Method and system for predicting drug-target interaction based on deep learning

Abstract

The invention relates to a drug-target interaction prediction method and system based on deep learning. The method comprises the following steps: determining an SMILES sequence of a drug molecule to be predicted and an amino acid sequence of protein; generating a drug molecular map, inputting the drug molecular map to a GPS layer, and extracting the feature representation of the drug by using GINE and multiple attention mechanisms; extracting the feature representation of the protein through an ESM-2 model; and performing cross attention mechanism feature fusion on the feature representation of the drug and the feature representation of the protein to obtain a fused feature, and predicting the binding probability of the drug and the protein based on the fused feature. The feature representation of the medicine is extracted by combining a GINE layer and a multi-head attention mechanism, and the feature representation of the protein is extracted by using the pre-trained ESM-2 model, so that the comprehensive extraction of the features can be realized, and the accuracy of the feature representation is improved; drug and protein features are effectively fused through a cross attention mechanism, and efficient and accurate prediction of drug-target interaction is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of drug target prediction, and in particular to a method and system for predicting drug-target interaction based on deep learning. Background Art

[0002] Drug-target interaction (DTI) prediction plays a key role in drug discovery. Although traditional in vitro experiments are reliable, their high cost and time-consuming development cycle limit their widespread use in large-scale data applications. In contrast, using computer methods to identify potential DTIs can effectively narrow the search range of candidate compounds, accelerate the drug discovery process and reduce development costs. In recent years, computer-assisted DTI prediction methods have received widespread attention and continued progress, bringing new research prospects and opportunities to the biomedical field.

[0003] In the field of drug-target interaction (DTI) prediction, computational methods are generally divided into three categories: ligand-based methods, structure-based methods, and chemical genomics-based methods. Ligand-based methods: This type of method relies on the similarity between drugs and known target protein ligands to predict their interactions. By comparing the similarity between new drugs and known ligands, their possible targets are inferred. This type of method is simple and easy to use and is suitable for situations where there are a large number of known ligands. Structure-based methods: This method uses the three-dimensional structural information of drugs and target proteins to simulate their interaction trends. Through molecular docking, molecular dynamics simulation and other means, the binding mode and affinity of drugs and targets are predicted. Chemical genomics-based methods: This method comprehensively utilizes multiple information of drugs and targets (such as genome sequences and drug chemical structures), usually combined with machine learning and deep learning techniques, to predict their interactions. It can learn complex relationship patterns from large-scale data and improve the accuracy and efficiency of predictions.

[0004] Although existing DTI prediction technologies have achieved certain success, they still have some significant limitations. Ligand-based methods rely on the number of known ligands of the target protein, and the similarity assumption is not always accurate; structure-based methods rely on the exact three-dimensional structure of the target protein and have high computational complexity; chemical genomics-based methods have model complexity, and traditional machine learning models may not be able to effectively learn complex concepts when processing large-scale data. The limitation of feature extraction is that many traditional models cannot fully utilize the potential structural information of the data when extracting drug and protein features, resulting in limited prediction performance. Summary of the invention

[0005] Based on this, in order to solve the above technical problems, a deep learning-based drug-target interaction prediction method and system are provided, which can achieve efficient and accurate prediction of drug-target interaction.

[0006] A method for predicting drug-target interaction based on deep learning, the method comprising:

[0007] Determine the SMILES sequence of the drug molecule to be predicted and the amino acid sequence of the protein;

[0008] Determine structural information of the drug analysis based on the SMILES sequence of the drug analysis, and generate a drug molecule graph according to the structural information;

[0009] The drug molecule graph is input into the GPS layer, and the feature representation of the drug is extracted using GINE and multi-head attention mechanism;

[0010] Inputting the amino acid sequence of the protein into the ESM-2 model to extract the characteristic representation of the protein;

[0011] The feature representation of the drug and the feature representation of the protein are cross-attentionally fused to obtain fused features, and the binding probability of the drug and the protein is predicted based on the fused features.

[0012] In one embodiment, the drug molecule graph is input into the GPS layer, and the feature representation of the drug is extracted using GINE and a multi-head attention mechanism, including:

[0013] Acquire node and edge features based on the drug molecule graph;

[0014] Inputting the drug molecule graph into the GPS layer to obtain an adjacency matrix of the drug molecule graph;

[0015] GINE is used to extract features based on the node and edge features and the adjacency matrix; and a multi-head attention mechanism is used to extract features based on the nodes.

[0016] In one embodiment, the amino acid sequence of the protein is input into the ESM-2 model to extract the characteristic representation of the protein, including:

[0017] Input the amino acid sequence of the protein into the ESM-2 model to obtain a high-level embedding representation of the protein sequence;

[0018] Based on the ESM-2 model, CNN is used to process the high-level embedding representation to extract the feature representation of the protein.

[0019] In one embodiment, based on the ESM-2 model, CNN is used to process the high-level embedding representation to extract the characteristic representation of the protein, including:

[0020] Based on the ESM-2 model, obtaining a protein embedding vector according to the high-level embedding representation;

[0021] The protein embedding vector is input into CNN and the feature representation of the protein is output.

[0022] In one embodiment, the feature representation of the drug and the feature representation of the protein are cross-attentionally fused to obtain fused features, including:

[0023] Based on the characteristic representation of the drug, determining the number of nodes and the embedding dimension of the drug characteristic graph;

[0024] Linearly transforming the characteristic representation of the drug, obtaining a drug query vector according to the number of nodes and the embedding dimension of the drug characteristic graph, and obtaining a drug key vector and a drug value vector based on the characteristic representation of the protein;

[0025] Inputting the feature representation of the protein into a linear layer, calculating a protein query vector, and obtaining a protein key vector and a protein value vector based on the feature representation of the drug;

[0026] Based on the drug query vector, drug key vector, drug value vector, protein query vector, protein key vector, and protein value vector, feature fusion is performed to obtain fused features.

[0027] In one embodiment, feature fusion is performed based on the drug query vector, drug key vector, drug value vector, protein query vector, protein key vector, and protein value vector to obtain fused features. The fused features include:

[0028] Calculating a drug attention matrix based on the drug query vector, the drug key vector, and the drug value vector; and calculating a protein attention matrix based on the protein query vector, the protein key vector, and the protein value vector;

[0029] Feature fusion is performed based on the attention matrix of the drug and the attention matrix of the protein to obtain fused features.

[0030] In one embodiment, feature fusion is performed based on the attention matrix of the drug and the attention matrix of the protein to obtain fused features, including:

[0031] Calculating a drug feature map based on the drug attention matrix, and calculating a protein feature map based on the protein attention matrix;

[0032] Connecting the drug feature map and the protein feature map in the channel dimension, and obtaining a target drug feature map and a target protein feature map through a linear layer;

[0033] The target drug feature map and the target protein feature map are downsampled by a global maximum pooling operation to obtain a joint feature representation as a fused feature.

[0034] In one embodiment, the method further comprises:

[0035] Inputting the joint feature representation into a deep learning-based classifier and mapping it to an output space through a fully connected layer in the classifier;

[0036] The AdamW optimizer is used to minimize the cross entropy loss function, and parameter optimization is performed based on the optimized cross entropy loss function.

[0037] In one embodiment, predicting the probability of drug-protein binding based on the fused features includes:

[0038] The fused features are passed through a fully connected network to predict the binding probability of the drug and the protein.

[0039] A deep learning-based drug-target interaction prediction system, the system comprising:

[0040] A sequence determination module is used to determine the SMILES sequence of the drug molecule to be predicted and the amino acid sequence of the protein;

[0041] A drug molecule graph determination module, used to determine the structural information of the drug analysis based on the SMILES sequence of the drug analysis, and generate a drug molecule graph according to the structural information;

[0042] A drug feature extraction module, used to input the drug molecular graph into the GPS layer, and extract the feature representation of the drug using GINE and multi-head attention mechanism;

[0043] A protein feature extraction module, used to input the amino acid sequence of the protein into the ESM-2 model to extract the feature representation of the protein;

[0044] The feature fusion and prediction module is used to perform cross-attention mechanism feature fusion on the feature representation of the drug and the feature representation of the protein to obtain fused features, and predict the binding probability of the drug and the protein based on the fused features.

[0045] The above-mentioned deep learning-based drug-target interaction prediction method and system extracts the feature representation of drugs by combining the GINE layer and the multi-head attention mechanism, and uses the pre-trained ESM-2 model to extract the feature representation of proteins, which can achieve comprehensive feature extraction and improve the accuracy of feature representation; through the cross-attention mechanism, the drug and protein features are effectively integrated to achieve efficient and accurate prediction of drug-target interactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A diagram of an application environment of a method for predicting drug-target interaction based on deep learning in one embodiment;

[0047] Figure 2 A framework diagram of a GPS-DTI model proposed in one embodiment;

[0048] Figure 3 A schematic diagram of a process of predicting a drug-target interaction based on deep learning in one embodiment;

[0049] Figure 4 is a structural block diagram of a drug-target interaction prediction system based on deep learning in one embodiment;

[0050] Figure 5 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0052] The drug-target interaction prediction method based on deep learning provided in the embodiments of the present application can be applied to Figure 1 In the application environment shown. Figure 1As shown, the application environment includes a computer device 110. The computer device 110 can determine the SMILES sequence of the drug molecule to be predicted and the amino acid sequence of the protein; the computer device 110 can determine the structural information of the drug analysis based on the SMILES sequence of the drug analysis, and generate a drug molecule graph based on the structural information; the computer device 110 can input the drug molecule graph into the GPS layer, and extract the feature representation of the drug using GINE and the multi-head attention mechanism; the computer device 110 can input the amino acid sequence of the protein into the ESM-2 model to extract the feature representation of the protein; the computer device 110 can perform cross-attention mechanism feature fusion of the feature representation of the drug and the feature representation of the protein to obtain the fused feature, and predict the binding probability of the drug and the protein based on the fused feature. Among them, the computer device 110 can be, but is not limited to, various personal computers, laptops, smart phones, robots, unmanned aerial vehicles, tablet computers and other devices.

[0053] In one embodiment, Figure 2 As shown, a prediction model for a prediction method of drug-target interaction based on deep learning is provided. The prediction model is a prediction model constructed using deep learning, namely, a GPS-DTI model, which can realize the prediction of drug-protein interaction.

[0054] like Figure 2 As shown in the figure, the SMILES sequence of the drug and the amino acid sequence of the protein can be used as input. For the protein part, the pre-trained ESM-2 model is used to obtain the high-level embedding of the protein sequence, and then these embeddings are input into the CNN to extract the feature representation of the protein; for the drug part, the drug molecular graph is first constructed and then passed to the GPS layer to obtain the feature representation of the drug. The GPS layer combines GINE and multi-head attention mechanisms to capture both local and global features in the drug molecular graph. Next, the extracted protein and drug features are fused using the cross-attention module to capture the interaction relationship between the drug and the target.

[0055] In this embodiment, if Figure 2The GPS-DTI model shown extracts drug features by combining the GINE layer and the multi-head attention mechanism, and uses the pre-trained ESM-2 model to extract high-level features of protein sequences, and then fuses drug and protein features through the cross-attention mechanism. The local feature extraction of the GINE layer is combined with the global information capture of the multi-head attention mechanism through an innovative architecture, and protein feature extraction is performed using a pre-trained model, thereby achieving comprehensive feature extraction. Finally, the drug and protein features are effectively fused through the cross-attention mechanism, achieving efficient and accurate prediction of drug-target interactions. Compared with the traditional single feature extraction method, the model's predictive performance is significantly improved, and excellent generalization ability is demonstrated, bringing new breakthroughs to the field of drug discovery.

[0056] In one embodiment, Figure 3 As shown, a method for predicting drug-target interaction based on deep learning is provided, comprising the following steps:

[0057] Step 302: determine the SMILES sequence of the drug molecule to be predicted and the amino acid sequence of the protein.

[0058] Computer equipment can perform data preparation, namely collecting SMILES sequences of drug molecules and amino acid sequences of proteins for which interaction prediction is required.

[0059] Step 304: determine the structural information of the drug analysis based on the SMILES sequence of the drug analysis, and generate a drug molecule graph according to the structural information.

[0060] The computer equipment can first construct the drug molecule graph, and generate molecules and, including atomic nodes and chemical bond edges, from the structural information of the drug molecules.

[0061] Step 306, input the drug molecule graph into the GPS layer, and use GINE and multi-head attention mechanism to extract the feature representation of the drug.

[0062] After constructing the drug molecular graph, it can be applied to the GPS layer, that is, the constructed drug molecular graph is input into the GPS layer, and the feature representation of the drug is extracted using GINE and multi-head attention mechanism.

[0063] Specifically, in one embodiment, a deep learning-based drug-target interaction prediction method provided may also include a process for extracting a feature representation of a drug, the specific process including: obtaining node and edge features based on a drug molecule graph; inputting the drug molecule graph into a GPS layer to obtain an adjacency matrix of the drug molecule graph; using GINE to perform feature extraction based on node and edge features and an adjacency matrix; and using a multi-head attention mechanism to perform feature extraction based on nodes.

[0064] For the feature representation extraction of drugs, 10-layer GPS can be used to extract features from drug molecular graphs. The feature representation extraction of drugs can be expressed as: Where A is the adjacency matrix of the graph, X (l) and E (l) are the node and edge features of the lth layer, GINE l is the GINE layer of the lth layer, ClobalAttn l It is the multi-head attention mechanism of the lth layer, and finally obtains the required drug feature representation F d =X (l+1) .

[0066] Step 308 , input the amino acid sequence of the protein into the ESM-2 model to extract the characteristic representation of the protein.

[0067] For protein feature representation extraction, we can first use the pre-trained ESM-2 model to encode the amino acid sequence of the protein to obtain the high-level embedding of the protein, and then feed the protein embedding output by ESM-2 into a 3-layer 1D-CNN to further extract the protein feature representation.

[0068] In one embodiment, a deep learning-based drug-target interaction prediction method provided may also include a process for extracting a characteristic representation of a protein, the specific process including: inputting the amino acid sequence of the protein into an ESM-2 model to obtain a high-level embedding representation of the protein sequence; based on the ESM-2 model, using CNN to process the high-level embedding representation to extract the characteristic representation of the protein.

[0069] In one embodiment, the process of extracting the feature representation of the protein may also include: obtaining a protein embedding vector based on the ESM-2 model according to the high-level embedding representation; inputting the protein embedding vector into the CNN, and outputting the feature representation of the protein.

[0070] Specifically, for protein feature extraction, we first obtain the high-level embedding representation E of the protein sequence by using the ESM-2 model. P ∈R l×d , where l is the length of the protein sequence and d is the embedding dimension; the obtained embedding representation is then processed using a three-layer CNN to extract higher-level features, and the output of each layer of CNN will be used as the input of the next layer until the required layer depth is reached; for the l-th layer of CNN, it updates the protein hidden representation The process can be expressed as in, and They are the learnable weight matrix and bias vector of the Lth layer CNN respectively; is the hidden representation of the protein at layer l; That is, the input of the first layer of CNN is the original protein embedding vector; the last layer obtains the feature representation of the required protein.

[0071] In step 310, the feature representation of the drug and the feature representation of the protein are cross-attentionally fused to obtain fused features, and the binding probability of the drug and the protein is predicted based on the fused features.

[0072] After obtaining the feature representation of the drug and the feature representation of the protein, the feature representations of the drug and protein can be input into the cross-attention module to fuse the features of the two by calculating the interaction weight between the protein and the drug.

[0073] In one embodiment, a deep learning-based drug-target interaction prediction method provided may also include a process of calculating each vector for feature fusion, and the specific process includes: based on the feature representation of the drug, determining the number of nodes and the embedding dimension of the drug feature graph; linearly transforming the feature representation of the drug, obtaining a drug query vector according to the number of nodes and the embedding dimension of the drug feature graph, and obtaining a drug bond vector and a drug value vector based on the feature representation of the protein; inputting the feature representation of the protein into a linear layer, calculating a protein query vector, and obtaining a protein bond vector and a protein value vector based on the feature representation of the drug; performing feature fusion based on the drug query vector, drug bond vector, drug value vector, protein query vector, protein bond vector, and protein value vector to obtain fused features.

[0074] In one embodiment, a deep learning-based drug-target interaction prediction method is provided that may further include a process of calculating an attention matrix for feature fusion, the specific process including: calculating a drug attention matrix based on a drug query vector, a drug bond vector, and a drug value vector; and calculating a protein attention matrix based on a protein query vector, a protein bond vector, and a protein value vector; performing feature fusion based on the drug attention matrix and the protein attention matrix to obtain fused features.

[0075] In another embodiment, a deep learning-based drug-target interaction prediction method provided may also include a feature fusion process, the specific process including: calculating a drug feature map based on the drug attention matrix, and calculating a protein feature map based on the protein attention matrix; connecting the drug feature map and the protein feature map in the channel dimension, and obtaining a target drug feature map and a target protein feature map through a linear layer; downsampling the target drug feature map and the target protein feature map through a global maximum pooling operation to obtain a joint feature representation as a fused feature.

[0076] When the feature representation of the drug and the feature representation of the protein are cross-attentionally fused, the following can be specifically included: Obtain drug query vector through linear transformation Where m d represents the number of nodes in the drug feature graph, and De is the embedding dimension. Drug key vector and drug value vector Protein feature map Get, where n p is the sequence length of the protein. Specifically, Here W i q , W i k , is the learned weight matrix, and is the dimension of each attention head, i ranges from 1 to heads, and heads is the number of attention heads.

[0077] The processing of protein features is similar to that of drug features. p is fed into a linear layer to compute the query vector for the protein Then use the drug feature map to generate the protein bond vector and protein value vector The specific calculation is: Among them, W i q , W i k , W i v is shared between drugs and proteins, which means they use the same weight matrix for linear transformation. By applying the softmax function, the attention matrices for drugs and proteins are calculated as follows: in Represents the dimensions of the drug and protein bond vectors.

[0078] The drug / protein feature map for each attention head is obtained by multiplying the drug / protein attention matrix of each attention head with the corresponding drug / protein value matrix. Specifically: Subsequently, the feature maps of all attention heads are concatenated in the channel dimension and passed through a linear layer to obtain the final drug and protein feature representation. and Z D =Concat(Z 1,D ,Z 2,D ,...,Z heads,D )×W Z , Z P =Concat(Z 1,P ,Z 2,P ,...,Z heads,P )×W Z ,in is the shared weight matrix.

[0079] Next, the obtained feature map is combined with the original feature map to obtain the final drug feature map and protein feature maps Finally, the drug and protein feature maps are downsampled through the global maximum pooling operation, and finally, the obtained results are connected to obtain the joint feature representation

[0080] In one embodiment, a deep learning-based drug-target interaction prediction method provided may also include a process of designing a prediction classifier, the specific process including: inputting the joint feature representation into the deep learning-based classifier, mapping it to the output space through the fully connected layer in the classifier; using the AdamW optimizer to minimize the cross entropy loss function, and performing parameter optimization based on the optimized cross entropy loss function.

[0081] In order to predict the probability of drug-target interaction (DTI), a deep learning-based classifier is designed. The classifier receives the joint feature representation f as input and maps it to the output space through a fully connected layer. The output layer contains a unit whose weight W and bias b can be learned through training. The calculation formula of DTI probability p is: p = Sigmoid (Wf + b).

[0082] The Sigmoid function converts the result of the linear combination into a probability value between 0 and 1. During the training phase, the AdamW optimizer can be used to minimize the cross entropy loss function with L2 regularization to optimize the model parameters. The loss function is defined as: Among them, yi is the true label of the i-th sample, pi is the corresponding predicted probability, θ is the set of all learnable parameters, and λ is the L2 regularization strength.

[0083] Next, you can choose the widely recognized evaluation indicators AUROC (i.e., area under the receiver operating characteristic curve), AUPRC (i.e., area under the precision-recall curve), and F1 as the evaluation criteria. AUC is the area under the receiver operating characteristic (ROC) curve, which measures the probability that the machine learning model will rank a random positive sample higher than a random negative sample.

[0084] Among them, AUPR is the area under the precision-recall curve, which evaluates the ability of a machine learning model to retrieve all positive samples (perfect recall) while avoiding predicting any negative samples as positive samples (perfect precision).

[0085] The definition of the F1 indicator is: Among them, true positives (TP) and true negatives (TN) represent the number of successfully identified drug-target interactions and non-interactions, respectively. False positives (FP) and false negatives (FN) represent the number of drug-target instances that are incorrectly identified as interacting and non-interacting, respectively.

[0086] Next, in one embodiment, a deep learning-based drug-target interaction prediction method provided may also include a prediction process, and the specific process may include: passing the fused features through a fully connected network to predict the binding probability of the drug and the protein.

[0087] The accuracy of feature representation is improved by encoding drug molecule graphs and protein sequences; the use of cross-attention modules enhances the model's ability to understand interaction patterns; and the fully connected network gives an accurate assessment of the probability of drug-protein binding.

[0088] In this application, a model named GPS-DTI is proposed for the first time. It extracts drug features by combining the GINE layer and the multi-head attention mechanism for the first time, and uses the pre-trained ESM-2 model to extract high-level features of protein sequences, and then fuses drug and protein features through the cross-attention mechanism. Compared with the prior art, this application has the following improvements:

[0089] Comprehensive feature extraction and fusion: GINE layers and multi-head attention mechanisms are used to extract drug features, which can more comprehensively capture the local and global structural information of drug molecular graphs compared to the single feature extraction methods in the prior art. The pre-trained ESM-2 model is used to extract high-level features of protein sequences, and the drug and protein features are fused through a cross-attention mechanism, which can more effectively capture the interdependence between drugs and proteins compared to the simple feature concatenation or weighted summation in the prior art.

[0090] Improve prediction performance: Through innovative architecture design, it overcomes the limitations of traditional methods in feature extraction and achieves efficient and accurate prediction of drug-target interactions; it can learn complex relationship patterns from large-scale data to improve prediction accuracy and efficiency.

[0091] It should be understood that, although the various steps in the above-mentioned flow chart are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the above-mentioned flow chart may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0092] In one embodiment, Figure 4 As shown, a drug-target interaction prediction system based on deep learning is provided, comprising: a sequence determination module 410, a drug molecule graph determination module 420, a drug feature extraction module 430, a protein feature extraction module 440 and a feature fusion and prediction module 450, wherein:

[0093] A sequence determination module 410 is used to determine the SMILES sequence of the drug molecule to be predicted and the amino acid sequence of the protein;

[0094] A drug molecule graph determination module 420 is used to determine the structural information of the drug analysis based on the SMILES sequence of the drug analysis, and generate a drug molecule graph according to the structural information;

[0095] The drug feature extraction module 430 is used to input the drug molecular graph into the GPS layer and extract the feature representation of the drug using GINE and multi-head attention mechanism;

[0096] The protein feature extraction module 440 is used to input the amino acid sequence of the protein into the ESM-2 model to extract the feature representation of the protein;

[0097] The feature fusion and prediction module 450 is used to perform cross-attention mechanism feature fusion on the feature representation of the drug and the feature representation of the protein to obtain fused features, and predict the binding probability of the drug and the protein based on the fused features.

[0098] In one embodiment, the drug feature extraction module 430 is also used to obtain node and edge features based on the drug molecule graph; input the drug molecule graph into the GPS layer to obtain the adjacency matrix of the drug molecule graph; use GINE to extract features based on node and edge features and the adjacency matrix; use a multi-head attention mechanism to extract features based on nodes.

[0099] In one embodiment, the protein feature extraction module 440 is also used to input the amino acid sequence of the protein into the ESM-2 model to obtain a high-level embedding representation of the protein sequence; based on the ESM-2 model, CNN is used to process the high-level embedding representation to extract the feature representation of the protein.

[0100] In one embodiment, the protein feature extraction module 440 is further used to obtain a protein embedding vector based on the ESM-2 model according to the high-level embedding representation; input the protein embedding vector into the CNN, and output the feature representation of the protein.

[0101] In one embodiment, the feature fusion and prediction module 450 is also used to determine the number of nodes and the embedding dimension of the drug feature graph based on the feature representation of the drug; linearly transform the feature representation of the drug, obtain the drug query vector according to the number of nodes and the embedding dimension of the drug feature graph, and obtain the drug key vector and the drug value vector based on the feature representation of the protein; input the feature representation of the protein into the linear layer, calculate the protein query vector, and obtain the protein key vector and the protein value vector based on the feature representation of the drug; perform feature fusion based on the drug query vector, the drug key vector, the drug value vector, the protein query vector, the protein key vector, and the protein value vector to obtain the fused features.

[0102] In one embodiment, the feature fusion and prediction module 450 is also used to calculate the drug attention matrix based on the drug query vector, the drug key vector, and the drug value vector; and to calculate the protein attention matrix based on the protein query vector, the protein key vector, and the protein value vector; and perform feature fusion based on the drug attention matrix and the protein attention matrix to obtain fused features.

[0103] In one embodiment, the feature fusion and prediction module 450 is also used to calculate a drug feature map based on the drug's attention matrix, and to calculate a protein feature map based on the protein attention matrix; connect the drug feature map and the protein feature map in the channel dimension, and obtain a target drug feature map and a target protein feature map through a linear layer; downsample the target drug feature map and the target protein feature map through a global maximum pooling operation to obtain a joint feature representation as a fused feature.

[0104] In one embodiment, the feature fusion and prediction module 450 is also used to input the joint feature representation into a deep learning-based classifier, and map it to the output space through the fully connected layer in the classifier; the AdamW optimizer is used to minimize the cross entropy loss function, and parameter optimization is performed based on the optimized cross entropy loss function.

[0105] In one embodiment, the feature fusion and prediction module 450 is also used to predict the binding probability of the drug and the protein by passing the fused features through a fully connected network.

[0106] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a prediction method for drug-target interaction based on deep learning is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0107] Those skilled in the art will understand that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0108] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps of a method for predicting drug-target interaction based on deep learning are implemented.

[0109] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of a method for predicting drug-target interactions based on deep learning are implemented.

[0110] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0111] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0112] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A method for predicting drug-target interactions based on deep learning, characterized in that: The method comprises: Determine the SMILES sequence of the drug molecule to be predicted and the amino acid sequence of the protein; Determine structural information of the drug analysis based on the SMILES sequence of the drug analysis, and generate a drug molecule graph according to the structural information; The drug molecule graph is input into the GPS layer, and the feature representation of the drug is extracted using GINE and multi-head attention mechanism; Inputting the amino acid sequence of the protein into the ESM-2 model to extract the characteristic representation of the protein; The feature representation of the drug and the feature representation of the protein are cross-attentionally fused to obtain fused features, and the binding probability of the drug and the protein is predicted based on the fused features.

2. The method for predicting drug-target interaction based on deep learning according to claim 1, characterized in that: The drug molecule graph is input into the GPS layer, and the feature representation of the drug is extracted using GINE and multi-head attention mechanism, including: Acquire node and edge features based on the drug molecule graph; Inputting the drug molecule graph into the GPS layer to obtain an adjacency matrix of the drug molecule graph; GINE is used to extract features based on the node and edge features and the adjacency matrix; and a multi-head attention mechanism is used to extract features based on the nodes.

3. The method for predicting drug-target interaction based on deep learning according to claim 1, characterized in that: The amino acid sequence of the protein is input into the ESM-2 model to extract the characteristic representation of the protein, including: Input the amino acid sequence of the protein into the ESM-2 model to obtain a high-level embedding representation of the protein sequence; Based on the ESM-2 model, CNN is used to process the high-level embedding representation to extract the feature representation of the protein.

4. The method for predicting drug-target interaction based on deep learning according to claim 3, characterized in that: Based on the ESM-2 model, CNN is used to process the high-level embedding representation to extract the feature representation of the protein, including: Based on the ESM-2 model, obtaining a protein embedding vector according to the high-level embedding representation; The protein embedding vector is input into CNN and the feature representation of the protein is output.

5. The method for predicting drug-target interaction based on deep learning according to claim 1, characterized in that: The feature representation of the drug and the feature representation of the protein are cross-attentionally fused to obtain fused features, including: Based on the characteristic representation of the drug, determining the number of nodes and the embedding dimension of the drug characteristic graph; Linearly transforming the characteristic representation of the drug, obtaining a drug query vector according to the number of nodes and the embedding dimension of the drug characteristic graph, and obtaining a drug key vector and a drug value vector based on the characteristic representation of the protein; Inputting the feature representation of the protein into a linear layer, calculating a protein query vector, and obtaining a protein key vector and a protein value vector based on the feature representation of the drug; Based on the drug query vector, drug key vector, drug value vector, protein query vector, protein key vector, and protein value vector, feature fusion is performed to obtain fused features.

6. The method for predicting drug-target interaction based on deep learning according to claim 5, characterized in that: Based on the drug query vector, drug key vector, drug value vector, protein query vector, protein key vector, and protein value vector, feature fusion is performed to obtain fused features, including: Calculating a drug attention matrix according to the drug query vector, the drug key vector, and the drug value vector; and calculating a protein attention matrix according to the protein query vector, the protein key vector, and the protein value vector; Feature fusion is performed based on the attention matrix of the drug and the attention matrix of the protein to obtain fused features.

7. The method for predicting drug-target interaction based on deep learning according to claim 6, characterized in that: Based on the attention matrix of the drug and the attention matrix of the protein, feature fusion is performed to obtain fused features, including: Calculating a drug feature map based on the drug attention matrix, and calculating a protein feature map based on the protein attention matrix; Connecting the drug feature map and the protein feature map in the channel dimension, and obtaining a target drug feature map and a target protein feature map through a linear layer; The target drug feature map and the target protein feature map are downsampled by a global maximum pooling operation to obtain a joint feature representation as a fused feature.

8. The method for predicting drug-target interaction based on deep learning according to claim 7, characterized in that: The method further comprises: Inputting the joint feature representation into a deep learning-based classifier and mapping it to an output space through a fully connected layer in the classifier; The AdamW optimizer is used to minimize the cross entropy loss function, and parameter optimization is performed based on the optimized cross entropy loss function.

9. The method for predicting drug-target interaction based on deep learning according to claim 1, characterized in that: Predicting the probability of drug-protein binding based on the fused features includes: The fused features are passed through a fully connected network to predict the binding probability of the drug and the protein.

10. A drug-target interaction prediction system based on deep learning, characterized in that: The system comprises: A sequence determination module is used to determine the SMILES sequence of the drug molecule to be predicted and the amino acid sequence of the protein; A drug molecule graph determination module, used to determine the structural information of the drug analysis based on the SMILES sequence of the drug analysis, and generate a drug molecule graph according to the structural information; A drug feature extraction module, used to input the drug molecular graph into the GPS layer, and extract the feature representation of the drug using GINE and multi-head attention mechanism; A protein feature extraction module, used to input the amino acid sequence of the protein into the ESM-2 model to extract the feature representation of the protein; The feature fusion and prediction module is used to perform cross-attention mechanism feature fusion on the feature representation of the drug and the feature representation of the protein to obtain fused features, and predict the binding probability of the drug and the protein based on the fused features.

Images