Drug-protein binding affinity prediction method
By introducing dynamic attention module and KAN neural network into the drug-protein binding affinity prediction model, the problem of insufficient feature extraction and lack of dynamic adjustment mechanism is solved, and more accurate and efficient multimodal data fusion is achieved, which improves the drug -Protein binding affinity predicted performance.
Patent Information
- Application Number
- CN202411158582.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-22
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-08-22
AI Technical Summary
The existing drug-protein binding affinity prediction methods have insufficient feature extraction and lack of dynamic adjustment mechanisms, which leads to ineffective multimodal data fusion, affecting the accuracy of prediction.
The affinity prediction model trained with multimodal drug-protein data input includes a feature extraction module, a dynamic attention module using memory module and a prediction module. The historical attention weight is stored through the memory matrix and dynamically adjusted the attention weight. , combined with KAN neural network and multi-layer perceptron, make full use of multimodal information for prediction.
Improve the performance of the prediction model, reduces the computational complexity and time cost, enables more flexibility in adjusting attention weights, adapts to new input data, enhances feature extraction ability and data context correlation, and improves drug-protein binding affinity prediction accuracy.
Smart Images

Figure CN118969065A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of machine learning and data mining, and particularly relates to a method for predicting drug-protein binding affinity. Background Art
[0002] A key step in drug discovery is predicting the binding affinity between a drug and its target protein. This process has traditionally relied on experimental methods, which are often time-consuming and costly. Therefore, developing efficient and accurate computational models to predict drug-protein binding affinity is crucial.
[0003] In recent years, with the development of deep learning technology, more and more researchers have applied it to the prediction of drug-protein interactions. These methods have significantly improved the accuracy of predictions by learning feature representations from large-scale data. Existing drug-protein binding affinity prediction methods mainly include machine learning-based and deep learning-based methods. Machine learning methods rely on manually designed features, which may not be comprehensive enough in complex molecular and protein interactions. Deep learning methods overcome the limitations of manually designed features by automatically extracting features, but most of them only use a single type of data and fail to fully utilize the multimodal information of drugs and proteins. In addition, existing deep learning methods usually use fixed attention weights during training and lack a dynamic adjustment mechanism, resulting in suboptimal performance of the model under different data conditions.
[0004] Therefore, the existing technology has problems such as insufficient feature extraction and lack of dynamic adjustment mechanism, which leads to ineffective multimodal data fusion. Summary of the Invention
[0005] To solve the above-mentioned problems in the prior art, the present invention adopts a drug-protein binding affinity prediction method, which is characterized by comprising: obtaining multimodal drug-protein data, inputting the multimodal drug-protein data into a trained affinity prediction model to obtain a predicted value of drug-protein affinity; the affinity prediction model comprises: a feature extraction module, a dynamic attention module using a memory module, and a prediction module; the training process of the affinity prediction model comprises:
[0006] S1: Obtain multimodal drug-protein data, input the multimodal drug-protein data into the feature extraction module, and obtain multimodal drug-protein features;
[0007] S2: splicing multimodal drug-protein features to obtain the initial comprehensive feature representation;
[0008] S3: Input the initial comprehensive feature representation into the dynamic attention module of the application memory module to obtain the final comprehensive feature;
[0009] S4: Input the final comprehensive features into the prediction module to obtain the prediction results;
[0010] S5: Calculate the loss function value based on the prediction results, and update the model parameters based on the loss function value. When the loss function value is minimized, the trained affinity prediction model is obtained.
[0011] The multimodal drug-protein data includes: a SMILES string of the drug, a molecular graph structure of the drug, and an amino acid sequence of the protein.
[0012] The feature extraction modules include: SMILES feature extraction module, molecular graph structure feature extraction module and amino acid sequence feature extraction module.
[0013] The SMILES feature extraction module includes: an embedding layer, a one-dimensional convolutional neural network, and an LSTM network. The SMILES feature extraction module processes the SMILES string of the drug by: converting the SMILES string into an integer code; inputting the integer-coded SMILES string into the embedding layer to obtain an embedding matrix; inputting the embedding matrix into the one-dimensional convolutional neural network to obtain local features; and inputting the local features into the LSTM network to obtain SMILES string features.
[0014] The molecular graph structure feature extraction module processes the molecular graph structure of the drug, including: representing the molecular graph structure of the drug as a molecular undirected graph G = (V, E), where V is a set of nodes, each node represents an atom in the molecular graph structure, and E is a set of edges, each edge represents a chemical bond in the molecular graph structure; initializing the feature vectors of each node and each edge; and using a graph convolutional network to process the feature vectors of the nodes and edges in the molecular undirected graph G to obtain the molecular graph structure features.
[0015] The amino acid sequence feature extraction module includes: an embedding layer, a one-dimensional convolutional neural network and an LSTM network; the amino acid sequence feature extraction module processes the amino acid sequence of the protein, including: converting the amino acid sequence of the protein into an integer code; inputting the integer-coded amino acid sequence into the embedding layer to obtain an embedding matrix; inputting the embedding matrix into the one-dimensional convolutional neural network to obtain local features; and inputting the local features into the LSTM network to obtain amino acid sequence features.
[0016] The dynamic attention module of the memory module includes a memory matrix, wherein the memory matrix includes attention weights of multiple historical time steps; the dynamic attention module of the memory module processes the initial comprehensive feature representation including:
[0017] Calculate the initial attention weight α of the initial comprehensive feature representation at time step t t ; According to the initial attention weight α t Update the memory matrix M at time step t-1 t-1 , get the memory matrix M at time step t t , according to the initial attention weight α t and memory matrix M t Calculate the final attention weight α at time step t t,final , according to the attention weight α t,final Adjust the initial comprehensive feature representation to obtain the final comprehensive feature.
[0018] Update the memory matrix M t-1 Including: If the memory matrix M t-1 The attention weights of T time steps have been stored, so in the memory matrix M t-1 Remove the attention weight of the earliest time step and add the attention weight α t ; Otherwise, directly in the memory matrix M t-1 Add attention weight α t ; Where T is the maximum dimension of the memory matrix.
[0019] Calculating the final attention weight at time step t includes: calculating the memory matrix M t The average of the attention weights of the historical time steps, the average value and the attention weight α t Perform weighted averaging to obtain the final attention weight at time step t.
[0020] The prediction module includes: a KAN neural network and a multi-layer perceptron; the prediction module processes the final comprehensive features by inputting the final comprehensive feature representation into the KAN neural network, inputting the output of the KAN neural network into the multi-layer perceptron, and obtaining a prediction result.
[0021] Beneficial effects:
[0022] 1. The present invention adds a memory module to the dynamic attention module to obtain an attention weight memory mechanism, and uses the attention weight memory mechanism to dynamically fuse multimodal features using historical attention weights, thereby improving the performance of the prediction model; 2. The present invention stores the historical attention weight distribution in the memory matrix, which makes the model more lightweight and reduces the computational complexity and time cost of model training; 3. The attention weight memory mechanism combines the historical attention weight distribution to more flexibly adjust the current attention weight distribution. Compared with storing other historical information, it can better retain and utilize the relevance of data context to adapt to new input data; 4. The present invention makes full use of the multimodal information of drugs and proteins for prediction, namely sequence data and graph structure data, thereby improving the performance of the prediction model; 5. The present invention introduces the KAN network to better capture the relationship between features and optimize the feature extraction capability of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 A flowchart of a method for predicting drug-protein binding affinity provided by an embodiment of the present invention;
[0024] Figure 2 This is a structural diagram of the affinity prediction model provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0025] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0026] like Figure 1 As shown, the present invention proposes a drug-protein binding affinity prediction method, comprising: obtaining multimodal drug-protein data; the multimodal drug-protein data comprising: a SMILES string of a drug, a molecular graph structure of a drug, and an amino acid sequence of a protein; inputting the multimodal drug-protein data into a trained affinity prediction model to obtain a predicted value of drug-protein affinity; the affinity prediction model comprising: a feature extraction module, a dynamic attention module applying a memory module, and a prediction module; the training process of the affinity prediction model comprises:
[0027] S1: Collect multimodal drug-protein data, input the drug's SMILES string, the drug's molecular graph structure, and the protein's amino acid sequence into the feature extraction module to obtain SMILES string features, molecular graph structure features, and amino acid sequence features;
[0028] S2: Concatenate SMILES string features, molecular graph structure features, and amino acid sequence features to obtain the initial comprehensive feature representation;
[0029] S3: Input the initial comprehensive feature representation into the dynamic attention module of the application memory module to obtain the final comprehensive feature;
[0030] S4: Input the final comprehensive features into the prediction module to obtain the prediction results;
[0031] S5: Calculate the loss function value based on the prediction results, and update the model parameters based on the loss function value. When the loss function value is minimized, the trained model is obtained.
[0032] like Figure 2 As shown, the feature extraction module includes: SMILES feature extraction module, molecular graph structure feature extraction module and amino acid sequence feature extraction module;
[0033] Furthermore, the SMILES feature extraction module includes: an embedding layer, a one-dimensional convolutional neural network, and an LSTM network; the processing of the drug SMILES string includes:
[0034] Step 1: Convert the SMILES string to integer encoding (Label encoding);
[0035] Step 2: Input the integer-encoded SMILES string into the embedding layer to convert the discrete characters into a continuous vector representation, namely the embedding matrix.
[0036] Step 3: Input the embedding matrix into a one-dimensional convolutional neural network (1D-CNN) to extract local features;
[0037] Step 4: Input the output of 1D-CNN into LSTM to capture the long-term dependencies in the sequence and obtain SMILES string features:
[0038] h t ,c t =LSTM(C t ,h t-1 ,c t-1 )
[0039] Among them, C t is the output of 1D-CNN, h t and c t are the hidden state and cell state of LSTM time step t, respectively, h t-1 and c t-1They are the hidden state and cell state of LSTM time step t-1 respectively. The final output of LSTM is the SMILES string feature F SMILES .
[0040] Furthermore, the amino acid sequence feature extraction module includes: an embedding layer, a one-dimensional convolutional neural network and an LSTM network; the amino acid sequence feature extraction module processes the amino acid sequence of the protein including;
[0041] Step 1: Encode the amino acid sequence of the protein as integers or labels; for protein sequences, 20 different categories are used to represent amino acids;
[0042] Step 2: Input the encoded amino acid sequence into the embedding layer to obtain the embedding matrix;
[0043]
[0044] Among them, L P is the length of the protein’s amino acid sequence, F e is the embedding dimension.
[0045] Step 3: Input the embedding matrix into the one-dimensional convolutional neural network 1D-CNN to extract local features;
[0046] Step 4: Input the features extracted by 1D-CNN into LSTM to capture the long-term dependencies of the sequence. The final output of LSTM is the protein amino acid sequence features.
[0047] Furthermore, the molecular graph structure feature extraction module processes the molecular graph structure of the drug including:
[0048] Step 1: Molecular Graph Representation: Represent the molecular graph structure of the drug as a molecular undirected graph G = (V, E), where V is a set of nodes representing atoms in the molecular graph structure, and E is a set of edges representing chemical bonds in the molecular graph structure.
[0049] Step 2: Feature Attribution: Initialize the feature vector of each node according to the type and attributes of the atom, and initialize the feature vector of each edge according to the type and attributes of the chemical bond;
[0050] Step 3: Use the graph convolutional network (GCN) to process the feature vectors of nodes and edges in the molecular undirected graph G to obtain the molecular graph structural features; the specific process is as follows:
[0051] Perform graph convolution on each node and update the node feature vector; the graph convolution operation can be expressed as:
[0052]
[0053] Among them, N(v) is the set of neighbor nodes of node v, w is the neighbor node of node v, c vw is the normalization coefficient, W (k) is the weight matrix of the k-th layer graph convolution, is the feature representation of node w output by the k-th layer graph convolution, is the feature representation of node v output by the k+1th layer graph convolution, b (k) is the bias term of the k-th layer graph convolution, and σ is the activation function;
[0054] Finally, the final feature representation of each node of the last layer of graph convolution output is obtained
[0055] The final feature representation is obtained by global pooling. Aggregate into a vector matrix of dimension d2×n to obtain the molecular graph structure feature F Graph :
[0056]
[0057] Furthermore, the multimodal feature fusion and splicing process includes:
[0058] The features of each modality are concatenated to form an initial comprehensive feature representation:
[0059] F concat =[F SMILES ||F Protein ||F Graph ]
[0060] in, Represents SMILES string features, Indicates the amino acid sequence characteristics of proteins, represents the molecular graph structural features, || represents the vector splicing operation, d1, d2, d3, and n are the dimensions of the above three features, and the initial comprehensive feature after splicing is expressed as F concat ∈R m×n , m=d1+d2+d3.
[0061] Furthermore, the dynamic attention block of the memory module is applied to process the initial comprehensive feature representation, including:
[0062] Step 1: Calculate the attention weight α of the initial comprehensive feature representation at the current time step t t , specifically including:
[0063] Calculate the attention score of the initial comprehensive feature representation at the current time step t:
[0064]
[0065] Among them, e t represents the attention score of the feature vector at the current time step t, Represents the weight matrix used to calculate the attention score, F concat,t represents the feature vector at the current time step t in the comprehensive feature representation, that is, the t-th feature vector of the comprehensive feature representation, b a is the bias term.
[0066] Use the Softmax function to convert the attention score of the current time step t into the attention weight of the current time step t:
[0067]
[0068] Among them, α t represents the attention weight of the feature vector at the current time step t, and m is the number of feature vectors represented by the initial comprehensive feature.
[0069] Step 2: According to the attention weight α at the current time step t t Update the memory matrix M of the previous time step t-1 t-1 , specifically including:
[0070] The memory matrix M records the attention weight distribution of the last T historical time steps, which is expressed as follows:
[0071] M∈R T×n
[0072] Here, n represents the dimension of the attention weight vector at each time step.
[0073] If the memory matrix M of the previous time step t-1 The attention weights of the last T time steps have been stored, so in the memory matrix M t-1 Remove the attention weight of the earliest time step and add the attention weight α of the current time step t , the formula is as follows:
[0074] M t =[α t ||M t-1,1:T-1 ]
[0075] Otherwise, directly in the memory matrix M t-1 Add the attention weight α of the current time step t Among them, M t represents the memory matrix of the current time step t, α trepresents the attention weight of the current time step, M t-1,1:T-1 Represents the result of removing the earliest record from the memory matrix of the previous time step.
[0076] Step 3: According to the attention weight α of the current time step t and memory matrix M t Calculate the final attention weight at time step t, adjust the initial comprehensive feature representation according to the final attention weight at time step t, and obtain the final comprehensive feature;
[0077] According to the current attention weight and the historical information in the memory matrix, the representation of each feature vector is dynamically adjusted. The specific steps are as follows:
[0078] The attention weight of the current time step is weighted averaged with the historical weight in the memory matrix to form the final attention weight. The specific formula is as follows:
[0079]
[0080] Among them, α t,final represents the final attention weight at time step t, γ represents the parameter that controls the ratio of the current attention weight to the historical weight, and its value range is [0,1], α t represents the attention weight of the current time step, M t,k represents the attention weight of the k-th historical time step in the memory matrix.
[0081] Use the final attention weights to adjust the initial comprehensive feature representation:
[0082]
[0083] Among them, F Attention represents the final comprehensive feature representation, α t,final represents the final attention weight of the feature vector at time step t, F Concat,t represents the t-th eigenvector in the comprehensive feature representation.
[0084] Furthermore, the prediction module includes: KAN neural network and multi-layer perceptron; the prediction module processes the final comprehensive features including:
[0085] Step 1: Input the final comprehensive feature representation into the KAN neural network;
[0086] KAN (Knowledge-Aware Neural Network) optimizes the model's feature extraction capabilities by introducing domain knowledge:
[0087] H KAN =KAN(F Attention )
[0088] Among them, H KAN To represent the features extracted by the KAN neural network, the structure of the KAN neural network may include multiple layers, such as graph convolution layers, graph attention layers, etc., to better capture the relationship between features.
[0089] Step 2: Input the output of the KAN neural network into the multilayer perceptron;
[0090] Multi-layer perceptron MLP is used to further process the feature representation H extracted by KAN KAN , and finally get the prediction results.
[0091] MLP consists of several fully connected layers and activation functions. The output of each layer can be expressed as:
[0092]
[0093] in, represents the output of the l-th layer MLP, W (l) represents the weight matrix of the lth layer, b (l) represents the bias term of the lth layer, and σ represents the activation function.
[0094] The final output layer of the MLP is
[0095]
[0096] Among them, O MLP Represents the final output of MLP, Y pred is the predicted drug-protein binding affinity value, W out represents the weight matrix of the output layer, b out represents the bias term of the output layer, and L represents the number of layers of the MLP.
[0097] Furthermore, the loss function L of the affinity prediction model MSE for;
[0098]
[0099] Among them, y i is the true drug-protein affinity value of the i-th sample, is the affinity value predicted by the model, and N is the number of samples.
[0100] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation plans of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for predicting drug-protein binding affinity, characterized in that: include: Obtain multimodal drug-protein data, input the multimodal drug-protein data into a trained affinity prediction model, and obtain a predicted value of drug-protein affinity; The affinity prediction model includes: a feature extraction module, a dynamic attention module using a memory module, and a prediction module; the training process of the affinity prediction model includes: S1: obtaining multimodal drug-protein data, inputting the multimodal drug-protein data into a feature extraction module, and obtaining multimodal drug-protein features; S2: splicing multimodal drug-protein features to obtain the initial comprehensive feature representation; S3: Input the initial comprehensive feature representation into the dynamic attention module of the application memory module to obtain the final comprehensive feature; S4: Input the final comprehensive features into the prediction module to obtain the prediction results; S5: Calculate the loss function value according to the prediction result, and update the model parameters according to the loss function value. When the loss function value is the minimum, the trained affinity prediction model is obtained.
2. A method for predicting drug-protein binding affinity according to claim 1, characterized in that: The multimodal drug-protein data includes: a SMILES string of a drug, a molecular graph structure of a drug, and an amino acid sequence of a protein.
3. A method for predicting drug-protein binding affinity according to claim 2, characterized in that: The feature extraction modules include: SMILES feature extraction module, molecular graph structure feature extraction module and amino acid sequence feature extraction module.
4. A method for predicting drug-protein binding affinity according to claim 3, characterized in that: The SMILES feature extraction module includes: an embedding layer, a one-dimensional convolutional neural network and an LSTM network; the SMILES feature extraction module processes the SMILES string of the drug including: converting the SMILES string into an integer code; inputting the integer-coded SMILES string into the embedding layer to obtain an embedding matrix; inputting the embedding matrix into the one-dimensional convolutional neural network to obtain local features; and inputting the local features into the LSTM network to obtain SMILES string features.
5. A method for predicting drug-protein binding affinity according to claim 3, characterized in that: The molecular graph structure feature extraction module processes the molecular graph structure of the drug, including: representing the molecular graph structure of the drug as a molecular undirected graph G = (V, E), where V is a node set, the nodes represent the atoms of the molecular graph structure, and E is an edge set, and the edges represent the chemical bonds of the molecular graph structure; initializing the feature vectors of each node and each edge; using a graph convolutional network to process the feature vectors of the nodes and edges in the molecular undirected graph G to obtain the molecular graph structure features.
6. A method for predicting drug-protein binding affinity according to claim 3, characterized in that: The amino acid sequence feature extraction module includes: an embedding layer, a one-dimensional convolutional neural network and an LSTM network; the amino acid sequence feature extraction module processes the amino acid sequence of the protein including: converting the amino acid sequence of the protein into an integer code; inputting the integer-coded amino acid sequence into the embedding layer to obtain an embedding matrix; inputting the embedding matrix into the one-dimensional convolutional neural network to obtain local features; and inputting the local features into the LSTM network to obtain amino acid sequence features.
7. A method for predicting drug-protein binding affinity according to claim 1, characterized in that: A dynamic attention module applying a memory module includes a memory matrix, wherein the memory matrix includes attention weights of multiple historical time steps; The dynamic attention module of the application memory module processes the initial comprehensive feature representation including: Calculate the initial attention weight λ of the initial comprehensive feature representation at time step t t ; According to the initial attention weight α t Update the memory matrix M at time step t-1 t-1 , get the memory matrix M at time step t t , according to the initial attention weight α t and memory matrix M t Calculate the final attention weight α at time step t t,final , according to the attention weight α t,final Adjust the initial comprehensive feature representation to obtain the final comprehensive feature.
8. A method for predicting drug-protein binding affinity according to claim 7, characterized in that: Update the memory matrix M t-1 Including: If the memory matrix M t-1 The attention weights of T time steps have been stored, so in the memory matrix M t-1 Remove the attention weight of the earliest time step and add the attention weight α t ; Otherwise, directly in the memory matrix M t-1 Add attention weight α t ; Where T is the maximum dimension of the memory matrix.
9. A method for predicting drug-protein binding affinity according to claim 7, characterized in that: Calculating the final attention weight at time step t includes: calculating the memory matrix M t The average of the attention weights in the historical time steps, the average and the attention weight α t Perform weighted averaging to obtain the final attention weight at time step t.
10. A method for predicting drug-protein binding affinity according to claim 1, characterized in that: The prediction module includes: a KAN neural network and a multi-layer perceptron; the prediction module processes the final comprehensive features by: inputting the final comprehensive feature representation into the KAN neural network, inputting the output of the KAN neural network into the multi-layer perceptron, and obtaining a prediction result.
Citation Information
Patent Citations
Drug-target intelligent recommendation method based on graph representation learning
CN117423378A
DTI prediction method based on hierarchical multi-modal self-attention graph neural network
CN118173161A
Drug-target affinity prediction method based on patch coding and mixed attention
CN118447916A
System and method for molecular reconstruction from molecular probability distributions
US20220198286A1
System and method for predicting biological activity of chemical or biological molecules and evidence thereof
US20230326545A1
Cited By
Glass hardness prediction method based on hierarchical prediction neural network
CN119314585A
Tunneling machine cutterhead torque calculation method, device and equipment and storage medium
CN119918383A
BBB repairing agent evaluation method and system based on ischemia reperfusion injury model
CN121366715A
