A compound-protein interaction prediction method fusing multi-view information

Abstract

The application provides a compound-protein interaction prediction method fusing multi-view information; firstly, a compound SMILES string is converted into a molecular graph structure by using an RDkit toolkit, and feature information of the compound is extracted by using a graph attention network; secondly, an amino acid long sequence of a protein is divided into subsequences, and biological information of the protein is mined by using a multilayer Transformer encoder; then, interactions of the compound and the protein are simulated from micro and macro views respectively, and interaction information of the two views is fused; finally, the fused multi-view interaction information is sent into a full connection neural network, the model is saved, and interactions of the compound and the protein are predicted; the application deeply mines feature information of the protein, effectively fuses interactions of the compound and the protein in multi-views, improves a prediction accuracy of the interactions of the compound and the protein, and greatly improves an efficiency of new drug research and development.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical artificial intelligence, and particularly relates to a compound-protein interaction prediction method fusing multi-view information. BACKGROUND

[0002] In the medical field and the global pharmaceutical industry, the development of new drugs is a complex, long and very low success rate process. In the case of limited efficacy of existing drugs, exploring new indications of existing drugs can greatly reduce the cost of new drug development and to some extent alleviate the problem of limited drug efficacy. In the traditional drug development process, one-third of the time and funds are spent on early compound action target screening and discovery, and how to quickly identify compound-protein interactions becomes a key task in the drug discovery process.

[0003] In recent years, with the significant increase in the amount of available biomedical data, greater data volume and more computing resources have enabled deep learning to be widely used in compound-protein interactions. In the compound-protein interaction prediction task, compounds and proteins are often input in the form of sequences, and deep learning algorithms superior to shallow machine learning are used to extract feature information of molecular compounds and protein sequences, such as convolutional neural networks, recurrent neural networks, etc. However, due to the complexity of biological features, these methods fail to mine biological features with explainability and deeper levels.

[0004] With the aid of deep learning technology, significant progress has been made in many aspects of drug discovery. Although deep learning has shown its powerful and efficient learning ability in the study of compound-protein interactions, current methods still have certain limitations. For example, proteins in compound-protein interactions are composed of sequences of 20 amino acids, and different arrangement orders represent biological genetic information, so the context information of the amino acid sequence represents the biological information of the protein and contains the mystery of life. However, most compound-protein interaction methods based on deep learning use classical convolutional neural networks and recurrent neural networks to extract protein information, which makes it difficult to extract complex biological information in the amino acid sequence; and in the interaction process of compounds and proteins, many current prediction methods directly splice the obtained compound and protein representation vectors to obtain the interaction vector representation of CPI, these methods only simulate the interaction between compounds and proteins, and lack the interaction information between substructures. SUMMARY

[0005] The present application aims to provide a multi-view information fusion compound-protein interaction prediction model, which predicts the interaction relationship between compounds and proteins, improves the verification efficiency, and reduces the verification cost.

[0006] To achieve the above object, the technical scheme of the present application is as follows: a compound-protein interaction prediction method fusing multi-view information, comprising:

[0007] Step 1: express the compound SMILES string as a molecular graph structure, and use a graph attention network to extract compound feature information;

[0008] Step 2: divide the protein sequence into amino acid subsequences, and use a multi-layer Transformer encoder to extract protein feature information;

[0009] Step 3: perform microscopic and macroscopic interaction on the extracted compound feature information and protein feature information, respectively, and fuse multi-view interaction information;

[0010] Step 4: input the fused compound-protein multi-view interaction information into a classification model for training, and then save the model;

[0011] Step 5: load the model, input the compound and protein information to be predicted, predict the compound-protein interaction, and output the prediction result.

[0012] Further, step 1 specifically comprises:

[0013] Step 1.1: use the RDKit toolkit to construct the compound SMILES string into a two-dimensional graph structure; the two-dimensional graph structure is represented as G={V,E}, wherein V is a set of atomic nodes, E is a set of edges, v i ∈V represents the i-th atom, e ij ∈E represents the edge between atom i and atom j; use a randomly generated vector to represent the initial vector of the atom wherein S c and D c are the number of atoms of the compound and the initial dimension of the vector, respectively;

[0014] Step 1.2: use a graph attention network to extract the feature information of the compound.

[0015] Further, step 1.2 specifically comprises:

[0016] Step 1.2.1: obtain the attention coefficient α of the neighbor node to the atom v i α ij represents the importance of node v j to node v i , which is calculated by the following formula:

[0017]

[0018] where W is a shared parameter matrix that linearly transforms the vertex features, || denotes concatenation of the transformed feature vectors, a(·) denotes mapping the concatenated feature vectors to a real number, N i is the neighbor node set of vertex v i , LeakyReLU(·) is a normalization representation;

[0019] Step 1.2.2: Update the latent vector representation of the node using the attention coefficient a as the final representation of vertex v i :

[0020]

[0021] where v i ' is the embedding representation of the fusion of the neighborhood information of node i, v j is the neighbor node of vertex v i , W is a parameter matrix, and σ(·) is an activation function;

[0022] Step 1.2.3: Use the multi-head attention mechanism to aggregate the output vectors of the K-layer graph attention network, and set K to 3:

[0023]

[0024] where, is the feature representation of an atom in a compound, v i '(K) is the output vertex vector of the aggregation of the K-layer graph attention network, is the attention coefficient between atom i and atom j in the k-th layer graph attention network, W k is the matrix parameter of the k-th layer graph attention network, n is the number of atoms in the compound, d is the dimension of the atomic feature representation, and σ(·) is an activation function.

[0025] Further, step 2 specifically includes:

[0026] Step 2.1: Referring to Figure 2 , divide the protein into overlapping n-gram amino acid subsequences, for example, when n is 3, the amino acid sequence "MNPAAEA…ATDSY" will be divided into "MNP", "NPA", "PAA", …, "ATD", "TDS", "DSY" subsequences;

[0027] Step 2.2: Represent the protein sequence as S = s1, s2,..., s p-n+1 , where s i is the i-th subsequence, and p is the total length of the amino acid sequence;

[0028] Step 2.3: Obtain word embedding representation using random initialization where m = n - p + 1, m is the number of sub-sequences after division, d x is the dimension of x embedding ;

[0029] Step 2.4: Word embedding representation x embedding Through n-layer Transformer encoder, the embedding representation of protein sub-sequence is extracted.

[0030] Further, the step 2.4 specifically comprises:

[0031] Step 2.4.1: Learn the importance of different positions of amino acid sequence through Attention mechanism:

[0032]

[0033] where Q = K = V = x embedding , h is the number of attention heads, d x is the dimension of x embedding , softmax(·) normalizes the scores so that the calculated values are all positive and the sum is 1;

[0034] Step 2.4.2: Capture the features of different dimensions of amino acids and learn the rich feature information of different spaces of amino acid sequence through multi-head mechanism:

[0035]

[0036] head h = Attention(QW i Q ,KW i K ,VW i V ), (6)

[0037] where is a learnable parameter matrix, || represents concatenating H heads;

[0038] Step 2.4.3: Each position word is individually processed through Position-wise Feed-Forward Network layer, first through ReLU activation function, then linear transformation, and finally the feature representation of amino acid sub-sequence is obtained

[0039] FFN(x) = max(0, xW1 + b1)W2 + b2, (7)

[0040] X residueFFN(MultiHead(x embedding )). (8)

[0041] where W1 and W2 are parameter matrices, and b1 and b2 are bias vectors.

[0042] Further, the step 3 specifically includes:

[0043] Step 3.1: In the microscopic view, extract the interactive features Pair of the compound atoms and the protein amino acids. micro ;

[0044] Step 3.2: In the macroscopic view, extract the interactive features Pair between the compound and the protein. macro ;

[0045] Step 3.3: Concatenate Pair obtained in step 3.1 and Pair obtained in step 3.2, to obtain the final compound-protein interaction feature representation Pair micro . macro . output .

[0046] Further, the step 3.1 specifically includes:

[0047] Step 3.1.1: Separate the potential features X atom and X residue of the compound and the protein through multi-layer perception, and convert them into the microscopic feature representations x c and x p of the compound and the protein:

[0048] x c = f(W c X atom +b), (9)

[0049] x p = f(W r X residue +b), (10)

[0050] where f(·) is a nonlinear activation function, b is a bias vector.

[0051] Step 3.1.2: Obtain the microscopic attention matrix of the compound and the protein

[0052] Α cp = f(W·(x c +x p )+b), (11)

[0053] where, is a parameter matrix, b is a bias vector;

[0054] Step 3.1.3: Generating microscopic Attention matrix of compound and protein by different average dimensions and A p ∈ m×d :

[0055] A c = sigmoid(MEAN(A cp , 2)), (12)

[0056] A p = sigmoid(MEAN(A cp , 1)), (13)

[0057] where MEAN(Input, dim) means the average operation of Input in given dimension dim, sigmoid(·) is an activation function.

[0058] Step 3.1.4: In order to preserve more original feature information of atoms and amino acid sub-sequences, the microscopic compound-protein pair is updated as:

[0059] Compoud = X compound · 0.5 + X compound ⊙ A c , (14)

[0060] Protein = X protein · 0.5 + X protein ⊙ A p , (15)

[0061] where ⊙ represents the dot product of vector elements.

[0062] Step 3.1.5: Concatenating the latent vector representation of compound and protein to obtain the microscopic view of compound-protein pair representation:

[0063] Pair micro = concat(Compound, Protein), (16)

[0064] where concat(·) represents the concatenation operation, Pair micro represents the microscopic view of compound-protein pair representation.

[0065] Further, the step 3.2 specifically comprises:

[0066] Step 3.2.1: Dimension reduction is performed on the representation of atoms and amino acid sub-sequences obtained in the encoding module to obtain the representation of compound and protein as the input z of AE2 network.(0,v) = x v (v = 1, 2):

[0067]

[0068] where n denotes the number of atoms of the compound, || denotes concatenation operation, and similarly we have:

[0069]

[0070] where m denotes the number of protein amino acid subsequences;

[0071] Step 3.2.2: Refer to Figure 3 , and z (0,v) = x v Input AE networks, AE networks are composed of L layers of nonlinear transformation, then the output of the 1th layer is:

[0072]

[0073] where W and b are the weight matrix and bias of the 1th layer in AE networks, a(·) denotes a nonlinear activation function, and the corresponding reconstruction is the output z (L,v) of the Lth layer.

[0074] Step 3.2.3: Take the low-dimensional representation as the latent vector representation H, which represents the macroscopic interaction features of the compound and the protein.

[0075] Step 3.2.4: Update the latent vector representation H in the Degradation networks, the input is g (0,v) = H, and the output of the 1th layer is:

[0076]

[0077] Similarly, W l dg and b are the weight matrix and bias of the 1th layer in Degradation networks,

[0078] Step 3.2.5: Obtain the macroscopic feature representation Pair macro of the compound and the protein.

[0079] Further, the step 4 specifically comprises:

[0080] The final compound-protein multi-view interaction representation Pair obtained in step 3.3 is obtained output And the label is input into the fully connected neural network, the model is trained with Cross-EntropyLoss, and the best model best_model is saved:

[0081]

[0082] Wherein, y is the true label value, is the predicted value.

[0083] Further, the step 5 specifically comprises:

[0084] Load the model best_model, input the compound-protein information in the verification data into the model, judge whether the compound and protein exist interaction relationship, and output the corresponding evaluation index.

[0085] The present application can achieve the following technical effects due to the use of the above technical solutions: the present application uses a deep learning model to deeply mine the structure information of the compound and the biological information of the protein, effectively enriching the feature information of the compound and the protein. The compound-protein interaction feature representation is extracted by combining micro and macro views, effectively enhancing the model prediction performance, improving the efficiency and accuracy of the compound-protein interaction, and having robustness in balanced and unbalanced data sets, greatly reducing the cost of drug research and development, and to a certain extent, alleviating the problem of limited drug efficacy. BRIEF DESCRIPTION OF DRAWINGS

[0086] Figure 1 It is a compound-protein interaction prediction method flow chart fusing multi-view information;

[0087] Figure 2 It is a protein feature information extraction flow chart;

[0088] Figure 3 It is an AE2-Nets structure diagram. DETAILED DESCRIPTION

[0089] The embodiments of the present application are implemented on the premise of the technical solutions of the present application, and detailed implementation methods and specific operation processes are given, but the protection scope of the present application is not limited to the following embodiments.

[0090] The present application will be described in detail below in conjunction with the embodiments, so that those skilled in the art can implement it after referring to the present application.

[0091] The embodiment takes Windows system as a development environment, takes Pycharm as a development platform, and takes Python as a development language, and adopts the compound-protein interaction prediction method fusing multi-view information to perform prediction of compound-protein interaction relationship.

[0092] As shown in the compound-protein interaction prediction method fusing multi-view information in the embodiment, the method comprises the following steps: Figure 1

[0093] SMILES string representation of the compound and sequence representation of the protein are acquired, two public data sets of Human and C.elegans are adopted, the data set is divided into a training set and a test set, the ratio is 8:2, the training set is divided into 10 parts using the Scikit-learn library, 9 parts are used as a training set and one part is used as a validation set in turn, and the average of 10 times is used as an estimate of the accuracy; the training set is used as input, the model is trained and the best model is saved; the test set is used as input, the saved model is loaded, the evaluation index and the prediction score Score of the compound and the protein are obtained, the evaluation index includes accuracy (ACC), AUC and AUPR;

[0094]

[0095]

[0096]

[0097] TP is a true positive, the number of positive classes correctly predicted as positive classes; FP is a false positive, the number of negative classes incorrectly predicted as positive classes; FN is a false negative, the number of positive classes incorrectly predicted as negative classes; TN is a true negative, the number of negative classes correctly predicted as negative classes. AUC is represented by the area under the ROC curve, and AUPR represents the relationship between precision and recall;

[0098] According to the above steps, the compound-protein relationship prediction effect of the application is compared with the GNN-CPI model, the TransformerCPI model, the Moltrans model and the BACPI model. As can be seen from Table 1, the method proposed in the application is obviously better than other methods in ACC, AUC and AUPR.

[0099] Table 1 Comparison of different model prediction results for compound-protein interaction

[0100]

[0101]

[0102] ​The foregoing description of specific exemplary embodiments of the application has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the application to the precise forms disclosed, and obviously many modifications and variations are possible in light of the above teaching. It is intended that the scope of the application be limited not with this detailed description, but rather by the claims appended hereto.

Claims

1. A compound-protein interaction prediction method fusing multi-view information, characterized in that, The application relates to a compound-protein interaction prediction method based on multi-view interaction information fusion. Step 1: compound SMILES string is expressed as a molecular graph structure, and a graph attention network is used to extract compound feature information; Step 2: the protein sequence is divided into amino acid subsequences, and a multi-layer Transformer encoder is used to extract protein feature information; Step 3: microscopic and macroscopic interactions are carried out on the extracted compound feature information and protein feature information, and multi-view interaction information is fused; Step 4: the fused compound-protein multi-view interaction information is sent into a classification model for training, and then the model is saved; Step 5: the model is loaded, the compound and protein information to be predicted are input, the compound-protein interaction is predicted, and the prediction result is output; The step 3 specifically comprises: Step 3.1: In micro view, extract interaction features of compound atoms and protein amino acids ; Step 3.2: Extracting interaction features between compounds and proteins in macroscopic view ; Step 3.3: The compound obtained from Step 3.1 and Step 3.2 is spliced to obtain the final compound-protein interaction feature representation ; The step 3.1 specifically comprises: Step 3.1.1: Potential features of the compound and protein and Transformed into microscopic feature representations of the compound and protein after passing through a multilayer perceptron and : (9) (10) wherein, is a non-linear activation function, , is a bias vector; Step 3.1.2: Obtaining the micro-attention matrix of the compound and the protein : (11) wherein, is a parameter matrix, is a bias vector; Step 3.1.3: Generating micro-Attention matrices for compounds and proteins by different average dimensions and : (12) (13) wherein, will be In a given dimension performing an averaging operation, is an activation function; Step 3.1.4: In order to retain more original feature information of atoms and amino acid subsequences, the microscopic compound-protein pair is updated based on: (14) (15) wherein denotes vector element-wise multiplication; Step 3.1.5: The latent vector representation of the compound and the protein is spliced to obtain the microscopic view compound-protein pair representation: (16) wherein, represents a splicing operation, represents compound-protein pair representation on a micro view.

2. The fusion multi-view information compound-protein interaction prediction method of claim 1, wherein, The step 1 specifically comprises: Step 1.1: Use the RDKit toolkit to construct the compound SMILES string into a two-dimensional graph structure; the two-dimensional graph structure is represented as follows: ,in, It is a set of atomic nodes. It is a set of edges. Indicates the first One atom, Represents atoms and atoms The edges between them; the initial vectors of the atoms are represented by randomly generated vectors. , ,in, and These represent the number of atoms in the compound and the initial dimension of the vector, respectively. Step 1.2: the feature information of the compound is extracted by using a graph attention network.

3. The fusion multi-view information compound-protein interaction prediction method of claim 2, wherein, Step 1.2 specifically comprises: Step 1.2.1: Acquire the attention coefficient of the neighbor node for the atom , represents the importance of the node for the node , which is calculated by the following formula:​ (1) wherein is a shared parameter matrix, performing a linear transformation on the vertex features, denotes concatenating the transformed feature vectors, denotes mapping the concatenated feature vectors to a real number, is a vertex set of neighbor nodes, performing a normalization representation; Step 1.2.2: Using attention coefficients Updating the latent vector representation of the node as the final representation of the vertex : (2) wherein, is a fusion node an embedded representation of neighborhood information, is a node a neighbor node of the node, is a parameter matrix, is an activation function; Step 1.2.3: the output vector of the K-layer graph attention network is aggregated by using a multi-head attention mechanism: (3) wherein, is a feature representation of an atom in a compound, is an output vertex vector of a polymeric K-layer graph attention network, is a K-th layer graph attention network, is an attention coefficient between an atom and an atom of a K-th layer graph attention network, is a number of atoms of a compound, is a dimension of an atom feature representation, is an activation function.

4. The fusion multi-view information compound-protein interaction prediction method of claim 3, wherein, The step 2 specifically comprises: Step 2.1: the protein is divided into overlapping n-gram amino acid subsequences; Step 2.2: Represent the protein sequence as follows ,in For the first Subsequences The total length of the amino acid sequence; Step 2.3: Obtain word embedding representation using random initialization where, , is the number of divided subsequences, is dimension of Step 2.4: Word Embedding Representation By Layer Transformer Encoder, extracting an embedding representation of the protein sequence.

5. The fusion multi-view information compound-protein interaction prediction method of claim 4, wherein, The step 2.4 specifically comprises: Step 2.4.1: the importance of different positions of the amino acid sequence is learned by using an Attention mechanism: (4) wherein, , , the number of attention heads, the dimension of the dimension of the scores are normalized so that the calculated values are all positive numbers and the sum is 1; Step 2.4.2: multi-head mechanism is used to capture the features of different dimensions of the amino acid, and the feature information of different spatial richness of the amino acid sequence is learned: (5) (6) wherein, , is a learnable parameter matrix, represents concatenating H pads; Step 2.4.3: Each position's word is individually passed through a Position-wise Feed-Forward Network layer, first through a ReLU activation function, then through a linear transformation, and finally to obtain the feature representation of the amino acid subsequence : (7) (8) wherein, and is a parameter matrix, and is a bias vector.

6. The fusion multi-view information compound-protein interaction prediction method of claim 5, wherein, The step 3.2 specifically comprises: Step 3.2.1 : Dimensionality reduction of the representations of the atomic and amino acid subsequences obtained in the encoding module, resulting in representations of the compounds and proteins as input to the AE2 network ):​ (17) wherein, represents the number of atoms of the compound, represents the splicing operation, and similarly, we get: (18) wherein, represents the number of protein amino acid subsequences; Step 3.2.2: Apply the L-layer non-linear transformation to the output of step 3.2.

1. The input to the AE networks, AE networks consists of L-layer non-linear transformation, then the output of the first layer is: The output of the layer is: (19) where and are the weight matrix and bias of the layer in the AE network, respectively, denotes a nonlinear activation function, , and the corresponding reconstruction is given by the output of the layer. Step 3.2.3: Reduce the representation of the low-dimensional representation as a latent vector representation representing macro-interaction features of compounds and proteins; Step 3.2.4: Update latent vector representations in Degradation networks , input is , the layer output is: (20) wherein, and is the weight matrix and bias of the layer in the Degradation networks, ; Step 3.2.5: Obtain macroscopic representation of compound and protein .

7. The fusion multi-view information compound-protein interaction prediction method of claim 6, wherein, The step 4 specifically comprises: The final compound-protein multi-view interaction representation obtained from step 3.3 and the label into a fully connected neural network, train the model using Cross-Entropy Loss, and save the best model best_model: (21) wherein, is the true label value, is the predicted value.

8. The fusion multi-view information compound-protein interaction prediction method of claim 7, wherein, The step 5 specifically comprises: The model best_model is loaded, the compound-protein information in the verification data is input into the model, whether the compound and the protein exist the interaction relationship is judged, and the corresponding evaluation index is output.

Images