Construction method of anti-cancer peptide generation model and training method thereof

Abstract

The application discloses a kind of anti-cancer peptide generation model training methods, comprising: obtaining general polypeptide sequence data, anti-cancer peptide sequence data and non-anti-cancer peptide sequence data, and according to amino acid glossary general polypeptide sequence data, anti-cancer peptide sequence data and non-anti-cancer peptide sequence data are digitally encoded, to obtain general polypeptide sequence data set, anti-cancer peptide sequence data set and non-anti-cancer peptide sequence data set, general polypeptide sequence data set, anti-cancer peptide sequence data set and non-anti-cancer peptide sequence data set are respectively divided according to training set and test set 8:2 ratio, to obtain general polypeptide sequence training set and test set, anti-cancer peptide sequence training set and test set and non-anti-cancer peptide sequence training set and test set, general polypeptide sequence training set is input into the generation module of anti-cancer peptide generation model, and the weight parameter and bias parameter of each layer in generation module are updated and optimized using back propagation algorithm, to obtain pre-trained generation module.

Description

Technical Field

[0001] The present invention belongs to the interdisciplinary technical field of computer artificial intelligence and anticancer peptide design, and more specifically, relates to a method for constructing an anticancer peptide generation model and a training method thereof. Background Art

[0002] Peptides offer high specificity, selectivity, safety, accessibility, and low immunogenicity. They are superior to the small molecules or proteins that have been the primary focus of previous drug design research, making them promising drug candidates for therapeutic and diagnostic applications in biotechnology. The design of peptides, particularly anticancer peptides, has evolved from wet labs to machine learning and now to deep generative models.

[0003] Existing methods for generating anticancer peptides are all based on learning from anticancer peptide sequence data, and are mainly implemented through the following three methods. The first is to use traditional machine learning models. First, a feature extractor is designed to process the data, and then anticancer peptides are generated through machine learning algorithms. Then, a predictor is used to predict the activity of the generated anticancer peptides. The second is to simply use a deep peptide generation model to directly fit the anticancer peptide data or add the distribution of peptide data with the properties required by the drug, and then use a predictor to predict the activity of the generated anticancer peptides. The third is to train an anticancer peptide generation model with controllable properties through some constraints. The constraints can be reinforcement learning or rejection sampling methods.

[0004] However, existing methods for generating anticancer peptides have some significant flaws: First, traditional machine learning techniques require a large amount of domain expertise to design feature extractors, which is limited in data processing, making the peptide generation process cumbersome; Second, existing deep peptide generation models generally only consider a specific activity or a small number of properties. After generation, they also need to use prediction models as auxiliary screening, and finally use experiments to screen peptide sequences that better meet the required drug properties. Such a peptide design framework does not achieve the ideal speed; Third, existing anticancer peptide generation models only consider sequence information, but not the structural information that has a decisive relationship with the physical and chemical properties of the peptide. This will result in the model being unable to learn the structural information that has a decisive relationship with the physical and chemical properties of the peptide. Summary of the Invention

[0005] In view of the above defects or improvement needs of the prior art, the present application provides a construction method of an anticancer peptide generation model and a training method thereof, which aims to solve the technical problems that the traditional machine learning technology needs a large amount of field professional knowledge to design a feature extractor, is limited in data processing, and makes the polypeptide generation process cumbersome, the technical problem that the existing polypeptide deep generation model cannot achieve an ideal speed, and the technical problem that the existing anticancer peptide generation model cannot learn the structural information which has a decisive relationship with the physical and chemical properties of the polypeptide because the structural information is not considered.

[0006] To achieve the above-mentioned purpose, according to one aspect of the present application, a construction method of an anticancer peptide generation model is provided, which comprises a generation module and a prediction module. The generation module is a sequence generation adversarial network (SeqGAN), and comprises two generators with the same structure and one discriminator. The structure of the generator is as follows:

[0007] The first layer is an embedding layer, which inputs a 20-dimensional polypeptide sequence index vector and outputs a 20*32 embedding matrix;

[0008] The second layer is a recurrent neural network (RNN) layer based on a gated recurrent unit (GRU), which inputs a 20*32 matrix at each time step and outputs a 1*32 hidden unit vector. Finally, 20 1*32 hidden unit vectors are outputted.

[0009] The third layer is a Softmax output layer, which inputs the hidden unit vector of the output at each time step and outputs a 1*20 vector through linear mapping and a Softmax activation function. Finally, a 20*20 matrix is outputted.

[0010] The other generator is a generator required by a push strategy, which has the same structure as the first generator.

[0011] The structure of the discriminator is as follows:

[0012] The first layer is an embedding layer, which inputs a 20*20 polypeptide sequence one-hot encoding and outputs a 20*64 embedding matrix using a 20*64 matrix.

[0013] The second layer is a convolution layer, which inputs a 20*64 embedding matrix. There are two kinds of convolution kernels with sizes of 2 and 3 and step lengths of 1. The number of the convolution kernels is 100 and 200, respectively. The output sizes of the feature matrices are 19*100 and 18*200, respectively.

[0014] The third layer is a rectified linear unit (RELU) layer, which inputs and outputs 19*100 and 18*200 feature matrices using a RELU nonlinear activation function.

[0015] The fourth layer is the pooling layer, which inputs 19*100 and 18*200 feature matrices, uses max pooling to select the maximum value in each feature vector, then concatenates the obtained 1*100 and 1*200 matrices, and outputs a 1*300 peptide sequence feature vector;

[0016] The fifth layer is the Highway layer, which inputs and outputs a 1*300 feature vector;

[0017] The sixth layer is a dropout layer that inputs and outputs a 1*300 feature vector;

[0018] The 7th layer is the Softmax output layer. It inputs a 1*300 feature vector, uses a 300*2 matrix, and obtains a 1*2 vector through linear mapping and Softmax activation function. Finally, it outputs a value through the Argmax function.

[0019] The prediction module is an anticancer peptide predictor based on graph neural network GNN, and its specific structure is as follows:

[0020] The first layer is the embedding layer, which inputs a molecular fingerprint vector of size N and outputs an embedding matrix of size N*50, where N is the length of the molecular fingerprint, i.e. the number of atoms;

[0021] The second layer is the GNN layer, which inputs an N*50 embedding matrix and an N*N adjacency matrix, uses linear mapping and ReLU activation function, feature matrix update function and solves the mean of each feature dimension, and outputs a 1*50 vector;

[0022] Layers 3, 4, and 5 are all GNN layers identical to layer 2, with the same inputs and outputs as layer 2.

[0023] The sixth layer is the ReLU activation layer, which concatenates the vectors output by layers 2-5 to obtain a 1*200 vector. After linear mapping and ReLU activation function, it outputs a 1*800 vector.

[0024] The 7th layer is a fully connected layer, which inputs a 1*800 vector and outputs a 1*2 vector.

[0025] According to another aspect of the present invention, a method for training an anticancer peptide generation model is provided, comprising the following steps:

[0026] (1) Obtaining general polypeptide sequence data, anticancer peptide sequence data, and non-anticancer peptide sequence data, and digitally encoding the general polypeptide sequence data, anticancer peptide sequence data, and non-anticancer peptide sequence data according to an amino acid vocabulary to obtain a general polypeptide sequence dataset, an anticancer peptide sequence dataset, and a non-anticancer peptide sequence dataset.

[0027] (2) The general polypeptide sequence dataset, anticancer peptide sequence dataset, and non-anticancer peptide sequence dataset obtained in step (1) are divided into a training set and a test set ratio of 8:2 to obtain a general polypeptide sequence training set and a test set, an anticancer peptide sequence training set and a test set, and a non-anticancer peptide sequence training set and a test set.

[0028] (3) inputting the general peptide sequence training set obtained in step (2) into the generation module of the anticancer peptide generation model, and using the back propagation algorithm to update and optimize the weight parameters and bias parameters of each layer in the generation module to obtain a pre-trained generation module;

[0029] (4) using the RDKit toolkit to convert the amino acid sequence formats of the anticancer peptide sequence data and the non-anticancer peptide sequence data obtained in step (1) into anticancer peptide molecular graph data and non-anticancer peptide molecular graph data respectively;

[0030] (5) using a one-dimensional Weissfeld-Lehmann 1-WL algorithm to encode the anticancer peptide molecular graph data and the non-anticancer peptide molecular graph data obtained in step (4) and set labels to obtain anticancer peptide molecular fingerprint data and non-anticancer peptide molecular fingerprint data, both of which are represented by numerical vectors and contain molecular fingerprints and adjacency matrices, respectively;

[0031] (6) The anticancer peptide molecular fingerprint data and the non-anticancer peptide molecular fingerprint data obtained in step (5) are divided into a training set and a test set at a ratio of 8:2 to obtain an anticancer peptide molecular fingerprint training set and a test set, and a non-anticancer peptide molecular fingerprint training set and a test set.

[0032] (7) inputting the anticancer peptide molecular fingerprint training set and the non-anticancer peptide molecular fingerprint training set obtained in step (6) into the prediction module of the anticancer peptide generation model, and using the back propagation algorithm to update and optimize the weight parameters and bias parameters of each layer in the prediction module to obtain an updated prediction module of the anticancer peptide generation model;

[0033] (8) The anticancer peptide sequence training set obtained in step (2) is input into the generation module of the anticancer peptide generation model obtained in step (3), and the weight parameters and bias parameters of each layer in the generation module of the anticancer peptide generation model are updated and optimized using the back propagation algorithm to obtain a fine-tuned generation module of the anticancer peptide generation model, thereby obtaining a trained anticancer peptide generation model.

[0034] Preferably, step (3) includes the following sub-steps:

[0035] (3-1) Input the numerical encoding vector of the general peptide sequence training set obtained in step (2) into the embedding layer of the generator of the generation module in the anticancer peptide generation model to obtain the embedding matrix x = {x1,…,x n}, where n represents the length of the peptide sequence, x1 represents the embedding vector of the first amino acid in a peptide sequence, and x n Represents the embedding vector of the nth amino acid in a polypeptide sequence;

[0036] (3-2) The embedding matrix x obtained in step (3-1) is converted into {x1,…,x n} is input to the GRU-based RNN layer of the generator in the generation module to obtain n hidden vectors {h1,…,h n};

[0037] where n hidden vectors {h1,…,h n} is obtained by sequentially embedding n vectors {x1,…,x n}Using the update function g, in the process of generating n hidden vectors, the hidden vector h at the tth time step is obtained t By embedding vector x t and the hidden vector h at the t-1th time step t-1 The calculation formula is:

[0038] h t =g(x t ,h t-1 )

[0039] where t∈[1,n], x t is the embedding vector at the tth time step, h t is the hidden vector at the tth time step, and g is the update function;

[0040] (3-3) The n hidden vectors {h1,…,h n} is input to the Softmax layer of the generator in the generation module to obtain a generated polypeptide sequence, which includes the generator output tokens of n time steps;

[0041] where the generator output token at each time step is calculated as:

[0042] y t =softmax(Wh t +b)

[0043] Where t∈[1,n], W is the weight matrix, b is the bias vector, h t is the hidden vector at the tth time step, y t represents the generator output token at the t-th time step, and softmax() represents the softmax activation function;

[0044] (3-4) Use the cross entropy loss of the generated polypeptide sequence obtained in step (3-3) and the general polypeptide sequence training set obtained in step (2) to perform gradient update on the generator of the generation module, and iteratively train the updated generator until the loss function is minimized, thereby obtaining the pre-trained generator G of the generation module θ ;

[0045] The cross entropy loss at the t-th time step is calculated as:

[0046]

[0047] Among them, y t is the generator output token at the t-th time step, is the token at position t of the peptide sequence in the training set, L t is the cross entropy loss at the t-th time step;

[0048] (3-5) using the generator of the generation module obtained in step (3-4) to sample a polypeptide sequence of fixed length;

[0049] (3-6) The general peptide sequence training set obtained in step (2) and the peptide sequence obtained in step (3-5) are sequentially input into the embedding layer, convolution layer, RELU layer, pooling layer, Highway layer and dropout layer of the discriminator in the generation module to obtain a feature matrix;

[0050] (3-7) Input the feature matrix obtained in step (3-6) into the Softmax layer of the discriminator in the generation module to obtain the discriminant score;

[0051] (3-8) Using the cross entropy loss of the discriminant score obtained in step (3-7), the discriminator of the generation module is gradient updated, and the updated discriminator is iteratively trained until the loss function is minimized, thereby obtaining the pre-trained discriminator D of the generation module;

[0052] The loss of the discriminator is calculated as follows:

[0053]

[0054] Among them, y i is the i-th polypeptide sequence in the general polypeptide sequence training set, is a peptide sequence in the generated peptide sequence set, D() is the discriminant score of the discriminator D, BCE() represents the cross entropy between the discriminant score and the true label 0 or 1, N represents the number of peptide sequences in the general peptide sequence training set, L D is the discriminator loss;

[0055] (3-9) The generator G of the generation module obtained in step (3-4) θ As the Roll-out generator G β use;

[0056] (3-10) Generator G of the generation module obtained in step (3-4) θ Generate a polypeptide sequence segment using the Roll-out generator G obtained in step (3-9) β Generate subsequent tokens for the polypeptide sequence segment to obtain a complete polypeptide sequence, and input the generated complete polypeptide sequence into the discriminator D of the generation module obtained in step (3-8) to obtain a sequence reward for the generated complete polypeptide sequence.

[0057] (3-11) Obtain the expected reward based on the sequence reward of the complete peptide sequence generated in step (3-10), and maximize the expected reward for the generator G obtained in step (3-4) θ The parameters of the gradient are updated to obtain the updated generator G of the generation module θ ;

[0058] (3-12) Generator G of the generation module updated using step (3-11) θ performing sampling to obtain a plurality of polypeptide sequence samples;

[0059] (3-13) Inputting the general peptide sequence training set obtained in step (2) and the peptide sequence samples sampled in step (3-12) into the discriminator of the generation module obtained in step (3-8), and repeating steps (3-6) to (3-8) once to obtain an updated generation module discriminator;

[0060] (3-14) The generator G of the generation module obtained in step (3-11) θ As the Roll-out generator G β Use the generator G of the generation module obtained in step (3-11) θ Generate peptide sequence segments using the Roll-out generator G β Generate subsequent tokens for the peptide sequence segment to obtain the complete peptide sequence, and input the generated complete peptide sequence into the discriminator D of the generation module obtained in step (3-13) to obtain the sequence reward of the generated complete peptide sequence;

[0061] (3-15) Obtain the expected reward based on the sequence reward of the complete polypeptide sequence generated in step (3-15), and maximize the expected reward for the generator G obtained in step (3-11) θ The parameters of the gradient are updated to obtain the updated generator G of the generation module θ ;

[0062] (3-16) Repeat steps (3-12) to (3-15) until the generator and discriminator in the generation module converge, thereby obtaining a generation module of a pre-trained anti-cancer peptide generation model.

[0063] Preferably, the calculation method to maximize the expected reward is as follows:

[0064]

[0065] Among them, J() is the objective function, is the expectation, θ is the generator G of the generation module θ Parameters, G θ (s t |S 1:t-1 ) generates a polypeptide sequence segment, is the sequence reward, S is G θ The generated peptide sequence; we hope to obtain the parameters θ of the generative model so that the generative model can be generated in S 1:t-1 Make the best choice to obtain a large reward R when generating the Tth (last) peptide sequence T , how to get the best choice depends on the sequence reward

[0066] Sequence Rewards is calculated as follows:

[0067]

[0068] Among them, s t It's G θ The tth token generated, T represents the length of the complete polypeptide sequence, S 1:t is the generator G θ The tth polypeptide sequence segment generated, Indicates that in G θ Generate peptide sequence segment S 1:t When Monte Carlo sampling is used, Based on the Roll-out generator G β Generate N complete polypeptide sequences of length T; t <T时, Denotes the discriminator D to G β The discrimination score of the generated complete peptide sequence, when t = T, Denotes the discriminator D to G θ The discrimination score of the generated complete peptide sequence;

[0069] Using gradient ascent, the generator G of the generation module θ The gradient of the objective function, i.e. the gradient update, is calculated as follows:

[0070]

[0071] θ←θ+αJ(θ)

[0072] Where α is the learning rate, T is the length of the complete polypeptide sequence, and S 1:t-1 is the generator G that generates the module θ The generated t-1th polypeptide sequence segment, It is a sequential reward. is the expectation, which can be approximated by the Monte Carlo sampling method. Represents the generator G θ Seek the derivative, is the objective function gradient.

[0073] Preferably, the node label and edge labels is calculated as follows:

[0074]

[0075] in, Indicates updating the label of node v at the tth iteration, Indicates updating the label of node u at the t-1th iteration, Indicates that the label of the edge between nodes u and v is updated in the tth iteration, N(v) represents the set of neighbor nodes of node v, and HASH() indicates that the hash function can be used to map all data groups obtained in the formula into unique integers.

[0076] Preferably, step (7) includes the following sub-steps:

[0077] (7-1) Inputting the anticancer peptide molecular fingerprint training set and the non-anticancer peptide molecular fingerprint training set obtained in step (6) into the embedding layer of the prediction module to obtain the embedding matrix H of the molecular fingerprint;

[0078] (7-2) Input the embedding matrix H and the corresponding adjacency matrix A obtained in step (7-1) into the GNN layer of the prediction module to obtain the feature vector;

[0079] (7-3) Concatenate the i feature vectors obtained in step (7-2), input them into the relu linear layer and the fully connected layer, and then pass them through the softmax layer to obtain the prediction vector;

[0080] (7-4) The prediction vector obtained in step (7-3) and the corresponding cross entropy loss of the data label obtained in step (5) are used to perform gradient update, and the updated prediction module is iteratively trained until the loss function is minimized, thereby obtaining the prediction module of the pre-trained anti-cancer peptide generation model.

[0081] Preferably, in step (7-2), the feature vector H passing through the i-th GNN layer i The update formula is as follows:

[0082] H i =H i-1 +A·relu(W GNN H i-1 )

[0083] Among them, i is the i-th GNN layer, relu (Rectified Linear Unit) is a nonlinear activation function; W GNN is the weight matrix of the GNN layer, and A is the neighbor matrix corresponding to the eigenvector;

[0084] In each GNN layer, the output is the sum of all feature vectors in n feature dimensions The mean on each same dimension of is calculated as follows:

[0085]

[0086] in represents the numerical representation of the kth atom of the feature vector in the i-th GNN layer, n is the number of dimensional features, H i Represents the feature vector at the i-th GNN layer.

[0087] Preferably, step (8) includes the following sub-steps:

[0088] (8-1) Input the numerical encoding vector of the anticancer peptide sequence training set obtained in step (2) into the embedding layer of the generator of the generation module in the anticancer peptide generation model to obtain the embedding matrix x = {x1,…,x n}, where n represents the length of the peptide sequence, x1 represents the embedding vector of the first amino acid in a peptide sequence, and x n Represents the embedding vector of the nth amino acid in a polypeptide sequence;

[0089] (8-2) The embedding matrix x obtained in step (8-1) is converted into {x1,…,x n} is input to the GRU-based RNN layer of the generator in the generation module to obtain n hidden vectors {h1,…,h n};

[0090] (8-3) The n hidden vectors {h1,…,h n} is input to the Softmax layer of the generator in the generation module to obtain a generated polypeptide sequence, which includes the generator output tokens of n time steps;

[0091] (8-4) The cross entropy loss of the generated polypeptide sequence obtained in step (8-3) and the anticancer peptide sequence training set obtained in step (2) is used to perform gradient update on the generator of the generation module, and the updated generator is iteratively trained until the loss function is minimized, thereby obtaining the pre-trained generator G of the generation module. θ ;

[0092] (8-5) using the generator of the generation module obtained in step (8-4) to sample a polypeptide sequence of fixed length;

[0093] (8-6) The anticancer peptide sequence training set obtained in step (2) and the peptide sequence obtained in step (8-5) are sequentially input into the embedding layer, convolution layer, RELU layer, pooling layer, Highway layer and discard layer of the discriminator in the generation module to obtain a feature matrix;

[0094] (8-7) Input the feature matrix obtained in step (8-6) into the Softmax layer of the discriminator in the generation module to obtain the discriminant score;

[0095] (8-8) Using the cross entropy loss of the discriminant score obtained in step (8-7), the discriminator of the generation module is gradient updated, and the updated discriminator is iteratively trained until the loss function is minimized, thereby obtaining the pre-trained discriminator D of the generation module;

[0096] (8-9) The generator G of the generation module obtained in step (8-4) θ Also serves as the Roll-out generator G β use;

[0097] (8-10) Using the generator G of the generation module obtained in step (8-4) θ Generate a polypeptide sequence segment using the Roll-out generator G obtained in step (8-9) β Generate subsequent tokens for the generated polypeptide sequence segments to obtain a complete polypeptide sequence, and input the generated complete polypeptide sequence into the discriminator D of the generation module obtained in step (8-8) to obtain a sequence reward for the generated complete polypeptide sequence;

[0098] (8-11) For the complete polypeptide sequence obtained in step (8-10), the molecular fingerprint corresponding to the polypeptide sequence is obtained by the calculation method of the above steps (4) and (5), and then the molecular fingerprint is input into the prediction module obtained in step (7) to obtain the structural reward of the generated complete polypeptide sequence data;

[0099] (8-12) Linearly combine the sequence reward of the complete polypeptide sequence obtained in step (8-10) and the structure reward of the complete polypeptide sequence obtained in step (8-11) to obtain a linear combination reward, and maximize the expected linear combination reward for the generator G obtained in step (8-4) θ The parameters of the gradient are updated to obtain the updated generator G of the generation module θ .

[0100] (8-13) Repeat steps (8-5) to (8-12) using the generator of the generation model updated in step (8-12) and the discriminator of the generation model updated in step (8-8) until the generation module converges, thereby obtaining a trained anti-cancer peptide generation model.

[0101] Preferably, in steps (8-12),

[0102] The linear combination reward is calculated as follows:

[0103]

[0104] Among them, s t It's G θ The tth token generated, S 1:t-1 It's G θ The generated t-1th polypeptide sequence segment, is the sequence reward for the complete peptide sequence, It is the structural reward of the molecular fingerprint of the complete polypeptide sequence, FP(S 1:t-1 ,s t ) represents the molecular fingerprint of the complete polypeptide sequence, and μ is a mixing constant less than 1;

[0105] Using gradient ascent, G θ The gradient of the objective function and the gradient update are calculated as follows:

[0106]

[0107] θ←θ+αJ(θ)

[0108] Where θ is the generator G θ Parameters, T is the length of the complete polypeptide sequence, α is the learning rate, s t It's G θ The tth token generated, S 1:t-1 It's G θ The generated t-1th polypeptide sequence segment, G θ (s t |S 1:t-1 ) generates a polypeptide sequence segment, is a linear combination reward, is the expectation, which can be approximated by the Monte Carlo sampling method. Represents the generator G θ Seek the derivative, is the objective function gradient.

[0109] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0110] 1. Since the present invention adopts step (3), it utilizes the sequence generative adversarial network and graph neural network in deep learning, so there is no need to manually design feature processors, and the model can automatically extract data features and update; in addition, it uses general peptide data to pre-train the generation module of the anti-cancer peptide generation model, so it can learn the sequence properties of general peptide data;

[0111] 2. Since the present invention adopts step (8), it uses the linear combination reward of the peptide generated by the anti-cancer peptide generation model provided by the discriminator in the generation module of the anti-cancer peptide generation model and the prediction module as feedback information, so that the generation module can learn the anti-cancer properties of the peptide and other drug-related properties from more abundant and diverse information.

[0112] 3. Since the present invention adopts steps (4) to (8), it uses the prediction module of the anticancer peptide generation model to learn the structural information of the anticancer peptide and then provides feedback information to the anticancer peptide generation model, so that the anticancer peptide generation model can use reinforcement learning to simultaneously integrate the sequence information and structural information of the anticancer peptide into the training, and learn the generation of anticancer peptide sequences with anticancer properties and polypeptide structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] Figure 1 This is an overall flow chart of the training method for the anticancer peptide generation model provided by the present invention;

[0114] Figure 2 This is the overall framework diagram of the anti-cancer peptide generation model provided by the present invention. DETAILED DESCRIPTION

[0115] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0116] The basic idea of ​​the present invention is to provide an innovative training method for an anti-cancer peptide generation model, which uses the generator of a generative adversarial network as a strategy in reinforcement learning, utilizes the policy gradient in reinforcement learning, and uses the linear combination of the scores of the generated peptides given by the prediction module based on the graph neural network and the discriminator in the generative adversarial network as a reward to guide the update of the generator. On the basis of learning the effective peptide grammar, it further learns the generation rules containing specific properties and structure-property relationships to achieve high-confidence de novo generation of anti-cancer peptides.

[0117] like Figure 2 As shown, the present invention provides a method for constructing an anticancer peptide generation model, which includes two parts: a generation module and a prediction module. The generation module is a sequence generative adversarial network (SeqGAN) comprising two generators with identical structures and one discriminator.

[0118] The generator structure is as follows:

[0119] The first layer is the embedding layer, which inputs a peptide sequence index vector of size 20 and outputs a 20*32 embedding matrix;

[0120] The second layer is a recurrent neural network (RNN) layer based on the gated recurrent unit (GRU). At each time step, it takes a 20*32 dimensional matrix as input and outputs a 1*32 hidden unit vector. Finally, it outputs 20 1*32 hidden unit vectors.

[0121] The third layer is the Softmax output layer, which takes as input the hidden unit vector of the output of each time step of 1*32. Through linear mapping and Softmax activation function, it outputs a 1*20 vector and finally a 20*20 matrix.

[0122] The other generator is the generator required for the roll-out strategy and has the same structure as the first generator.

[0123] In the discriminator, the first layer is the embedding layer, which inputs a 20*20 peptide sequence one-hot encoding, uses a 20*64 matrix, and outputs a 20*64 embedding matrix;

[0124] The second layer is the convolution layer, which inputs a 20*64 embedding matrix. There are two types of convolution kernels, with sizes of 2 and 3, a stride of 1, and the number of kernels is 100 and 200 respectively. The output feature matrices are 19*100 and 18*200 respectively.

[0125] The third layer is the rectified linear unit (RELU) layer, which uses the RELU nonlinear activation function to input and output 19*100 and 18*200 feature matrices;

[0126] The fourth layer is the pooling layer, which inputs 19*100 and 18*200 feature matrices, uses max pooling to select the maximum value in each feature vector, then concatenates the obtained 1*100 and 1*200 matrices, and outputs a 1*300 peptide sequence feature vector;

[0127] The fifth layer is the Highway layer, which inputs and outputs a 1*300 feature vector;

[0128] The sixth layer is a dropout layer that inputs and outputs a 1*300 feature vector;

[0129] The 7th layer is the Softmax output layer. It inputs a 1*300 feature vector, uses a 300*2 matrix, and obtains a 1*2 vector through linear mapping and Softmax activation function. Finally, it outputs a value through the Argmax function.

[0130] The prediction module is an anticancer peptide predictor based on Graph Neural Network (GNN), and its specific structure is as follows:

[0131] The first layer is the embedding layer, which inputs a molecular fingerprint vector of size N and outputs an embedding matrix of size N*50, where N is the length of the molecular fingerprint, i.e. the number of atoms;

[0132] The second layer is the GNN layer, which inputs an N*50 embedding matrix and an N*N adjacency matrix, uses linear mapping and ReLU activation function, feature matrix update function and solves the mean of each feature dimension, and outputs a 1*50 vector;

[0133] Layers 3, 4, and 5 are all GNN layers identical to layer 2, with the same inputs and outputs as layer 2.

[0134] The sixth layer is the ReLU activation layer, which concatenates the vectors output by layers 2-5 to obtain a 1*200 vector. After linear mapping and ReLU activation function, it outputs a 1*800 vector.

[0135] The 7th layer is a fully connected layer, which inputs a 1*800 vector and outputs a 1*2 vector.

[0136] The advantage of the above-mentioned anticancer peptide generation model is that it uses a sequence-generated adversarial network based on reinforcement learning to solve the backpropagation problem of discrete data and enable the evaluation of the intermediate states of generation; at the same time, the graph neural network is used as a structure predictor to provide feedback to the generation model. Since the structure predictor ensures that the peptide structure meets the anti-cancer or other drug properties, the generation model can learn the relevant properties of anticancer peptides from richer and more diverse information, thereby improving the efficiency of anticancer peptide design and development.

[0137] The generator is used as a strategy in reinforcement learning, and the policy gradient in reinforcement learning is utilized. The linear combination of the scores of the generated peptides by the prediction module based on the graph neural network and the discriminator in the generative adversarial network is used as a reward to guide the update of the generator. On the basis of learning the effective peptide grammar, the generation rules containing specific properties and structure-property relationships are further learned to achieve the de novo generation of anti-cancer peptides with high confidence.

[0138] like Figure 1 As shown, it is an overall flow chart of the training method of the anti-cancer peptide generation model of the present invention, and the training method includes the following steps:

[0139] (1) Obtaining general polypeptide sequence data, anticancer peptide sequence data, and non-anticancer peptide sequence data, and digitally encoding the general polypeptide sequence data, anticancer peptide sequence data, and non-anticancer peptide sequence data according to an amino acid vocabulary to obtain a general polypeptide sequence dataset, an anticancer peptide sequence dataset, and a non-anticancer peptide sequence dataset.

[0140] (2) The general polypeptide sequence dataset, anticancer peptide sequence dataset, and non-anticancer peptide sequence dataset obtained in step (1) are divided into a training set and a test set ratio of 8:2 to obtain a general polypeptide sequence training set and a test set, an anticancer peptide sequence training set and a test set, and a non-anticancer peptide sequence training set and a test set.

[0141] (3) inputting the general peptide sequence training set obtained in step (2) into the generation module of the anticancer peptide generation model, and using the back propagation algorithm to update and optimize the weight parameters and bias parameters of each layer in the generation module to obtain a pre-trained generation module;

[0142] This step is specifically as follows:

[0143] (3-1) Input the numerical encoding vector of the general peptide sequence training set obtained in step (2) into the embedding layer of the generator of the generation module in the anticancer peptide generation model to obtain the embedding matrix x = {x1,…,x n}, where n represents the length of the peptide sequence, x1 represents the embedding vector of the first amino acid in a peptide sequence, and x n Represents the embedding vector of the nth amino acid in a polypeptide sequence;

[0144] (3-2) The embedding matrix x obtained in step (3-1) is converted into {x1,…,x n} is input to the GRU-based RNN layer of the generator in the generation module to obtain n hidden vectors {h1,…,h n};

[0145] Specifically, n hidden vectors {h1,…,h n} is obtained by sequentially embedding n vectors {x1,…,x n}Using the update function g, in the process of generating n hidden vectors, the hidden vector h at the tth time step is obtained t By embedding vector x t and the hidden vector h at the t-1th time step t-1 The calculation formula is:

[0146] h t =g(x t ,h t-1 )

[0147] where t∈[1,n], x t is the embedding vector at the tth time step, h t is the hidden vector at the tth time step, and g is the update function;

[0148] (3-3) The n hidden vectors {h1,…,h n} is input to the Softmax layer of the generator in the generation module to obtain a generated polypeptide sequence, which includes the generator output tokens of n time steps;

[0149] Specifically, the generator output token at each time step is calculated as:

[0150] y t =softmax(Wh t +b)

[0151] Where t∈[1,n], W is the weight matrix, b is the bias vector, h t is the hidden vector at the tth time step, y t represents the generator output token at the t-th time step, and softmax() represents the softmax activation function;

[0152] (3-4) Use the cross entropy loss of the generated polypeptide sequence obtained in step (3-3) and the general polypeptide sequence training set obtained in step (2) to perform gradient update on the generator of the generation module, and iteratively train the updated generator until the loss function is minimized, thereby obtaining the pre-trained generator G of the generation module θ;

[0153] Specifically, the cross entropy loss at the t-th time step is calculated as:

[0154]

[0155] Among them, y t is the generator output token at the t-th time step, is the token at position t of the peptide sequence in the training set, L t is the cross entropy loss at the t-th time step;

[0156] The advantage of the above sub-steps (3-2) to (3-4) is that the generation module of the anti-cancer peptide generation model converges faster through pre-training;

[0157] (3-5) using the generator of the generation module obtained in step (3-4) to sample a polypeptide sequence of fixed length;

[0158] (3-6) The general peptide sequence training set obtained in step (2) and the peptide sequence obtained in step (3-5) are sequentially input into the embedding layer, convolution layer, RELU layer, pooling layer, Highway layer and dropout layer of the discriminator in the generation module to obtain a feature matrix;

[0159] (3-7) Input the feature matrix obtained in step (3-6) into the Softmax layer of the discriminator in the generation module to obtain the discriminant score;

[0160] (3-8) Using the cross entropy loss of the discriminant score obtained in step (3-7), the discriminator of the generation module is gradient updated, and the updated discriminator is iteratively trained until the loss function is minimized, thereby obtaining the pre-trained discriminator D of the generation module;

[0161] Specifically, the loss of the discriminator is calculated as follows:

[0162]

[0163] Among them, y i is the i-th polypeptide sequence in the general polypeptide sequence training set, is a peptide sequence in the generated peptide sequence set, D() is the discriminant score of the discriminator D, BCE() represents the cross entropy between the discriminant score and the true label 0 or 1, N represents the number of peptide sequences in the general peptide sequence training set, L D is the discriminator loss;

[0164] (3-9) The generator G of the generation module obtained in step (3-4) θ As the Roll-out generator G βuse;

[0165] (3-10) Generator G of the generation module obtained in step (3-4) θ Generate a polypeptide sequence segment using the Roll-out generator G obtained in step (3-9) β Generate subsequent tokens for the peptide sequence segment to obtain the complete peptide sequence, and input the generated complete peptide sequence into the discriminator D of the generation module obtained in step (3-8) to obtain the sequence reward of the generated complete peptide sequence;

[0166] Specifically, the core idea is to use the policy gradient method in reinforcement learning to select the best strategy and adjust the strategy; this method generates the generator G of the module θ As a strategy, the generator G of the generation module θ The goal is to maximize the expected reward, which is calculated as follows:

[0167]

[0168] Among them, J() is the objective function, is the expectation, θ is the generator G of the generation module θ Parameters, G θ (s t |S 1:t-1 ) generates a polypeptide sequence segment, is the sequence reward, S is G θ The generated peptide sequence; we hope to obtain the parameters θ of the generative model so that the generative model can be generated in S 1:t-1 Make the best choice to obtain a large reward R when generating the Tth (last) peptide sequence T , how to get the best choice depends on the sequence reward

[0169] The discriminant score of the discriminator for the complete peptide sequence As the generator G θ Generate sequence rewards for peptide sequence segments In form, it is an action-value function, that is, the value of the state obtained after taking action a starting from state s;

[0170] In the generator G θ In the process of generating peptides, in order to evaluate the action-value of the intermediate state, the Monte Carlo search algorithm is used to explore the intermediate token s t and the subsequent polypeptide sequence S of the intermediate token t:T The possibility of the generator G θ The generated peptide sequence segment S 1:t Based on the Roll-out generator G βGenerate a complete peptide sequence; in order to increase the accuracy of peptide sequence evaluation, the sequence reward adopts the average value of the discrimination scores of the generated N complete peptide sequences. is calculated as follows:

[0171]

[0172] Among them, s t It's G θ The tth token generated, T represents the length of the complete polypeptide sequence, S 1:t is the generator G θ The tth polypeptide sequence segment generated, Indicates that in G θ Generate peptide sequence segment S 1:t When Monte Carlo sampling is used, Based on the Roll-out generator G β Generate N complete polypeptide sequences of length T; t <T时, Denotes the discriminator D to G β The discrimination score of the generated complete peptide sequence, when t = T, Denotes the discriminator D to G θ The discrimination score of the generated complete peptide sequence;

[0173] (3-11) Obtain the expected reward based on the sequence reward of the complete peptide sequence generated in step (3-10), and maximize the expected reward for the generator G obtained in step (3-4) θ The parameters of the gradient are updated to obtain the updated generator G of the generation module θ ;

[0174] Specifically, using gradient ascent, the generator G of the generation module θ The gradient of the objective function, i.e. the gradient update, is calculated as follows:

[0175]

[0176] θ←θ+αJ(θ)

[0177] Where α is the learning rate, T is the length of the complete polypeptide sequence, and S 1:t-1 is the generator G that generates the module θ The generated t-1th polypeptide sequence segment, It is a sequential reward. is the expectation, which can be approximated by the Monte Carlo sampling method. Represents the generator G θ Seek the derivative, is the objective function gradient;

[0178] The sub-steps (3-9) to (3-11) have the advantage that the back propagation problem of discrete data is solved by reinforcement learning, and the intermediate sequence of the polypeptide sequence generation process can be evaluated.

[0179] (3-12) The generator G of the generation module updated by step (3-11) θ Sampling to obtain a plurality of polypeptide sequence samples;

[0180] (3-13) The general polypeptide sequence training set obtained in step (2) and the polypeptide sequence samples sampled in step (3-12) are input into the discriminator of the generation module obtained in step (3-8), and steps (3-6)-(3-8) are repeated once to obtain an updated generation module discriminator;

[0181] (3-14) The generator G of the generation module obtained in step (3-11) θ is used as a Roll-out generator G β Using the generator G of the generation module obtained in step (3-11) θ to generate a polypeptide sequence segment, using the Roll-out generator G β to generate a subsequent token for the polypeptide sequence segment to obtain a complete polypeptide sequence, and inputting the generated complete polypeptide sequence into the discriminator D of the generation module obtained in step (3-13) to obtain a sequence reward of the generated complete polypeptide sequence;

[0182] Specifically, the calculation process in this step is the same as that in step (3-10), which will not be repeated here;

[0183] (3-15) According to the sequence reward of the generated complete polypeptide sequence obtained in step (3-15), an expected reward is obtained, and the parameters of the generator G θ obtained in step (3-11) are updated by gradient to obtain an updated generator G θ of the generation module.

[0184] Specifically, the calculation in this step is the same as that in step (3-11), which will not be repeated here;

[0185] (3-16) Repeat steps (3-12) to (3-15) until the generator and discriminator in the generation module converge, thereby obtaining a pre-trained anticancer peptide generation model generation module.

[0186] Specifically, the generator and discriminator used in the repeated steps are updated in the latest step, and the purpose is to let the generator and discriminator be adversarially trained;

[0187] The advantage of this step is that the generation module of the anti-cancer peptide generation model is pre-trained with the general polypeptide training set, so that the model learns the grammar rules of general polypeptides and generates more reasonable polypeptide sequences.

[0188] (4) The amino acid sequence format in the anti-cancer peptide sequence data and the non-anti-cancer peptide sequence data obtained in step (1) is respectively converted into anti-cancer peptide molecular graph data and non-anti-cancer peptide molecular graph data using the RDKit toolkit.

[0189] Specifically, the molecular graph representation includes different molecular data information, such as atoms, chemical bonds and coordinates. Denoted as G = (V, E), where V is the set of nodes, E is the set of edges, and G(v) represents the set of all adjacent nodes of node v ∈ V. In the molecular graph, u ∈ V represents an atom, and (u, v) ∈ E represents a chemical bond between atom u and atom v. The edge (u, v) and the edge (v, u) are the same.

[0190] (5) The anti-cancer peptide molecular graph data and the non-anti-cancer peptide molecular graph data obtained in step (4) are encoded using a one-dimensional Weisfeiler-Lehman (1-Weisfeiler-Lehman, 1-WL) algorithm and labeled to obtain anti-cancer peptide molecular fingerprint data and non-anti-cancer peptide molecular fingerprint data (both of which are represented by numerical vectors and contain molecular fingerprints and adjacency matrices, respectively);

[0191] Specifically, the 1-WL algorithm gives each node in the molecular graph data a label, and constantly aggregates the information of adjacent nodes and edges to update the label of the node until the label no longer changes; wherein in each iteration process, the node label and the edge label are calculated as follows:

[0192]

[0193] wherein, represents the label of node v updated in the tth iteration, represents the label of node u updated in the (t-1)th iteration, represents the label of the edge between nodes u and v updated in the tth iteration, N(v) represents the neighbor node set of node v, and HASH() represents that all data groups obtained in the formula can be mapped into unique integers by using a hash monomorphism function;

[0194] The anti-cancer peptide molecular fingerprint data is marked as 1 and set as a positive sample, and the non-anti-cancer peptide molecular fingerprint data is marked as 0 and set as a negative sample;

[0195] The advantages of the above steps (4) to (5) are that the molecular structure has a decisive relationship with the physical and chemical properties of the polypeptide, and the molecular fingerprint can effectively represent the structure of the polypeptide.

[0196] (6) The anticancer peptide molecular fingerprint data and the non-anticancer peptide molecular fingerprint data obtained in step (5) are divided into a training set and a test set at a ratio of 8:2 to obtain an anticancer peptide molecular fingerprint training set and a test set, and a non-anticancer peptide molecular fingerprint training set and a test set.

[0197] (7) inputting the anticancer peptide molecular fingerprint training set and the non-anticancer peptide molecular fingerprint training set obtained in step (6) into the prediction module of the anticancer peptide generation model, and using the back propagation algorithm to update and optimize the weight parameters and bias parameters of each layer in the prediction module to obtain an updated prediction module of the anticancer peptide generation model;

[0198] This step is specifically as follows:

[0199] (7-1) Inputting the anticancer peptide molecular fingerprint training set and the non-anticancer peptide molecular fingerprint training set obtained in step (6) into the embedding layer of the prediction module to obtain the embedding matrix H of the molecular fingerprint;

[0200] (7-2) Input the embedding matrix H and the corresponding adjacency matrix A obtained in step (7-1) into the GNN layer of the prediction module to obtain the feature vector;

[0201] Specifically, the feature vector H after the i-th GNN layer i The update formula is as follows:

[0202] H i =H i-1 +A·relu(W GNN H i-1 )

[0203] Among them, i is the i-th GNN layer, relu (Rectified Linear Unit) is a nonlinear activation function; W GNN is the weight matrix of the GNN layer, and A is the neighbor matrix corresponding to the eigenvector;

[0204] In each GNN layer, the output is the sum of all feature vectors in n feature dimensions The mean on each same dimension of is calculated as follows:

[0205]

[0206] in represents the numerical representation of the kth atom of the feature vector in the i-th GNN layer, n is the number of dimensional features, H i Represents the feature vector at the i-th GNN layer;

[0207] (7-3) Concatenate the i feature vectors obtained in step (7-2), input them into the relu linear layer and the fully connected layer, and then pass them through the softmax layer to obtain the prediction vector;

[0208] (7-4) The prediction vector obtained in step (7-3) and the corresponding cross entropy loss of the data label obtained in step (5) are used to perform gradient update, and the updated prediction module is iteratively trained until the loss function is minimized, thereby obtaining the prediction module of the pre-trained anti-cancer peptide generation model.

[0209] The advantage of this step is that the prediction module is used to learn the structural information of the anticancer peptide and then provide feedback information to the anticancer peptide generation model, so that the generation module can learn the anticancer properties of the peptide and other drug-related properties from richer and more diverse information.

[0210] (8) The anticancer peptide sequence training set obtained in step (2) is input into the generation module of the anticancer peptide generation model obtained in step (3), and the weight parameters and bias parameters of each layer in the generation module of the anticancer peptide generation model are updated and optimized using the back propagation algorithm to obtain a fine-tuned generation module of the anticancer peptide generation model, thereby obtaining a trained anticancer peptide generation model.

[0211] This step is specifically as follows:

[0212] (8-1) Input the numerical encoding vector of the anticancer peptide sequence training set obtained in step (2) into the embedding layer of the generator of the generation module in the anticancer peptide generation model to obtain the embedding matrix x = {x1,…,x n}, where n represents the length of the peptide sequence, x1 represents the embedding vector of the first amino acid in a peptide sequence, and x n Represents the embedding vector of the nth amino acid in a polypeptide sequence;

[0213] (8-2) The embedding matrix x obtained in step (8-1) is converted into {x1,…,x n} is input to the GRU-based RNN layer of the generator in the generation module to obtain n hidden vectors {h1,…,h n};

[0214] The specific calculation has been explained in step (3-2) and will not be repeated here;

[0215] (8-3) The n hidden vectors {h1,…,h n} is input to the Softmax layer of the generator in the generation module to obtain a generated polypeptide sequence, which includes the generator output tokens of n time steps;

[0216] The specific calculation has been explained in step (8-2) and will not be repeated here;

[0217] (8-4) The cross entropy loss of the generated polypeptide sequence obtained in step (8-3) and the anticancer peptide sequence training set obtained in step (2) is used to perform gradient update on the generator of the generation module, and the updated generator is iteratively trained until the loss function is minimized, thereby obtaining the pre-trained generator G of the generation module. θ ;

[0218] The specific calculation has been explained in step (8-2) and will not be repeated here;

[0219] (8-5) using the generator of the generation module obtained in step (8-4) to sample a polypeptide sequence of fixed length;

[0220] (8-6) The anticancer peptide sequence training set obtained in step (2) and the peptide sequence obtained in step (8-5) are sequentially input into the embedding layer, convolution layer, RELU layer, pooling layer, Highway layer and discard layer of the discriminator in the generation module to obtain a feature matrix;

[0221] (8-7) Input the feature matrix obtained in step (8-6) into the Softmax layer of the discriminator in the generation module to obtain the discriminant score;

[0222] (8-8) Using the cross entropy loss of the discriminant score obtained in step (8-7), the discriminator of the generation module is gradient updated, and the updated discriminator is iteratively trained until the loss function is minimized, thereby obtaining the pre-trained discriminator D of the generation module;

[0223] (8-9) The generator G of the generation module obtained in step (8-4) θ Also serves as the Roll-out generator G β use;

[0224] (8-10) Using the generator G of the generation module obtained in step (8-4) θ Generate a polypeptide sequence segment using the Roll-out generator G obtained in step (8-9) β Generate subsequent tokens for the generated polypeptide sequence segments to obtain a complete polypeptide sequence, and input the generated complete polypeptide sequence into the discriminator D of the generation module obtained in step (8-8) to obtain a sequence reward for the generated complete polypeptide sequence;

[0225] Specifically, the Monte Carlo search algorithm and the calculation of the sequence reward used in this step have been explained in (3-10) and will not be repeated here;

[0226] (8-11) For the complete polypeptide sequence obtained in step (8-10), the molecular fingerprint corresponding to the polypeptide sequence is obtained by the calculation method of the above steps (4) and (5), and then the molecular fingerprint is input into the prediction module obtained in step (7) to obtain the structural reward of the generated complete polypeptide sequence data;

[0227] Specifically, to obtain a molecular fingerprint, first the peptide sequence Decode the amino acid sequence into an amino acid sequence through a dictionary, then use the RDKit toolkit to convert the amino acid sequence into a molecular graph in mol format, and then use the 1-WL algorithm to encode the molecular graph data to obtain a molecular fingerprint; the detailed information and calculations have been explained in steps (4) and (5) and will not be repeated here;

[0228] The calculation method of structural rewards is the same as that of sequence rewards;

[0229] (8-12) Linearly combine the sequence reward of the complete polypeptide sequence obtained in step (8-10) and the structure reward of the complete polypeptide sequence obtained in step (8-11) to obtain a linear combination reward, and maximize the expected linear combination reward for the generator G obtained in step (8-4) θ The parameters of the gradient are updated to obtain the updated generator G of the generation module θ ;

[0230] Specifically, the linear combination reward is calculated as follows:

[0231]

[0232] Among them, s t It's G θ The tth token generated, S 1:t-1 It's G θ The generated t-1th polypeptide sequence segment, is the sequence reward for the complete peptide sequence, It is the structural reward of the molecular fingerprint of the complete polypeptide sequence, FP(S 1:t-1 ,s t ) represents the molecular fingerprint of the complete polypeptide sequence, and μ is a mixing constant less than 1;

[0233] Using gradient ascent, G θ The gradient of the objective function and the gradient update are calculated as follows:

[0234]

[0235] θ←θ+αJ(θ)

[0236] Where θ is the generator G θParameters, T is the length of the complete polypeptide sequence, α is the learning rate, s t It's G θ The tth token generated, S 1:t-1 It's G θ The generated t-1th polypeptide sequence segment, G θ (s t |S 1:t-1 ) generates a polypeptide sequence segment, is a linear combination reward, is the expectation, which can be approximated by the Monte Carlo sampling method. Represents the generator G θ Seek the derivative, is the objective function gradient;

[0237] The advantage of this sub-step is that it uses reinforcement learning to simultaneously integrate the sequence information and structural information of the anticancer peptide into the training of the anticancer peptide sequence generation model, so that the anticancer peptide generation model learns the generation of anticancer peptide sequences with anticancer properties and polypeptide structures.

[0238] (8-13) Repeat steps (8-5) to (8-12) using the generator of the generation model updated in step (8-12) and the discriminator of the generation model updated in step (8-8) until the generation module converges, thereby obtaining a trained anti-cancer peptide generation model.

[0239] Test results

[0240] The performance of the anti-cancer peptide generation model was evaluated using the Bilingual Evaluation Understudy (BLUE). Based on the statistical language model-based algorithm N-Gram, the evaluation metrics were categorized into BLUE-1, BLUE-2, BLUE-3, and BLUE-4. The anti-cancer peptide generation model described in the present invention was named strGAN. Table 1 compares strGAN with three baseline models: RNN, SeqGAN, and VAE. It can be seen that strGAN achieved comparable performance to the baselines. Furthermore, when the score of the structure-based property prediction module was added as a reward (μ = 0.5), and each contributed a certain proportion to the discriminator in the GAN, the performance of the anti-cancer peptide generation model was better than when only the discriminator results were used as a reward (μ = 1) or only the property prediction module results were used as a reward (μ = 0).

[0241] Table 1

[0242] BLUE-1 BLUE-2 BLUE-3 BLUE-4 RNN 0.879 0.903 0.736 0.367 SeqGAN 0.853 0.921 0.724 0.358 VAE 0.947 0.957 0.768 0.375 strGAN(μ=0) 0.913 0.902 0.745 0.457 strGAN (μ=0.5) 0.934 0.928 0.779 0.429 strGAN(μ=1) 0.920 0.913 0.730 0.436

[0243] Compared to other models, this method overcomes the shortcomings of current deep learning methods for peptide generation, such as insufficient utilization of structural information, low sampling efficiency, and cumbersome screening steps. By employing policy gradients in reinforcement learning, using the graph neural network predictor and the GAN's own discriminator as reward functions, it trains a generative model that can generate high-confidence peptides, achieving the effect of automatic peptide screening. Because the structure predictor guarantees certain specific chemical properties of the peptide, the generator can produce peptides with more stable properties, easier synthesis, and druggability, thereby improving the efficiency of therapeutic peptide design and development.

[0244] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for constructing an anticancer peptide generation model, comprising a generation module and a prediction module; wherein the generation module is a sequence generative adversarial network (SeqGAN), and comprises two generators with identical structures and a discriminator, characterized in that: The generator structure is as follows: The first layer is the embedding layer, which inputs a peptide sequence index vector of size 20 and outputs a 20*32 embedding matrix; The second layer is a recurrent neural network (RNN) layer based on the gated recurrent unit (GRU). At each time step, the input is a 20*32 dimensional matrix, the output is a 1*32 hidden unit vector, and finally the output is 20 1*32 hidden unit vectors. The third layer is the Softmax output layer, which takes as input the hidden unit vector of the output of each time step of 1*32. Through linear mapping and Softmax activation function, it outputs a 1*20 vector and finally a 20*20 matrix. The other generator is the one needed to launch the strategy, and has the same structure as the first generator; The structure of the discriminator is as follows: The first layer is the embedding layer, which inputs a 20*20 peptide sequence one-hot encoding, uses a 20*64 matrix, and outputs a 20*64 embedding matrix; The second layer is the convolution layer, which inputs a 20*64 embedding matrix. There are two types of convolution kernels, with sizes of 2 and 3, a stride of 1, and the number of kernels is 100 and 200 respectively. The output feature matrices are 19*100 and 18*200 respectively. The third layer is the linear rectification function RELU layer, which uses the RELU nonlinear activation function to input and output 19*100 and 18*200 feature matrices; The fourth layer is the pooling layer, which inputs 19*100 and 18*200 feature matrices, uses max pooling to select the maximum value in each feature vector, then concatenates the obtained 1*100 and 1*200 matrices, and outputs a 1*300 peptide sequence feature vector; The fifth layer is the Highway layer, which inputs and outputs a 1*300 feature vector; The sixth layer is a dropout layer that inputs and outputs a 1*300 feature vector; The seventh layer is the Softmax output layer. It takes a 1*300 feature vector as input, uses a 300*2 matrix, and uses linear mapping and the Softmax activation function to obtain a 1*2 vector. Finally, it outputs a value through the Argmax function. The prediction module is an anticancer peptide predictor based on graph neural network GNN, and its specific structure is as follows: The first layer is the embedding layer, which inputs a molecular fingerprint vector of size N and outputs an embedding matrix of size N*50, where N is the length of the molecular fingerprint, i.e. the number of atoms; The second layer is the GNN layer, which inputs an N*50 embedding matrix and an N*N adjacency matrix, uses linear mapping and ReLU activation function, feature matrix update function and solves the mean of each feature dimension, and outputs a 1*50 vector; Layers 3, 4, and 5 are all GNN layers identical to layer 2, with the same inputs and outputs as layer 2. The sixth layer is the ReLU activation layer, which concatenates the vectors output by layers 2-5 to obtain a 1*200 vector. After linear mapping and ReLU activation function, it outputs a 1*800 vector. The 7th layer is a fully connected layer, which inputs a 1*800 vector and outputs a 1*2 vector.

2. A training method for an anticancer peptide generation model, characterized in that: The following steps are involved: (1) obtaining general polypeptide sequence data, anticancer peptide sequence data, and non-anticancer peptide sequence data, and digitally encoding the general polypeptide sequence data, anticancer peptide sequence data, and non-anticancer peptide sequence data according to an amino acid vocabulary to obtain a general polypeptide sequence dataset, an anticancer peptide sequence dataset, and a non-anticancer peptide sequence dataset; (2) dividing the general peptide sequence dataset, anticancer peptide sequence dataset, and non-anticancer peptide sequence dataset obtained in step (1) into a training set and a test set ratio of 8:2 to obtain a general peptide sequence training set and a test set, an anticancer peptide sequence training set and a test set, and a non-anticancer peptide sequence training set and a test set; (3) inputting the general peptide sequence training set obtained in step (2) into the generation module of the anticancer peptide generation model, and using the back propagation algorithm to update and optimize the weight parameters and bias parameters of each layer in the generation module to obtain a pre-trained generation module; (4) using the RDKit toolkit to convert the amino acid sequence formats of the anticancer peptide sequence data and the non-anticancer peptide sequence data obtained in step (1) into anticancer peptide molecular graph data and non-anticancer peptide molecular graph data respectively; (5) using a one-dimensional Weissfeld-Lehmann 1-WL algorithm to encode the anticancer peptide molecular graph data and the non-anticancer peptide molecular graph data obtained in step (4) and set labels to obtain anticancer peptide molecular fingerprint data and non-anticancer peptide molecular fingerprint data, both of which are represented by numerical vectors and contain molecular fingerprints and adjacency matrices, respectively; (6) Dividing the anticancer peptide molecular fingerprint data and the non-anticancer peptide molecular fingerprint data obtained in step (5) into a training set and a test set at a ratio of 8:2 to obtain an anticancer peptide molecular fingerprint training set and a test set, and a non-anticancer peptide molecular fingerprint training set and a test set; (7) inputting the anticancer peptide molecular fingerprint training set and the non-anticancer peptide molecular fingerprint training set obtained in step (6) into the prediction module of the anticancer peptide generation model, and using the back propagation algorithm to update and optimize the weight parameters and bias parameters of each layer in the prediction module to obtain an updated prediction module of the anticancer peptide generation model; (8) The anticancer peptide sequence training set obtained in step (2) is input into the generation module of the anticancer peptide generation model obtained in step (3), and the weight parameters and bias parameters of each layer in the generation module of the anticancer peptide generation model are updated and optimized using the back propagation algorithm to obtain a fine-tuned generation module of the anticancer peptide generation model, thereby obtaining a trained anticancer peptide generation model.

3. The training method for the anticancer peptide generation model according to claim 2, characterized in that: Step (3) includes the following sub-steps: (3-1) Input the numerical encoding vector of the general peptide sequence training set obtained in step (2) into the embedding layer of the generator of the generation module in the anticancer peptide generation model to obtain the embedding matrix x = {x1,…,x n }, where n represents the length of the peptide sequence, x1 represents the embedding vector of the first amino acid in a peptide sequence, and x n Represents the embedding vector of the nth amino acid in a polypeptide sequence; (3-2) The embedding matrix x obtained in step (3-1) is converted into {x1,…,x n } is input to the GRU-based RNN layer of the generator in the generation module to obtain n hidden vectors {h1,…,h n }; where n hidden vectors {h1,…,h n } is obtained by sequentially embedding n vectors {x1,…,x n }Using the update function g, in the process of generating n hidden vectors, the hidden vector h at the tth time step is obtained t By embedding vector x t and the hidden vector h at the t-1th time step t-1 The calculation formula is: h t =g(x t ,h t-1 ) where t∈[1,n], x t is the embedding vector at the tth time step, h t is the hidden vector at the tth time step, and g is the update function; (3-3) The n hidden vectors {h1,…,h n } is input to the Softmax layer of the generator in the generation module to obtain a generated polypeptide sequence, which includes the generator output tokens of n time steps; where the generator output token at each time step is calculated as: y t =softmax(Wh t +b) Where t∈[1,n], W is the weight matrix, b is the bias vector, h t is the hidden vector at the tth time step, y t represents the generator output token at the t-th time step, and softmax() represents the sofmtax activation function; (3-4) Use the cross entropy loss of the generated polypeptide sequence obtained in step (3-3) and the general polypeptide sequence training set obtained in step (2) to perform gradient update on the generator of the generation module, and iteratively train the updated generator until the loss function is minimized, thereby obtaining the pre-trained generator G of the generation module θ ; The cross entropy loss at the t-th time step is calculated as: Among them, y t is the generator output token at the t-th time step, is the token at position t of the peptide sequence in the training set, L t is the cross entropy loss at the t-th time step; (3-5) using the generator of the generation module obtained in step (3-4) to sample a polypeptide sequence of fixed length; (3-6) The general peptide sequence training set obtained in step (2) and the peptide sequence obtained in step (3-5) are sequentially input into the embedding layer, convolution layer, RELU layer, pooling layer, Highway layer and dropout layer of the discriminator in the generation module to obtain a feature matrix; (3-7) Input the feature matrix obtained in step (3-6) into the Softmax layer of the discriminator in the generation module to obtain the discriminant score; (3-8) Using the cross entropy loss of the discriminant score obtained in step (3-7), the discriminator of the generation module is gradient updated, and the updated discriminator is iteratively trained until the loss function is minimized, thereby obtaining the pre-trained discriminator D of the generation module; The loss of the discriminator is calculated as follows: Among them, y i is the i-th polypeptide sequence in the general polypeptide sequence training set, is a peptide sequence in the generated peptide sequence set, D() is the discriminant score of the discriminator D, BCE() represents the cross entropy between the discriminant score and the true label 0 or 1, N represents the number of peptide sequences in the general peptide sequence training set, L D is the discriminator loss; (3-9) The generator G of the generation module obtained in step (3-4) θ As the Roll-out generator G β use; (3-10) Generator G of the generation module obtained in step (3-4) θ Generate a polypeptide sequence segment using the Roll-out generator G obtained in step (3-9) β Generate subsequent tokens for the peptide sequence segment to obtain the complete peptide sequence, and input the generated complete peptide sequence into the discriminator D of the generation module obtained in step (3-8) to obtain the sequence reward of the generated complete peptide sequence; (3-11) Obtain the expected reward based on the sequence reward of the complete peptide sequence generated in step (3-10), and maximize the expected reward for the generator G obtained in step (3-4) θ The parameters of the gradient are updated to obtain the updated generator G of the generation module θ ; (3-12) Generator G of the generation module updated using step (3-11) θ performing sampling to obtain a plurality of polypeptide sequence samples; (3-13) Inputting the general peptide sequence training set obtained in step (2) and the peptide sequence samples sampled in step (3-12) into the discriminator of the generation module obtained in step (3-8), and repeating steps (3-6) to (3-8) once to obtain an updated generation module discriminator; (3-14) The generator G of the generation module obtained in step (3-11) θ As the Roll-out generator G β Use the generator G of the generation module obtained in step (3-11) θ Generate peptide sequence segments using the Roll-out generator G β Generate subsequent tokens for the peptide sequence segment to obtain the complete peptide sequence, and input the generated complete peptide sequence into the discriminator D of the generation module obtained in step (3-13) to obtain the sequence reward of the generated complete peptide sequence; (3-15) Obtain the expected reward based on the sequence reward of the complete polypeptide sequence generated in step (3-15), and maximize the expected reward for the generator G obtained in step (3-11) θ The parameters of the gradient are updated to obtain the updated generator G of the generation module θ ; (3-16) Repeat steps (3-12) to (3-15) until the generator and discriminator in the generation module converge, thereby obtaining a generation module of a pre-trained anti-cancer peptide generation model.

4. The training method for the anticancer peptide generation model according to claim 3, characterized in that: The calculation for maximizing expected reward is as follows: Among them, J() is the objective function, is the expectation, θ is the generator G of the generation module θ Parameters, G θ (s t |S 1:t-1 ) generates a polypeptide sequence segment, is the sequence reward, S is G θ The generated peptide sequence; we hope to obtain the parameters θ of the generative model so that the generative model can be generated in S 1:t-1 Make the best choice to obtain a large reward R when generating the Tth polypeptide sequence T ; Sequence Rewards is calculated as follows: where s t is the t-th token generated by G θ , T represents the length of the complete polypeptide sequence, and S 1:t is the t-th polypeptide sequence segment generated by the generator G θ . denotes that when generating the polypeptide sequence segment S θ by G 1:t , N complete polypeptide sequences of length T are generated using Monte Carlo sampling and the Roll-out generator G based on the polypeptide sequence segment; when t < T, β denotes the discrimination score of the discriminator D for the complete polypeptide sequence generated by G , and when t = T, β denotes the discrimination score of the discriminator D for the complete polypeptide sequence generated by G . θ ​ Using gradient ascent, the generator G of the generation module θ The gradient of the objective function, i.e. the gradient update, is calculated as follows: Where α is the learning rate, T is the length of the complete polypeptide sequence, and S 1:t-1 is the generator G that generates the module θ The generated t-1th polypeptide sequence segment, It is a sequential reward. is the expectation, which is approximated by the Monte Carlo sampling method. Represents the generator G θ Seek derivation, is the objective function gradient.

5. The method for training an anticancer peptide generation model according to any one of claims 2 to 4, characterized in that: Node Labels and edge labels is calculated as follows: in, Indicates updating the label of node v at the tth iteration, Indicates updating the label of node u at the t-1th iteration, Indicates that the label of the edge between nodes u and v is updated in the tth iteration, N(v) represents the set of neighbor nodes of node v, and HASH() represents the use of a hash function to map all data groups obtained in the formula into unique integers.

6. The training method for the anticancer peptide generation model according to claim 5, characterized in that: Step (7) includes the following sub-steps: (7-1) Inputting the anticancer peptide molecular fingerprint training set and the non-anticancer peptide molecular fingerprint training set obtained in step (6) into the embedding layer of the prediction module to obtain the embedding matrix H of the molecular fingerprint; (7-2) Input the embedding matrix H and the corresponding adjacency matrix A obtained in step (7-1) into the GNN layer of the prediction module to obtain the feature vector; (7-3) Concatenate the i feature vectors obtained in step (7-2), input them into the relu linear layer and the fully connected layer, and then pass them through the softmax layer to obtain the prediction vector; (7-4) The prediction vector obtained in step (7-3) and the corresponding cross entropy loss of the data label obtained in step (5) are used to perform gradient update, and the updated prediction module is iteratively trained until the loss function is minimized, thereby obtaining the prediction module of the pre-trained anti-cancer peptide generation model.

7. The training method for the anticancer peptide generation model according to claim 6, characterized in that: In step (7-2), the feature vector H passing through the i-th GNN layer i The update formula is as follows: A i =H i-1 +A·relu(W GNN A i-1 ) Among them, i is the i-th GNN layer, relu (Rectified Linear Unit) is a nonlinear activation function; W GNN is the weight matrix of the GNN layer, and A is the neighbor matrix corresponding to the eigenvector; In each GNN layer, the output is the sum of all feature vectors in n feature dimensions The mean on each same dimension of is calculated as follows: in represents the numerical representation of the kth atom of the feature vector in the i-th GNN layer, n is the number of dimensional features, H i Represents the feature vector at the i-th GNN layer.

8. The training method for the anticancer peptide generation model according to claim 7, characterized in that: Step (8) includes the following sub-steps: (8-1) Input the numerical encoding vector of the anticancer peptide sequence training set obtained in step (2) into the embedding layer of the generator of the generation module in the anticancer peptide generation model to obtain the embedding matrix x = {x1,…,x n }, where n represents the length of the peptide sequence, x1 represents the embedding vector of the first amino acid in a peptide sequence, and x n Represents the embedding vector of the nth amino acid in a polypeptide sequence; (8-2) The embedding matrix x obtained in step (8-1) is converted into {x1,…,x n } is input to the GRU-based RNN layer of the generator in the generation module to obtain n hidden vectors {h1,…,h n }; (8-3) The n hidden vectors {h1,…,h n } is input to the Softmax layer of the generator in the generation module to obtain a generated polypeptide sequence, which includes the generator output tokens of n time steps; (8-4) The cross entropy loss of the generated polypeptide sequence obtained in step (8-3) and the anticancer peptide sequence training set obtained in step (2) is used to perform gradient update on the generator of the generation module, and the updated generator is iteratively trained until the loss function is minimized, thereby obtaining the pre-trained generator G of the generation module. θ ; (8-5) using the generator of the generation module obtained in step (8-4) to sample a polypeptide sequence of fixed length; (8-6) The anticancer peptide sequence training set obtained in step (2) and the peptide sequence obtained in step (8-5) are sequentially input into the embedding layer, convolution layer, RELU layer, pooling layer, Highway layer and discard layer of the discriminator in the generation module to obtain a feature matrix; (8-7) Input the feature matrix obtained in step (8-6) into the Softmax layer of the discriminator in the generation module to obtain the discriminant score; (8-8) Using the cross entropy loss of the discriminant score obtained in step (8-7), the discriminator of the generation module is gradient updated, and the updated discriminator is iteratively trained until the loss function is minimized, thereby obtaining the pre-trained discriminator D of the generation module; (8-9) The generator G of the generation module obtained in step (8-4) θ Also serves as the Roll-out generator G β use; (8-10) Using the generator G of the generation module obtained in step (8-4) θ Generate a polypeptide sequence segment using the Roll-out generator G obtained in step (8-9) β Generate subsequent tokens for the generated polypeptide sequence segments to obtain a complete polypeptide sequence, and input the generated complete polypeptide sequence into the discriminator D of the generation module obtained in step (8-8) to obtain a sequence reward for the generated complete polypeptide sequence; (8-11) For the complete polypeptide sequence obtained in step (8-10), the molecular fingerprint corresponding to the polypeptide sequence is obtained by the calculation method of the above steps (4) and (5), and then the molecular fingerprint is input into the prediction module obtained in step (7) to obtain the structural reward of the generated complete polypeptide sequence data; (8-12) Linearly combine the sequence reward of the complete polypeptide sequence obtained in step (8-10) and the structure reward of the complete polypeptide sequence obtained in step (8-11) to obtain a linear combination reward, and maximize the expected linear combination reward for the generator G obtained in step (8-4) θ The parameters of the generator G are updated to obtain the updated generator module θ ; (8-13) Repeat steps (8-5) to (8-12) using the generator of the generation model updated in step (8-12) and the discriminator of the generation model updated in step (8-8) until the generation module converges, thereby obtaining a trained anti-cancer peptide generation model.

9. The training method for the anticancer peptide generation model according to claim 8, characterized in that: In steps (8-12), The linear combination reward is calculated as follows: Among them, s t It's G θ The tth token generated, S 1:t-1 It's G θ The generated t-1th polypeptide sequence segment, is the sequence reward for the complete peptide sequence, It is the structural reward of the molecular fingerprint of the complete polypeptide sequence, FP(S 1:t-1 ,s t ) represents the molecular fingerprint of the complete polypeptide sequence, and μ is a mixing constant less than 1; Using gradient ascent, G θ The gradient of the objective function and the gradient update are calculated as follows: Where θ is the generator G θ Parameters, T is the length of the complete polypeptide sequence, α is the learning rate, s t It's G θ The tth token generated, S 1:t-1 It's G θ The generated t-1th polypeptide sequence segment, G θ (s t |S 1:t-1 ) generates a polypeptide sequence segment, is a linear combination reward, is the expectation, which is approximated by the Monte Carlo sampling method. Represents the generator G θ Seek derivation, is the objective function gradient.

Images