Drug target affinity prediction method based on sequence modality and graph modality

Abstract

The invention provides a drug target affinity prediction method based on a graph mode and a sequence mode, and aims to solve the problems of single-mode limitation and local and overall structure capture of an existing graph neural network and realize multi-mode feature fusion. The method comprises the following steps: data preprocessing: converting a protein sequence and a molecular SMILES sequence into a Graph form; extracting features of a graph mode and a sequence mode, and respectively designing feature extraction modules; multi-modal feature fusion: carrying out feature fusion through a specific module; and carrying out regression prediction by using a deep neural network (DNN). In the model training, the prediction performance is improved by optimizing parameters and structures, and comparison experiment verification is performed, so that the reliability is shown, and the application potential in drug research and development is indicated.

Description

Technical Field

[0001] The present invention relates to the biological field and the computer field Background Art

[0002] Drug target affinity prediction plays a key role in drug design and optimization. By predicting the affinity between drug molecules and target structures, drug developers can specifically design and optimize the structure of drug molecules to improve their affinity and efficacy. The improvement in the accuracy and reliability of drug target affinity prediction enables drug developers to more efficiently screen candidate drug molecules with higher affinity, thereby reducing experimental costs and improving screening efficiency. This computationally assisted method has brought revolutionary changes to the process of drug discovery and optimization and accelerated the progress of drug research and development.

[0003] In the early stages of drug development, the selection of drug targets is usually based on experience and trial and error. However, this method is time-consuming and expensive, and cannot fully utilize the large amount of bioinformatics and cheminformatics data. Therefore, researchers began to use computational models and algorithms to predict the interaction between drugs and targets.

[0004] Molecular docking is one of the commonly used methods for predicting drug-target affinity. It simulates the binding process between drug molecules and target molecules and predicts the strength of affinity through calculation. The molecular docking method can predict the interaction mode and binding ability between drugs and targets based on the structural information of drugs and targets, thereby evaluating the affinity of drugs. In addition, the construction and simulation of the three-dimensional structure of drug molecules is also an important means of predicting drug-target affinity. Through computational models and algorithms, the spatial configuration and stability of drug molecules in the target structure can be predicted, thereby predicting the affinity between them.

[0005] With the rapid development of technologies such as machine learning and deep learning, these methods have also been widely used in drug-target affinity prediction. By training the model and using known drug-target interaction data, the association between drug structure and target characteristics can be learned and used to predict unknown drug-target interactions.

[0006] At present, deep learning-based DTIs prediction methods are mainly divided into sequence-based methods and graph-based methods. In sequence-based methods, one-dimensional sequence embedding of drug compounds and proteins is used as input. SMILES character sequences are used to characterize compound molecules, and then the SMILES character sequences and protein sequences are simply converted into one-hot encoding, and then CNN or RNN is used to extract features from them. However, using SMILES character sequences to characterize compound molecules cannot fully contain the physical and chemical properties of molecules and the geometric structure of molecules. CNN can extract local features of amino acids, and then stack multiple layers of CNN to compress the dimension of feature vectors to achieve the purpose of extracting global information. However, this may lose some global information. RNN may experience gradient vanishing when extracting long-distance sequence features, and thus cannot take into account the relationship between amino acid residues with longer distances in the sequence. And the one-dimensional representation of proteins has no structural representation. In graph-based methods, drug molecule graphs are used as drug representations. In order to make full use of the structural information of proteins, contact graphs can be used to construct protein graphs, and amino acids can be used as nodes of protein graphs. However, current graph-based methods treat drugs and proteins as independent graphs, and their features are usually extracted only through GNN. It is worth noting that the node features in the drug and protein graphs are just artificially selected atomic and amino acid properties, which makes the respective DTI models lack generalization.

[0007] Among graph-based methods, shallow DTIs prediction models are well developed and have shown good performance on various datasets. However, there are still a series of problems, such as GNNs with a small number of layers are not enough to capture the overall structure of the compound. In order to capture the global structure of the graph, the convolutional layers of the graph should be stacked deeply. Specifically, in order to capture the structure composed of many k-hop neighborhood-based networks, k graph convolutional layers should be stacked. However, due to the problems of over-smoothing and gradient vanishing, it is currently not feasible to build a deep structure of GNN. Therefore, most state-of-the-art (SOTA) GNN models do not exceed 3 or 4 layers in depth. Methods to overcome the depth limitation of GNNs to better capture the global structure of compounds and improve the model to more accurately preserve and utilize the local structural information of molecules. This is expected to further improve the performance and generalization ability of graph-based DTIs prediction models.

[0008] Regardless of whether it is a sequence-based method or a graph-based method, most deep learning models of drug-target interactions simply connect the feature vectors of compounds and proteins to simulate their interactions, which lacks interpretability. These methods ignore the contribution of intermolecular interactions between atoms in compounds and amino acids between proteins to affinity. Summary of the invention

[0009] The purpose of the present invention is to solve three problems existing in the prediction of drug protein affinity, namely, the limitation of single modality, the existing graph neural network capturing the local and overall structure of the compound, and the completion of multi-modal feature fusion.

[0010] In order to solve the above three problems, the present invention proposes a drug target affinity prediction method based on graph modality and sequence modality. The concept of the present invention is: first, the graph format with spatial structure and one-dimensional sequence information are simultaneously fed into the model through preprocessing of drugs and proteins. Then, the feature learning of drug and protein modalities is completed through graph modality feature extraction and sequence modality feature extraction. Finally, the features of the drug target pair are fused and input into the DNN to predict the affinity value of the drug target.

[0011] According to the above inventive concept, the present invention adopts the following technical solutions:

[0012] A. Data preprocessing: Through data preprocessing, the protein sequence and the SMILES sequence of the molecule are converted into a graph format that can be recognized by the model, and the sequence is compiled so that the computer can process the sequence.

[0013] B. Feature extraction of graph modality: Design feature extraction modules of graph modality according to the properties of molecular graph and protein graph respectively.

[0014] C. Feature extraction of sequence mode: Since the data input is all sequence, the feature extraction module used in this part is the same.

[0015] D. Multimodal feature fusion: Since the features extracted by different feature extraction modules are not interpretable if they are simply spliced ​​to simulate their interaction, a module needs to be designed to fuse them.

[0016] E. Use deep neural network (DNN) for regression prediction: Integrate the multimodal features of drug targets and apply DNN for regression prediction of drug target affinity.

[0017] While the model is being trained, its performance is optimized to ensure that the model is highly reliable and effective in predicting molecular-protein interactions. By continuously adjusting the parameters and structure of the model, as well as optimizing the loss function during training, the performance of the model can be improved and made more adaptable to specific prediction tasks. This comprehensive approach helps ensure that the model achieves better prediction results in the fields of drug discovery and bioinformatics. And comparative experiments were conducted with multiple prediction networks. At the same time, the proposed model was applied to the case of high-affinity drug repurposing of Alzheimer's disease-related proteins. The drugs screened from the FDA-approved drug library partially overlap with those in the anti-Alzheimer's drug library. This case shows that the model in this article has a certain degree of reliability in drug development, and provides a use case for applying deep learning models for predicting affinity to the search for high-affinity drugs for actual diseases. It also proves that the model in this article can help drug developers find high-affinity drugs or lead compounds for diseases, and provide a reference. The specific process is as follows Figure 1 shown.

[0018] Compared with the prior art, the present invention has the following outstanding substantive features and significant advantages: Since the amount of information contained in a single modality data is limited, the present invention proposes to use sequence and structure modality data to represent and learn drugs and proteins respectively. Specifically, for the sequence encoder, the Transformer architecture is used for processing. And a pre-training model is used, so that even when the amount of data is small, this method can learn high-quality sequence features. For Graph data, since different two data use different graph feature extraction modules, this helps to transmit different types of interaction information, thereby further obtaining graph node features with higher discrimination. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is an overall flow chart of the present invention

[0020] Figure 2 It is an architecture diagram of the drug-target affinity prediction method based on sequence modality and graph modality of the present invention.

[0021] Figure 3 yes Figure 2 The detailed description of the CGMS module is shown in the figure.

[0022] Figure 4 yes Figure 2 Detailed description of the Elem-feature Fusion module.

[0023] Figure 5 This is the experimental result of the model on the Davis dataset.

[0024] Figure 6 This is the experimental result of the model on the Kiba dataset. DETAILED DESCRIPTION

[0025] The present invention is described in detail below with reference to the accompanying drawings.

[0026] In the present invention, the machine running the experiment is a server cluster with an A100 graphics card, and the model is developed using the Pytorch development framework.

[0027] A. Data preprocessing.

[0028] For the SMILES encoding of molecules, the SMILES sequence is converted into a three-dimensional molecular graph structure through the chemical development library Rdkit, and the three-dimensional molecular graph structure is converted into a Graph format through the PyG library, which facilitates the subsequent encoder to extract information. The dimension of the initial vector is 88 dimensions.

[0029] For the amino acid sequence of a protein, the Grpah of the protein is a weighted protein graph constructed with residues as nodes, residue interactions as edges, and probability values ​​as edge features. In order to handle longer protein sequences (more than 1200 residues), a phase and splicing strategy is used to construct a contact map. The specific operation is as follows: The entire sequence is divided into multiple fixed-length L sequences with a step size of L / 2. The contact map of each subsequence is predicted in turn by the ESM-2 model. The probability matrix generated by each subsequence is spliced, and the overlapping parts are averaged. Algorithm 1 describes the specific construction process of the contact map. In addition, each residue has an initial feature vector. The information in this vector includes residue type, polarity, hydrophobicity, weight, group dissociation constant, etc. The dimension of the initial vector is 33 dimensions.

[0030] B. Feature extraction of graph modality.

[0031] For feature learning of drug molecular graphs. In this paper, for molecular graphs, we use a combination of local neural networks and global neural networks to complete graph neural network feature learning.

[0032] Add centrality encoding to the node description. This encoding is a key measure of the importance of a graph node. Generally, this centrality encoding is degree information. The larger the degree of a node, the more edges it has connected to other nodes, and the more important such a node is. The formula is as follows.

[0033]

[0034] Among them, z+ ,z - ∈R d is a learnable vector determined by the in-degree and out-degree of the node. By using centrality encoding in the input, attention can capture node importance signals in query and key. Therefore, the model can capture semantic relevance and node importance in the attention mechanism.

[0035] The GINE model is used to consider the local features of compound molecules. The basic idea is to update the node representation by iteratively aggregating the neighbor information of the node to capture the global structure of the graph. Since GNNs cannot complete very deep models, this model is still used to extract the local features of the compound. The specific implementation formula is as follows

[0036]

[0037] A variant of Transformer is introduced to extract the global features of compounds, mainly by modifying the attention mechanism. Two additional encodings are introduced on top of the attention mechanism, namely spatial encoding and edge encoding.

[0038] There are two advantages of using spatial encoding. First, the receptive field of GNN is limited to neighbors, while Spatial Encoding provides global information, and each node can pay attention to all other nodes in the graph. Second, by using b φ(i,j) , each node in the model can adaptively participate in all other nodes based on the structural information of the graph. φ(i,j) is a learnable scalar, which is given by φ(v i ,v j )Decide.

[0039] In many graph tasks, the information on the edges plays a very important role. However, the Transformer model designed for processing sequence data does not have the ability to capture the information on the edges, because the concept of "edges" does not exist in sequence data. Therefore, the information on the edges can be introduced into the attention mechanism as a weight bias. That is, when calculating the correlation between two nodes, the weighted sum of the edge features on the shortest path between the two nodes is used as the attention bias, where the weight is learnable. It can be expressed as formula (3).

[0040]

[0041] in, Is the shortest path SP (i,j) The characteristics of the nth edge in , is the corresponding learnable weight vector.

[0042] Finally, add the spatial encoding and side information encoding to A ij As shown below.

[0043]

[0044] The above steps are integrated and named CGMS module. The specific process is as follows Figure 3 shown.

[0045] For feature learning of protein graphs, we use a Graph Attention Network (GAT) with a gated skipping mechanism. GAT can dynamically adjust weights according to the importance of neighboring nodes to capture more meaningful neighborhood information. This is particularly important in protein graphs because different amino acid residues may contribute differently to the target task. In addition, protein graphs are usually heterogeneous, that is, there are different properties between nodes (amino acid residues) and edges (interactions). GAT can better handle this heterogeneity through a multi-head attention mechanism and capture the complex relationships between different types of nodes and edges. Stacking multiple GNN layers is necessary because of the need to aggregate distant neighbor information. However, graph neural networks often suffer from gradient vanishing and node feature degradation when stacking multiple layers. To address these problems, we introduce a gated skipping mechanism in the representation learning of each hidden layer. By adjusting the forgetting and update rates, this mechanism can fuse the states of different hidden features. As the depth of the model increases, each node is able to aggregate the information carried by remote nodes and retain the unique features of the node itself.

[0046] The following formula describes the graph attention network with gated skip connection mechanism.

[0047] e ij =a(WX i ||WX j )

[0048]

[0049]

[0050] Z i =sigmoid(W 1 P i (l+1) +W 2 P i (l) +b)

[0051] P i (l+1) =Z i P i(l+1) +(1-Z i ) i (l)

[0052] Among them, X i is the feature vector of node i, is the set of adjacent nodes of node i, E ij represents the attention coefficient between node i and node j, α represents the normalized attention coefficient, P i is the hidden layer feature of node i, σ is the nonlinear activation function, and a,W are the learnable weight matrices. i (l) and P i (l+1) represents the lth and l+1th layers of node i, W 1 and W 2 is the trainable matrix, b is the bias, z i Represents the learning ratio coefficient that retains the previous hidden layer information. The sigmoid activation function ensures that the learning ratio coefficient falls between [0-1].

[0053] C. Feature learning of sequence modality.

[0054] A deep bidirectional Transformer encoder is used to model molecular compounds and output high-quality feature embedding. Specifically, a variant of Transformer is used, that is, a Liner transformation is first performed on the input tensor, and then two basic Transformers are added to complete the operation of the compound sequence mode. At the same time, for the one-dimensional amino acid sequence representation of proteins, the sequence mode operation of molecular compounds can also be used.

[0055] D. Multimodal feature fusion.

[0056] For dual-modal fusion, this paper also built a module named Elem-feature Fusion, which first splices the two modules and then sends them to a lightweight attention module. At the same time, the original two features are sent to the Linear layer and the activation layer respectively, and then clicked with the values ​​calculated by the Attention Block, and then a residual operation is performed. Finally, the two processed features are added to complete the feature fusion of two different modes. Figure 4 shown.

[0057] Through the above operations, we obtained the drug characterization E d and protein characterization p, the two representations are feature-fused, and the present invention constructs a Cross-Attention module in the model. The Cross-Attention module establishes a connection based on the attention mechanism between the compound feature embedding and the protein feature embedding to better simulate the combination of compounds and proteins. The attention score of each atom in the compound constitutes the attention vector of the compound, and the attention score of each amino acid in the protein constitutes the attention vector of the protein. This enables the model in this paper to pay attention to different atoms of the compound and different amino acids of the protein according to the attention score, so that the model can automatically capture atoms in the compound and amino acid residues in the protein with higher attention scores, and can predict the function of the compound-protein interaction site. This improves the predictive performance of the model and enhances the interpretability from the perspective of intermolecular interactions.

[0058] E. Use deep neural network (DNN) for regression prediction.

[0059] The features of the fused drug-target pairs are fed into a deep neural network for regression prediction of drug-target affinity.

[0060] Reference Figure 5 ,6 They show the comparison between the present invention and other models. The experimental results show that the drug protein affinity prediction method based on graph mode and sequence mode proposed in the present invention is better than other mainstream affinity methods in various performances.

[0061] In summary, the drug protein affinity prediction method based on graph modality and sequence modality of the present invention extracts graph structure data from drug and protein sequence data, and screens potential drug binding sites based on protein structure, thereby establishing a high-precision drug target affinity prediction framework. The framework can extract drug target pair features from multimodal data, and effectively reduce the noise impact of non-interaction regions through multimodal information fusion. It can be widely used in the prediction of protein target affinity, and has great potential for guiding drug discovery.

Claims

1. A method for predicting drug-target affinity, comprising the following steps: A. Preprocess the drug and target data and encode small molecule compounds and protein sequences into graph data; and Convert the sequence into a form that can be understood by a computer; B. Introduce position coding and structure coding to encode and represent the graphic data of drugs and targets; C.Graph modality feature extraction; D. Sequence modal feature extraction; E. Use the multimodal fusion module to fuse drug and target features; F. Use deep neural network (DNN) for regression prediction.

2. The method according to claim 1, wherein the preprocessing step comprises preprocessing the SIMLES of the molecular compound. The steps of data preprocessing are as follows: For the SMILES encoding of molecules, the SMILES sequence is converted into a three-dimensional molecular graph structure through the chemical development library Rdkit, and the three-dimensional molecular graph structure is converted into a Graph format through the PyG library, which facilitates the subsequent encoder to extract information. The dimension of the initial vector is 88 dimensions.

3. The method according to claim 1, wherein the preprocessing step includes data encoding the protein sequence. The data preprocessing steps are as follows: For the amino acid sequence of a protein, the Grpah of a protein is a weighted protein graph constructed with residues as nodes, residue interactions as edges, and probability values ​​as edge features. In addition, each residue has an initial feature vector, which contains information such as residue type, polarity, hydrophobicity, weight, group dissociation constant, etc. The dimension of the initial vector is 33 dimensions.

4. According to the method described in claim 1, the local information of the drug graph is extracted using the GINE neural network; the global information of the drug graph is extracted using the GraphTransformer model; and the information of the target graph is extracted using the gated skip link graph attention mechanism network.

5. According to the method of claim 1, the sequence information of drugs and targets is extracted using a Transformer model.

6. According to the method of claim 1, the vectors of drug features and protein features extracted by the multimodal fusion module graph modality and sequence modality are fused.

7. According to the method of claim 1, the features of the fused drug-target pair are fed into a deep neural network (DNN) for regression prediction of drug-target affinity.

Images