Drug target affinity prediction method, electronic equipment and computer readable storage medium

Abstract

The invention discloses a drug target affinity prediction method, electronic equipment and a computer readable storage medium, the method comprises the following steps: feature extraction is carried out on input small molecule SMILES and protein sequences, and the small molecule adopts 10 molecular fingerprints of RDK, Topological, MACCS, AtomPair, ECFP4, FCFP4, FCFP6, Avalon, Layered and Pattern to construct mixed fingerprint features; compressing the high-dimensional molecular fingerprint features to vector dimensions consistent with protein features through linear mapping to realize feature balance, and splicing to form a fusion feature vector; and performing nonlinear interaction and expression enhancement on the fusion features by adopting a multi-expert hybrid module, and finally outputting an affinity prediction value between the small molecule and the protein through a linear regression layer. Through fusion of multi-source chemical fingerprints and deep protein pre-training representation, higher feature expression ability and stronger model generalization are realized; the complexity of the model is reduced through feature mapping and a lightweight multi-expert hybrid model, so that the model has better stability and expandability.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of artificial intelligence and drug discovery, and particularly relates to a drug target affinity prediction method, an electronic device and a computer readable storage medium. BACKGROUND

[0002] Drug-target affinity prediction (DTA) is one of the core tasks in drug design, aiming to predict the binding strength of drugs and targets using molecular structure and protein sequence information. Existing methods mainly include double-tower models based on convolutional networks (such as DeepDTA) and graph neural networks (such as GraphDTA and DeepCDA). These methods have limited representation ability, and the molecular features often use only a single type of fingerprint or graph feature. At the same time, although sequence models such as Transformer have strong interaction ability, they have significant time and space complexity problems. In addition, the feature dimensions of drugs and proteins differ greatly, and direct concatenation can cause the model to rely too much on high-dimensional drug features, weakening the role of protein representation. SUMMARY

[0003] The key problem to be solved by the application is that the existing drug target affinity prediction model does not fully model the molecular side information and lacks protein feature semantics. The application provides a drug target affinity prediction method, an electronic device and a computer readable storage medium.

[0004] To achieve the above technical purposes, the application adopts the following technical solutions: An affinity prediction scheme based on hybrid molecular fingerprints and protein pre-training representation, comprising: (1) inputting a small molecule SMILES sequence and a protein amino acid sequence; (2) using a molecular fingerprint algorithm to extract 10 types of molecular fingerprint features, namely RDK, Topological, MACCS, AtomPair, ECFP4, FCFP4, FCFP6, Avalon, Layered and Pattern, from the small molecule SMILES, and concatenating to obtain a molecular hybrid fingerprint feature vector with a length of 16896 ; (3) using an ESM2 pre-training model (facebook / esm2_t6_8M_UR50D) to extract features from the protein sequence, to obtain a protein feature vector with a length of 320 ; (4) the molecular features are mapped to a projection feature with a length of 320 through a linear layer , and are concatenated with the protein features Concatenate to form a joint distribution feature with a length of 640 dimensions; (5) Input the 640-dimensional joint distribution feature into a multi-expert hybrid module to obtain the final fusion feature; (6) Output the final fusion feature to 1 dimension through a linear layer to obtain the affinity prediction value of the final substance and the target.

[0005] Further, the specific process of step (2) is: For an input small molecule SMILES, use RDKit tools to convert it to a molecule type, and then use the built-in molecular fingerprint extraction method to extract 10 different molecular fingerprints: After concatenating all molecular fingerprint vectors, a molecular hybrid fingerprint feature vector with a dimension of 16896 can be obtained :

[0006] Further, the specific process of step (3) is: For the input protein sequence, use the pre-trained ESM2 model to extract the sequence feature, and extract the last layer of the returned result and perform maximum pooling calculation:

[0007] Among them represents the sequence feature output of the ESM model, represents the number of amino acids of the protein, represents the protein feature after maximum pooling in dimension .

[0008] Further, the specific process of step (4) is: Introduce a learnable linear mapping layer parameter matrix and a bias term , and convert the molecular fingerprint feature space vector to the protein sequence feature space consistent with to realize dimension matching and balanced feature contribution:

[0009] Among them is the converted molecular fingerprint feature vector consistent with .

[0010] Further, the specific process of step (5) is: Construct a hybrid module containing experts to the features obtained before and Feature cross-computation is performed, and each expert consists of a shallow network with a ReLU activation function:

[0011]

[0012] in This indicates the combined distribution characteristics of proteins and small molecules. This represents the feature transformation of the i-th expert. This represents the output of the i-th expert. Then, the outputs of the k experts are concatenated and processed through a routing network. Merge the results:

[0013] in d represents the final fused feature, and d represents the dimension of the fused feature.

[0014] Furthermore, the specific process of step (6) is as follows: Introducing an output linear layer Information in d-dimensional space Transform back into 1D space to obtain the predicted affinity value:

[0015] in This represents the final predicted affinity value.

[0016] Furthermore, we feed the prior data obtained from the experiment into the model for training, and use the error between the actual affinity value and the predicted affinity value as the model's minimization objective, with HuberLoss as the loss function, to optimize and learn all the parameters of the entire model through the stochastic gradient descent algorithm.

[0017] Furthermore, .

[0018] Furthermore, .

[0019] An electronic device includes a memory and a processor, wherein the memory stores a computer program, characterized in that, when the computer program is executed by the processor, the processor enables the processor to implement any of the above-mentioned drug target affinity prediction methods.

[0020] A computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements a drug target affinity prediction method as described in any one of the preceding claims. Beneficial effects

[0021] The present application innovatively fuses 10 different types of molecular fingerprints (RDK, Topological, MACCS, AtomPair, ECFP4, FCFP4, FCFP6, Avalon, Layered, Pattern) to form a hybrid molecular representation covering multiple dimensions of chemical properties such as molecular topology, substructure patterns, bond connectivity, and ring system characteristics. This representation has more comprehensive expert knowledge representation ability than single fingerprint or single graph neural network feature, which helps to significantly improve the accuracy and chemical semantic integrity of small molecule structure representation. Secondly, the present application introduces the ESM2 protein pre-training model to extract protein sequence features. Compared with traditional CNN, BiLSTM or simple statistical features, this model learns more rich evolutionary semantic information through unsupervised training of a large number of protein sequences, has better generalization ability, stability and expression ability for unknown protein sequences, thereby significantly improving the quality of protein side information in the affinity prediction process. Finally, the present application cites a multi-expert hybrid model to perform deep nonlinear interaction on the fused protein small molecule features, realizes lightweight and high-precision drug target affinity prediction, and is suitable for efficient deployment in large-scale drug screening tasks. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is the overall architecture diagram described in the embodiments of the present application. DETAILED DESCRIPTION

[0023] The embodiments of the present application will be described in detail below. The embodiments are based on the technical solutions of the present application, and give detailed implementation methods and specific operation processes, and further explain and describe the technical solutions of the present application.

[0024] The present embodiment provides a drug target affinity prediction method, as shown in Figure 1 , including the following steps: I. Input of the model Formally, for a protein sequence, we represent it as a string of amino acid letters, for example:

[0025] We call the ESM2 model to embed the sequence to get :

[0026] where represents a 6-layer, 320 hidden size Transformer network pre-trained by UR50D. Then we perform maximum pooling calculation on the sequence features extracted by the ESM2 model:

[0027] where denotes the sequence feature output of the ESM model, denotes the number of amino acids of the protein, denotes the protein feature after max-pooling over dimension .

[0028] For input small molecules, we directly convert them into molecule type using RDKit tool, and then extract 10 different molecular fingerprints using the built-in molecular fingerprint extraction method: After concatenating all the molecular fingerprint vectors, we can get the molecular hybrid fingerprint feature vector with dimension sum of 16896 :

[0029] To balance the dimensions of and , we introduce a learnable linear mapping layer parameter matrix and bias term , and convert the vector in the molecular fingerprint feature space to the protein sequence feature space consistent with :

[0030] where is the converted molecular fingerprint feature vector consistent with .

[0031] II. Interaction of protein and drug molecule features We construct a hybrid module containing experts to perform feature cross calculation on drug molecule features and protein features , and each expert is composed of a shallow network with ReLU activation function:

[0032]

[0033] where denotes the joint distribution feature of the protein and small molecule, denotes the feature transformation of the i-th expert, denotes the result output by the i-th expert. Then the outputs of the k experts are concatenated, and the results are combined through the routing network :

[0034] wherein represents the final fused feature, d represents the dimension of the fused feature.

[0035] III. Affinity prediction Directly introduce output linear layer Convert the information of d-dimensional space back to 1-dimensional space as the predicted affinity value:

[0036] wherein represents the final predicted affinity value. And combine the real affinity value Use Huber Loss as the loss function:

[0037] wherein respectively represent the method predicted and the real affinity value between the current input drug and target, is a hyperparameter in the loss function, which is set to 1.0 here, to control the learning curve.

[0038] IV. Experimental verification In order to verify the effectiveness of using the present application (hereinafter referred to as FPR_ESM_MLP) for drug target affinity prediction and the superior performance compared with other methods, this part evaluates the performance of FPR_ESM_MLP through extensive experiments. We respectively carried out affinity regression experiments on 2 classic affinity prediction data sets Davis and Kiba, and took CI and MSE as evaluation indexes, wherein CI represents the monotonicity score evaluation after sorting the predicted affinity according to the size relationship of the real affinity, and MSE represents the error square of the predicted affinity and the real affinity. We selected DeepDTA, GraphDTA, SGC_SAGE, DeepCDA, DeepDTAGen and other graph neural network, convolutional neural network or Transforemr based double tower model or sequence model as the experimental baseline model, and in order to ensure the fairness of the experimental comparison, we all used the same data division to obtain the training set, validation set and test set, the training set was used for model training and the best model on the validation set was saved during the training process, and finally the results of the model on the test set were recorded and recorded in the following table.

[0039] Table 1 Performance comparison of FPR_MSE_MLP and other baseline models From Table 1, it is not difficult to see that FPR_ESM_MLP has achieved good performance on Davis and Kiba datasets, which shows that FPR_ESM_MLP has strong drug target affinity prediction ability and generalization, especially in the MSE index, which is better than the previous optimal method SGC_SAGE on the two datasets, and the prediction error is reduced by 5.29% and 6.30% respectively. Since the regression task directly optimizes the MSE loss, it also fully demonstrates the superiority and robustness of FPR_ESM_MLP in regression problems.

[0040] The above embodiments are preferred embodiments of the present application, and those skilled in the art can also make various transformations or improvements on the basis of this, and these transformations or improvements should belong to the scope of protection required by the present application without departing from the patent concept of the present application.

Claims

1. A method for predicting drug target affinity, characterized in that, Includes the following steps: (1) Input the small molecule SMILES sequence and the protein amino acid sequence; (2) Ten molecular fingerprint features, namely RDK, Topological, MACCS, AtomPair, ECFP4, FCFP4, FCFP6, Avalon, Layered, and Pattern, were extracted from the small molecules SMILES using a molecular fingerprint algorithm and then concatenated to obtain a molecular hybrid fingerprint feature vector with a length of 16896. ; (3) Use the ESM2 pre-trained model to extract features from the protein sequence to obtain a protein feature vector of length 320. ; (4) The molecular characteristics Projected features are obtained by mapping to 320 dimensions through a linear layer. and with protein characteristics The splicing forms a joint distribution feature with a length of 640 dimensions; (5) The joint distribution features of the 640 dimensions The input is fed into a multi-expert hybrid module to obtain the final fused features; (6) The final fusion features are output to 1D through a linear layer to obtain the predicted affinity between the drug and the target.

2. A method for predicting drug target affinity according to claim 1, characterized in that, The molecular fingerprint vectors in step (2) were all calculated using the RDKit tool.

3. A method for predicting drug target affinity according to claim 1, characterized in that, The protein feature extraction in step (3) is obtained by max pooling the last hidden state of the ESM2 pre-trained model: in This represents the sequence feature output of the ESM model. This indicates the number of amino acids in a protein. Indicates in dimension The protein characteristics after maximal pooling.

4. A method for predicting drug target affinity according to claim 1, characterized in that, The parameter matrix of the linear mapping layer in step (4) Bias term To calculate And achieve dimensionality matching and balanced feature contributions: .

5. A method for predicting drug target affinity according to claim 1, characterized in that, The multi-expert hybrid module in step (5) includes There are 10 experts, each consisting of a linear layer network activated using ReLU: in This indicates the combined distribution characteristics of proteins and small molecules. This represents the feature transformation of the i-th expert. This represents the output of the i-th expert, followed by concatenation of the outputs of k experts, and then processed through a routing network. Merge the results: in d represents the final fused feature, and d represents the dimension of the fused feature.

6. A method for predicting drug target affinity according to claim 1, characterized in that, The linear layer in step (6) is defined as an output projection layer from d-dimensional space to 1-dimensional space. : .

7. A method for predicting drug target affinity according to any one of claims 1 to 6, characterized in that, The Huber Loss is used as the loss function during model training, and its definition is as follows: in These represent the predicted affinity values ​​between the drug and the target, respectively, for the current input. This is a hyperparameter in the loss function used to control the learning curve.

8. A method for predicting drug target affinity according to claim 6, characterized in that... 。 9. A method for predicting drug target affinity according to claim 6, characterized in that... 。 10. A method for predicting drug target affinity according to claim 7, characterized in that... δ is set to 1.0, and the optimization algorithm used is AdamW with a learning rate of 3e−4.

11. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, which, when executed by the processor, implements a drug target affinity prediction method according to any one of claims 1 to 10.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements a drug target affinity prediction method according to any one of claims 1 to 10.

Images