Drug-target interaction prediction method based on graph neural network block

Abstract

The invention discloses a drug-target interaction prediction method based on graph neural network blocks. Accurate discrimination of the interaction between the drug and the target is a key stage of drug discovery, and the core of prediction of the interaction between the drug and the target is reasonable molecular characterization learning. The previous research has proved that GNN is effective in the aspect of drug feature coding. However, the GNN-based methods cannot effectively give consideration to local substructure characteristics of drug molecules and overall structure attributes of the molecules. For the current difficulty, a framework named GNNBlockDTI is provided. According to the GNNBlock DTI, multiple layers of GNNs are combined into a GNNBlock unit, and a feature enhancement strategy is introduced, so that a substructure mode of a drug map in a local range is captured. And filtering redundant information among the GNNBlocks by using a gating unit mechanism so as to obtain effective overall structure information of the medicine. According to the method, the global and local structure attributes of the drug molecules can be effectively coded, and the accuracy of drug-target interaction prediction is improved.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics, relates to a drug-target prediction link in a drug discovery process, and specifically relates to a drug-target interaction prediction method based on a graph neural network block. Background Art

[0002] The essence of drug action is to interact with biological target proteins to regulate their physiological functions. A key step in drug development is to screen compounds that interact with targets as candidate drugs. Virtual screening through computational methods can speed up the drug development cycle and reduce costs. In the field of machine learning, drug-target interaction prediction is usually regarded as a binary classification task, and its focus is on the feature encoding of drug and target input data. Drugs are usually represented as molecular graph data and feature encoded through graph neural networks (GNNs). In molecular graph representation learning, although shallow GNN-based models can identify substructure features within a local range, they cannot capture implicit structural patterns within a larger range. At the same time, when the pooling layer is used to read the entire graph representation from the node features, the local feature representation will suffer from severe information loss, resulting in the lack of consideration of the overall properties of the molecular graph in the obtained graph representation. Although the use of a deep GNN architecture can expand the receptive field of the model to cover the entire molecular graph, it cannot identify key structural information that highlights molecular features. In addition, due to the excessive number of message transmissions between nodes in the deep GNN model, the representations of all nodes in the graph will gradually become similar and become constants containing all node information. This over-smoothing of features will now cause the node features to be too single, thus seriously affecting the performance of graph representation. Summary of the invention

[0003] The purpose of this invention is to address the existing need for fully encoding drug and target representations, and propose a method that can retain the local substructure features of drug molecule graphs while embedding the overall properties of molecular graphs. We propose a new concept of graph neural network block (GNNBlock), which explicitly extracts structural feature properties in drug molecule graphs by constructing GNNBlocks that focus on substructures. In order to address the problems of information loss and information redundancy in the GNN architecture, feature enhancement strategies and gating units are also introduced to process and screen the key information of each layer of GNNBlock. Previous related methods have been unable to effectively balance the local substructure features and overall structural properties of drug molecules. This article will propose a new general framework to obtain reasonable drug molecule representations for accurate drug-target interaction predictions.

[0004] The present invention proposes a drug-target interaction prediction method, which is based on a graph neural network block and includes the following steps: Step S1, constructing a GNNBlock graph feature extraction module of a multi-layer GNN fused with gated logic, inputting the drug molecule feature representation into the graph feature extraction module, and finally reading the entire graph representation from the node feature matrix through pooling; Step S2, construct different feature extraction modules to process the composition sequence input and contact map input of the protein respectively, and send the amino acid composition sequence representation and protein residue contact map obtained by processing the protein sequence based on different pre-trained models into the above modules. Finally, the sequence representation and graph representation of the protein are fused, and the protein representation is obtained after processing by the maximum pooling layer (MaxPool).

[0005] Step S3, sending the drug representation obtained in step S1 and the protein sequence representation obtained in step S2 to a multi-layer perceptron classifier to obtain an interaction prediction value; In step S1, the drug molecular feature extraction process includes the following steps: First, the drug is represented as a molecular graph, whose nodes are atoms and edges are chemical bonds. The node features are obtained from the atomic features, including the atomic symbol, atomic charge number, atomic connection number, whether it is aromatic (0 for no, 1 for yes), and whether it is on a ring (0 for no, 1 for yes).

[0006] The drug molecular graph is then input into the GNNBlock module that combines multiple layers of GNN to capture the structural features in the local range of the molecular graph.

[0007] Then, a feature enhancement strategy of expanding first and then refining is adopted. The node features are upgraded in dimension in the last layer of GNNBlock, and then the expanded features are refined and reduced in dimension. The output of GNNBlock is re-encoded into a high-dimensional embedding to retain the effective information learned by the model.

[0008] Applying the gating strategy in GRU (Gated Recurrent Unit), the reset gate unit and update gate unit are set to filter and control the output of each layer of GNNBlock. Among them, the reset gate unit determines whether to retain the features of the previous layer, and the update gate unit determines the feature update ratio of this layer: R = sigmoid(H(L)*w_r1 + H(L+1)*w_r2 + b_r) Z = sigmoid(H(L)*w_z1 + H(L+1)*w_z2 + b_z) H(L+1) = tanh(H(L+1)*w_h1 + ((R*H(L))*w_h2) + b_h) Where H(L+1) represents the node feature vector of the L+1th layer GNNBlock. Sigmoid(·) and tanh(·) represent activation functions. W and b represent learnable parameters. Integrate the above modules to build a complete framework based on GNNBlock H(0) = GNNBlock(3)(X) H(L+1) = GNNBlock(n)(H(L)) H(L+1) = FeatureEnhance(H(L)) H(L+1) = GatedUnit( H(L), H(L+1) ) Where H(L+1) represents the node feature vector of the L+1th layer. X represents the initial atomic feature input. FeatureEnhance(·) and GatedUnit(·) represent the feature enhancement strategy and the gated unit mechanism, respectively.

[0009] The maximum pooling layer (MaxPool) is used to read the entire graph representation from the node feature matrix.

[0010] In step S2, the protein joint feature extraction process includes the following steps: First, we use the Bert-based protein pre-trained language model ProtBert to obtain the amino acid sequence representation, and use the Transformer-based protein sequence model ESM-1b to predict the protein residue contact map and obtain the protein graph network representation. Then the amino acid sequence representation and protein graph network representation are fed into the feature extraction modules of the 1DCNN-based framework and the weighted graph convolutional neural network (WGCN)-based framework, respectively, to obtain a high-dimensional joint representation of the protein. In the amino acid sequence extraction module, CNN layers with different window sizes are set to encode proteins at multiple scales. The intermediate vectors obtained from each CNN layer are fused and encoded to obtain the final representation of the protein sequence. H(L+1) = CNN( H(L) ) H_hat(L+1) = FC( H(L+1) ) Where H(L+1) represents the protein feature vector of the L+1th layer, H(0) is the initial input feature, FC(·) is the fully connected layer, and H_hat(L+1) represents the intermediate vector of the L+1th layer.

[0011] For the N-layer 1DCNN module, the protein representation is: P(seqs) = ( H_hat(1) || H_hat(2) || · · · || H_hat(N) ) (·||·) represents a vector concatenation operation.

[0012] In the protein graph network representation feature extraction module, a weighted graph convolutional neural network is used to represent proteins as a weighted connection graph with amino acids as nodes to obtain a protein graph representation; The sequence representation and graph representation of the protein are further fused element by element in the form of addition, and then the protein feature matrix is ​​processed by the maximum pooling layer (MaxPool) to obtain the protein representation.

[0013] The beneficial effects of the present invention are described as follows: The present invention provides a drug-target interaction prediction method based on graph neural network blocks. Traditional manual wet experiments to screen candidate drugs are not only costly and slow, but also cannot guarantee the success rate of the experiment. The present invention can effectively accelerate the verification cycle of drug-target interactions, reduce the cost of experimental trial and error, and guide and improve the success rate of wet experiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the specific implementation of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the specific implementation or the prior art description. Some specific embodiments of the present invention will be described in detail in an exemplary but not restrictive manner with reference to the drawings. The same reference numerals in the drawings indicate the same or similar components or parts. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings: Figure 1 This is the framework flow chart of the GNNBlockDTI model, GNNBlock module and GU gating logic module proposed in the present invention Figure 2 The AUC-ROC and PR-ROC performances of the different component layers N and L of GNNBlock in the GNNBlockDTI model Figure 3 The AUC-ROC and PR-ROC performance of different GNN types in the GNNBlockDTI model Figure 4 It is the effect of T-SNE dimension reduction and the vector difference result of all nodes in each graph after stacking GNNBlock, which is used to demonstrate the interpretability of the present invention; The performance of the present invention on three independent data sets is shown in the following table:

[0015] The ranking of the candidate drugs of the present invention on the three targets Q15825, P49773 and P30542 is as follows:

[0016] There are real corresponding drugs (Label is 1) among the top ten candidate drugs with predicted probability.

Claims

1. A drug-target interaction prediction method based on graph neural network blocks, characterized in that: The specific steps of the present invention are as follows: Step S1, constructing a GNNBlock graph feature extraction module of a multi-layer GNN fused with gated logic, inputting the drug molecule feature representation into the graph feature extraction module, and finally reading the entire graph representation from the node feature matrix through pooling; Step S2, constructing feature extraction modules based on CNN and GCN to process the composition sequence input and contact map input of the protein respectively, and sending the amino acid composition sequence representation and protein residue contact map obtained by processing the protein sequence based on different pre-trained models to the above modules, and finally fusing the sequence representation and map representation of the protein, and obtaining the protein representation after processing by the maximum pooling layer; Step S3, sending the drug representation obtained in step S1 and the protein sequence representation obtained in step S2 to a multi-layer perceptron classifier to obtain an interaction prediction value; In step S1, the drug molecular feature extraction process includes the following steps: First, the drug is represented as a molecular graph, whose nodes are atoms and edges are chemical bonds. The node features are obtained from the atomic features, including the atomic symbol, atomic charge number, atomic connection number, whether it is aromatic, and whether it is on a ring; The drug molecule graph is then input into the GNNBlock module that combines multiple layers of GNN to capture the structural features in the local range of the molecule graph. Then, a feature enhancement strategy of first expansion and then refinement is adopted. In the last layer of GNNBlock, the node features are upgraded in dimension, and then the expanded features are refined and reduced in dimension. The output of GNNBlock is re-encoded into a high-dimensional embedding to retain the effective information learned by the model. Applying the gating strategy in GRU, the reset gate unit and update gate unit are set to filter and control the output of each layer of GNNBlock; the reset gate unit determines whether to retain the features of the previous layer, and the update gate unit determines the feature update ratio of this layer: R = sigmoid(H(L)*w_r1 + H(L+1)*w_r2 + b_r) Z = sigmoid(H(L)*w_z1 + H(L+1)*w_z2 + b_z) H(L+1) = tanh(H(L+1)*w_h1 + ((R*H(L))*w_h2) + b_h) Where H(L+1) represents the node feature vector of the L+1th layer GNNBlock, Sigmoid(·), tanh(·) represent activation functions, and W, b represent learnable parameters; Combining the above modules, we build a complete framework based on GNNBlock: H(0) = GNNBlock(3)(X) H(L+1) = GNNBlock(n)(H(L)) H(L+1) = FeatureEnhance(H(L)) H(L+1) = GatedUnit( H(L), H(L+1) ) Where H(L+1) represents the node feature vector of the L+1th layer; X represents the initial atomic feature input, GNNBlock(n) represents the GNNBlock composed of n layers of GNN; FeatureEnhance(·) and GatedUnit(·) represent the feature enhancement strategy and the gated unit mechanism respectively; first, the structural features are extracted from the initial atomic features to obtain H(0), and then the node feature matrix is ​​obtained after being processed by multiple GNNBlocks containing feature enhancement strategies and gated unit mechanisms. Finally, the maximum pooling layer is used to read the entire image representation from the node feature matrix; In step S2, the protein joint feature extraction process includes the following steps: First, we use the Bert-based protein pre-trained language model ProtBert to obtain the amino acid sequence representation, and use the Transformer-based protein sequence model ESM-1b to predict the protein residue contact map and obtain the protein graph network representation. Then the amino acid sequence representation and protein graph network representation are fed into the feature extraction modules of the 1DCNN-based framework and the weighted graph convolutional neural network-based framework, respectively, to obtain a high-dimensional joint representation of the protein. In the amino acid sequence extraction module, CNN layers with different window sizes are set to encode proteins at multiple scales. The intermediate vectors obtained from each CNN layer are fused and encoded to obtain the final representation of the protein sequence. H(L+1) = CNN( H(L) ) H_hat(L+1) = FC( H(L+1) ) Where H(L+1) represents the protein feature vector of the L+1th layer, H(0) is the initial input feature, FC(·) is the fully connected layer, and H_hat(L+1) represents the intermediate vector of the L+1th layer; For the N-layer 1DCNN module, the protein representation is: P(seqs) = ( H_hat(1) || H_hat(2) || · · · || H_hat(N) ) (·||·) represents a vector concatenation operation; In the protein graph network representation feature extraction module, a weighted graph convolutional neural network is used to represent proteins as a weighted connection graph with amino acids as nodes to obtain a protein graph representation; The sequence representation and graph representation of the protein are further fused element by element in the form of addition, and then the protein feature matrix is ​​processed by the maximum pooling layer to obtain the protein representation.

Images