Antigen affinity prediction method and system based on deep learning

Abstract

The present invention belongs to the field of artificial intelligence, and discloses a method and a prediction model for antigen affinity based on deep learning. The method comprises: (1) obtaining a training data set of oligomer-immune molecule binding structure, including affinity; (2) for each group of oligomer-immune molecules, the sequence of the oligomer, the monomer position of the oligomer and the immune molecule are respectively represented as high-dimensional vectors, and the three high-dimensional vectors are fused into the vector of the binding structure of the oligomer-immune molecule; (3) using the vector and affinity data of the binding structure of the oligomer-immune molecule in the training data set to train a deep neural network, and establish an oligomer-immune molecule affinity prediction model; (4) inputting the vector of the binding structure of the oligomer-immune molecule to be tested into the antigen affinity prediction model, and using the trained deep neural network to predict the affinity of the oligomer-immune molecule binding to be tested.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and more specifically, the present invention relates to a method and system for predicting antigen affinity based on deep learning. Background Art

[0002] Antigenic Determinant (AD) refers to a special chemical group in an antigen molecule that determines the specificity of an antigen. Antigens bind to the antigen receptors on the surface of corresponding lymphocytes through antigenic epitopes, thereby activating lymphocytes and causing immune responses. The nature, number and spatial configuration of antigenic epitopes determine the specificity of antigens. The size of antigenic epitopes is suitable for the antigen binding site of the corresponding antibody. Generally, a polypeptide epitope contains 5 to 6 amino acid residues; a polysaccharide epitope contains 5 to 7 monosaccharides; and a nucleic acid hapten epitope contains 6 to 8 nucleotides. The specificity of an antigenic epitope is determined by all the residues that make it up, but some of these residues play a greater role than other residues when binding to antibodies. These residues are called immunodominant groups.

[0003] T cell epitopes are immunogenic polypeptide fragments that must be processed by antigen presenting cells into small peptide molecules and then combined with major histocompatibility complex (MHC) molecules to be recognized by T cells. T cell antigen receptors (T Cell Receptors, TCR) can only recognize small polypeptide molecules of about 10 to 20 amino acids. These epitopes are composed of amino acids connected in sequence and mainly exist in the hydrophobic region of antigen molecules. They are called linear epitopes or sequence epitopes. Since T cells can only recognize processed epitopes, they generally do not recognize conformational epitopes of natural antigens.

[0004] There are many methods for analyzing antigen epitopes, including chemical cleavage, bio-enzymatic hydrolysis, nuclear magnetic resonance spectroscopy (NMR), surface plasmon resonance (SPR), hybrid peptide cleavage, peptide library construction, theoretical measurement, etc. At present, with the development of computer technology, especially the popularization of artificial intelligence technology, direct screening and prediction of antigen epitopes from biological big data has become the antigen epitope analysis method with the highest throughput, lowest cost and shortest cycle.

[0005] As for T cell epitopes, the conventional approach for theoretical prediction is to extract many key features, such as antigen peptide sequence, antigen peptide length, antigen peptide expression, antigen peptide thermal stability, etc., by studying the data of MHC molecule-antigen peptide binding, thereby forming a machine learning model, combined with high-throughput sequencing technology, to make theoretical predictions of potential antigen peptides of unknown proteins or genes. At present, many companies at home and abroad have successively developed antigen prediction software, such as netMHCpan software developed by the Technical University of Denmark, EDGE software developed by Gritstone Oncology, and EPIP software developed by Wuhan BGI Genentech Biotechnology Co., Ltd.

[0006] With the continuous development and mutual promotion of modern bioinformatics, molecular biology and molecular immunology, epitope research and its application have made great progress and shown great application potential. The application of antigen epitopes is mainly reflected in three aspects, namely disease diagnosis, vaccine development and disease treatment.

[0007] In disease diagnosis, the key to the efficiency of epitope diagnostic methods is sensitivity and specificity. Antigenic epitopes are the basic units that stimulate immune responses. Single epitopes stimulate specific single immune responses, while multiple epitopes often stimulate multiple mixed immune responses, thereby producing nonspecific mixed antibodies, sensitized lymphocytes or effector factors. Therefore, epitope peptide research in disease diagnosis focuses on the selection of specific epitope peptides to achieve better diagnostic efficacy.

[0008] In vaccine development, conventional vaccines contain a large number of antigenic epitopes, including protective epitopes, inhibitory epitopes or ineffective epitopes. Only when the immune response induced by protective epitopes is dominant can the vaccine achieve the desired protective effect. Therefore, in previous research on antiviral vaccines, how to obtain protein epitopes with strong immunogenicity, strong sequence conservation, and key role in virus invasion is often a technical difficulty.

[0009] In disease treatment, epitope-induced immune responses are highly specific and targeted, and can be used for immunotherapy of tumors, infectious diseases, and autoimmune diseases. Immunotherapy activates cytotoxic responses to antigens by enhancing the patient's own immune system, and has been proven to be an effective strategy in recent years. This strategy utilizes a variety of cell surface antigens formed by various mutant proteins cleaved by intracellular proteasomes. These peptides bind to HLA molecules to form peptide-HLA complexes that are presented to T cell receptors (TCRs). If TCRs can recognize these peptide-HLA complexes, cytotoxic T lymphocytes (CTLs) will be activated. These cells are a subdivision of white blood cells and are a specific T cell that specializes in secreting various cytokines to participate in immune effects. They have a killing effect on certain antigenic substances such as viruses and tumor cells, and together with natural killer cells, they constitute an important line of defense for the body's antiviral and antitumor immunity. The first step in cytotoxic T lymphocyte therapy is to predict the binding affinity of antigens to HLA molecules. The rapidly progressing neoantigen therapy in the field of tumor treatment is a good application of epitopes in the field of disease treatment. Many foreign companies have conducted research on the use of neoantigens to treat malignant tumors and have entered the clinical trial stage, such as BioNTech, Neon Therapeutics, Gritstone Oncology, etc.

[0010] There are currently four methods for predicting the affinity between antigens and HLA molecules, namely structure-based methods, machine learning-based methods, position weight matrix (PSSM)-based methods, and combined methods. Among them, the machine learning-based method predicts the affinity of peptide binding by learning a high-dimensional classification plane based on known binding peptide and non-binding peptide information. The machine learning method can accurately predict the affinity of peptides for specific HLA alleles, such as HLA-A*0201, HLA-A*0101, and HLA-B*0702[1,2], and is therefore often used in many studies[3-5]. The most mainstream machine learning algorithm is the pan-specific affinity algorithm, which takes both the amino acid sequence of the HLA molecule and the amino acid sequence of the peptide as input and obtains the affinity prediction output through the machine learning algorithm. The most commonly used algorithm in the industry is NetMHCPan[6]. This method uses a 34-bit amino acid pseudo sequence to characterize the HLA molecule, then pre-processes the pseudo sequence and the short peptide sequence, and uses the pre-processing results as input features. The peptide-HLA molecule affinity prediction is then obtained through the BP neural network model (BackPropagation, BP). This method models each peptide-HLA molecule pair as a unique input sequence, so that the mapping relationship between the sequence and the peptide-HLA molecule affinity value can be learned through the model. This approach is called the pan-specific training strategy (Pan Model) in the industry. NetMHCPan's work shows that a single BP model is not effective. In order to obtain the best results, NetMHCPan usually aggregates and learns a large number of models at the same time.

[0011] In recent years, many affinity prediction models based on deep learning have emerged in the industry. Among them, the more representative algorithm is DeepSeqPan[7]. This method uses two independent convolutional networks to learn peptide and HLA features respectively, and then predicts the affinity output through neural networks after merging the features. Similar ideas are used in MHCSeqNet[8] and AI-MHC[9]. MHCSeqNet uses gated recurrent units (GRU) instead of convolutional neural networks to learn multiple peptide data of variable lengths. AI-MHC processes different peptides through padding, thereby using more efficient convolutional neural networks to achieve variable-length input learning. Similar work includes ACME

[10] . The difference from DeepSeqPan is that the concatenation of each HLA calculation layer will concatenate the peptide representation vector. In addition to using natural language processing technology, convMHC

[11] uses an image processing-like idea to model the affinity model: the physical and chemical data of the location of peptides and HLA molecules are established as multiple two-dimensional data matrices, and a two-dimensional convolutional neural network is used to predict the affinity model.

[0012] The main shortcomings of the existing technology are: 1) There is a lack of effective ways to model and express peptide-HLA complexes. In the above methods, [6-10] use independent neural networks to learn the sequence features of peptides and HLA molecules, and then directly splice them together. The model design does not take into account the complex structure that will be produced once the peptide and typing are presented; 2) The peptide sequence is short, and the direct use of deep networks has poor results. Most related works use shallow networks. [6] A large number of traditional neural network aggregations are used to make up for the shortcomings of weak learning ability of a single model; [7,9,10] all use shallow convolutional networks to learn peptide features; 3) Some methods do not consider the diversity of peptide length.

[11] Although deep networks can be used for peptide learning, the design can only consider fixed length and cannot achieve variable length; the same problem occurs in [6].

[0013] Therefore, the present invention requires an improved affinity prediction model based on deep learning. Summary of the invention

[0014] Based on the problems existing in the prior art, the object of the present invention is to provide an improved affinity prediction model based on deep learning and a construction method thereof.

[0015] Therefore, in a first aspect, the present invention provides a method based on deep learning antigen affinity, the method comprising:

[0016] (1) obtaining a training data set of oligomer-immune molecule binding structures, wherein the training data set includes the affinity of the oligomer-immune molecule binding;

[0017] (2) for each group in the training data set of the oligomer to be tested and the immune molecule and the binding structure of the oligomer-immune molecule, the sequence of the oligomer, the monomer position of the oligomer and the immune molecule are represented as high-dimensional vectors respectively, and the three high-dimensional vectors are fused into a vector of the binding structure of the oligomer-immune molecule;

[0018] (3) using the vector and affinity data of the oligomer-immune molecule binding structure in the training data set to train a deep neural network and establish an oligomer-immune molecule affinity prediction model;

[0019] (4) Inputting the vector of the oligomer-immune molecule binding structure to be tested into the antigen affinity prediction model, and using the trained deep neural network to predict the affinity of the oligomer-immune molecule binding to be tested.

[0020] In one embodiment, the oligomer is selected from the group consisting of: a polypeptide, a polysaccharide, and a nucleic acid hapten.

[0021] In one embodiment, the oligomers are vectors of variable length.

[0022] In one embodiment, the immune molecule is a major histocompatibility complex molecule, such as a human leukocyte antigen, including but not limited to class I and class II human leukocyte antigens.

[0023] In one embodiment, the affinity of the oligomer-immune molecule binding is expressed as the IC of the oligomer-immune molecule binding. 50 value.

[0024] In one embodiment, in (2), a monomer vector representation that can reflect the relationship between any two monomers is obtained through word embedding.

[0025] In one embodiment, the input oligomer sequence in (2) is (x1, x2, ..., x n ), where x i is the amino acid character, x i ∈{A,C,D,...,Y}, the mapping result is to convert x i Transformed into an m-dimensional vector z i ∈R m , R is the real number space, the sequence (x1, x2, ..., x n ) is transformed into (z1, z2, ..., z n ).

[0026] In one embodiment, in (2), the monomer position characterization vector is: the label of each amino acid position of the polypeptide is mapped to a high-dimensional continuous numerical space: the input is unified as a vector (1, 2, ..., n), and each position i is mapped into a vector p by a fully connected neural network i ∈R m , R is the real number space, that is, (1, 2, ..., n) and the output is (p1, p2, ..., p n ).

[0027] In one embodiment, the high-dimensional vectors of the oligomer sequence and the monomer positions of the oligomer in (2) are added to form a characterization vector for the oligomer.

[0028] In one embodiment, the input of the immune molecule vector in (2) is (y1, y2, ..., y k ), y i ∈{A,C,D,...,Y}, the output is (z′1,z′2,...,z′ n ), z′ i ∈R m , mapping from the k-dimensional raw amino acid vector to an n×m-dimensional vector with the same format as the oligomer.

[0029] In one embodiment, in (2), the characterization vector of the oligomer and the vector of the immune molecule are subjected to tensor multiplication, tensor addition or attention operation.

[0030] In one embodiment, the deep neural network in (3) includes a deep convolutional layer and a multi-layer fully connected layer, wherein the deep convolutional layer is used to extract the feature vector of the oligomer-immune molecule in the training data set, which is input into the multi-layer fully connected layer, and the multi-layer fully connected layer maps the feature vector to an affinity value, and obtains the network parameters through a back-propagation algorithm.

[0031] In one embodiment, for each deep convolutional layer, the input is copied three times, two of which are learned through different convolutional networks, and one of the convolution outputs is subjected to a Sigmoid normalization operation to obtain two vectors A and B, which are bitwise multiplied to obtain a residual A×B; the other copy is added to the residual of A×B.

[0032] In one embodiment, a global pooling layer θ is added after the last convolutional layer: R n×p →R p , p is the output dimension of the last layer, and the feature vector is obtained.

[0033] In one embodiment, in the multi-layer fully connected layer, the feature vector obtained by the deep convolutional layer is mapped to the affinity value by the fully connected layer, as follows:

[0034] 1) Perform a linear transformation on the output feature vector x of the deep convolutional layer and use it as the input vector y=Wx+b of the fully connected layer;

[0035] 2) Use the linear rectification function to perform nonlinear transformation on the linear transformation result:

[0036]

[0037] 1) and 2) constitute a layer of mapping network, and the affinity prediction results are output through multiple layers of mapping networks.

[0038] In one embodiment, in (4), the trained deep neural network includes a deep convolutional layer and a multi-layer fully connected layer, wherein the deep convolutional layer is used to extract the feature vector of the oligomer-immune molecule to be tested, input the feature vector into the multi-layer fully connected layer, and predict the affinity of the oligomer-immune molecule binding to be tested.

[0039] In a second aspect, the present invention provides a method for establishing an antigen affinity prediction model based on deep learning, the method comprising establishing an oligomer-immune molecule affinity prediction model through (1)-(3) of the method of the first aspect of the present invention.

[0040] In a third aspect, the present invention provides an antigen affinity prediction model established using the method of the second aspect of the present invention, wherein the antigen affinity prediction model comprises: a data acquisition module, an input vector establishment module and a model establishment module.

[0041] The data acquisition module is used to acquire a data set of the oligomer-immune molecule binding structure, wherein the data set includes the affinity of the oligomer-immune molecule binding;

[0042] The input vector establishment module is used to use the data set of the oligomer-immune molecule binding structure to map the sequence of the oligomer, the monomer position of the oligomer and the immune molecule into high-dimensional vectors respectively, and fuse the three high-dimensional vectors into the vector of the oligomer-immune molecule binding structure;

[0043] The model building module is used to train a deep neural network using the vector and affinity data of the oligomer-immune molecule binding structure to establish an antigen affinity prediction model.

[0044] In one embodiment, in the model building module, the deep neural network includes a deep convolutional layer and a multi-layer fully connected layer, the deep convolutional layer is used to extract the feature vector of the oligomer-immune molecule in the training data set, and input the multi-layer fully connected layer, the multi-layer fully connected layer maps the feature vector to an affinity value, and obtains the network parameters through a back propagation algorithm.

[0045] In one embodiment, for each deep convolutional layer, the input is copied three times, two of which are learned through different convolutional networks, and one of the convolution outputs is subjected to a Sigmoid normalization operation to obtain two vectors A and B, which are bitwise multiplied to obtain a residual A×B; the other copy is added to the residual of A×B.

[0046] In one embodiment, a global pooling layer θ is added after the last convolutional layer: R n×p →R p ,p is the output dimension of the last layer, and the feature vector is obtained.

[0047] In one embodiment, in the multi-layer fully connected layer, the feature vector obtained by the deep convolutional layer is mapped to the affinity value by the fully connected layer, as follows:

[0048] 1) Perform a linear transformation on the output feature vector x of the deep convolutional layer and use it as the input vector y=Wx+b of the fully connected layer;

[0049] 2) Use the linear rectification function to perform nonlinear transformation on the linear transformation result:

[0050]

[0051] 1) and 2) constitute a layer of mapping network, and the affinity prediction results are output through multiple layers of mapping networks.

[0052] The present invention uses deep learning technology to model the composite structure of oligomers and immune molecules, thereby enhancing the expressiveness of the deep learning model; and realizes a deep model of short sequences of oligomers by improving the deep learning technology, thereby enhancing the complexity and learning ability of a single neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 The overall structure of the antigen affinity prediction model according to one embodiment of the present invention is shown.

[0054] Figure 2 The input network structure of the antigen affinity prediction model according to one embodiment of the present invention is shown.

[0055] Figure 3 The computational network structure of the antigen affinity prediction model according to one embodiment of the present invention is shown. DETAILED DESCRIPTION

[0056] The two most important design steps of deep learning technology are the input feature expression layer design (Embedding Layer) and the deep network structure design of the computing layer. The present invention takes into account the design of the expression layer of the complex, and realizes the deep design of the polypeptide computing layer through effective means, thereby improving the learning ability of the model. Therefore, the model fusion learning amount required is smaller than that of the solution commonly used in the industry, netMHCPan, which improves the efficiency of the process; and the present invention is not limited to fixed-length sequences, and has a wider range of application scenarios.

[0057] In the present invention, the deep residual network model is applicable to the prediction of affinity, antigen presentation and immunogenicity models of all polypeptide sequences; it is also applicable to other short gene sequence scenarios, including but not limited to MHC molecule sequence analysis, DNA high-throughput sequencing fragment analysis, etc. The length of the polypeptide sequence is generally 9 or 10 amino acids, and can also be other lengths.

[0058] In the present invention, tensor addition and attention operations can also be used to integrate the vector of peptides and the vector of immune molecules into one vector. The attention operation changes the tensor product of peptides and typing into an inner product operation, and then nonlinearly normalizes it to calculate the attention weight matrix and multiply it with the peptide / typing tensor. However, tensor addition is not as effective as tensor multiplication; attention operations are common in natural language processing models, such as Google's BERT

[12] . It is also possible to stereo-bind peptides and HLA molecule sequences, or splice peptides and HLA molecules.

[0059] In the present invention, the method for establishing an antigen affinity prediction model based on deep learning includes: (1) obtaining a training data set of oligomer-immune molecule binding structures, wherein the training data set includes the affinity of the oligomer-immune molecule binding; (2) for each group in the training data set of the oligomer-immune molecule binding structure, the sequence of the oligomer, the monomer position of the oligomer and the immune molecule are respectively represented as high-dimensional vectors, and the three high-dimensional vectors are fused into a vector of the oligomer-immune molecule binding structure; (3) using the vector and affinity data of the oligomer-immune molecule binding structure in the training data set to train a deep neural network to establish an oligomer-immune molecule affinity prediction model.

[0060] Preferably, the deep neural network includes a deep convolutional layer and a multi-layer fully connected layer, the deep convolutional layer is used to extract the feature vector of the oligomer-immune molecule in the training data set, input the feature vector into the multi-layer fully connected layer, the multi-layer fully connected layer maps the feature vector into an affinity value, and obtains the network parameters through a back propagation algorithm.

[0061] In an example, in multiple convolutional layers, each convolutional neural network learns the residual of the previous layer: the i-th layer input is recorded as X, and the i-th layer convolutional network is conv i, then the output of the i-th layer is X+conv i (X); that is, the input is copied into three copies, two of which are learned through different convolutional networks, and one of the convolution outputs is subjected to Sigmoid normalization to obtain two vectors A and B, which are then multiplied bitwise to obtain the residual (A×B); the other copy is added to the residual of A×B.

[0062] In one example, in multiple fully connected layers, the feature vectors learned by the computational network are mapped to affinity values ​​through the fully connected network as follows:

[0063] 1) Perform a linear transformation on the input of the mapping network (i.e., the output of the computing network) to transform the input vector x of the mapping layer into y=Wx+b;

[0064] 2) Use the linear rectification function (relu) to perform nonlinear transformation on the linear transformation result:

[0065]

[0066] 1) and 2) form a mapping network, which outputs the affinity prediction results after multiple layers.

[0067] The present invention is exemplified below by the modeling and application of polypeptide-HLA molecules. The polypeptide-HLA molecule dataset used in the present invention is from IEDB public data.

[0068] In the present invention, the pan-specific model (Pan Model) needs to map the polypeptide and HLA molecule sequence text into a unique vector, so that the neural network model can establish a functional mapping relationship between the vector and the prediction index; in addition, the mapping process needs to model the binding structure between the polypeptide and the HLA molecule.

[0069] The main body of the deep neural network model of the present invention consists of three parts. Figure 1 :A neural network consists of three parts: input network, calculation network and mapping network. In the model training process, the network parameters need to be obtained through the training process. The training process uses the vector data of the binding structure of oligomers-immune molecules with known affinity to obtain the network parameters through the back propagation algorithm; in the test / prediction process, the vector of the binding structure of oligomers-immune molecules to be predicted is input into the input network, and the predicted affinity value is output through the calculation network and mapping network. The following is a specific function introduction.

[0070] Input network:

[0071] The function of the input network is to build a high-dimensional vector of peptides and HLA molecules through a neural network, model the peptide-HLA molecule binding structure, and achieve pan-specific input of the model. Different peptides + HLA molecules are mapped to different vectors. The purpose of high dimensionality is to fully capture the subtle differences between different input sequences. Here, the inventor inputs a pseudo sequence of HLA molecules, which has a higher accuracy rate.

[0072] Specific design reference Figure 2 Each amino acid molecule in the peptide sequence and HLA molecule sequence is mapped into a higher-dimensional vector A by a fully connected network; the peptide position is marked as 1 to the peptide length, and the position sequence is mapped to the same dimension through another network; the vector is expanded into a tensor in the graph and then tensor operations are performed to obtain the final input tensor that mixes the peptide and HLA molecule information. Figure 2 The example is 15 peptide length and 128-dimensional mapping output dimension. Figure 2 The fully connected circle with 4 at the top and 3 at the bottom represents "each amino acid molecule is mapped from the character space of A to Z to a high-dimensional continuous numerical space". This is a commonly used representation of a fully connected layer. The upper circle represents the fully connected input dimension, i.e., the amino acid character space, and the lower circle represents the fully connected output dimension, i.e., a higher-dimensional vector. Taking a peptide of length 15 as an example, each position is mapped from an amino acid character to a continuous high-dimensional vector, where the high-dimensional dimension is also exemplified as m. 15×128×1 is the process of expanding a vector into a tensor, the purpose of which is to complete the tensor addition and tensor multiplication in the figure. The principle of tensor operation is to use the last two dimensions as matrices for operation, and the first dimensions as batch size. After adding two tensors of 15×128×1, the result is still 15×128×1, and after multiplying with a tensor of 15×1×128, the result is 15×128×128.

[0073] The language model [12-15] can be used to embed words in text sequences, and the original input data can be mapped into three high-dimensional vectors through a neural network. Here, we take the peptide length n, the HLA molecule length k, and the output dimension mapping to m as an example:

[0074] 1) Peptide representation vector Through a fully connected neural network, the original peptide sequence is regarded as a variable-length vector, and each amino acid molecule in the vector is mapped from the character space of A to Z to a high-dimensional continuous numerical space. The input peptide sequence is (x1, x2, ..., x n ), where x i is the amino acid character, x i ∈{A,C,D,...,Y}, the mapping result is to convert x i Transformed into an m-dimensional vector z i ∈R m, m is a parameter that needs cross-validation debugging, R is a real number space, the sequence (x1, x2, ..., x n ) is transformed into (z1, z2, ..., z n );

[0075] 2) Peptide position representation vector: Through a fully connected neural network, the label of each amino acid position of the peptide is mapped to a high-dimensional (e.g., m-dimensional) continuous numerical space: the input is unified as a vector (1, 2, ..., n), and each position i is mapped into a vector p by the fully connected neural network i ∈R m , R is the real number space, that is, (1, 2, ..., n) and the output is (p1, p2, ,.., p n );

[0076] 3) HLA molecule representation vector: The HLA molecule corresponding to each peptide also needs to be mapped to a high-dimensional continuous numerical space. Considering that each bit of the HLA molecule and the peptide is an amino acid molecule and has the same meaning, the neural network model used for mapping here is the same as that of the peptide representation vector. The input of the HLA molecule vector is (y1, y2, ..., y k ), y i ∈{A,C,D,...,Y}, the output is (z′1,z′2,...,z′ n ), z′ i ∈R m , mapping the k-dimensional original amino acid vector to an n×m-dimensional vector with the same peptide format, the process is divided into the following steps:

[0077] a) Through the neural network in 1) (y1, y2, ..., y k ) is mapped to (y′1, y′2, ..., y′ k ), y′ j ∈R m ;

[0078] b) For (y′1, y′2,...,y′ k ) Take the average to get Copy n copies to get (z′1, z′2, ..., z′ n ),

[0079] There are two mechanisms that can be used to fuse the above three vectors. One is to add the three vectors similar to the traditional language model. However, genome sequences are different from languages. Peptides and HLA molecules have spatial binding structures that need to be captured by more complex mechanisms. Therefore, another mechanism can be used:

[0080] 1) Calculate the peptide characterization vector A = (z1+p1, z2+p2, ..., z n +p n );

[0081] 2) For each element a of A i , perform tensor expansion (Kronecker product): R m →R m×1 ;

[0082] 3) Let the HLA molecule representation vector be B, and for each element b of B i , perform tensor expansion (Kronecker product): R m →R 1×m ;

[0083] 4) Perform bitwise tensor product on A and B in:

[0084]

[0085] 5) Yes Perform a flatten operation. That is, for each element in n bits Convert the result of formula (1) into m 2 dimensional vector (a i1 b i1 , ..., a i1 b im , ..., a im b i1 , ..., a im b im ). This vector is the final input of the computing network.

[0086] The peptides and HLA molecules are mapped to the same high-dimensional space through a neural network, and the three-dimensional structure mapping of the peptide and HLA molecule sequences is generated through tensor multiplication, providing input for a high-efficiency pan-specific affinity model.

[0087] For training data, the computational network and the mapping network need to be combined as a whole to complete the learning process. Through the deep convolutional neural network of the computational network and the fully connected network of the mapping network, the vector and affinity value of the binding structure of the oligomer-immune molecule are input into the computational network and the mapping network for training, completing the learning of the computational network and mapping network parameters.

[0088] Computational Network:

[0089] The computational network includes a multi-layer convolutional neural network, which is used to effectively extract short sequence features through a deep convolutional neural network to obtain a feature vector. Because the input peptide and HLA molecule are very short (the peptide corresponding to type I HLA is usually less than 15 bits, and the peptide corresponding to type II HLA is usually less than 26 bits), low-resolution data will lead to overfitting of the deep convolutional neural network, making the deep convolutional neural network not very effective. In order to increase the complexity of the deep convolutional neural network to enhance the learning ability and improve the effect of using the deep convolutional neural network, a residual neural network is used in combination with a gated linear unit (Gated Linear Unit) activation (see

[15] ) to significantly enhance the generalization ability of the deep convolutional neural network (e.g., at least 5 layers or more, preferably 10 layers or more, such as more than 15 layers, or even more than 20 layers), improve the complexity, and enhance the learning ability of the algorithm. For specific design references, Figure 3 . The input is copied three times, two copies are learned through different convolutional networks, and one of the convolution outputs is subjected to Sigmoid normalization to obtain two vectors A and B. The residual is obtained by bitwise multiplication (A×B in the figure); the other copy is added to the residual of A×B. By adopting the above residual network design ( Figure 3 ), each layer of convolutional neural network learns the residual of the previous layer: the i-th layer input is recorded as X, and the i-th layer convolutional network is conv i , then the output of the i-th layer is X+conv i (X). By learning the residual instead of the input itself, the overfitting effect caused by adding layers to the convolutional neural network is effectively reduced. The final calculation network includes, for example, a 10-layer convolutional neural network, and uses gated linear units as the activation mechanism. The overall process is as follows:

[0090] 1) For input X = (X1, X2, ..., X n ), input two convolutional neural networks conv A and conv B , the output dimensions of the two convolutional neural networks are consistent with the input X;

[0091] 2) Calculate the two convolutional neural network outputs conv A (X) and conv B (X), perform a bitwise Sigmoid mapping σ: R → [0, 1] on the output of one of the convolutional neural networks. For example, for conv B :

[0092]

[0093] 3) Bitwise multiplication conv A (X) i and σ(conv B (X)i )get(conv A (X)1×σ(conv B (X)1), conv A (X)2×σ(conv B (X)2),...,conv A (X) n ×σ(conv B (X) n )), and X=(X1,X2,...,X n ) performs vector addition to obtain the output of the computational layer, which is then passed to the next layer after passing through the Dropout layer;

[0094] Considering that the length of peptides is variable in practice, the output length after passing through all convolutional layers is also variable. A global pooling layer θ is added after the last convolutional neural network layer: R n×p →R p , p is the output dimension of the last layer, and the feature vector is obtained.

[0095] For the deep convolutional neural network model of amino acid sequences, the generalization of the deep convolutional residual network structure is improved by learning the residual output of each layer of the convolutional neural network instead of the output entity design, so that a single model can achieve good accuracy, thereby significantly improving the overall process efficiency.

[0096] Mapping the network:

[0097] The mapping network consists of multiple layers, whose function is to map the feature vectors learned by the computational network into affinity values ​​through a fully connected network. The overall process is as follows:

[0098] 1) Perform a linear transformation on the input of the mapping network (i.e., the feature vector of the output of the computing network) to transform the input vector x of the mapping network into y=Wx+b;

[0099] 2) Use the linear rectification function (relu) to perform nonlinear transformation on the linear transformation result:

[0100]

[0101] 1) and 2) constitute a layer of mapping network, and the affinity prediction results are output through multiple layers of mapping networks.

[0102] For test / prediction data, the peptides to be tested / predicted + HLA molecules are input into the deep convolutional neural network of the trained computational network, and the feature vectors of oligomers-immune molecules are extracted according to the parameters of the network. The mapping network and the parameters of the mapping network are input, and the multi-layer fully connected layers of the mapping network map the feature vectors into affinity values. In one example, the affinity value can be IC50 .

[0103] The effect of the antigen affinity prediction model of the present invention is verified by the following examples.

[0104] Example

[0105] The algorithm does not require graphics card resources, but the corresponding software environment needs to be configured:

[0106] Python: 3.6.5

[0107] Anaconda: 4.5.4

[0108] Tensorflow: 1.3.0

[0109] Cuda:8.0

[0110] Cudnn: 6.0

[0111] When the inventor uses it, by executing the prediction script provided by this software: python predict.py file path to be predicted, the prediction result of each experimental data will be generated. The file format to be predicted needs to be organized into the "mhc, peptide" format. The inventor put an example file in the code, test_batch.csv in the data folder, as follows:

[0112]

[0113] The test data is a pMHC (9, 10mer) affinity quantitative data set verified by the MHC ligand assay method of the IEDB database (http: / / www.iedb.org / ), including 113 HLA typings such as A0101, A0201, and A2402. Using this script will generate a predict.csv file in the result folder. This file records the peptide, typing type, and IC predicted by the software for each sample. 50 The results generated by the IEDB data test are as follows:

[0114]

[0115] The inventors compared the algorithm results of the present invention with those of NetMHCPan on this data. 50It is converted into the peptide positive and negative properties which are more concerned in actual research and development. Therefore, the evaluation indicators used are the area under the ROC curve (AUC) and f1 score commonly used in machine learning classification problems. The model of the present invention only uses one neural network model, while NetMHCPan uses 200 model fusion learning [netpan3 (2)]. The comparison results are as follows:

[0116]

[0117]

[0118] The results show that the present invention uses a single model to achieve the accuracy of NetMHCPan using 200 models. At the same time, the execution efficiency of the model of the present invention is very high. On the NVIDIA p6000 graphics card, the method of the present invention takes about 30 minutes to train 110,000 data and takes about 1 minute to predict 24,000 data in the test file test_batch.csv. The polypeptide length distribution of this set of data is 8 to 11 bits, and the HLA distribution of the training set and the test set is inconsistent.

[0119] Through this embodiment, the method of the present invention has the following four advantages: 1) By designing a short peptide-HLA molecule complex characterization layer and a short peptide deep residual network, the method of the present invention improves the affinity prediction accuracy of a single neural network model; 2) By improving the accuracy of a single model, the method of the present invention also reduces the execution time overhead of the model. On the inventor's machine (NVIDIA p6000 graphics card), the training time for 116559 data only takes half an hour; 3) The method of the present invention allows the input polypeptide to have a variable length; 4) Because the present invention develops a pan-specific algorithm, it can be used for new typing, that is, rare typing that has never been seen in the previous training set, and can also be predicted by the algorithm of the present invention.

[0120] References:

[0121] [1] Nielsen M, Lundegaard C, Worning P, et al. Reliable prediction of T-cell epitopes using neural networks with novel sequence representations. Protein Sci 2003;12:1007-17.

[0122] [2]Zhang GL,Ansari HR,Bradley P,et al.Machine learning competition inimmunology-prediction of HLA class I binding peptides.J Immunol Meth 2011;374:1-4.

[0123] [3]Carreno BM,Magrini V,Becker-Hapak M et al.Cancer immunotherapy.Adendritic cell vaccine increases the breadth and diversity of melanomaneoantigen-specifc T cells.Science 2015;348L803-8.

[0124] [4]Walter S,Weinschenk T,Stenzl A,et al.Multipeptide immune responseto cancer vaccine IMA901 after single-dose cyclophosphamide associates withlonger patient survival.Nat Med 2012;18:1254-61.

[0125] [5]Yadav M,Jhunjhunwala S,Phung QT,et al.Predicting immunogenictumour mutations by combining mass spectrometry and exome sequencing.Nature2014;515:572-76.

[0126] [6]Nielsen,M.,Andreatta,M.NetMHCpan-3.0;improved prediction ofbinding to MHC class I molecules integrating information from multiplereceptor and peptide length datasets.Genome Med 8,33(2016)doi:10.1186 / s13073-016-0288-x.

[0127] [7]Liu,Z.,Cui,Y.,Xiong,Z.et al.DeepSeqPan,a novel deep convolutionalneural network model for pan-specific class I HLA-peptide binding affinityprediction.Sci Rep 9,794(2019)doi:10.1038 / s41598-018-37214-1.

[0128] [8]PhloyPhisut,P.,Pomputtapong,N.,Sriswasdi,S.et al.MHCSeqNet:a deepneural network model for universal MHC binding prediction.BMC Bioinformatics20,270(2019)doi:10.1186 / s12859-019-2892-4.

[0129] [9]Sidhom,J.-W.a.P.D.a.B.A.AI-MHC:an allele-integrated deep learningframework for improving Class I&Class II HLA-binding predictions.bioRxiv,p.318881(2018).

[0130]

[10] Yan Hu,Ziqiang Wang,Hailin Hu,FangPing Wan,Lin Chen,YuanpengXiong,Xiaoxia Wang,Dan Zhao,Weiren Huang,Jianyang Zeng,ACME:pan-specificpeptide-MHC class I binding prediction through attention-based deep neuralnetworks,Bioinformatics.

[0131]

[11] Han,Youngmahn&Kim,Dongsup.(2017).Deep convolutional neuralnetworks for pan-specific peptide-MHC class I binding prediction.BMCBioinformatics.18.585.10.1186 / s12859-017-1997-x.

[0132]

[12] Devlin J,Chang M W,Lee K,et al.Bert:Pre-training of deepbidirectional transformers for language understanding[J].arXiv preprintarXiv:1810.04805,2018.

[0133]

[13] Vaswani A,Shazeer N,Parmar N,et al.Attention is all you need[C] / / Advances in neuralinformation processing systems.2017:5998-6008.

[0134]

[14] Gehring J,Auli M,Grangier D,et al.Convolutional sequence tosequence learning[C] / / Proceedings of the 34th International Conference onMachine Learning-Volume 70.JMLR.org,2017:1243-1252.

[0135]

[15] Dauphin Y N,Fan A,Auli M,et al.Language modeling with gatedconvolutional networks[C] / / Proceedings of the 34th International Conferenceon Machine Leaming-Volume 70.JMLR.org,2017:933-941.

Claims

1. A method for predicting the affinity of oligomers to immune molecules based on deep learning, the method comprising: (1) obtaining a training data set of oligomer-immune molecule binding structures, wherein the training data set includes the affinity of the oligomer-immune molecule binding; (2) for each group in the training data set of the oligomer to be tested and the immune molecule and the binding structure of the oligomer-immune molecule, the sequence of the oligomer, the monomer position of the oligomer and the immune molecule are represented as high-dimensional vectors respectively, and the three high-dimensional vectors are fused into a vector of the binding structure of the oligomer-immune molecule; Wherein, the vector for fusing the three high-dimensional vectors into the oligomer-immune molecule binding structure includes: Adding the oligomer sequence and the high-dimensional vector of the monomer position of the oligomer to form a characterization vector of the oligomer; Performing tensor multiplication, tensor addition or attention operation on the characterization vector of the oligomer and the vector of the immune molecule to obtain a vector of the binding structure of the oligomer-immune molecule; (3) using the vectors and affinity data of the oligomer-immune molecule binding structure in the training data set to train a deep neural network to establish an oligomer-immune molecule affinity prediction model, wherein the deep neural network uses a residual neural network combined with a gated linear unit activation; The deep neural network includes a deep convolution layer and a multi-layer fully connected layer, wherein the deep convolution layer is used to extract the feature vector of the oligomer-immune molecule in the training data set and input it into the multi-layer fully connected layer, which maps the feature vector into an affinity value and obtains the network parameters through a back propagation algorithm; (4) inputting the vector of the oligomer-immune molecule binding structure to be tested into the antigen affinity prediction model, and using the trained deep neural network to predict the affinity of the oligomer-immune molecule binding to be tested; The oligomer is selected from the group consisting of: a polypeptide and a nucleic acid hapten.

2. The method according to claim 1, wherein the immune molecule is a major histocompatibility complex molecule.

3. The method according to claim 2, wherein the major histocompatibility complex molecule is human leukocyte antigen.

4. The method according to claim 3, wherein the human leukocyte antigen is class I human leukocyte antigen or class II human leukocyte antigen.

5. According to the method described in any one of claims 1 to 4, in (2), a monomer vector representation that can reflect the relationship between any two monomers is obtained through word embedding.

6. The method according to any one of claims 1 to 4, wherein the oligomer is a polypeptide, and the oligomer sequence input in (2) is (x1, x2, ..., x n ), where x i is the amino acid character, x i ∈{A,C,D,...,Y}, the mapping result is to convert x i Transformed into an m-dimensional vector z i ∈R m , R is the real number space, the sequence (x1, x2, ..., x n ) is transformed into (z1, z2, ..., z n ).

7. The method according to claim 6, wherein in (2) the monomer position characterization vector is: the label of each amino acid position of the polypeptide is mapped to a high-dimensional continuous numerical space: the input is unified into a vector (1, 2, ..., n), and each position i is mapped into a vector p by a fully connected neural network i ∈R m , R is the real number space, that is, (1, 2, ..., n) and the output is (p1, p2, ..., p n ).

8. The method according to any one of claims 1 to 4, wherein in (2) the input of the immune molecule vector is (y1, y2, ..., y k ), y i ∈{A,C,D,...,Y}, the output is (z′1,z′2,...,z′ n ), z′ i ∈R m , mapping from the k-dimensional raw amino acid vector to an n×m-dimensional vector with the same format as the oligomer.

9. According to the method described in any one of claims 1 to 4, for each deep convolutional layer, the input is copied three times, two of which are learned through different convolutional networks, and one of the convolution outputs is subjected to Sigmoid normalization to obtain two vectors A and B, and the residual A×B is obtained by bitwise multiplication; the other copy is added to the residual of A×B.

10. According to the method of claim 9, a global pooling layer θ is added after the last convolutional layer: R n×p →R p , p is the output dimension of the last layer, and the feature vector is obtained.

11. The method according to claim 10, wherein in the multi-layer fully connected layer, the feature vector obtained by the deep convolutional layer is mapped to the affinity value by the fully connected layer, and the method is as follows: 1) Perform a linear transformation on the output feature vector x of the deep convolutional layer and use it as the input vector y=Wx+b of the fully connected layer; 2) Use the linear rectification function to perform nonlinear transformation on the linear transformation result: 1) and 2) constitute a layer of mapping network, and the affinity prediction results are output through multiple layers of mapping networks.

12. According to the method of claim 10 or 11, in (4), the trained deep neural network includes a deep convolutional layer and a multi-layer fully connected layer, the deep convolutional layer is used to extract the feature vector of the oligomer-immune molecule to be tested, input it into the multi-layer fully connected layer, and predict the affinity of the oligomer-immune molecule binding to be tested.

13. A method for establishing an oligomer-immune molecule affinity prediction model based on deep learning, the method comprising establishing an oligomer-immune molecule affinity prediction model through (1)-(3) of any one of the methods of claims 1-12; the oligomer is selected from: polypeptides and nucleic acid haptens.

14. An antigen affinity prediction model established by the method of claim 13, wherein the antigen affinity prediction model comprises: Data acquisition module, input vector building module and model building module, The data acquisition module is used to acquire a data set of the oligomer-immune molecule binding structure, wherein the data set includes the affinity of the oligomer-immune molecule binding; The input vector establishment module is used to use the data set of the oligomer-immune molecule binding structure to map the sequence of the oligomer, the monomer position of the oligomer and the immune molecule into high-dimensional vectors respectively, and fuse the three high-dimensional vectors into the vector of the oligomer-immune molecule binding structure; The model building module is used to train a deep neural network using the vector and affinity data of the oligomer-immune molecule binding structure to establish an antigen affinity prediction model.

Images