Cross-domain information fusion method, device and software product for drug discovery

Abstract

The application discloses a cross-domain information fusion method, equipment and software product for drug discovery, adopts a drug target interaction prediction model, and performs zero-sample interaction prediction on drug and target data for input drug and target data information; the drug target interaction prediction model comprises a target protein encoder, a drug encoder, a first fusion encoder, a second fusion encoder and a multilayer perceptron; the protein encoder is used for feature extraction of a protein; the drug encoder is used for drug feature extraction; the first fusion encoder is used for fusion of drug features and target protein features; the second fusion encoder is used for fusion of protein features and protein features; and the multilayer perceptron is used for drug target interaction classification of fusion features of different fusion encoders. The application adopts an innovative double-channel framework model comprehensive DTI prediction, and improves the accuracy and robustness of the result.

Description

Technical Field

[0001] The present invention belongs to the fields of computer science, biochemistry and deep learning technology, and relates to a cross-domain information fusion method, equipment and software product, and specifically to a cross-domain information fusion method, equipment and software product for drug discovery, which is used to improve the generalization performance of drug response prediction models. Background Art

[0002] Drug-protein interaction prediction is a crucial part of the drug development process, including small molecule drug-target interactions (DTI) and large molecule drug-protein interactions. Protein-protein interaction (PPI) is one of the research areas studying the interaction between large molecule drugs and proteins. DTI and PPI are specific applications of protein representation learning in various fields of drug development. The information contained in protein representations differs in the two application areas. First, the DTI field is committed to detecting the information contained in smaller, specific regions of the protein structure (i.e., local information), such as binding pockets and activity. Second, the PPI field focuses on capturing the overall structure and biological function of the protein (i.e., global information), including protein size and tertiary structure. In addition, local and global protein information are interdependent and affect the biological function of the protein at the same time. Therefore, integrating protein information from different fields can generate a more comprehensive protein representation from multiple perspectives and improve the prediction accuracy of drug-protein interactions.

[0003] Small molecule DTI prediction plays a crucial role in drug discovery because small molecule drugs offer advantages such as high chemical stability and low production costs. When the training data contains drugs and targets of the same type as those in the test set, these drugs and targets are considered "known." Conversely, if they do not appear in the training data, they are considered "novel" (zero-shot learning). The fields of computational biology and medicine are experiencing rapid growth, leading to the continuous identification and exploration of new targets. At the same time, drugs represent only a tiny fraction of the vast molecular universe. Therefore, the prediction of novel DTIs is more closely related to the practical applications of drug development engineering.

[0004] Despite significant progress in DTI prediction for known drugs and targets, the current task of predicting interactions between new drugs remains very challenging. For example, targets (i.e., disease-associated proteins) contain highly complex and diverse information. In addition, traditional methods do not contain global protein information, thereby limiting the knowledge gained from protein characterization. These challenges severely limit the performance of existing DTI prediction methods for new drugs and targets. For example, in Nguyen's work, their model based on graph neural networks (GNNs) achieved a prediction performance of over 90% in the area under the receiver operating characteristic curve (AUROC) for known drugs and targets. However, for new drugs and targets, the AUROC dropped to at least 60%. Therefore, it is important to obtain comprehensive and diverse information about target proteins. Summary of the Invention

[0005] In order to solve the technical problems existing in the drug-target interaction prediction method in the prior art, the present invention proposes a cross-domain information fusion method, device and software product for drug discovery.

[0006] In a first aspect, the present invention provides a cross-domain information fusion method for drug discovery, characterized by: using a drug-target interaction prediction model to perform zero-sample interaction prediction on the input drug and target data information;

[0007] The drug-target interaction prediction model includes a target protein encoder, a drug encoder, a first fusion encoder, a second fusion encoder and a multilayer perceptron;

[0008] The protein encoder is used for protein feature extraction; the drug encoder is used for drug feature extraction; the first fusion encoder is used for fusing drug features with target protein features; the second fusion encoder is used for fusing protein features with protein features; the multilayer perceptron is used for classifying drug-target interactions based on the fusion features of different fusion encoders.

[0009] Preferably, the target protein encoder comprises three identical Transformer encoder layers, the number of attention heads of the Transformer encoder is 8, an activation layer is added to the sixth Transformer encoder layer, and the activation function is ReLU;

[0010]

[0011] Among them, σ represents the activation function, which is the ReLU function. represents the learnable weight matrix in the Transformer encoder, represents the hidden protein representation of layer l, and X p represents the embedded representation of the protein, θ p Indicates the maximum sequence length of the protein, D p represents the embedding dimension of the protein.

[0012] Preferably, the drug encoder comprises three graph convolution layers, an activation layer is added after the third convolution layer, and the activation function is a ReLU function;

[0013]

[0014] Where σ represents the activation function, which is the ReLU function; GCN is the drug encoder layer; and Represent the learnable weight matrix and bias matrix of the learnable layer l, respectively. represents the adjacency matrix of the molecular graph, represents the node representation of the lth layer, and X d represents the embedded representation of drug molecules, θ d The maximum number of nodes representing drug molecules, D d represents the embedding dimension of the protein.

[0015] Preferably, the first fusion encoder and the second fusion encoder each include a bilinear interaction graph and a pooling layer, the bilinear interaction graph includes two attention heads, the embedding size is 168, and the pooling window size of the pooling layer is 3;

[0016] The bilinear interaction graph is used to obtain pairwise interaction information:

[0017]

[0018] in, and Represent the learnable weight matrices corresponding to drugs and targets, 1∈R N is a fixed 1 matrix, ° represents the Hadamard product; is a learnable weight matrix, σ represents the activation function, represents the hidden layer protein and drug representations generated by the third layer Transformer encoder layer and drug encoder layer GCN; N represents the number of atoms in the drug molecule, and K represents the dimension of the embedded feature;

[0019] Then apply bilinear pooling to the B matrix,

[0020]

[0021] Among them, U kand V k represents the weight value of the kth column, f k ′ is the kth element of the f′ vector; M and N represent the number of amino acids in the protein and the number of atoms in the drug molecule, respectively; B i,j is the element in the i-th row and j-th column of the B matrix; express The i-th column of the matrix, express The jth column of the matrix;

[0022] The feature map f in the first fusion encoder dt for:

[0023] f dt =SumPool(f′,s);

[0024] Where SumPool(·) is a one-dimensional non-overlapping pooling operation, s represents the one-dimensional non-overlapping pooling step size; the feature map f in the second fusion encoder pp Acquire in the same way;

[0025] The final fusion representation f=f is obtained by comparing the discrimination of the two channel features. dt -f pp .

[0026] Preferably, the multilayer perceptron includes three fully connected layers, the input of which is the difference between the output features of two fusion encoders, the feature dimension is 256, and the dimensions of the fully connected layers are 256, 512, and 256 respectively; the output is a numerical value used to represent the probability of interaction between the drug and the target.

[0027] Preferably, the drug-target interaction prediction model is a trained model;

[0028] During training, drugs i and target protein t i The feature representation process is as follows:

[0029] P i =Φ drug (d i ),T i =Φ Φarget (t i );

[0030] Among them, i∈[0,N], N represents the total number of samples in the training data set, Φ drug and Φ target represents drug encoder and protein encoder, P i and T i It is a drug i and target t icharacteristics;

[0031] The first fusion encoder Φ f1 The encoded features are:

[0032] F1=Φ f1 ([P i ,T i ]);

[0033] The second fusion encoder Φ f2 The encoded features are:

[0034] F2=Φ f2 ([T i ,T i ]);

[0035] Among them, F1,F2∈R N×m , R represents the real number field, and m represents the dimension of the eigenvector.

[0036] Preferably, the drug-target interaction prediction model is a trained model;

[0037] During training, the goal of the multilayer perceptron is to minimize the binary cross entropy loss:

[0038]

[0039] Among them, θ represents all learnable weight matrices and bias vectors, y i represents the true labels of all drug-target pairs, p i The model predicts the interaction between drug-target pairs, and λ is a hyperparameter used for L2 regularization.

[0040] In a second aspect, the present invention provides a cross-domain information fusion device for drug discovery, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the cross-domain information fusion method for drug discovery when executing the program.

[0041] In a third aspect, the present invention provides a software product, comprising a computer program, which, when executed by a processor, implements the cross-domain information fusion method for drug discovery.

[0042] Compared with the prior art, the beneficial effects of the present invention include:

[0043] (1) The present invention adopts a cross-domain (i.e., PPI and DTI) information fusion strategy to integrate local and global protein information to obtain more comprehensive protein information and improve the accuracy of drug-target interaction prediction;

[0044] (2) This invention combines the PPI and traditional DTI modules to establish two distinct channels; using a dual-channel approach to obtain comprehensive protein information to improve the prediction performance of new drugs and targets. This method obtains global protein information missing from the DTI module in the PPI channel, obtaining richer protein information;

[0045] (3) The two channels of the present invention share all protein encoder parameters, combining cross-domain protein information and enhancing protein characterization;

[0046] (4) The present invention maps drug-target pairs into a new representation space, thereby increasing the distance between samples with different labels. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The technical solution of the present invention is further illustrated below using embodiments and specific implementation methods. In addition, some drawings are also used in the process of illustrating the technical solution. For those skilled in the art, other drawings and the intention of the present invention can be obtained based on these drawings without making any creative efforts.

[0048] Figure 1 The original diagram of the method of the present invention;

[0049] Figure 2 This is a diagram showing the difference between the principles of the present invention and the existing method;

[0050] Figure 3 Comparison of the results of various methods with respect to known drugs and targets in the examples of the present invention;

[0051] Figure 4 Comparison of visualization results of novel drugs and targets in examples of the present invention with those of DrugBAN. DETAILED DESCRIPTION

[0052] In order to facilitate ordinary technicians in this field to understand and implement the present invention, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0053] This embodiment provides a cross-domain information fusion method for drug discovery, which uses a drug-target interaction prediction model to perform zero-sample interaction prediction on the input drug and target data information;

[0054] The drug-target interaction prediction model includes a target protein encoder, a drug encoder, a first fusion encoder, a second fusion encoder and a multilayer perceptron;

[0055] The target protein encoder is used for protein feature extraction; the drug encoder is used for drug feature extraction; the first fusion encoder is used for fusing drug features with target protein features; the second fusion encoder is used for fusing protein features with protein features; the multilayer perceptron is used for classifying drug-target interactions based on the fusion features of different fusion encoders.

[0056] Please see Figure 1 In one embodiment, the target protein encoder comprises three identical Transformer encoder layers, the number of attention heads of the Transformer encoder is 8, an activation layer is added to the sixth Transformer encoder layer, and the activation function is ReLU;

[0057]

[0058] Among them, σ represents the activation function, which is the ReLU function. represents the learnable weight matrix in the Transformer encoder, represents the hidden protein representation of layer l, and X p represents the embedded representation of the protein, θ p Indicates the maximum sequence length of the protein. In this example, θ p =1200, D p represents the embedding dimension of the protein.

[0059] In one embodiment, the drug encoder comprises three graph convolution layers, an activation layer is added after the third convolution layer, and the activation function is a ReLU function;

[0060]

[0061] Where σ represents the activation function, which is the ReLU function; GCN is the drug encoder layer; and Represent the learnable weight matrix and bias matrix of the learnable layer l, respectively. represents the adjacency matrix of the molecular graph, represents the node representation of the lth layer, and X d represents the embedded representation of drug molecules, θ d Indicates the maximum number of nodes of drug molecules. In this embodiment, θ d =290, D d represents the embedding dimension of the protein.

[0062] In one embodiment, the first fusion encoder and the second fusion encoder each include a bilinear interaction graph and a pooling layer, the bilinear interaction graph includes two attention heads, the embedding size is 168, and the pooling window size of the pooling layer is 3;

[0063] The bilinear interaction graph is used to obtain pairwise interaction information:

[0064]

[0065] in, and Represent the learnable weight matrices corresponding to drugs and targets, 1∈R N is a fixed 1 matrix, ° represents the Hadamard product; is a learnable weight matrix, σ represents the activation function, and the ReLU activation function is used in this embodiment; represents the hidden layer protein and drug representations generated by the third layer Transformer encoder layer and the drug encoder layer GCN; N represents the number of atoms in the drug molecule, K represents the dimension of the embedded feature, and in this example, k = 256;

[0066] Then apply bilinear pooling to the B matrix,

[0067]

[0068] Among them, U k and V k represents the weight value of the kth column, f k ′ is the kth element of the f′ vector; M and N represent the number of amino acids in the protein and the number of atoms in the drug molecule, respectively; B i,j is the element in the i-th row and j-th column of the B matrix; express The i-th column of the matrix, express The jth column of the matrix;

[0069] The feature map f in the first fusion encoder dt for:

[0070] f dt =SumPool(f′,s);

[0071] Where SumPool(·) is a one-dimensional non-overlapping pooling operation, s represents the one-dimensional non-overlapping pooling step size; the feature map f in the second fusion encoder pp Acquire in the same way;

[0072] The final fusion representation f=f is obtained by comparing the discrimination of the two channel features. dt -f pp .

[0073] In one embodiment, the multilayer perceptron includes three fully connected layers, the input of which is the difference between the output features of two fusion encoders, the feature dimension is 256, and the dimensions of the fully connected layers are 256, 512, and 256 respectively; the output is a numerical value used to represent the probability of interaction between the drug and the target.

[0074] In one embodiment, the drug-target interaction prediction model is a trained model;

[0075] During training, drugs i and target protein t i The feature representation process is as follows:

[0076] P i =Φ drug (d i ),T i =Φ target (t i );

[0077] Among them, i∈[0,N], N represents the total number of samples in the training data set, Φ drug and Φ target represents drug encoder and protein encoder, P i and T i It is a drug i and target t i characteristics;

[0078] The first fusion encoder Φ f1 The encoded features are:

[0079] F1=Φ f1 ([P i ,T i ]);

[0080] The second fusion encoder Φ f2 The encoded features are:

[0081] F2=Φ f2 ([T i ,T i ]);

[0082] Among them, F1,F2∈R N×m , R represents the real number field, and m represents the dimension of the eigenvector.

[0083] In one embodiment, the drug-target interaction prediction model is a trained model;

[0084] During training, the goal of the multilayer perceptron is to minimize the binary cross entropy loss:

[0085]

[0086] Among them, θ represents all learnable weight matrices and bias vectors, y i represents the true labels of all drug-target pairs, p i The model predicts the interaction between drug-target pairs, and λ is a hyperparameter used for L2 regularization.

[0087] Please see Figure 2 , which is a diagram showing the difference between the principles of the present invention and the existing method;

[0088] It can be seen from the figure that the traditional method is limited to extracting local information of proteins and ignores the global information of proteins, which makes the traditional method's prediction effect unsatisfactory in zero-sample scenarios.

[0089] This embodiment also provides a cross-domain information fusion device for drug discovery, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the cross-domain information fusion method for drug discovery is implemented.

[0090] This embodiment also provides a software product, including a computer program, which implements the cross-domain information fusion method for drug discovery when executed by a processor.

[0091] When the model cannot obtain more effective protein features, it can no longer effectively distinguish new drug target pairs in the representation space. The core of this invention is to make full use of the global information of proteins for feature enhancement and construct channels in different fields for multi-angle perception of protein information.

[0092] To verify the effectiveness of the cross-domain information fusion method for drug discovery, this experiment conducted extensive experiments on three drug-target interaction datasets: BindingDB, BioSNAP, and human. In order to study the effectiveness and generalization of the model in complex scenarios, this experiment adopted different dataset partitioning methods for two different application scenarios. For known drugs and targets, there is overlap between the drug and protein types in the training set and the test set. On the other hand, for new drugs and targets, the drug and target types in the training set and the test set do not overlap. Therefore, the drug-target pairs in the test dataset are completely novel to the model. This experiment widely selected 5 representative drug-target interaction prediction methods based on deep learning models such as CNN, GNN and Transformer, and compared the effects between the models of the core method of the present invention:

[0093] (1) DeepConv-DTI (convolutional neural network encoding drug and protein sequences);

[0094] (2) GraphDTA (Graph Convolutional Neural Network Encoding Drugs);

[0095] (3) MolTrans (Transformer encoding drugs and targets);

[0096] (4) HyperAttentionDTI (convolutional neural network encoding drug and protein sequences + attention mechanism);

[0097] (5) DrugBAN (graph convolutional neural network encoding drug molecules);

[0098] In order to explore the prediction performance of the cross-domain information fusion method for drug discovery for known drugs and targets, this example designed an experiment for comparison.

[0099] Please see Figure 3 , which is a schematic diagram of the results of the present invention using known drugs and targets on three data sets;

[0100] In this scenario, each experimental dataset is randomly divided into a training set, a validation set, and a test set in a ratio of 7:1:2. On this basis, this experiment compares the proposed method with other baseline methods. Figure 3 Comparison results for the BindingDB, BioSNAP, and human datasets are presented. The proposed method generates optimal and suboptimal AUROC scores for the BindingDB and BioSNAP datasets. This can be attributed to its more efficient capture of drug-target interaction features. However, the predictive performance of the proposed method for the human dataset is insufficient. This may be due to the smaller number of samples in the human dataset and sample fluctuations resulting in a difference of approximately 1% between the proposed method and the optimal method. Figure 3 (b) Reports the distances between the positive and negative drug-target pair features for 100 targets in the BindingDB dataset. Figure 3 (b) It can be seen that the proposed method can improve the distance between positive and negative sample features better than the baseline, with the average distance increased by 11.3%.

[0101] Overall, all methods performed well in this context, particularly on the human dataset, where all methods achieved AUROCs greater than 97%, indicating a relatively easy prediction task. The availability of known drugs and targets is particularly valuable for drug repurposing, where large drug libraries can be evaluated against known proteins implicated in the disease of interest.

[0102] In order to explore the prediction performance of the cross-domain information fusion method for drug discovery for novel drugs and targets, this example designed an experiment for comparison.

[0103] Please see Table 1 below, which shows the results of the present invention (SiamDTI) on three datasets for novel drugs and targets;

[0104] Table 1

[0105]

[0106]

[0107] In this scenario, the training and test data do not contain the same drugs and targets, making it impossible to rely on known drug and target features when predicting the test data. However, this scenario is suitable for large-scale virtual drug screening of novel targets, as almost all molecules in the drug screening library are novel, such as those generated by diffusion models. Results show that in this scenario, due to the limited information available, all methods perform worse than in the known drug and target scenario. In contrast, our proposed method achieves state-of-the-art performance in terms of AUROC and AUPRC metrics. On the BindingDB dataset, our proposed method improves AUROC and AUPRC by 7.6% and 13.7%, respectively, compared to DrugBAN. On the BioSNAP dataset, our proposed method improves AUROC and AUPRC by 9.7% and 15.2%, respectively, compared to DrugBAN. Furthermore, our proposed method improves AUROC and AUPRC by 5.7% and 4.6%, respectively, compared to the next-best method. Experimental results demonstrate that our proposed method can effectively extract features of novel drugs and targets, resulting in superior performance in this scenario and confirming its enhanced generalization capability.

[0108] Please see Figure 4 , which is a visualization comparison result diagram of the present invention and the comparative method DrugBAN method;

[0109] As can be seen from the figure, whether in the known drug target scenario or the unknown drug target scenario, the method of the present invention has a larger spacing between samples with different labels than the comparative method DrugBAN, and samples with the same label are more clustered.

[0110] The present invention proposes a cross-domain information fusion method for drug-target interaction prediction. The core method of the present invention is to use cross-domain supervised learning for feature enhancement to obtain more useful target protein information. Its innovative dual-channel framework provides a comprehensive DTI prediction method, improving the accuracy and robustness of the results. By combining the PPI channel to integrate more useful target protein information, the present invention demonstrates the value of the previously unexplored dual-channel information fusion strategy in DTI prediction tasks. Experimental results show that the prediction accuracy of the present invention for new drugs and targets is higher than that of other DTI prediction methods. Compared with the remaining methods, the maximum improvement of the present invention is 29.1%.

[0111] It should be understood that the embodiments described above are only some of the embodiments of the present invention, rather than all of the embodiments. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention may be arbitrarily combined with each other to form a feasible technical solution. Such combination is not restricted by the order of steps and / or structural composition mode, but must be based on the ability of ordinary technicians in this field to implement it. When the combination of technical solutions is mutually inconsistent or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0112] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of protection of the patent of the present invention. Under the guidance of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested by the present invention shall be based on the attached claims.

Claims

1. A cross-domain information fusion method for drug discovery, characterized by: Adopting the drug-target interaction prediction model, we can perform zero-sample interaction prediction on the input drug and target data. The drug-target interaction prediction model includes a target protein encoder, a drug encoder, a first fusion encoder, a second fusion encoder and a multilayer perceptron; The first fusion encoder and the second fusion encoder each include a bilinear interaction graph and a pooling layer, the bilinear interaction graph includes two attention heads, the embedding size is 168, and the pooling window size of the pooling layer is 3; The bilinear interaction graph is used to obtain pairwise interaction information: in, and Represent the learnable weight matrices corresponding to drugs and targets, 1∈R N is a fixed 1 matrix, It means Hadamard; Is a learnable weight matrix, σ represents the activation function; represents the hidden layer protein and drug representations generated by the third layer Transformer encoder layer and drug encoder layer GCN; N represents the number of atoms in the drug molecule, and K represents the dimension of the embedded feature; Then apply bilinear pooling to the B matrix, Among them, U k and V k represents the weight value of the kth column, f k ′ is the kth element of the f′ vector; M and N represent the number of amino acids in the protein and the number of atoms in the drug molecule, respectively; B i,j is the element in the i-th row and j-th column of the B matrix; express The i-th column of the matrix, express The jth column of the matrix; The feature map f in the first fusion encoder dt for: f dt = SumPool(f′,s); Where SumPool(·) is a one-dimensional non-overlapping pooling operation, s represents the one-dimensional non-overlapping pooling step size; the feature map f in the second fusion encoder pp Acquire in the same way; The final fusion representation f=f is obtained by comparing the discrimination of the two channel features. dt -f pp ; The target protein encoder is used for protein feature extraction; the drug encoder is used for drug feature extraction; the first fusion encoder is used for fusing drug features with target protein features; the second fusion encoder is used for fusing protein features with protein features; the multilayer perceptron is used for classifying drug-target interactions using the fusion features of the first fusion encoder and the second fusion encoder.

2. The cross-domain information fusion method for drug discovery according to claim 1, characterized in that: The target protein encoder includes three identical Transformer encoder layers, the number of attention heads of the Transformer encoder is 8, and an activation layer is added to the sixth Transformer encoder layer, and the activation function is ReLU; Among them, σ represents the activation function, which is the ReLU function. represents the learnable weight matrix in the Transformer encoder, represents the hidden protein representation of layer l, and X p represents the embedded representation of the protein, θ p Indicates the maximum sequence length of the protein, D p represents the embedding dimension of the protein.

3. The cross-domain information fusion method for drug discovery according to claim 1, characterized in that: The drug encoder includes three graph convolution layers, an activation layer is added after the third convolution layer, and the activation function is a ReLU function; Where σ represents the activation function, which is the ReLU function; GCN is the drug encoder layer; and Represent the learnable weight matrix and bias matrix of the learnable layer l, respectively. represents the adjacency matrix of the molecular graph, represents the node representation of the lth layer, and X d represents the embedded representation of drug molecules, θ d The maximum number of nodes representing drug molecules, D d represents the embedding dimension of the protein.

4. The cross-domain information fusion method for drug discovery according to claim 1, characterized in that: The multilayer perceptron includes three fully connected layers. The input is the difference between the output features of two fusion encoders. The feature dimension is 256, and the dimensions of the fully connected layers are 256, 512, and 256 respectively. The output is a numerical value used to represent the probability of interaction between the drug and the target.

5. The cross-domain information fusion method for drug discovery according to any one of claims 1 to 4, characterized in that: The drug-target interaction prediction model is a trained model; During training, drugs i and target protein t i The feature representation process is as follows: P i =Φ drug (d i ),T i =Φ target (t i ); Among them, i∈[0,N], N represents the total number of samples in the training data set, Φ drug and Φ target represents drug encoder and protein encoder, P i and T i It is a drug i and target t i characteristics; The first fusion encoder Φ f1 The encoded features are: F1=Φ f1 ([P i ,T i ]); The second fusion encoder φ f2 The encoded features are: F2=Φ f2 ([T i ,T i ]); Among them, F1,F2∈R N×m , R represents the real number field, and m represents the dimension of the eigenvector.

6. The cross-domain information fusion method for drug discovery according to any one of claims 1-2, characterized in that: The drug-target interaction prediction model is a trained model; During training, the goal of the multilayer perceptron is to minimize the binary cross entropy loss: Among them, θ represents all learnable weight matrices and bias vectors, y i represents the true labels of all drug-target pairs, p i The model predicts the interaction between drug-target pairs, and λ is a hyperparameter used for L2 regularization.

7. A cross-domain information fusion device for drug discovery, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the cross-domain information fusion method for drug discovery according to any one of claims 1 to 6 is implemented.

8. A software product comprising a computer program, characterized in that When the computer program is executed by a processor, the cross-domain information fusion method for drug discovery according to any one of claims 1 to 6 is implemented.

Images