A virus and host protein interaction prediction method based on a densely connected convolutional network

Abstract

The application is a virus and host protein interaction prediction method based on a densely connected convolutional network, comprising the following steps: step 1, preprocessing data; step 2, using a Skip-Gram model in Word2Vec and One-Hot encoding to obtain embedding vectors of amino acids; step 3, constructing a densely connected convolutional network for feature extraction according to a word embedding matrix of the protein; step 4, obtaining average RPM-PSSM feature descriptors through a PSSM feature matrix; step 5, constructing a fully connected neural network for feature extraction according to the feature descriptors; step 6, performing feature fusion through a concatenate operation, and finally using the trained model for prediction. The application performs feature extraction on different features, contains protein evolution information and physicochemical properties, local key information and long-range dependence in the sequence, and better improves the prediction accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of bioinformatics, and particularly relates to a virus and host protein-protein interaction prediction method based on a densely connected convolutional network. BACKGROUND

[0002] Different types of viruses can cause viral infection or viral disease. For treatment methods, specific vaccines and drugs are mainly developed. In the process of viral infection of the host, viruses transport their own genetic material into the host cell nucleus through interspecies protein interaction, control host cell metabolism, and destroy host cell function. Therefore, identifying PPI between viruses and hosts helps to understand the mechanism of viral infection and design the development of antiviral drugs and vaccines. SUMMARY

[0003] The purpose of the application is to provide a virus and host protein-protein interaction prediction method based on a densely connected convolutional network, which can improve the accuracy of virus host protein-protein interaction prediction.

[0004] The method adopted by the application is: a virus and host protein-protein interaction prediction method based on a densely connected convolutional network, comprising the following steps:

[0005] Step 1: data preprocessing: for virus host protein-protein interaction data, protein sequence information is preprocessed to remove outliers and missing values.

[0006] Step 2: use the Skip-Gram model in Word2Vec and One-Hot encoding to obtain the embedding vector of amino acids:

[0007] Step 2.1: all protein sequence information is processed into protein sequence simulation documents, and amino acids are simulated as words;

[0008] Step 2.2: use the trained Skip-Gram model in Word2Vec to obtain the embedding vector of 20 amino acids, wherein the Skip-Gram model is to obtain the co-occurrence similarity of amino acids, and the formula is as follows:

[0009]

[0010]

[0011] wherein, Y A is the negative log-likelihood loss function result, T is the given protein sequence size, c is the context size, the subsequence of the given protein sequence T is taken as the center, the length is 2c+1, and a represents the amino acid that constitutes the protein sequence, a t t ​It is the first part of the insertion of the t-th amino acid, a t+j It is the first part of the embedding of adjacent amino acids, p is the probability defined as Softmax, n is the negative sampling size, and a' k It is not related to a in the same context t Co-occurring negative samples;

[0012] Step 2.3: Based on the dipole and volume of the amino acid side chains, the 20 amino acids are grouped into seven groups, namely {AGV}, {C}, {FILP}, {MSTY}, {HNQW}, {DE}, and {KR}. One-Hot encoding is used, and the groups are concatenated with the amino acid vectors from Step 2.2 to form a 20×40 amino acid matrix.

[0013] Step 3: Based on the word embedding matrix of the protein, construct a densely connected convolutional network for feature extraction:

[0014] Step 3.1: Define DCNNBlock1 and DCNNBlock2 blocks to construct a densely connected convolutional network (Dense-net). The kernel variations and convolution formulas within the dense blocks are as follows:

[0015]

[0016]

[0017]

[0018] Where nb_filter is the convolution kernel, M c It is the shared weight of each convolutional layer, b c It is the bias vector, and h() is the activation function. It is the latent vector generated after passing through the convolutional layer, F con It is the result of concatenating the output of the previous convolutional layer with the output of the current layer;

[0019] Step 3.2: After obtaining a 3D tensor through six convolutional layers, design a self-attention module to mine long-range dependencies. The formula is as follows:

[0020] Q i =W Q X i ,K i =W K X i V i =W V X i (6)

[0021] α i =p(z=i|X,Q)=softmax(s(X)i ,Q)) (7)

[0022]

[0023]

[0024]

[0025] Q i For query volume, K i V is the key value. i For the value, W Q W K W V It is of size d∈R 2D The parameter weight matrix, z∈[1,n] represents the index position of the selected information, s(X i Q) is the attention scoring function, and c is the attention score. ij It is a key vector and query volume (Q) j It is calculated based on the correlation in a specific space. This is the scaling factor.

[0026] Step 4: Obtain the average RPM-PSSM feature descriptor using the PSSM feature matrix:

[0027] Step 4.1: Data needs to be processed before using the POSSUM online web server. The sequence length must be between 50 and 5000; otherwise, POSSUM cannot be used.

[0028] Step 4.2: When using the POSSUM online web server, users upload protein sequence files in FASTA format through the client web interface. POSSUM calculates the RPM-PSSM descriptor in the server background and returns the descriptor results to the user via email after the job is completed.

[0029] Step 5: Construct a fully connected neural network based on the feature descriptors for feature extraction:

[0030] Step 5.1: Design a multi-layer DNN with fully connected layers, using the RPM-PSSM feature descriptor as input, with the following formula:

[0031] o1=α(W1E+b1) (11)

[0032] o i =α(W i o i-1 +b i ),2≤i≤M (12)

[0033] where o i represents the output generated by the i-th fully connected layer, W i and b i represent the parameter matrix and bias term respectively, and alpha represents the RELU activation function that can realize nonlinear transformation, and M represents the number of fully connected layers. The output generated by the last fully connected layer is used as the final RPM-PSSM feature.

[0034] Step 6: Feature fusion is performed using the concatenate operation, and finally the trained model is used for prediction:

[0035] Step 6.1: Perform element-wise multiplication between the input tensor list of the extracted viral protein sequence features and the host protein sequence features, and the same processing is performed on the RPM-PSSM feature, and the formula is as follows:

[0036] P seq = P human ⊙ P virus (13)

[0037] P rpm_pssm = P humax_pssm ⊙ P virus_pssm (14)

[0038] P seq_pssm = P seq ⊙ P rpm_pssm (15)

[0039] The last three features P seq_pssm , P seq , P rpm_pssm are fused to obtain a 192-dimensional feature vector P com .

[0040] Step 6.2: Construct a multi-layer perception (MLP) to P com Use softmax to realize prediction, and output the prediction result:

[0041] p(y) = softmax(w*P com +b) (16).

[0042] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0043] Compared with the prior art, the present application has the following beneficial effects:

[0044] 1. The present application considers the amino acid residue embedding information and protein sequence evolution information of the protein sequence, and fuses the sequence features and evolution information features together, thereby increasing the effectiveness of protein feature extraction;

[0045] 2. The protein sequence information is input into a densely connected convolutional network, all layers are connected under the premise of ensuring maximum information transmission between layers in the network, better use of sequence characteristics improves the correctness of interaction prediction;

[0046] 3. The densely connected convolutional network extracts local key information of the sequence, and then uses the self-attention mechanism to extract remote dependency relationship of the sequence and focuses on the information with greater influence, while considering local information, remote information and focus information, which better improves the correctness of interaction prediction; BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 is a model flowchart;

[0048] Figure 2 is the ROC curve of the performance comparison between the embodiment of the application and other methods under 5-fold cross-validation. DETAILED DESCRIPTION

[0049] The application will be further described below in combination with the drawings and examples.

[0050] The method adopted by the application is a virus and host protein interaction prediction method based on a densely connected convolutional network, comprising the following steps:

[0051] Step 1: Preprocessing the data: preprocessing the virus host protein-protein interaction data and protein sequence information, removing outliers and missing values.

[0052] Step 2: Use the Skip-Gram model in Word2Vec and One-Hot encoding to obtain the embedding vector of amino acid:

[0053] Step 2.1: Process all protein sequence information into protein sequence simulation documents, and simulate amino acids as words;

[0054] Step 2.2: Use the trained Skip-Gram model in Word2Vec to obtain the embedding vector of 20 kinds of amino acids, wherein the Skip-Gram model is to obtain the co-occurrence similarity of amino acids, and the formula is as follows:

[0055]

[0056]

[0057] wherein, Y A is the negative log-likelihood loss function result, T is the given protein sequence size, c is the context size, the subsequence of the given protein sequence T is taken as the amino acid a tis the first part embedding of the tthamino acid, a t is the first part embedding of the adjacent amino acid, p is the probability defined as Softmax, n is the negative sampling size, a' t+j is the first part embedding of the adjacent amino acid, p is the probability defined as Softmax, n is the negative sampling size, a' k is the negative sample that does not co-occur with a t in the same context;

[0058] Step 2.3: According to the dipole and volume of the side chain of the amino acid, 20 kinds of amino acids are grouped into seven groups, namely {AGV}, {C}, {FILP}, {MSTY}, {HNQW}, {DE} and {KR}, encoded using One-Hot, and spliced with the vector of amino acid in step 2.2 to form a 20x40 amino acid matrix.

[0059] Step 3: According to the word embedding matrix of the protein, a densely connected convolutional network is constructed for feature extraction:

[0060] Step 3.1: Define DCNNBlock1 block and DCNNBlock2 block to construct a densely connected convolutional network (Dense-net), and the convolution kernel changes and convolution formula in the dense block are as follows:

[0061]

[0062]

[0063]

[0064] where nb_filter is the convolution kernel, M c is the shared weight of each layer of convolution, b c is the bias vector, h() is the activation function, is the latent vector generated after the convolution layer, F con is the splicing result of the output of the previous layer of convolution and the output of the current layer;

[0065] Step 3.2: After obtaining a 3D tensor through six layers of convolution, a self-attention module is designed to mine long-range dependencies, and the formula is as follows:

[0066] Q i = W Q X i , K i = W K X i , V i = W V X i (6)

[0067] α i = p(z = i | X, Q) = softmax(s(X i , Q)) (7)

[0068]

[0069]

[0070]

[0071] where Q i is the query, K i is the key, V i is the value, W Q , W K , W V are parameter weight matrices of size d e R 2D , z e [1, n] represents the index position of the selected information, s(X i , Q) is the attention scoring function, and the attention score c ij is the key vector and the query (Q j ) are calculated in a certain spatial correlation, is a proportionality factor.

[0072] Step 4: Obtain the average RPM-PSSM feature descriptor through the PSSM feature matrix:

[0073] Step 4.1: Before using the POSSUM online Web server, the data needs to be processed, and the sequence length must be between 50 and 5000, otherwise POSSUM cannot be used;

[0074] Step 4.2: There are two methods for using the POSSUM online Web server. Method 1: Through the client Web interface, users can upload protein sequence files in FASTA format, and the number of protein sequences in the file is not limited in size. POSSUM can calculate the RPM-PSSM descriptor in the server background, and the descriptor result is returned to the user by email after the job is completed. Method 2: Directly upload the original protein sequence on the client side, but the number of sequence cannot be more than 500, and the result is also obtained by email.

[0075] Step 5: Construct a fully connected neural network according to the feature descriptor to extract features:

[0076] Step 5.1: Design a multi-layer DNN with a fully connected layer, using the RPM-PSSM feature descriptor as input, whose formula is as follows:

[0077] o1=α(W1E+b1) (11)

[0078] o i =α(W i o i-1 +b i ),2≤i≤M (12)

[0079] Among them o i W represents the output of the i-th fully connected layer. i and b i Let represent the parameter matrix and bias term, respectively; α represent the ReLU activation function that enables nonlinear transformation; and M represent the number of fully connected layers. The output of the last fully connected layer is used as the final RPM-PSSM feature.

[0080] Step 6: Perform feature fusion using the concatenate operation, and finally use the trained model for prediction:

[0081] Step 6.1: Multiply the extracted viral protein sequence features and the host protein sequence features element-wise in the input tensor list. Perform the same processing on the RPM-PSSM features. The formula is as follows:

[0082] P seq =P human ⊙P virus (13)

[0083] P rpm_pssm =P human_pssm ⊙P virus_pssm (14)

[0084] P seq_pssm =P seq ⊙P rpm_pssm (15)

[0085] The last three features P seq_pssm P seq P rpm_pssm Feature fusion yields a 192-dimensional feature vector P. com ;

[0086] Step 6.2: Construct a Multilayer Perceptron (MLP) to analyze P com Use softmax to perform prediction and output the prediction results:

[0087] p(y) = softmax(w*P) com +b) (16).

[0088] Example 1:

[0089] like Figure 1 As shown, the specific embodiment of the present invention is as follows:

[0090] Step 1: First, 33300 protein-protein interaction pairs were obtained from the VirHostNet 2.0 database, including 1046 viral proteins and 7459 host proteins. After removing duplicates, abnormalities and missing data, there were 29940 protein-protein interaction pairs left. The sequence information of viral and host proteins was obtained by querying the UniProt database. Proteins with sequence lengths not between 50 and 5000 were removed, as well as the corresponding protein-protein interaction pairs.

[0091] Step 2: All protein sequence information was processed into protein sequence simulation documents. The amino acids in the sequence were separated by spaces. The trained Word2Vec Skip-Gram model was used to obtain the embedding vector representation of 20 amino acids. Each amino acid formed a unique vector representation. According to the dipole and volume of the side chain of the amino acid, the 20 amino acids were grouped into seven groups, and the One-Hot encoding was used to obtain the unique vector of the 20 amino acids. Then the amino acids were spliced into a 20x40 amino acid matrix, and finally the amino acid matrix was used to represent the 2D matrix of the protein sequence.

[0092] Step 3: Define Dense Layer() function, DCNNBlock1 block and DCNNBlock2 block to build dense connected convolutional network Dense-net, forming a six-layer convolutional network. The Dense Layer() function is mainly composed of convolutional layers, which can be called by DCNNBlock1 block and DCNNBlock2 block to form dense blocks. DCNNBlock1 is set as a dense convolutional block with constant nb_filter convolutional kernels, and DCNNBlock2 is set as a dense convolutional block with half convolutional kernels. After convolution, the self-attention module is used to mine long-range dependencies, and finally a 64-dimensional vector is obtained after global average pooling (GAP).

[0093] Step 4: Use the POSSUM online Web server to input protein sequences in raw or FASTA format to obtain PSSM matrices of different sizes and RPM-PSSM feature descriptors generated by averaging in local regions.

[0094] Step 5: Design a multi-layer DNN with a fully connected layer, using the RPM-PSSM feature descriptor as input, which is input to the designed multi-layer DNN, whose formula is as follows:

[0095] o1=α(W1E+v1) (7)

[0096] o i =α(W i o i-1 +bi ),2≤i≤M (8)

[0097] Among them o i W represents the output produced by the i-th fully connected layer. i and b i Let represent the parameter matrix and bias term, respectively; α represent the ReLU activation function that enables nonlinear transformation; and M represent the number of fully connected layers. The last fully connected layer outputs the final RPM-PSSM feature.

[0098] Step 6: Multiply the extracted viral protein sequence features and the host protein sequence features using the input tensor list (element-by-element). Perform the same processing on the RPM-PSSM features. The formula is as follows:

[0099] P seq =P human ⊙P virus (13)

[0100] P rpm_pssm =P human_pssm ⊙P virus_pssm (14)

[0101] P seq_pssm =P seq ⊙P rpm_pssm (15)

[0102] The last three features P seq_pssm P seq P rpm_pssm Feature fusion yields a 192-dimensional feature vector P. com Construct a multilayer perceptron (MLP) to study P com Use softmax to perform prediction and output the prediction results:

[0103] p(y) = softmax(w*P) com +b) (16)

[0104] Verification of the effectiveness of this invention:

[0105] The performance of this invention was evaluated using 5-fold cross-validation. In 5-fold cross-validation, all viral host protein-protein interaction data were randomly divided into 5 parts. One part was used as the test set each time, and the remaining 4 parts were used as the training set. The area under the ROC curve (AUC) was used to quantitatively evaluate the performance of this invention. To prevent randomness, the validation results were the average of 100 5-fold cross-validations. This invention was compared with other methods, and the results are as follows: Figure 2 As shown, this invention achieved the highest AUC value, AUC = 0.99. Verification results demonstrate that the performance of this invention is superior to other methods.

Claims

1. A method for predicting viral and host protein-protein interactions based on a densely connected convolutional network, characterized in that, Comprising the following steps: Step 1: Preprocessing data; Step 2: Using the Skip-Gram model in Word2Vec and One-Hot encoding to obtain the embedding vector of amino acids; Step 3: According to the word embedding matrix of the protein, a densely connected convolutional network is constructed for feature extraction; Step 3.1: Define DCNNBlock1 block and DCNNBlock2 block to construct a densely connected convolutional network (Dense-net), and the convolution kernel changes and convolution formula in the dense block are as follows: where nb_filter is the convolution kernel, is the shared weight of each layer of convolution, is the bias vector, h() is the activation function, is the latent vector generated after the convolution layer, is the splicing result of the previous layer of convolution output and the current layer output; Step 3.2: After obtaining a 3D tensor through six layers of convolution, a self-attention module is designed to mine long-range dependencies, and the formula is as follows: wherein is the query, is the key value, is the value, , , is a parameter weight matrix of size , denotes the index position of the selected information, is an attention scoring function, the attention score is computed from the key vector and the query in a certain spatial correlation, is a scaling factor; Step 4: Obtain the average RPM-PSSM feature descriptor through the PSSM feature matrix; Step 5: According to the feature descriptor, a fully connected neural network is constructed for feature extraction; Step 6: Feature fusion is performed by concatenate operation, and finally the trained model is used for prediction; Step 6.1: The extracted viral protein sequence features and host protein sequence features are input into the element-wise product of the input tensor list, and the RPM-PSSM features are also processed in the same way, and the formula is as follows: The last three features , , Feature fusion, get 192-dimensional feature vector ; Step 6.2: Build a multi-layer perceptron, MLP, to predict Using Implement the prediction, output the prediction result: 。 2. The virus-host protein-protein interaction prediction method based on dense connection convolutional network according to claim 1, wherein, Said step 1 is specifically to preprocess the virus-host protein-protein interaction data and protein sequence information, and remove outliers and missing values. 3.The virus-host protein-protein interaction prediction method based on dense connection convolutional network according to claim 1, characterized in that Said step 2 is specifically: Step 2.1: Process all protein sequence information into protein sequence simulation documents, and simulate amino acids as words; Step 2.2: Use the trained Skip-Gram model in Word2Vec to obtain the embedding vector of 20 amino acids, where the Skip-Gram model is used to obtain the co-occurrence similarity of amino acids, and the formula is as follows: where YAis the negative log-likelihood loss function result, T is the given protein sequence size, c is the context size, the sub-sequence of the given protein sequence T, in amino acid is centered, length is 2c+1, a denotes the amino acid, is the first part embedding of the t-th amino acid, is the first part embedding of the adjacent amino acid, p is the probability defined as Softmax, n is the negative sampling size, is the negative sample that does not co-occur with in the same context; Step 2.3: According to the dipole and volume of the side chain of the amino acid, 20 kinds of amino acids are grouped into seven groups, namely {AGV}, {C}, {FILP}, {MSTY}, {HNQW}, {DE} and {KR}, and One-Hot is used for encoding., and the vector of amino acid in step 2.2 is spliced to form a 20x40 amino acid matrix.

4. The virus-host protein interaction prediction method based on dense connection convolutional network according to claim 1, characterized in that Said step 4 is specifically: Step 4.1: Before using the POSSUM online Web server, the data needs to be processed, and the sequence length must be between 50 and 5000, otherwise it cannot be used POSSUM; Step 4.2: When using the POSSUM online Web server, through the client Web interface, the user uploads the protein sequence file in FASTA format, and the POSSUM calculates the RPM-PSSM descriptor in the server background. After the job is completed, the descriptor result is returned to the user by email.

5. The method of claim 1, wherein the method is based on a dense connection convolutional network. Said step 5 is specifically: Step 5.1: Design a multi-layer DNN with a fully connected layer, use the RPM-PSSM feature descriptor as input, and the formula is as follows: wherein represents the output generated by the i-th fully connected layer, and represent the parameter matrix and the bias term, respectively, represents a RELU activation function that can implement a non-linear transformation, M represents the number of fully connected layers; the output generated by the last fully connected layer is used as the final RPM-PSSM feature.

Images