second feature extraction module is not specifically limited in the present embodiment, and can be selected and adjusted according to actual needs or possible needs.

[0099] As a preferred implementation, referring to Figure 3 As shown, the feature fusion module mainly includes a full connection (FC) module (FC_linear), and of course the main structure of the feature fusion module is not specifically limited in the embodiment, and can be selected and adjusted according to actual needs or possible needs.

[0100] Specifically, referring to Figure 3 As shown, the input data peptide_input, pseudo_input and Expression_input correspond to the digitalized input after preprocessing of the antigen sequence, the data after digitalization of the pseudo sequence and the normalized expression data respectively.

[0101] For the input peptide_input and pseudo_input, only encoding is performed at each site, and no feature processing is performed. It is assumed that each input peptide_input and pseudo_input after preprocessing is and wherein a i and b i are integers, respectively corresponding to the numbers obtained by encoding the peptide sequence and the pseudo sequence through an encoding dictionary alphabet_dict. The number of each site is changed into a (1, 1) vector, and then the input peptide_input and pseudo_input are vectors with dimensions of (n p , 1) and (n s , 1) respectively.

[0102] In the embodiment, a word embedding method is used to initialize the vector of the amino acid site. Specifically, a certain amino acid in the sequence space is mapped or embedded into another numerical vector space through a certain method, so that each amino acid can also be changed into a multi-dimensional vector. In the present application, each amino acid is mainly mapped into a one-dimensional vector through a Word2Vec module in a gensim toolkit, wherein the Word2Vec module is a simple neural network module, and also needs data training. The input is (n s , 1) format data, and the output is (n s , n emb ) format data, wherein n emb is the length of the feature vector of each amino acid after one-dimensional expansion to high dimension, that is, the vector of each site is (1, n emb ).

[0103] To ensure that the word embedding vectors have strong contextual meaning, it is necessary to train the Word2Vec model in advance using an antigen presentation-related training dataset (the Word2Vec model is a module in the entire algorithm model, in which parameters are involved in training and parameter optimization). This allows the model to learn the meaning expressed by the sequence. After the Word2Vec model is trained, the embedding vectors obtained by the input sequence through the Word2Vec model have richer meaning. The feature vectors between different amino acids are not completely independent and have a certain degree of correlation. The specific degree of correlation depends on the contextual data.

[0104] In terms of model structure, refer to Figure 3 As shown, the feature data from the two embedding channels and the Expression_input are concatenated together using a concat operation to form a single input feature data. Assume the dimensions of the data from the two embedding channels are (n... p ,n emb ) and (n s ,n emb The Expression_input dimension is (1,1), and the second dimension is not n. emb Since they cannot be concatenated together using the `concat` operation, the solution adopted in this invention is to copy `Expression_input` n times. emb The result is a set of parts with dimensions (1, n). emb The data is 0, and then concatenated using the concat operation to obtain data with dimensions (n). p +n s +1,n emb That is, in the first dimension, the three inputs are concatenated end-to-end using a concat operation to obtain n. p +n s +1 dimension cascaded input data.

[0105] In this embodiment, a one-dimensional CNN neural network structure with stronger feature extraction capabilities is used to extract features from the sequence, referencing... Figure 3 As shown, Figure 3 In this example, `conv 1d` represents a 1D convolution operation. To extract richer features, this embodiment selects three convolution kernels of different sizes for the convolution operation, with convolution sizes of 1, 5, and 9, and corresponding padding of 0, 2, and 4, respectively, ensuring that the sequence length remains unchanged before and after convolution. After the three-channel convolution operation is completed, the resulting sequences have dimensions of (n... p +n s +1,cn1), (n p +n s +1,cn5) and (n p +n s +1,cn9) of the output feature dimension data, wherein cn1, cn5 and cn9 respectively correspond to the number of 1, 5 and 9 convolution kernels, that is, the length of the expanded feature vector. After obtaining three outputs, the concat operation is performed again to concatenate the operations in the second dimension to obtain (n p +n s +1,cn) dimension data, wherein cn is equal to cn1+cn5+cn9.

[0106] After extracting the deep features through the convolution operation, the local and overall sequence needs to be considered. In the network structure designed in the embodiment, the feature data obtained after the convolution operation is subjected to the RNN network layer to perform ordered fusion of the front and rear feature vectors. The RNN type network structure that can be selected mainly includes general rnn, LSTM, BiLSTM, GRU and BiGRU, etc. The embodiment selects BiGRU, an RNN type network structure that considers forward and reverse and has high running efficiency, to extract features. Compared with other types of RNN structures, BiGRU can guarantee efficiency while having the best algorithm effect.

[0107] In order to make the feature extraction and fusion more sufficient, two layers of BiGRU structure are designed in the model. After the two layers of BiGRU structure, the fused feature dimension is (n p +n s +1,rn), wherein rn is the number of hidden layer units in the BiGRU structure and is also the second dimension size of the output feature.

[0108] After the above convolution feature extraction and ordered feature fusion, the next step is to use the fully connected network layer FC to perform dimension mapping, that is, all features are fused to output two-dimensional features according to certain rules. However, considering that the BiGRU operation may lose some important convolution features, the embodiment uses the concat operation to concatenate the convolution features and the BiGRU features to obtain feature data with an output of (n p +n s +1,rn+cn). Finally, the feature fusion and dimension reduction operation of the FC layer obtains output data of (1, 2). The algorithm model of the embodiment is to predict the presentation probability of the input sequence, that is, to output a score value between 0 and 1. Therefore, the output result with a dimension of (1, 2) needs to be subjected to softmax calculation first, wherein the softmax function is as follows:

[0109]

[0110] Where i is each element in the output result (1, 2), the probability value of the output result after the transformation of the function is between 0-1, and then the value of the second dimension is taken to obtain a score value between 0-1.

[0111] Step 104, performing a presentation prediction process by the model to predict the presentation probability corresponding to the antigen sequence training data.

[0112] In this step, the antigen sequence training data is subjected to a presentation prediction process by the antigen presentation prediction model to predict the presentation probability corresponding to the antigen sequence training data.

[0113] As a preferred embodiment, in this step, the antigen sequence training data is subjected to amino acid encoding and vector mapping processing by the feature encoding module of the antigen presentation prediction model; the antigen sequence training data after the amino acid encoding and vector mapping processing is subjected to polypeptide sequence one-dimensional convolution feature extraction processing by the first feature extraction module of the antigen presentation prediction model; the antigen sequence training data after the polypeptide sequence one-dimensional convolution feature extraction processing is subjected to polypeptide sequence ordered feature extraction processing by the second feature extraction module of the antigen presentation prediction model; and the antigen sequence training data after the polypeptide sequence ordered feature extraction processing is subjected to feature fusion processing by the feature fusion module of the antigen presentation prediction model.

[0114] Specifically, after the model is created and initialized, each parameter involved in the model is a randomized parameter, the model does not have any expression ability, and the output result of the model after accepting the input data has no actual meaning. Therefore, the model needs to be trained to have a certain meaning in the output result.

[0115] The model can be understood as a function, and the model training is to determine the parameters of the model by a certain method using the preprocessed training data. Once the parameter values are determined, the training is completed, the input is substituted into the model, and the meaningful numerical value is output. In the present application, the output label in the training data represents the presentation ability of the antigen data, so after training a large number of antigen inputs and output labels corresponding to the presentation ability of the antigen inputs, the model learns the mapping relationship between the antigen input data and the output label by a certain method. Once the training is completed, the model parameters are determined, and the model has the ability to predict the presentation ability of the input antigen data.

[0116] Step 105, training the model according to the predicted presentation probability and the presentation probability training data.

[0117] In this step, the antigen presentation prediction model is trained according to the predicted presentation probability and the presentation probability training data.

[0118] Step 106, testing the trained model according to the test data.

[0119] In this step, the antigen sequence test data is input into the trained antigen presentation prediction model to test the trained antigen presentation prediction model to obtain the trained and tested antigen presentation prediction model.

[0120] Specifically, the model testing, i.e., the prediction ability of the input antigen data presentation, can directly use the preprocessed test data described in the embodiment to test the model.

[0121] The following describes the sub-module structure of the algorithm and the corresponding input and output in more detail by implementing the algorithm model with specific experimental data.

[0122] Suppose the experimental input data is composed of two groups of data for batch processing, and the data is:

[0123] peptide_input(2,12)—>[[‘L’,‘D’,‘E’,‘A’,‘L’,‘Q’,‘W’,‘H’,‘L’,‘L’,‘L’,‘L’],

[0124] [‘D’,‘E’,‘A’,‘L’,‘Q’,‘W’,‘H’,‘D’,‘L’,‘L’,‘L’,‘L’]]

[0125] pseudo_input(2,35)—>[['Y','Y','S','E','Y','R','N','I','Y','A','Q','T','D','E','S','N','L','Y',

[0126] 'L','R','Y','N','F','Y','T','W','A','V','L','T','Y','T','W','Y','A'],

[0127] ['Y','Y','S','E','Y','R','N','I','Y','A','Q','T','D','E','S','N','L','Y',

[0128] 'L','R','Y','N','F','Y','T','W','A','V','L','T','Y','T','W','Y','A']]

[0129] Expression_input(2,1)—>[[[3.6525]],[[3.6525]]]

[0130] 1: The first network layer into which the experimental input character matrix data enters is the embedding layer, which is mainly composed of the Word2Vec model. Its role is to encode the input peptide_input and pseudo_input character matrix into a digital matrix x_pep and x_pse:

[0131] peptide_input(2,12)—>[[9,2,3,0,9,13,18,6,0,0,0,0],[2,3,0,9,13,18,6,2,0,0,0,0]]

[0132] pseudo_input(2,35)—>[[19,19,15,3,19,14,11,7,19,0,13,16,2,3,15,11,9,19,9,14,19,11,4,19,16,18,0,17,9,16,19,16,18,19,0],[19,19,15,3,19,14,11,7,19,0,13,16,2,3,15,11,9,19,9,14,19,11,4,19,16,18,0,17,9,16,19,16,18,19,0]]

[0133] After encoding, the digital matrix data x_pep and x_pse and another input are concatenated through the concat operation to become a concatenated matrix data X, which has a dimension of (2,48,25), where 25 is the feature vector dimension n described in the previous section emb The processed experimental data is:

[0134]

[0135] 2: After the concat operation, the data is subjected to the first step of feature extraction through three parallel convolution structures, each consisting of a convolution layer with a kernel size of 1*3, 1*5, and 1*9. The input feature channel number of the convolution operation is n emb , and the output channel is 64, i.e., cn1, cn5, and cn9 in the previous section are all 64. After the concatenated matrix data X is subjected to the three parallel convolution structure operations, the dimension is (2,48,64), and then the concat operation is performed to obtain the output feature matrix X_conved, which has a dimension of (2,48,192). The experimental data is shown below:

[0136]

[0137] 3: After the initial feature extraction, the convolution feature matrix X_conved is obtained, and then the second step of feature extraction is performed by the network layer composed of two layers of BiGRU. After the input experimental data X_conved is processed by two layers of BiGRU, the output feature matrix X_BiGRUed with a dimension of (2, 48, 192) is obtained, where 192 is the feature dimension described in the previous section rn , that is, BiGRU does not change the size of the feature dimension. Then, X_BiGRUed is concatenated with X_conved to obtain the final extracted feature matrix data X_fea, which has a dimension of (2, 48, 384). The experimental data X_fea is as follows:

[0138]

[0139] 4: After the feature extraction, the feature matrix data X_fea is obtained, which is then processed by the last fully connected layer for the final feature fusion to obtain the fused output matrix X_out, which has a dimension of (2, 2). The experimental data X_out is as follows:

[0140] [[3.4756, -3.5319],

[0141] [[4.6604, -4·6849]]

[0142] After the softmax function processing in the previous section, the final prediction score y = [[9.99098e-01, 9.04278e-04], [9.99912e-01, 8.73592e-05]] is obtained. Each group of data is a two-dimensional vector, where the first dimension represents the probability of not presenting, and the second dimension represents the probability of presenting. To obtain the presentation probability, the second dimension of each group of data is selected, and the final presentation score of the two groups of data is [9.04278e-04, 8.73592e-05].

[0143] As another embodiment, as shown in Figure 4 , the embodiment also provides an antigen presentation prediction method, mainly including the following steps:

[0144] Step 201, obtaining antigen sequence data;

[0145] Step 202, inputting the antigen sequence data into the trained antigen presentation prediction model, specifically inputting the antigen sequence data into the antigen presentation prediction model trained by the antigen presentation prediction model training method as described above;

[0146] Step 203, performing presentation prediction processing on the antigen sequence data by the trained antigen presentation prediction model;

[0147] Step 204, output the predicted antigen sequence data corresponding to the presentation probability.

[0148] The following specifically explains the specific operation environment and use scenario of the training method and prediction method of the antigen presentation prediction model provided in this embodiment.

[0149] Optional environment description:

[0150] In this embodiment, the required computing resources can be selected as GPU (graphics processing unit) or CPU (central processing unit). Since the use of GPU is relatively efficient, GPU is preferred. This embodiment is developed based on python (computer programming language) under the ubuntu20.04 environment, so the corresponding running environment needs to be configured. The specific optional environment configuration requirements are as follows:

[0151]

[0152]

[0153] After the environment is installed, the experimental personnel can directly execute the following code running command in the engineering directory to test their own data:

[0154] python predict.py--test_csv_path test_input.csv--pred_save_pathpredict_result.csv

[0155] Among them, predict.py is the code python code file, and two command parameters test_csv_path and pred_save_path need to be passed in the command, which respectively represent the csv format polypeptide antigen data file path to be tested and the prediction result saving path.

[0156] Among them, the input data sample is as follows:

[0157] peptide expmhc

[0158] LIEDHFDVT 14.5525 B4002

[0159] IQVVGVETD 0.00125 B4002

[0160] TKTQKIRLL 0.02625 B4002

[0161] GISLQKKLQ 41.2275 B4002

[0162] LKAKIINIC 0.005 B4002

[0163] SPIIILNLV 10.195 B4002

[0164] TPSERPTAQ 7.33625 B4002

[0165] As shown above, the input file contains polypeptide antigen data peptide, corresponding expression exp and corresponding typing mhc, the third column mhc indicates which type the polypeptide belongs to, and is not necessary data for algorithm input.

[0166] The algorithm model input also needs the pseudo sequence corresponding to the peptide. This data does not need to be input by the user, and the program will query the corresponding pseudo sequence according to the input peptide sequence corresponding to the mhc in the data preprocessing process (each peptide sequence has a corresponding mhc, and the mhc can query the corresponding pseudo sequence, and different mhc may have the same pseudo sequence; the input mhc is only convenient for data analysis, such as ensuring the balance of various mhc data in the training set, testing various mhc effects, and actually the mhc does not belong to the model input), and the query dictionary is saved locally.

[0167] If the pseudo sequence cannot be queried, the input sequence is determined as invalid sequence data. That is, when the mhc species in the experimental data set is limited, the pseudo is also limited, and the trained model may not query the pseudo sequence in the later use, at which time the peptide sequence corresponding to the pseudo sequence is determined as invalid sequence.

[0168] After preparing the test data, the command can be executed to obtain the presentation score of each polypeptide sequence, and the result is saved in the predict_result.csv file. The result sample is shown below:

[0169] peptide score mhc predict

[0170] LIEDHFDVT 0.0212 B4002 0

[0171] IQVVGVETD 0.013 B4002 0

[0172] TKTQKIRLL 0.0426 B4002 0

[0173] GISLQKKLQ 0.0082 B4002 0

[0174] LKAKIINIC 0.0255 B4002 0

[0175] SPIIILNLV 0.0058 B4002 0

[0176] TPSERPTAQ 0.0027 B4002 0

[0177] RALLNLHTD 0.0123 B4002 0

[0178] The second column is the presentation probability of each polypeptide, and the fourth column is 0 or 1, that is, 1 if the probability is greater than 0.5, otherwise 0, representing whether the antigen polypeptide can be presented.

[0179] The training method and prediction method of the antigen presentation prediction model provided in the embodiment effectively establish a neural network algorithm model based on antigen sequences by using deep learning technology. After training a large amount of antigen sequence training data and corresponding presentation probability data representing the presentation ability, the model can mine the deep features of the input antigen sequence and learn the one-to-one correspondence between the deep features and the label. After the model completes automatic optimization of parameters, the model has the prediction and decision-making ability for input antigen sequence data, thereby effectively solving the problem of poor prediction of new antigen presentation ability and improving the accuracy and efficiency of predicting antigen presentation ability.

[0180] The training method and prediction method of the antigen presentation prediction model provided in the embodiment propose to use the word embedding method commonly used in NLP to initialize the vector of the amino acid site for the input sequence coding problem. This coding method is highly applicable in the corresponding scene of the data, which can make the amino acid initialization vector expression more meaningful.

[0181] The training method and prediction method of the antigen presentation prediction model provided in the embodiment consider using a one-dimensional CNN neural network structure with stronger feature extraction capability to extract the features of the sequence for the sequence site deep feature extraction problem.

[0182] The training method and prediction method of the antigen presentation prediction model provided in the embodiment use the RNN module, a neural network technology for sequence modeling with memory function, to build the model in order to consider the correlation of the sites in the sequence and the integrity of the sequence.

[0183] The training method and prediction method of the antigen presentation prediction model provided in the embodiment can ensure accurate prediction of antigen presentation ability while maintaining high efficiency. The embodiment can process 100,000 samples per second by using GPU acceleration, which is very efficient. Moreover, the model involved in the embodiment only occupies about 3MB, which greatly reduces the occupied space.

[0184] Figure 5 This is a schematic diagram of the structure of an electronic device according to this embodiment. The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the training method or antigen presentation prediction method of the antigen presentation prediction model as described in the above embodiment. Figure 5 The electronic device 30 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0185] like Figure 5 As shown, the electronic device 30 can be manifested as a general-purpose computing device, such as a server device. The components of the electronic device 30 may include, but are not limited to: at least one processor 31, at least one memory 32, and a bus 33 connecting different system components (including memory 32 and processor 31).

[0186] Bus 33 includes a data bus, an address bus, and a control bus.

[0187] The memory 32 may include volatile memory, such as random access memory (RAM) 321 and / or cache memory 322, and may further include read-only memory (ROM) 323.

[0188] The memory 32 may also include a program / utility 325 having a set (at least one) of program modules 324, including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0189] The processor 31 executes various functional applications and data processing by executing computer programs stored in the memory 32, such as the training method or antigen presentation prediction method of the antigen presentation prediction model as disclosed in the above embodiments.

[0190] Electronic device 30 can also communicate with one or more external devices 34 (e.g., keyboard, pointing device, etc.). This communication can be performed via input / output (I / O) interface 35. Furthermore, the model-generating device 30 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 36. Figure 5 As shown, network adapter 36 communicates with other modules of model-generated device 30 over bus 33. It should be appreciated that other hardware and / or software modules can be used in conjunction with model-generated device 30, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (redundant array of independent disks) systems, tape drives, and data backup storage systems, etc. as are known to those skilled in the art.

[0191] It should be noted that although several units / modules or sub-units / modules of an electronic device are mentioned in the foregoing detailed description, such division is merely exemplary and not mandatory. Indeed, according to embodiments of the present disclosure, features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, features and functions of one unit / module described above can be further divided into multiple units / modules.

[0192] The embodiments also provide a computer readable storage medium, having stored thereon a computer program, the program being executed by a processor to implement the steps in the training method of an antigen presentation prediction model or the antigen presentation prediction method in the above embodiments.

[0193] More specifically, the readable storage medium can include, but is not limited to, a portable disc, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0194] In possible embodiments, the present disclosure can also be implemented in the form of a program product, which includes program codes for causing a terminal device to execute the steps in the training method of an antigen presentation prediction model or the antigen presentation prediction method in the above embodiments when the program product is executed on the terminal device.

[0195] The program codes for executing the present disclosure can be written in any combination of one or more programming languages, and can be executed entirely on the user device, partly on the user device and partly on a remote device, or entirely on a remote device.

[0196] Although the specific embodiments of the present disclosure are described above, those skilled in the art should understand that this is only an illustration, and the protection scope of the present disclosure is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of the present disclosure, and such changes and modifications all fall within the protection scope of the present disclosure.< / eos> < / unk>

Claims

1. A method for training an antigen presentation prediction model, characterized in that, The method comprises the following steps: acquiring antigen sequence training data and presentation probability training data for representing the presentation probability of the antigen sequence; inputting the antigen sequence training data into an antigen presentation prediction model to be trained; performing presentation prediction processing on the antigen sequence training data by using the antigen presentation prediction model to predict the presentation probability corresponding to the antigen sequence training data, wherein the antigen sequence training data is subjected to amino acid encoding and vector mapping processing by using a feature encoding module of the antigen presentation prediction model; and training the antigen presentation prediction model according to the predicted presentation probability and the presentation probability training data. The antigen sequence training data comprises a polypeptide sequence composed of amino acids. The step of performing presentation prediction processing on the antigen sequence training data by using the antigen presentation prediction model to predict the presentation probability corresponding to the antigen sequence training data further comprises the following steps: performing polypeptide sequence one-dimensional convolution feature extraction processing on the antigen sequence training data subjected to amino acid encoding and vector mapping processing by using a first feature extraction module of the antigen presentation prediction model; performing polypeptide sequence ordered feature extraction processing on the antigen sequence training data subjected to polypeptide sequence one-dimensional convolution feature extraction processing by using a second feature extraction module of the antigen presentation prediction model.

2. The training method according to claim 1, wherein the antigen sequence training data further comprises a short peptide sequence composed of amino acids and a pseudo sequence corresponding to the short peptide; and / or The presentation probability training data comprises a label value.

3. The training method according to claim 1, wherein the first feature extraction module comprises three parallel convolution structures.

4. The training method according to claim 3, wherein the three parallel convolution structures comprise one convolution layer with a convolution kernel size of 1*1, 1*5 and 1*9, respectively.

5. The training method according to claim 1, wherein the step of performing presentation prediction processing on the antigen sequence training data by using the antigen presentation prediction model to predict the presentation probability corresponding to the antigen sequence training data further comprises the following step: performing feature fusion processing on the antigen sequence training data subjected to polypeptide sequence ordered feature extraction processing by using a feature fusion module of the antigen presentation prediction model.

6. The training method according to claim 1, wherein before the step of inputting the antigen sequence training data into the antigen presentation prediction model to be trained, the training method further comprises the following step: performing preprocessing on the acquired antigen sequence training data, wherein the preprocessing comprises at least one of invalid data filtering processing, string sequence digitization processing and normalization processing.

7. The training method according to claim 1, further comprising the following steps: acquiring antigen sequence test data; inputting the antigen sequence test data into the trained antigen presentation prediction model to test the trained antigen presentation prediction model, so as to obtain the trained and tested antigen presentation prediction model.

8. A method of antigen presentation prediction, characterized in that, The method comprises the following steps: acquiring antigen sequence data; inputting the antigen sequence data into an antigen presentation prediction model trained by the training method of the antigen presentation prediction model according to any one of claims 1-7; performing presentation prediction processing on the antigen sequence data by the trained antigen presentation prediction model; and outputting the predicted presentation probability corresponding to the antigen sequence data.

9. A method of antigen presentation prediction, characterized in that, comprising: obtaining antigen sequence data; inputting the antigen sequence data into an antigen presentation prediction model; performing presentation prediction processing on the antigen sequence data by the antigen presentation prediction model to predict the presentation probability corresponding to the antigen sequence data, wherein the antigen sequence data is subjected to amino acid encoding and vector mapping processing by a feature encoding module of the antigen presentation prediction model; and outputting the predicted presentation probability corresponding to the antigen sequence data; the antigen sequence data comprises a polypeptide sequence composed of amino acids; the step of performing presentation prediction processing on the antigen sequence data by the antigen presentation prediction model to predict the presentation probability corresponding to the antigen sequence data further comprises: performing polypeptide sequence one-dimensional convolution feature extraction processing on the antigen sequence data subjected to amino acid encoding and vector mapping processing by a first feature extraction module of the antigen presentation prediction model; performing polypeptide sequence ordered feature extraction processing on the antigen sequence data subjected to polypeptide sequence one-dimensional convolution feature extraction processing by a second feature extraction module of the antigen presentation prediction model.

10. The antigen presentation prediction method of claim 9, wherein the antigen sequence data further comprises a short peptide sequence composed of amino acids and a pseudo sequence corresponding to the short peptide.

11. The antigen presentation prediction method of claim 9, wherein the step of performing presentation prediction processing on the antigen sequence data by the antigen presentation prediction model to predict the presentation probability corresponding to the antigen sequence data further comprises: performing feature fusion processing on the antigen sequence data subjected to polypeptide sequence ordered feature extraction processing by a feature fusion module of the antigen presentation prediction model.

12. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the training method of the antigen presentation prediction model according to any one of claims 1-7 or the antigen presentation prediction method according to any one of claims 8-11.

13. A computer readable medium having stored thereon computer instructions, wherein, The computer instructions, when executed by the processor, implement the training method of the antigen presentation prediction model according to any one of claims 1-7 or the antigen presentation prediction method according to any one of claims 8-11.

Images