Prediction of drug-target binding affinity based on time sequence feature fusion of unified message

Abstract

The invention provides a method for predicting and classifying drug-target binding affinity based on time sequence feature fusion of a unified message, and belongs to the field of medical prediction. The method comprises the following steps: converting an SMILES sequence of a drug into a subgraph form through RDKit, and expressing a protein sequence by one-hot coding; independently learning the expression of a drug molecular map and a protein sequence through a specific feature extraction module; after feature learning is completed, the model adopts a TFAM fusion network, information of two different features can be comprehensively considered, and the learned features are fused respectively; finally, the affinity value of the output drug target is estimated through several completely connected layers, and the architecture design of the whole model is as shown in Figure 1. Experiments carried out on two common data sets (Davis and KIBA) show that the method provided by the invention has better performance than a previous drug target affinity prediction method.

Description

TECHNICAL FIELD

[0001] The application provides a drug-target binding affinity prediction classification method based on the fusion of timing characteristics of unified messages, and belongs to the field of medical prediction. BACKGROUND

[0002] Drug development is an expensive and time-consuming process. It is estimated that it takes an average of $2.6 to $2.8 billion to develop a new drug, and it can take up to 17 years to pass FDA approval. At present, about 70 FDA-approved drugs are being studied to explore whether they can be repurposed to treat COVID-19. In the field of drug discovery and drug repurposing, predicting drug-target affinity (DTA) has become a key issue. The identification of drug-target protein interactions is crucial for drug discovery and is a hot topic of current research. However, the traditional method of confirming drug target affinity based on biological hypotheses and trial-and-error is inefficient and resource-intensive. With the advancement of computational methods and technologies, especially the application of machine learning in the field of chemical and biological research, computer-aided methods have great potential for drug discovery. These methods can significantly reduce the search space and shorten the drug discovery process. To address this challenge, researchers have developed computational methods based on protein amino acid sequences and compound SMILES, which aim to quickly and cost-effectively predict DTA. Using these computational methods, DTA prediction has become easier and more effective, helping to quickly identify potential drug-target interactions and thus accelerating the drug discovery process. This approach not only overcomes the scarcity of structural information, but also reduces the dependence on domain expert knowledge.

[0003] In the process of building a deep learning model, key steps include data preprocessing, drug feature extraction, protein feature extraction, and prediction. Although existing technologies have made some achievements in drug-target affinity prediction, there are still many challenges. For example, the interaction between features may be ignored in the feature fusion process, and there are deficiencies in the extraction of compound and target features. These problems need to be addressed through further research and technological innovation. SUMMARY

[0004] The purpose of the present application is to solve the above problems of the prior art, and propose a method for predicting drug-target binding affinity based on unified message timing feature fusion (TUMP-DTA). The features of the drug are extracted by the unified message passing model (UniMP) to solve the problem of discarding observed labels. Then, in order to enrich the representation of proteins, we use multi-dimensional protein sequences as input, embed and integer encode them, and use the effective channel attention (ECA) of the deep convolutional neural network to more effectively improve the protein features. Finally, a dual-phase feature fusion network (TFAM) is proposed, which enhances dual-phase features through channel and spatial attention, and is particularly suitable for change detection. Through sufficient experiments, it is proved that the prediction performance of the present application is better than the most advanced drug-target affinity prediction method, and has better generalization performance.

[0005] A method for predicting drug-target binding affinity based on unified message timing feature fusion, as shown in the overall structure diagram Figure 1 , comprising the following steps:

[0006] S1. The drug molecule sequence to be processed is embedded and integer coded and converted into a molecular graph using RDKit;

[0007] S2. The drug molecule graph obtained in S1 is used for feature extraction using the unified message passing model;

[0008] S3. The protein sequence to be processed is used as input, embedded and integer coded;

[0009] S4. The protein sequence coding obtained in S3 is used for protein feature extraction using the effective channel (ECA) of the deep convolutional neural network;

[0010] S5. The drug features and protein features obtained in S2 and S3 are integrated through a dual-phase feature fusion network;

[0011] S6. The features learned in S5 are estimated through several fully connected layers to output drug-target affinity values.

[0012] Preferably, the process of converting the drug molecule sequence into a molecular graph in S1 is as follows: in the molecular graph, the information on the node includes: atom symbol, number of adjacent atoms, number of adjacent hydrogen atoms, atom hidden value, and whether the atom is located in an aromatic ring structure. The 2D undirected graph of the drug can be described as G D =(V D ,E D ), where V D is a set of atom nodes represented by a feature vector as an atom node set, E Dis the edge set represented by the eigenvector. Here each atom is regarded as a node, and the edge between nodes represents the chemical bond relationship between atoms. Wherein is the i-th node, is the edge between node i and node j, A D ∈R N×N represents the adjacency matrix, when , it represents that there is an edge between node i and node j, otherwise, it represents that there is no chemical bond connection between the two nodes.

[0013] Preferably, the feature extraction step of S2 is: first, UniMP propagates features and labels simultaneously through the Graph Transformer network, which allows the model to utilize node label information during both training and inference, thereby improving classification accuracy. Second, UniMP introduces a masked label prediction method, which randomly masks a portion of the labels during training to force the model to learn how to make predictions from unmasked labels. This method helps to reduce overfitting to training data and improves the model's generalization ability to unseen data. Third, through the masked label prediction strategy, UniMP avoids the label leakage problem during training, that is, the model does not rely too much on label information during the training phase, but learns to infer the complete label from partial information. In addition, since UniMP combines the advantages of GNN and LPA, it can capture the complex relationships and patterns between nodes in the graph and provide richer node representations, which is an attribute that other GNN models do not have. GNN constructs a mapping multi-layer transformation and propagates node features X, which is expressed as follows:

[0014]

[0015] In the formula: σ is the activation function, W l is the trainable weight in the l-th layer, HL is the node representation layer in the l-th layer, H L is equal to the input feature X, and fout is the output layer of the final prediction Y.

[0016] Next, we introduce the label propagation algorithm (LPA) to iteratively propagate label information in the graph, please perform the following operations:

[0017]

[0018] In the formula: wherein is the initial label matrix, which is one-hot hot encoding vector for labeled nodes and zero vector for unlabeled nodes. To realize the simultaneous propagation of features and labels, we use the UniMP model. UniMP adopts the Graph Transformer structure combined with multi-head attention mechanism to process node features and label embeddings:

[0019]

[0020] where, are query and key vectors, e c,ij is edge feature, is attention weight, and represents the combination of multi-head attention mechanism. To further fuse feature and label information, we introduce a label embedding technique to convert partially observed labels from one-hot encoding to dense vectors similar to node features:

[0021]

[0022] where is label embedding vector, is the combination of node features and label embeddings as propagation information. To avoid the leakage of label information during training, we adopt the mask label prediction strategy. During training, we randomly mask a part of the labels and try to predict these masked labels:

[0023]

[0024] where is the partially masked label matrix, is the masked label, is the set of nodes with masked labels.

[0025] Preferably, S3 said protein sequence to be processed as input, embedding and integer encoding process: for protein sequence, we use one-hot encoding to represent protein, one-hot encoding is to use Nbit register to encode N states. The mathematical expression of a protein is:

[0026] p={p1,p2,…,p i …},p i ∈N P (6)

[0027] where p i represents the i-th amino acid, N P represents the amino acid set containing 25 common amino acids. Each type of amino acid is represented as an integer according to its related letter symbol.

[0028] In addition, we define the maximum length of protein sequence as 1000. Then we map the protein integer sequence to 128-dimensional vectors through the embedding layer:

[0029] X p = Embedding(P) (7)

[0030] Preferably, the position S4 uses the effective channel (ECA) of the deep convolutional neural network to perform the protein feature extraction step on the protein sequence encoding obtained by S3, which includes using ECA to perform protein sequence representation learning, which is widely used in feature extraction for image classification tasks. Specifically, after embedding the protein sequence for representation, first perform initial feature extraction through a convolutional module, then use an ECA module to extract global and local features respectively, and finally obtain a protein representation vector. The module produces channel attention through fast one-dimensional convolution, and the kernel size can be adaptively determined through nonlinear mapping of the channel dimension. This method can enhance the effectiveness of various deep CNN configurations. Starting from the convolutional layer:

[0031]

[0032] The core idea of the ECA module is to capture the dependency between channels through one-dimensional convolution. Compared with traditional attention mechanisms, the ECA module avoids complex dimension reduction and dimension elevation processes, thereby achieving high efficiency and lightweight characteristics. The ECA module first adaptively calculates the kernel size k of one-dimensional convolution according to the number of channels. The formula for calculating the kernel size k is as follows:

[0033]

[0034] The formula is used to calculate the kernel size k of one-dimensional convolution, where C represents the number of input feature channels, and b is a hyperparameter. The absolute value is evaluated and rounded to the nearest odd number to keep the kernel size odd. After obtaining the kernel size k, the ECA module performs one-dimensional convolution on the input features, thereby learning the importance of each channel relative to other channels. This method can be represented by the following formula:

[0035]

[0036] This formula means that the input feature in is converted to the output feature out through one-dimensional convolution operation (kernel size k). Conv1D k represents one-dimensional convolution operation with kernel size k.

[0037] Preferably, the step of integrating the drug features and protein features obtained from S2 and S3 through the dual temporal feature fusion network in S5 is as follows: the network identifies the key parts of the features through channel and spatial attention mechanisms, and highlights the important interactions between the dual temporal portal features using temporal information. In the channel branch, the input dual temporal features are passed through a global pooling operation across the spatial dimension to aggregate the spatial information. The aggregation process can be described as follows:

[0038]

[0039] where S c represents the aggregated spatial features, Avg(·) and Max(·) represent the average pooling and maximum pooling across the spatial dimension, respectively. The aggregated spatial features are passed to two 1-D convolutions (same as in the ECA module) to determine the dual temporal channel weights of the input dual temporal features. The two channel weights can be represented as:

[0040] W c1 ,W c2 = Conv1(S c ), Conv2(S c ) (12)

[0041] where Wc1 and Wc2 represent the dual temporal channel weights, and Conv1(·) and Conv2(·) represent the one-dimensional convolution operation. Subsequently, a Softmax function is applied to the dual temporal channel weights to ensure that their sum equals unity. This comparison determines which of the dual temporal channel weights is more important, highlighting the more critical aspects between the dual temporal features along the channel axis. The formulation of Softmax is as follows:

[0042]

[0043] Wc1 and Wc2 represent the output dual temporal channel weights. Finally, the dual temporal weights are multiplied by the dual temporal features and merged, effectively integrating the dual temporal features. The resulting output can be represented as:

[0044]

[0045] where Output represents the fused features. Since the sum of the dual temporal scale weights equals 1, the useful parts between the dual temporal stage features are retained and the useless parts are discarded, thereby achieving effective feature fusion.

[0046] Preferably, the step of estimating the output drug-target affinity value of the features learned by S5 through several fully connected layers in S6 is as follows: The key of this layer is to be able to integrate and refine the feature representation generated by the pre-sequence layer to output accurate prediction results. One of the significant advantages of the fully connected layer is that it can reduce the sensitivity of feature position to classification decision, thereby enhancing the robustness of the entire network. In the last stage of the model, a structure composed of three fully connected networks is adopted, which is specially designed to accurately output the predicted affinity score at the end of the network. This design not only ensures the sufficient integration of information, but also provides the model with the ability to generate highly accurate predictions. The network finally outputs the value of the binding affinity:

[0047] Y = FNN (Output) (15) BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 is the technical flowchart of the present application

[0049] Figure 2 is the overall structure diagram of the TFUM-DTA of the present application

[0050] Figure 3 is the drug molecule construction diagram of the present application

[0051] Figure 4 is the UniMP structure diagram of the present application

[0052] Figure 5 is the TFAM architecture structure diagram of the present application DETAILED DESCRIPTION

[0053] In order to make the purpose, technical scheme and advantages of the present application more clear, the technical scheme in the present application will be described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0054] A unified message-based time sequence feature fusion method for drug-target binding affinity prediction classification, comprising the following steps:

[0055] S1. The sequence of the drug molecule to be processed is embedded and integer coded and converted into a molecular graph using RDKit;

[0056] S2. The drug molecule graph obtained in S1 is used for feature extraction using a unified message passing model;

[0057] S3. The protein sequence to be processed is embedded and integer coded as input;

[0058] S4. The protein sequence obtained in S3 is encoded and protein features are extracted using the efficient channel (ECA) of a deep convolutional neural network;

[0059] S5. Integrate the drug and protein features obtained from S2 and S3 through a dual-temporal feature fusion network;

[0060] S6. The features learned in S5 are used to estimate the output drug-target affinity value through several fully connected layers.

[0061] Dataset preparation:

[0062] The experiments in this invention were conducted using two publicly available datasets. The Davis dataset tested the interactions between 72 kinase inhibitors and 442 kinases, expressed in ρKk. d The dissociation constant (K) in the logarithmic space domain d The binding affinity is used as the output value. KIBA data includes binding affinity for 2116 kinase inhibitors and 229 kinases, which is the biological activity of kinase inhibitors measured by the KIBA method, which takes into account different inhibitor potency indicators, such as K+. i ,K d And IC50.

[0063] Parameters and Experiments:

[0064] The experimental model parameters of this invention are updated using the Adam optimization algorithm, with the learning rate set to 0.0005 and each training batch containing 512 samples. To find the optimal model configuration, we performed five-fold cross-validation on the Davis dataset to tune the hyperparameters. Once we determined the optimal hyperparameter settings, we applied them to all other datasets without changing them. These experiments were conducted in a Linux operating system environment running Python 3.6.2.

[0065] The TFUM-DTA network of this invention was tested and evaluated together with various classification networks on the Davis and KIBA datasets. The final results are shown in Tables 1-3.

[0066] As shown in Table 1, the TFUM-DTA network proposed in this invention improves the CI by 1.2% on the Davis dataset compared to the best baseline model. It increased by 0.2%.

[0067] Table 1. Results of different methods on the Davis dataset.

[0068]

[0069] As can be seen from Table 2, compared with the best result of the baseline model, the MSE of TFUM-DTA is reduced by 0.2% on the KIBA dataset, improved by 0.3%.

[0070] Table 2 results of different methods on Davis dataset

[0071]

[0072] Table 3 ablation experiment results on Davis

[0073]

[0074] A series of ablation experiments were conducted in this study to verify the effectiveness of the UniMP, ECA-Net and TFAM modules in the drug target affinity prediction task. The experimental results are shown in Table 3, and the introduction of the UniMP module significantly improves the AUC and precision of the model.

[0075] The comparative experiment results from the two datasets show that the TFUM-DTA network proposed in the present application has better performance in precision, and is superior to several other classification networks. The robustness of the model is improved, thereby improving the overall performance of the model in drug target affinity prediction.

[0076] The above examples are only used to illustrate the technical solutions of the present application, and not to limit it. Although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can still be modified, or some or all of the technical features can be replaced by equivalents, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A unified message-based temporal feature fusion for prediction of drug-target binding affinity, characterized in that, The prediction network comprises: S1. Embedding and integer encoding the sequence of the drug molecule to be processed and converting it into a molecular graph using RDKit; S2. Extracting features of the drug molecular graph obtained in S1 using a unified message passing model; S3. Embedding and integer encoding the protein sequence to be processed as input; S4. Extracting protein features using the effective channel (ECA) of a deep convolutional neural network for the protein sequence encoding obtained in S3; S5. Integrating the drug features and protein features obtained in S2 and S3 through a dual-phase feature fusion network; S6. Estimating the output drug-target affinity value through several fully connected layers for the features learned in S5. 2.The unified message-based timing feature fusion method for predicting drug-target binding affinity according to claim 1, wherein, The drug molecule sequence to be processed is embedded and integer coded as described in S1 and converted into a molecular graph using RDKit as follows: In the molecular graph, the information on the node includes: atomic symbol, number of adjacent atoms, number of adjacent hydrogen atoms, atomic implicit value, and whether the atom is located in an aromatic ring structure, and the 2D undirected graph of the drug can be described as G D =(V D ,E D ), where V D is the atomic node set represented by the feature vector as the atomic node set, E D is the edge set represented by the feature vector as the edge set. Here, each atom is regarded as a node, and the edge between the nodes represents the chemical bond relationship between the atoms. Where V i D ∈V D is the ith node, is the edge between node i and node j, A D ∈R N×N represents the adjacency matrix, when , it means that there is an edge between node i and node j, otherwise, it means that there is no chemical bond connection between the two nodes. 3.The unified message-based timing feature fusion method for predicting drug-target binding affinity according to claim 1, wherein, The step S2 of extracting features of the drug molecular graph obtained in S1 using a unified message passing model comprises: firstly, UniMP simultaneously propagates features and labels through a Graph Transformer network, which allows the model to utilize label information during both training and inference, thereby improving the accuracy of classification. Secondly, UniMP introduces a masked label prediction method, which forces the model to learn how to make predictions from unmasked labels by randomly masking a portion of the labels during training. This method helps to reduce overfitting to training data and improves the model's generalization ability to unseen data. Thirdly, through the masked label prediction strategy, UniMP avoids the label leakage problem during training, i.e., the model does not excessively rely on label information during the training phase, but learns to infer the complete label from partial information. The calculation formula is as follows: wherein, is a query and key vector, e c,ij is an edge feature, is an attention weight, represents a merge of multi-head attention mechanisms. 4.The unified message-based timing feature fusion method for predicting drug-target binding affinity according to claim 1, wherein, The step S3 of embedding and integer encoding the protein sequence to be processed as input comprises: for the protein sequence, we use one-hot encoding to represent the protein, which is to use an Nbit register to encode N states. The mathematical expression of a protein is: p = {p1, p2,..., p i ...}, p i ∈ N P (6) where p i represents the i-th amino acid, N P represents the set of 25 common amino acids. Each amino acid type is represented as an integer according to its associated letter symbol. Furthermore, we define the maximum length of a protein sequence to be 1000. We then map the protein integer sequences to 128-dimensional vectors through an embedding layer: X p = Embedding(P) 5.The unified message-based timing feature fusion method for predicting drug-target binding affinity according to claim 1, wherein, The step S4 of extracting protein features using the effective channel (ECA) of a deep convolutional neural network for the protein sequence encoding obtained in S3 comprises: using ECA for protein sequence feature learning, which is widely used in image classification tasks. Specifically, after embedding the protein sequence for representation, first extract initial features through a convolution module, then use the ECA module to extract global and local features respectively, and finally obtain the protein representation vector. The module produces channel attention through fast one-dimensional convolution, and the kernel size can be adaptively determined through nonlinear mapping of the channel dimension. This method can enhance the effectiveness of various deep CNN configurations. The calculation formula is as follows: The core idea of the ECA module is to capture the dependency between channels through one-dimensional convolution. Compared with traditional attention mechanisms, the ECA module avoids complex dimension reduction and dimension increase processes, thereby achieving high efficiency and lightweight characteristics. The ECA module first adaptively calculates the kernel size k of one-dimensional convolution according to the number of channels. The formula for calculating the kernel size is as follows: The formula is used to calculate the kernel size k of one-dimensional convolution, where C represents the number of input feature channels, and b is a hyperparameter. The absolute value is evaluated and rounded to the nearest odd number to keep the kernel size odd. After obtaining the kernel size k, the ECA module performs one-dimensional convolution on the input features, thereby learning the importance of each channel relative to other channels. This approach can be represented by the following formula: This formula means that the input feature in is converted to the output feature out by a one-dimensional convolution operation (kernel size k), Conv1D k represents a one-dimensional convolution operation with kernel size k. 6.The unified message-based timing feature fusion method for predicting drug-target binding affinity according to claim 1, wherein, The step of integrating the drug features and protein features obtained from S2 and S3 through the dual-time feature fusion network in S5 is as follows: The network identifies the key parts of the features through channel and spatial attention mechanisms and highlights the important interactions between the dual-time portal features using temporal information. In the channel branch, the dual-time features are passed through a global pooling operation across the spatial dimension to aggregate spatial information. The aggregation process can be described as follows: where S c denotes the aggregated spatial features, Avg(·) and Max(·) denote average and max pooling across spatial dimensions, respectively. The aggregated spatial features are passed to two 1-D convolutions (same as in the ECA module) to determine the dual-time channel weights for the input dual-time features. The two channel weights can be represented as: W c1 ,W c2 = Conv1(S c ), Conv2(S c ) where Wc1 and Wc2 represent the dual-time channel weights, and Conv1(·) and Conv2(·) represent one-dimensional convolution operations. Subsequently, a Softmax function is applied to the dual-time channel weights to ensure that their sum equals unity. This comparison determines which of the dual-time channel weights is more important, highlighting the more critical aspects between the dual-time features along the channel axis. The formula for Softmax is as follows: Wc1 and Wc2 represent the output dual-time channel weights. Finally, the dual-time weights are multiplied by the dual-time features and combined, effectively integrating the dual-time features. The resulting output can be represented as: where Output represents the fused features. Since the sum of the dual-time scale weights equals 1, the useful parts between the dual-time scale features are retained, and the useless parts are discarded, thereby achieving effective feature fusion. 7.The unified message-based timing feature fusion method for predicting drug-target binding affinity according to claim 1, wherein, The step of estimating the output drug-target affinity value by passing the features learned in S5 through several fully connected layers in S6 is as follows: The key to this layer is to effectively integrate and refine the feature representations generated by the pre-sequence layer to output accurate prediction results. One significant advantage of fully connected layers is that they can reduce the sensitivity of feature positions to classification decisions, thereby enhancing the robustness of the entire network. In the final stage of the model, a structure composed of three fully connected layers is adopted, which is specifically designed to accurately output the predicted affinity score at the end of the network. This design not only ensures sufficient integration of information but also provides the model with the ability to generate highly accurate predictions. The network finally outputs the value of the binding affinity: Y = FNN (Output).

Images