An antibody sequence and structure co-design method based on a flow model

Abstract

The application discloses an antibody sequence structure collaborative design method based on a flow model, and comprises the following steps: S1, selecting a structural antibody data set as a general antibody data; S2, obtaining amino acid sequence information and structure information; S3, encoding the sequence information and the structure information into a graph neural network, updating the neural network, and obtaining amino acid coding; S4, inputting the amino acid coding into a first flow model for transformation, obtaining an amino acid type, inputting the predicted amino acid type into the graph neural network, updating the graph neural network, and obtaining new amino acid coding; S5, inputting the new amino acid coding into a second flow model for transformation, and obtaining three-dimensional coordinate information of the amino acid; and S6, performing a loss function on the amino acid type and the three-dimensional coordinate information and a true value, minimizing the loss function, and optimizing generated model parameters.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of antibody sequence structure prediction, and more particularly to an antibody sequence structure collaborative design method based on a flow model. BACKGROUND

[0002] At present, the traditional machine learning method is for a specific data set, generally with a large amount of data, and attempts to learn a predictive model or a generative model, so that the model can model the probability distribution in the data set, thereby obtaining accurate prediction or generating samples conforming to the characteristics of the data set. A method commonly used in generative models maps data set samples with complex distribution to hidden space with simple distribution through reversible, differentiable and continuous nonlinear transformation, so that the model can better model the probability distribution in the data set. The variable region is the position of the antibody binding to the antigen, and the composition and arrangement of the amino acids determine the antigen binding specificity of the antibody. There are some regions where amino acids can be frequently combined, which are called complementarity determining regions (CDR regions), and the CDR regions determine the idiotypic (anti-idiotypic antibody expression) of the antibody. Antibody design mainly targets the amino acid sequence of the CDR region.

[0003] However, existing machine learning is mostly focused on the study of predictive models, and the study of generative models is relatively insufficient. In actual problems, it is often necessary to use machine learning to generate new samples. How to effectively model the data distribution of the data set and generate reliable samples requires more research on generative models. The most commonly used generative models (generative adversarial networks and variational autoencoders) have their own shortcomings. The measure of the generative adversarial network is learned by data-driven learning, which is similar to a black box system and has low interpretability, and has the characteristics of difficult training; the flow model is in the stage of emerging development, and the current application is limited to computer vision tasks (face image generation, etc.), and the application in actual scenes is still relatively lacking; the current machine learning protein design method is mainly divided into two types: sequence design and structure-based sequence design. Sequence design considers the amino acid sequence of the protein in the data set but does not consider the protein structure, which is similar to the natural language processing method of directly generating sequences, and lacks the use of protein structure information. Structure-based sequence design takes the structure of the protein as the input condition to generate the amino acid sequence, which can utilize the protein structure information to a certain extent, but it is difficult to generate new structures. In addition, antibody design and protein design are slightly different, antibody design generates the amino acid sequence of the CDR region, and maintains the constancy of the framework region (FR region), and there is currently a lack of machine learning design method for antibodies.

[0004] Therefore, it is an urgent problem for those skilled in the art to design a generative model based on a flow model in combination with the sequence and structure of the antibody. SUMMARY

[0005] Therefore, the application provides an antibody sequence and structure co-design method based on a flow model.

[0006] To achieve the above object, the application adopts the following technical scheme:

[0007] An antibody sequence and structure co-design method based on a flow model comprises the following steps:

[0008] S1, selecting a structure antibody data set as a general antibody data;

[0009] S2, obtaining amino acid sequence information and structure information;

[0010] S3, encoding the sequence information and structure information into a graph neural network, updating the neural network, and obtaining amino acid encoding;

[0011] S4, inputting the amino acid encoding into a first flow model for transformation to obtain amino acid types, inputting the predicted amino acid types into the graph neural network for updating to obtain new amino acid encoding;

[0012] S5, inputting the new amino acid encoding into a second flow model for transformation to obtain three-dimensional coordinate information of the amino acid;

[0013] S6, performing a loss function on the amino acid types and three-dimensional coordinate information and a true value, minimizing the loss function, and optimizing the generative model parameters.

[0014] Preferably, the step S2 specifically comprises that the amino acid sequence information is composed of types of a plurality of amino acids connected head to tail, and the structure information is composed of three-dimensional coordinates of C, C α , and N atoms of a plurality of amino acids connected head to tail.

[0015] Preferably, the step S3 specifically comprises:

[0016] establishing a graph structure and encoding the sequence information and structure information, obtaining the structure information of the amino acid sequence, and using C αThe three-dimensional coordinates of atoms represent the spatial positions of their corresponding amino acids. Each amino acid is considered a node in the graph structure. Adjacent amino acid nodes in the sequence are connected by edges. In addition, for each amino acid, the six nearest amino acids are also considered its adjacent nodes, and the latter six are connected to the former by edges. After the graph structure is established, the torsion angles φ, ψ, and ω of each amino acid position on the protein backbone are obtained using structural information and spatial geometry knowledge. The torsion angle information processed by sine and cosine is encoded as node features in the graph structure.

[0017] v i =(cos φ i sin φ i cos ψi, sin ψ i cos ω i sin ω i )

[0018] In the diagram structure, the spatial positions of two adjacent amino acids are both determined by their corresponding C. α The atoms are represented by three-dimensional coordinates, thus encoding the Euclidean distance as an edge feature in the graph structure:

[0019] e ij =(||x i，α -x j，α ||)

[0020] Where, x i,α For the i-th amino acid, C α The three-dimensional coordinates of the atom, where the j-th amino acid is adjacent to the i-th amino acid in the graph structure;

[0021] The graph neural network is updated L times based on the graph structure and its corresponding node and edge features. The update formula for each time is:

[0022]

[0023] in, The node remains hidden from the last update. E(s j ) represents the encoding associated with amino acid type; e i,j The features of the edges; FFN(·) is a two-layer fully connected network; It is a 20-dimensional row vector;

[0024] The update of the above graph neural network can obtain the hidden state of each node (amino acid), which can be expressed as:

[0025]

[0026] in, The graph structure constructed above; MPN(·) represents the update method of the graph neural network.

[0027] Preferably, step S4 specifically includes:

[0028] By transforming the flow model to predict the amino acid type of the new point, the hidden state of the (t+1)th amino acid is obtained. Perform a 16-layer flow model transformation on it:

[0029]

[0030] Among them, the predicted amino acid types One-hot encoding; σ k ,μ k These are the learnable parameters for the transformation of the k-th layer flow model.

[0031] Preferably, step S5 specifically includes:

[0032] s t+1 Substitute to establish graph structure The new hidden state of each point is obtained by updating the graph neural network:

[0033]

[0034] The network parameters of MPN(·) are different from those in step S3; It is a 9-dimensional row vector;

[0035] After obtaining the new hidden state of each point, perform a 16-layer flow model transformation on it:

[0036]

[0037] Among them, predicting the C, C of each amino acid in the main chain. α ,N atom three-dimensional coordinates σ k ,μ k These are the learnable parameters for the transformation of the k-th layer flow model.

[0038] Preferably, step S6 specifically includes:

[0039] Training the generative model: Step S1 obtains a general antibody dataset; steps S2-S5 integrate sequence and structural information, and use a graph neural network to predict the type of each amino acid. t and each amino acid C, C on the main chain α The three-dimensional coordinates of the N atom A loss function is constructed by comparing the data with the corresponding real antibodies in the dataset; for amino acid types (sequence information), cross-entropy is selected as the loss function.

[0040]

[0041] where T is the length of the sequence set artificially; is the i-th dimension of the one-hot encoding of the t-th amino acid in the dataset; is the i-th dimension of the one-hot encoding of the predicted t-th amino acid;

[0042] For three-dimensional coordinates, the least squares is selected as the loss function:

[0043]

[0044] where y i,C ,y i,α ,y i,N is the true value of the three-dimensional coordinates;

[0045] The total loss function of the generation model is:

[0046] Loss = L sep + L struct

[0047] After the loss function is constructed, back propagation is used to minimize the loss function Loss by gradient descent to update the parameters of the generation model and realize the training of the model.

[0048] Training of the classifier: select a specific antibody dataset for a certain antigen, and pre-train a classifier. For this classifier, the input is the antibody sequence structure and antigen sequence structure, and the output is the prediction of whether the antigen antibody can specifically bind. For the sequence and structure of the antibody, the hidden state of each amino acid is obtained by using the graph structure creation and graph neural network processing method:

[0049]

[0050] The hidden state of each amino acid is summed, and the average is taken for the sequence length, to obtain the hidden state of :

[0051]

[0052] For the amino acid sequence and structure of the antigen, the hidden state of each amino acid is obtained by using the above graph structure creation and graph neural network processing method:

[0053]

[0054] The hidden state of each amino acid is summed, and the average is taken for the sequence length, to obtain the hidden state of :

[0055]

[0056] After obtaining the hidden state of the antibody and the antigen, whether the antibody and the antigen can be combined is predicted by a two-layer neural network:

[0057] pred = Sigmoid(FFN([h antibody , h antigen ]))

[0058] wherein, the Sigmoid(·) function outputs a probability value; the FFN(·) is a two-layer neural network, and outputs a scalar; the [·] is a vector splicing operation; the cross-entropy is taken as a loss function of the true value and the predicted value, and the classifier is trained;

[0059] After obtaining the classifier, an FR region is given, and the antibody amino acid sequence is generated by using the generated model trained in the previous step; for a specific sequence and structure of the antigen, the generated antibody amino acid sequence and structure are predicted by the classifier, if the antigen-antibody binding possibility obtained by the prediction is high, the generated antibody is added to the specific antibody data set of the antigen; the generated antibody amino acid sequence is compared with the antibody sequence in the antibody data set, a loss function is obtained, and the model is fine-tuned by using gradient descent to minimize the loss function;

[0060] The specificity binding probability of the generated antibody is scored by using the above classifier, and the higher the probability is, the stronger the specificity binding ability of the antibody and the antigen is, and the reinforcement learning encourages the generation of the antibody sequence with a high score, and the model is fine-tuned according to the antibody sequence.

[0061]

[0062]

[0063] wherein, the theta is a parameter of the generated model; the epsilon is a given learning rate; the pred(s) is a prediction probability of the classifier; the p θ (s) is a probability of the generated model generating a given sequence.

[0064] According to the technical scheme, compared with the prior art, the antibody sequence structure collaborative design method based on a flow model is provided; by specializing the protein sequence generation to the antibody sequence generation, the sequence and the structure of the antibody are input into the graph neural network, and then the sequence and the structure are generated by using the graph neural network, more accurate data distribution can be obtained, and the generated antibody sequence and structure are closer to the real situation. BRIEF DESCRIPTION OF DRAWINGS

[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute a part of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of the provided drawings.

[0066] Figure 1 The drawings are overall flow schematic diagrams provided by the present application.

[0067] Figure 2 The drawings are partial schematic diagrams of amino acid sequences provided by the present application.

[0068] Figure 3 The drawings are schematic diagrams of antibody amino acid sequence structure generation provided by the present application. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0070] The embodiments of the present application disclose an antibody sequence structure collaborative design method based on a flow model, comprising the following steps:

[0071] S1, selecting a structural antibody data set as a general antibody data;

[0072] S2, obtaining amino acid sequence information and structure information;

[0073] S3, encoding the sequence information and structure information into a graph neural network, updating the neural network, and obtaining amino acid encoding;

[0074] S4, inputting the amino acid encoding into a first flow model for transformation to obtain amino acid types, inputting the predicted amino acid types into the graph neural network for updating to obtain new amino acid encoding;

[0075] S5, inputting the new amino acid encoding into a second flow model for transformation to obtain three-dimensional coordinate information of the amino acid;

[0076] S6, performing loss function on the amino acid types and three-dimensional coordinate information and the true value, minimizing the loss function, and optimizing the generated model parameters.

[0077] To further optimize the above technical solution, step S2 specifically comprises: the amino acid sequence information is composed of types of a plurality of amino acids connected head to tail, and the structure information is composed of three-dimensional coordinates of C, C α and N atoms of a plurality of amino acids connected head to tail.

[0078] To further optimize the above technical solution, step S3 specifically comprises:

[0079] The sequence information and the structure information are encoded by establishing a graph structure, after the structure information of the amino acid sequence is obtained, the three-dimensional coordinates of C α atoms are used to represent the spatial positions of the corresponding amino acids, each amino acid is regarded as a node in the graph structure, and adjacent amino acid nodes in the sequence are connected by edges, in addition, for each amino acid, the six amino acids closest to it are also regarded as its adjacent nodes, and the latter six are connected to the former by edges, after the establishment of the graph structure, the torsion angles φ, ψ and ω of each amino acid position on the protein main chain are obtained by using the structure information and spatial geometry knowledge, and the torsion angle information processed by cosine and sine is encoded as the node features in the graph structure:

[0080] v i =(cos φ i , sin φ i , cos ψ i , sin ψ i , cos omega i , sin omega i )

[0081] The spatial positions of two adjacent amino acids in the graph structure are represented by the three-dimensional coordinates of their corresponding C α atoms, therefore, the Euclidean distance is encoded as the edge features in the graph structure:

[0082] e ij =(||x i，α -x j，α ||)

[0083] wherein x i,α is the three-dimensional coordinates of the C α atom in the i-th amino acid, and the j-th amino acid is adjacent to the i-th amino acid in the graph structure.

[0084] The update of the graph neural network is based on the above graph structure and its corresponding node and edge features, and the graph neural network is updated L times, and the update formula of each time is:

[0085]

[0086] wherein h is the node hidden state of the last update, E(s j ) is the code related to the type of amino acid; e i,j is the feature of the edge; FFN(·) is a two-layer fully connected network; is a 20-dimensional row vector;

[0087] The update of the above graph neural network can obtain the hidden state of each node (amino acid), which is expressed as:

[0088]

[0089] wherein, is the graph structure constructed above; MPN(·) represents the update mode of the graph neural network.

[0090] To further optimize the above technical solution, step S4 specifically comprises:

[0091] The amino acid type of the new point is predicted through the flow model transformation, and the hidden state of the (t+1) th amino acid is obtained The 16-layer flow model transformation is performed on it:

[0092]

[0093] wherein, the predicted amino acid type is one-hot encoding; sigma k , mu k are the learnable parameters of the kth layer flow model transformation.

[0094] To further optimize the above technical solution, step S5 specifically comprises:

[0095] Substitute s t+1 into the graph structure The new hidden state of each point is obtained by using the update mode of the graph neural network:

[0096]

[0097] wherein, the network parameters of MPN(·) are different from those of step S3; is a 9-dimensional row vector;

[0098] After obtaining the new hidden state of each point, the 16-layer flow model transformation is performed on it:

[0099]

[0100] wherein, the predicted three-dimensional coordinates of the C, C α and N atoms corresponding to each amino acid on the main chain sigma k , mu k are the learnable parameters of the kth layer flow model transformation.

[0101] To further optimize the above technical solutions, step S6 specifically comprises:

[0102] Training of the generation model: step S1 obtains the antibody general data set, and steps S2-S5 integrate sequence and structure information, and utilize a graph neural network to predict the type s of each amino acid t and the three-dimensional coordinates of each amino acid C, C α , and N atoms on the main chain Compared with the corresponding real antibodies in the data set, a loss function is constructed; for the amino acid type (sequence information), cross entropy is selected as the loss function:

[0103]

[0104] Wherein, T is a sequence length set artificially; is the i-th dimension of the one-hot encoding of the corresponding t-th amino acid in the data set; is the i-th dimension of the one-hot encoding of the predicted t-th amino acid;

[0105] For the three-dimensional coordinates, least squares is selected as the loss function:

[0106]

[0107] Wherein y i,C , y i,α , y i,N is the three-dimensional coordinate of the true value;

[0108] The total loss function of the generation model is:

[0109] Loss = L sep + L struct

[0110] After the loss function is constructed, back propagation is performed, the gradient descent is utilized to minimize the loss function Loss, the parameters of the generation model are updated, and the training of the model is realized;

[0111] Training of the classifier: a specific antibody data set for a certain antigen is selected, and a classifier is pre-trained; for the classifier, the input is the antibody sequence structure and the antigen sequence structure, and the output is the prediction of whether the antigen antibody can specifically bind; for the sequence and structure of the antibody, the hidden state of each amino acid is obtained by utilizing the processing mode of the graph structure creation and the graph neural network:

[0112]

[0113] The hidden state of each amino acid is summed, and the sequence length is averaged to obtain the hidden state of ​

[0114]

[0115] For the amino acid sequence and structure of the antigen, the hidden state of each amino acid is obtained by using the above graph structure creation and graph neural network processing method:

[0116]

[0117] The hidden state of each amino acid is summed and averaged over the sequence length to obtain the hidden state of

[0118]

[0119] After obtaining the hidden state of the antibody and antigen, whether they can bind is predicted by a two-layer neural network:

[0120] pred=Sigmoid(FFN([h antibody ,h antigen ]))

[0121] Where, Sigmoid(·) function outputs probability value; FFN(·) is a two-layer neural network, outputting a scalar; [·] is a vector splicing operation; cross-entropy is used as the loss function of true value and predicted value to train the classifier.

[0122] After obtaining this classifier, given a FR region, the antibody amino acid sequence is generated by using the generated model trained in the previous step; for a specific sequence and structure of the antigen, the generated antibody amino acid sequence and structure are predicted by the classifier, if the predicted antigen antibody binding probability is high, the generated antibody is added to the specific antibody data set for a certain antigen; the generated antibody amino acid sequence is compared with the antibody sequence in the antibody data set to obtain the loss function, which is back propagated to minimize the loss function by gradient descent to fine-tune the model.

[0123] The specificity binding probability of the generated antibody is scored by using the above classifier, the higher the probability, the stronger the specificity binding ability of the antibody and the antigen, reinforcement learning encourages the generation of antibody sequences with high scores, and the model is fine-tuned based on this:

[0124]

[0125]

[0126] Where, θ is the parameter of the generated model; ∈ is the given learning rate; pred(s) is the prediction probability of the classifier; p θ (s) is the probability of the generated model generating a given sequence.

[0127] ​As shown in Figure 2 1, the part selected by the black frame represents the position corresponding to one residue in the antibody amino acid sequence; 2, the carbon atom marked with subscript 1 is marked as C α (alpha carbon), the carbon atom marked with subscript 2 is marked as C; 3, the C α connected to each amino acid residue is different, thus determining the type of amino acid; 4, in one residue, there are three corresponding three-dimensional coordinates for two carbon atoms and one nitrogen atom; 5, φ, ψ, ω respectively represent the torsion angles between N-C α ,C α -C, C-N

[0128] The advantages of the present application are:

[0129] The sequence and structure of the antibody are cooperatively designed. The protein sequence generation is specialized to the antibody sequence generation in this model. The sequence and structure of the antibody are input into the graph neural network together, and the graph neural network is used to generate the sequence and structure respectively.

[0130] The flow model is used to model the data. The flow model can obtain more accurate data distribution, so that the generated antibody sequence and structure are closer to the real situation.

[0131] The fine-tuned model is more targeted to specific problems. After the model is trained with a general antibody data set, the model has the ability to generate antibodies, but the specificity is not strong. The model is fine-tuned with an antibody data set for a specific antigen, so that the model can be quickly applied to practical situations and generate antibodies with specificity.

[0132] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0133] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for structure co-design of antibody sequence based on flow model, characterized in that, The method comprises the following steps: S1, selecting a structural antibody data set as a general antibody data; S2, obtaining amino acid sequence information and structure information; S3, encoding the sequence information and structure information into a graph neural network, updating the neural network, and obtaining amino acid encoding; S4, inputting the amino acid encoding into a first flow model for transformation to obtain an amino acid type, inputting the predicted amino acid type into the graph neural network for updating to obtain new amino acid encoding; S5, inputting the new amino acid encoding into a second flow model for transformation to obtain three-dimensional coordinate information of the amino acid; S6, performing a loss function on the amino acid type and three-dimensional coordinate information and a true value, minimizing the loss function, and optimizing the generated model parameters; The step S3 specifically comprises: A graph structure is constructed and sequence and structural information are encoded. Once the structural information of the amino acid sequence is obtained, it is used... The three-dimensional coordinates of atoms are used to represent the spatial positions of their corresponding amino acids. Each amino acid is considered a node in the graph structure. Adjacent amino acid nodes in the sequence are connected by edges. In addition, for each amino acid, the six nearest amino acids are also considered its adjacent nodes, and the latter six are connected to the former by edges. After the graph structure is established, the torsion angle of each amino acid position on the protein backbone is calculated using structural information and spatial geometry. The torsion angle information processed by sine and cosine is encoded as node features in the graph structure: ； The spatial positions of two adjacent amino acids in a graph structure are both represented by their corresponding three-dimensional coordinates of atoms, so the Euclidean distance is encoded as an edge feature in a graph structure: ； in, For the i-th amino acid The three-dimensional coordinates of the atom, where the j-th amino acid is adjacent to the i-th amino acid in the graph structure; The update of the graph neural network is based on the above graph structure and its corresponding node and edge features, and the graph neural network is updated L times, and the update formula of each time is: ； wherein, is the hidden state of the node for the last update, ; is the encoding related to the amino acid type; is the feature of the edge; is a two-layer fully connected network; is a 20-dimensional row vector; The update of the above graph neural network can obtain the hidden state of each amino acid node, which is expressed as: ； wherein, ; is the graph structure constructed above; represents the update manner of the graph neural network; The step S4 specifically comprises: The amino acid type of the new point is predicted through the flow model transformation, and the hidden state of the t+1th amino acid is obtained , and 16-layer flow model transformation is performed ； wherein the predicted amino acid type is one-hot encoded; is the learned parameter of the k-th layer flow model transformation; The step S5 specifically comprises: Substituting into the graph structure , the new hidden state of each point is obtained by updating the graph neural network: ； wherein The network parameters of step S3 are different from those of step S2; is a 9-dimensional row vector; After obtaining the new hidden state of each point, a 16-layer flow model transformation is performed: ； wherein the predicted main chain of each amino acid corresponds to atomic three-dimensional coordinates ; is the learned parameter for the k-th layer flow model transformation; The step S6 specifically comprises: Training of the generative model: Step S1 obtains the antibody general dataset, steps S2-S5 integrate sequence and structure information, and use graph neural network to predict the type of each amino acid and the three-dimensional coordinates of each amino acid on the main chain atom ; compare with the real antibody corresponding in the dataset, construct the loss function; for the amino acid type, select cross entropy as the loss function: ； wherein T is a sequence length set artificially; is the i-th dimension of the one-hot encoding of the t-th amino acid in the data set; is the i-th dimension of the one-hot encoding of the t-th amino acid in the prediction; For three-dimensional coordinates, the least squares is selected as the loss function: ； wherein a three-dimensional coordinate that is a true value; The total loss function of the generated model is: ； After the loss function is constructed, the gradient descent is used to minimize the loss function Loss, and the parameters of the generated model are updated to realize the training of the model; Training of the classifier: select a specific antibody dataset for an antigen, pre-train a classifier, for which the input is the antibody sequence structure and antigen sequence structure, and the output is the prediction of whether the antigen antibody can specifically bind; for the sequence and structure of the antibody, the hidden state of each amino acid is obtained by using the processing method of graph structure creation and graph neural network: ; The hidden state of each amino acid is summed and averaged over the sequence length to obtain the hidden state of the sequence: : ； For the amino acid sequence and structure of the antigen, the above graph structure creation and graph neural network processing method are used to obtain the hidden state of each amino acid: ； The hidden state of each amino acid is summed and averaged over the sequence length to obtain the hidden state of the sequence: : ； After obtaining the hidden states of the antibody and the antigen, whether they can be combined is predicted through two-layer neural networks: ； wherein, a function outputs a probability value; is a two-layer neural network that outputs a scalar; is a vector concatenation operation that trains the classifier using cross-entropy as a loss function between true and predicted values; After obtaining the classifier, an FR region is given, and the antibody amino acid sequence is generated by using the generated model trained in the last step; for a specific sequence and structure of the antigen, the generated antibody amino acid sequence and structure are predicted by the classifier, if the antigen-antibody binding probability is high, the generated antibody is added to the specific antibody data set of the antigen; the generated antibody amino acid sequence is compared with the antibody sequence in the antibody data set to obtain a loss function, and the gradient descent is used to minimize the loss function to fine-tune the model; The specificity binding probability of the generated antibody is scored by using the above classifier, and the higher the probability is, the stronger the specificity binding ability of the antibody and the antigen is, and the reinforcement learning encourages the generation of the antibody sequence with high score, and the model is fine-tuned based on this, and the fine-tuning method is: ； ； wherein, are parameters of the generative model; is a given learning rate; is a predicted probability of the classifier; generates a probability of a given sequence for the generative model.

2. The method of claim 1, wherein the method is characterized by, The step S2 specifically comprises: the amino acid sequence information is composed of types of a plurality of amino acids connected head to tail, and the structure information is composed of three-dimensional coordinates of a plurality of atoms connected head to tail. atoms connected head to tail.

Images