Method and device for predicting activity of multi-view molecular pairs based on twin neural network

Abstract

The present invention discloses a method and device for predicting the activity of multi-view molecule pairs based on twin neural networks, including: constructing a molecule pair activity data set containing a protein target; extracting the initial embedding representations of two molecules and the initial embedding representations of the protein target respectively; constructing a multi-view molecule pair activity prediction model with a twin neural network architecture; optimizing the parameters of the multi-view molecule pair activity prediction model for pre-training tasks and fine-tuning tasks using the activity data of each pair of molecule pairs; and predicting the activity of molecule pairs using the multi-view molecule pair activity prediction model after parameter optimization. The method and device obtain the intramolecular interaction representation and the protein-ligand interaction representation through the multi-view molecule pair activity prediction model, and use the twin neural network to learn the activity differences between molecule pairs, and finally achieve efficient and accurate prediction of molecular activity, so as to better guide the optimization of lead compounds.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence drug technology, and specifically relates to a method and device for predicting the activity of multi-view molecular pairs based on twin neural networks. Background Art

[0002] New drug development is a highly professional, long-term, costly and high-risk project. Its process includes the determination of drug targets, discovery of lead compounds, optimization of lead compounds, determination of preclinical candidate compounds, clinical research, approval and marketing, and post-marketing monitoring. As the starting point of the entire new drug development process, lead compounds only have preliminary activity and need to be transformed into lead compounds with higher biological activity and better drugability using lead compound optimization methods. In the process of lead compound optimization, the properties of the drug molecule itself and the way it interacts with the target are particularly important and need to be continuously measured and verified.

[0003] The rapid development of deep learning technology has brought new possibilities for the optimization of lead compounds. By building a model to learn a large amount of molecular activity data, the model can accurately predict the activity of unknown molecules, efficiently explore the chemical structure space, and screen out modified lead compound molecules with higher activity. For example, the patent application with publication number CN115331750A discloses a method for predicting the activity of new target compounds based on deep learning. By building CNN, BiLSTM and other models to learn the characterization of molecules, the prediction of molecular activity is achieved. However, this method does not consider the interaction between protein targets and molecules, and the prediction accuracy is insufficient. For example, the patent application with publication number CN114822717A discloses a drug molecule processing method based on artificial intelligence. By performing activity prediction processing on multiple candidate drug molecules and target proteins, the activity information of each candidate drug molecule is obtained. However, this method simply merges protein features and molecular features by splicing or adding them together, without taking advantage of the activity differences between candidate molecules, and losing part of the information and training efficiency. Summary of the invention

[0004] In view of the above, the purpose of the present invention is to provide a multi-view molecule pair activity prediction method and device based on a twin neural network to achieve activity prediction of molecules.

[0005] To achieve the above-mentioned purpose of the invention, the embodiment provides a method for predicting the activity of a multi-view molecule pair based on a twin neural network, comprising the following steps:

[0006] Constructing a molecule pair activity data set including a protein target, wherein each pair of molecule pair activity data includes activity data of one molecule, activity data of another molecule, and a protein target;

[0007] Extracting initial embedding representations of the two molecules and initial embedding representations of the protein target respectively;

[0008] Constructing a multi-view molecule pair activity prediction model with a twin neural network architecture, including a multi-view feature enhancement extraction module and an activity prediction module, wherein the multi-view feature enhancement extraction module is used to perform feature enhancement extraction of local views and global views based on the initial embedding representation of each molecule and the initial embedding representation of the protein target to obtain local features and global features corresponding to each molecule, and the activity prediction module is used to predict the activity difference of two molecules based on the local features and global features corresponding to the two molecules, or predict the activity of another molecule when the activity of one molecule is known;

[0009] Utilizing the activity data of each pair of molecule pairs, the parameters of the pre-training task and the fine-tuning task of the multi-view molecule pair activity prediction model are optimized, wherein the pre-training task is to predict the activity difference between two molecules, and the fine-tuning task includes predicting the activity of another molecule based on the known activity of one molecule;

[0010] The molecule pair activity prediction was performed using the multi-view molecule pair activity prediction model after parameter optimization.

[0011] Preferably, the multi-view feature enhancement extraction module includes a local view module and a global view module, wherein the local view module is used to extract local features of intramolecular interactions based on an initial embedding representation of each molecule, and the global view module is used to extract global features of interactions between molecules and proteins based on an initial embedding representation of each molecule and an initial embedding representation of a protein target.

[0012] Preferably, the local view module comprises a plurality of local feature extraction layers, each of which comprises a self-attention layer and a first nonlinear layer, and the initial embedding representation of each molecule sequentially passes through the self-attention layer and the first nonlinear layer to extract the local features of the internal interactions;

[0013] The global view module includes global feature extraction layers with the same number of layers as the local feature extraction layers. Each global feature extraction layer includes a cross-attention layer and a second nonlinear layer, and also shares the self-attention layer in the local feature extraction layer. The initial embedding representation of each molecule is cross-attended with the initial embedding representation of the protein target in the cross-attention layer after the features are extracted by the self-attention layer to extract the global contextual information of the interaction between the molecular features and the protein, and then mapped by the second nonlinear layer to obtain the global features.

[0014] Preferably, the method further includes: performing normalized weighted processing on the local features and global features extracted by the multi-view feature extraction module and then inputting the processed features into the activity prediction module for activity prediction.

[0015] Preferably, in the activity prediction module, the activity difference between two molecules is predicted based on the local features corresponding to the two molecules and the global features, including:

[0016] The difference between the local features of the two molecules is input into the multi-layer perceptron, and the difference between the global features of the two molecules is also input into the multi-layer perceptron, which calculates and outputs the activity difference between the two molecules.

[0017] Preferably, in the activity prediction module, predicting the activity of another molecule when the activity of one molecule is known comprises:

[0018] Based on the difference in the activities of the two molecules obtained, if the activity of one molecule is known, the activity of the other molecule is the sum of the known activity of the one molecule and the difference in the activities of the two molecules.

[0019] Preferably, supervised learning is used for parameter optimization of the multi-view molecule pair activity prediction model. For pre-training tasks, a loss function is constructed based on the difference between the predicted activity difference and the actual activity difference for parameter optimization. For fine-tuning tasks, a loss function is constructed based on the difference between the predicted activity of another molecule and the actual activity of another molecule for parameter optimization.

[0020] To achieve the above-mentioned purpose of the invention, the embodiment of the present invention further provides a multi-view molecular pair activity prediction device based on a twin neural network, comprising:

[0021] A data construction unit, which is used to construct a molecule pair activity data set containing a protein target, each pair of molecule pair activity data includes first molecule activity data, second molecule activity data, and the protein target;

[0022] An embedding representation unit, which is used to extract initial embedding representations of two molecules and initial embedding representations of protein targets respectively;

[0023] A model building unit, which is used to build a multi-view molecule pair activity prediction model with a twin neural network architecture, including a multi-view feature enhancement extraction module and an activity prediction module, wherein the multi-view feature enhancement extraction module is used to perform feature enhancement extraction of local views and global views based on an initial embedding representation of each molecule and an initial embedding representation of a protein target to obtain local features and global features corresponding to each molecule, and the activity prediction module is used to predict the activity difference between two molecules based on the local features and global features corresponding to the two molecules, or to predict the activity of another molecule when the activity of one molecule is known;

[0024] A parameter optimization unit, which is used to optimize the parameters of the pre-training task and the fine-tuning task of the multi-view molecule pair activity prediction model using the activity data of each pair of molecule pairs, wherein the pre-training task is to predict the activity difference between two molecules, and the fine-tuning task includes predicting the activity of another molecule based on the known activity of one molecule;

[0025] The reasoning application unit is used to perform molecule pair activity prediction using the multi-view molecule pair activity prediction model after parameter optimization.

[0026] To achieve the above-mentioned purpose of the invention, an embodiment also provides a computing device, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned multi-view molecular pair activity prediction method based on twin neural network.

[0027] To achieve the above-mentioned purpose of the invention, an embodiment further provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the above-mentioned multi-view molecular pair activity prediction method based on a twin neural network.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] Through the multi-view molecular pair activity prediction model, we can obtain the intramolecular interaction characterization and the protein-ligand interaction characterization, and use the twin neural network to learn the activity differences between molecular pairs, ultimately achieving efficient and accurate prediction of molecular activity, which can be used to better guide the optimization of lead compounds. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0031] Figure 1 is a flow chart of a method for predicting activity of multi-view molecular pairs based on a twin neural network provided in an embodiment;

[0032] Figure 2 is a schematic diagram of the structure of the multi-view molecule pair activity prediction model provided in the embodiment;

[0033] Figure 3 is a schematic diagram of the structure of a multi-view feature enhancement extraction module provided in an embodiment;

[0034] Figure 4 is a prediction flow chart of a multi-view molecule pair activity prediction model provided in an embodiment;

[0035] Figure 5 It is a schematic diagram of the structure of a multi-view molecular pair activity prediction device based on a twin neural network provided in an embodiment. DETAILED DESCRIPTION

[0036] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0037] like Figure 1 As shown, the multi-view molecular pair activity prediction method based on a twin neural network provided in the embodiment includes the following steps:

[0038] S1, Construction of a molecular pair activity dataset containing protein targets.

[0039] In the embodiment, the molecule pair activity data set includes multiple molecule pair activity data, wherein each molecule pair activity data includes one molecule activity data, another molecule activity data, and a protein target. The format of each pair of molecule pair activity data is [molecule 1, molecule 2, protein target, activity 1, activity 2], wherein molecule 1 and its corresponding activity 1 constitute one molecule activity data, and molecule 2 and its corresponding activity 2 constitute another molecule activity data. Each molecule in the data is represented by a SMILES string, the protein is represented by its UniprotID number in the UniProt database, and the activity is represented in the format of the negative logarithm value pKi of the inhibition constant. The specific construction process includes:

[0040] Data including protein target data, small ligand molecule data, and corresponding activity categories and activity values ​​were obtained from the BindingDB database. The original data were filtered to retain only data with affinity type (affinity type) as inhibition constant (Ki), and the negative logarithm of the inhibition constant (pKi) was calculated. The ligand molecules were combined in pairs under the same protein target to obtain a molecular pair activity data set containing protein target data.

[0041] In particular, by screening the data with molecular Tanimoto similarity>0.6 and pKi>2, we can obtain the activity cliff subset of the molecular pair activity data, which is used for model training and evaluation of its performance when dealing with compound pairs with similar structures but significant differences in activity. The constructed molecular pair activity dataset containing protein target data can be used not only for the activity prediction of molecular pairs, but also for protein-molecule interaction prediction and activity cliff problems.

[0042] S2, extracts the initial embedding representations of the two molecules and the initial embedding representation of the protein target, respectively.

[0043] In the embodiment, the pre-trained protein feature extraction model and the molecule feature extraction model are used to extract the initial embedding representation of the protein and the initial embedding representation of the molecule respectively. Taking the pre-trained UniMol model as an example, the frozen UniMol model is used to extract the initial embedding representation of the protein and the initial embedding representation of the molecule, including the UniMol protein model and the UniMol molecule model, and the specific process is:

[0044] Obtain protein target 3D data, where the protein target 3D data includes the name of the protein target (TargetName) and its number (UniprotID) in the UniProt database. Use UniprotID to obtain the 3D protein structure predicted by the AlphaFold2 model from the DeepMind company's website, and download and save it in the format of a .pdb file; obtain molecular 3D data, specifically generate the 3D conformation of the molecule based on the SMILES string of the molecule using the RDKit library; use the weight file pocket_pre_220816.pt obtained by UniMol training on 3M protein data as the pre-trained weight of the UniMol protein model to extract the initial embedding representation of the protein, and use the weight file mol_pre_no_h_220816.pt obtained by UniMol training on 209M molecular 3D conformation data as the pre-trained weight of the UniMol molecular model to extract the initial embedding representation of the molecule. The extracted initial embedding representation of the molecule consists of the overall representation of the molecule and the representation of atoms other than hydrogen atoms.

[0045] S3, construct a multi-view molecular pair activity prediction model with a Siamese neural network architecture.

[0046] like Figure 2 As shown, the constructed multi-view molecule pair activity prediction model with a twin neural network architecture includes a multi-view feature enhancement extraction module and an activity prediction module, wherein the multi-view feature enhancement extraction module is used to perform feature enhancement extraction of local views and global views based on the initial embedding representation of each molecule and the initial embedding representation of the protein target to obtain the local features and global features corresponding to each molecule, and the activity prediction module is used to predict the activity difference between two molecules based on the local features and global features corresponding to the two molecules, or to predict the activity of another molecule when the activity of one molecule is known.

[0047] like Figure 3 As shown, the multi-view feature enhancement extraction module includes a local view module and a global view module, wherein the local view module is used to extract local features of intramolecular interactions based on the initial embedding representation of each molecule, and the global view module is used to extract global features of interactions between molecules and proteins based on the initial embedding representation of each molecule and the initial embedding representation of the protein target.

[0048] Specifically, the local view module includes N local feature extraction layers, each of which includes a self-attention layer and a first nonlinear layer. The initial embedding representation of each molecule or the output of the previous local feature extraction layer passes through the self-attention layer and the first nonlinear layer in sequence to extract the local features of internal interactions.

[0049] The global view module includes global feature extraction layers with the same number of layers as the local feature extraction layers, that is, it includes N global feature extraction layers, each of which includes a cross-attention layer and a second nonlinear layer, and also shares the self-attention layer in the local feature extraction layer, that is, the global feature extraction layer also includes a self-attention layer, and the self-attention layer shares the weight parameters of the local feature extraction layer. The initial embedding representation of the molecule or the output of the previous global feature extraction layer is cross-attended with the initial embedding representation of the protein target in the cross-attention layer after feature extraction by the self-attention layer to extract the global context information of the interaction between the molecular features and the protein, and then mapped by the second nonlinear layer to obtain the global features.

[0050] In an embodiment, both the first nonlinear layer and the second nonlinear layer may adopt a feedforward neural network.

[0051] Optionally, in the embodiment, the local features and global features extracted by the multi-view feature extraction module are subjected to normalized weighted processing and then input into the activity prediction module for activity prediction. The specific process is as follows:

[0052] Local features output by the local view module , first normalize the global features on the feature dimension (i.e. the last dimension) to obtain the normalized post-processing result , and then normalize the result on the sequence length dimension (i.e. the intermediate dimension) Use the softmax function to perform normalized exponential processing to obtain the weight sequence ,in is the i-th dimension feature, n is the number of dimensions, express The corresponding weight sequence is calculated as follows:

[0053] ;

[0054] ;

[0055] The weight sequence and Perform weighted summation to obtain the final local features representing intramolecular interactions .

[0056] Global features output by the global view module The final global features are obtained by normalizing and weighting in the same way as the local features. .

[0057] In the embodiment, Figure 4 As shown, in the activity prediction module, the activity difference between the two molecules is predicted based on the local features corresponding to the two molecules and the global features, including: the difference between the local features of the two molecules Input to the multi-layer perceptron MLP, the difference between the global features of the two molecules The same input is sent to the multi-layer perceptron MLP, which calculates and outputs the activity difference between the two molecules. ,Right now ,in and Represent the local features of molecule 1 and molecule 2 respectively, and Represent the global features of molecule 1 and molecule 2 respectively;

[0058] Predicting the activity of another molecule when the activity of one molecule is known, comprising: based on the difference in the activities of the two molecules obtained, predicting the activity of one molecule , then the activity of another molecule is the sum of the difference between the activity of one known molecule and the activity of two molecules, that is, ;

[0059] In an embodiment, the MLP may be a multi-layer perceptron including a linear layer, a Relu activation layer, and a linear layer.

[0060] S4, uses the activity data of each pair of molecule pairs to optimize the parameters of the multi-view molecule pair activity prediction model for pre-training tasks and fine-tuning tasks.

[0061] In the embodiment, the parameters of the multi-view molecule pair activity prediction model are optimized through two-stage training based on a pre-training task and a fine-tuning task, wherein the pre-training task is to predict the activity difference between two molecules, and the fine-tuning task includes predicting the activity of another molecule based on the known activity of one molecule.

[0062] When optimizing parameters, supervised learning is used. For pre-training tasks, a loss function is constructed based on the difference between the predicted activity difference and the actual activity difference to optimize parameters. For fine-tuning tasks, a loss function is constructed based on the difference between the predicted activity of another molecule and the actual activity of another molecule to optimize parameters. Specifically, each loss function can use the root mean square error RMSE.

[0063] S5, use the multi-view molecule pair activity prediction model with optimized parameters to perform molecule pair activity prediction.

[0064] The multi-view molecule pair activity prediction model with optimized parameters can be used for molecule pair activity prediction and realize activity prediction and analysis of unknown molecules, including activity difference prediction of two molecules and activity prediction of another unknown molecule based on the activity of known molecules. When data is input in the format of [molecule 1, molecule 2, protein target], the multi-view molecule pair activity prediction model will predict the activity difference and output the relative activity difference of the molecule pair. When data is input in the format of [molecule 1, molecule 2, protein target, activity 1], the multi-view molecule pair activity prediction model will predict the activity of the unknown molecule, that is, output the activity 2 of molecule 2.

[0065] The method of the present invention is an efficient and accurate method for predicting the activity of molecular pairs. It is not only applicable to the prediction of the activity of molecular pairs, but can also be extended to the prediction of the activity and relative activity of other biological macromolecule pairs. Its flexibility and generalization enable it to meet different drug discovery needs.

[0066] like Figure 5 As shown, the embodiment also provides a multi-view molecule pair activity prediction device based on a twin neural network, including a data construction unit 51, an embedding representation unit 52, a model construction unit 53, a parameter optimization unit 54, and an inference application unit 55, wherein the data construction unit 51 is used to construct a molecule pair activity data set containing a protein target; the embedding representation unit 52 is used to extract the initial embedding representations of two molecules and the initial embedding representations of the protein target respectively; the model construction unit 53 is used to construct a multi-view molecule pair activity prediction model with a twin neural network architecture; the parameter optimization unit 54 is used to optimize the parameters of the multi-view molecule pair activity prediction model for pre-training tasks and fine-tuning tasks using the activity data of each pair of molecule pairs; the inference application unit 55 is used to use the multi-view molecule pair activity prediction model after parameter optimization to predict the activity of molecule pairs.

[0067] It should be noted that when the multi-view molecular pair activity prediction device based on the twin neural network provided in the above embodiment performs molecular pair activity prediction, the division of the above-mentioned functional modules or units should be used as an example to illustrate that the above-mentioned functions can be assigned to different functional modules or units as needed, that is, the internal structure of the terminal or server is divided into different functional modules or units to complete all or part of the functions described above. In addition, the multi-view molecular pair activity prediction device based on the twin neural network provided in the above embodiment belongs to the same concept as the multi-view molecular pair activity prediction method embodiment based on the twin neural network. The specific implementation process is detailed in the multi-view molecular pair activity prediction method embodiment based on the twin neural network, which will not be repeated here.

[0068] Based on the same inventive concept, an embodiment further provides a computing device, including a memory and one or more processors, wherein executable code is stored in the memory, and when the one or more processors execute the executable code, the method for predicting the activity of multi-view molecular pairs based on a twin neural network is implemented, specifically including the following steps:

[0069] S1, constructing a molecular pair activity dataset containing protein targets;

[0070] S2, extracts the initial embedding representations of the two molecules and the initial embedding representations of the protein target, respectively;

[0071] S3, building a multi-view molecular pair activity prediction model with a Siamese neural network architecture;

[0072] S4, using the activity data of each pair of molecule pairs to perform parameter optimization of the pre-training task and fine-tuning task for the multi-view molecule pair activity prediction model;

[0073] S5, use the multi-view molecule pair activity prediction model with optimized parameters to perform molecule pair activity prediction.

[0074] The computing device provided in the embodiment, in addition to the processor and memory, also includes hardware required for other services such as internal bus, network interface, memory, etc. at the hardware level. The memory is a non-volatile memory, and the processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the multi-view molecular pair activity prediction method based on twin neural network described in S1-S5 above. Of course, in addition to software implementation, the present invention does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc., that is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0075] Based on the same inventive concept, an embodiment further provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the above-mentioned multi-view molecular pair activity prediction method based on a twin neural network is implemented, which specifically includes the following steps:

[0076] S1, constructing a molecular pair activity dataset containing protein targets;

[0077] S2, extracts the initial embedding representations of the two molecules and the initial embedding representations of the protein target, respectively;

[0078] S3, building a multi-view molecular pair activity prediction model with a Siamese neural network architecture;

[0079] S4, using the activity data of each pair of molecule pairs to perform parameter optimization of the pre-training task and fine-tuning task for the multi-view molecule pair activity prediction model;

[0080] S5, use the multi-view molecule pair activity prediction model with optimized parameters to perform molecule pair activity prediction.

[0081] In the embodiment, computer-readable media include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data.

[0082] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-view molecular pair activity prediction method based on twin neural networks, characterized in that: The following steps are involved: Constructing a molecule pair activity data set including a protein target, wherein each pair of molecule pair activity data includes activity data of one molecule, activity data of another molecule, and a protein target; Extracting initial embedding representations of the two molecules and initial embedding representations of the protein target respectively; Constructing a multi-view molecule pair activity prediction model with a twin neural network architecture, including a multi-view feature enhancement extraction module and an activity prediction module, wherein the multi-view feature enhancement extraction module is used to perform feature enhancement extraction of local views and global views based on the initial embedding representation of each molecule and the initial embedding representation of the protein target to obtain local features and global features corresponding to each molecule, and the activity prediction module is used to predict the activity difference of two molecules based on the local features and global features corresponding to the two molecules, or predict the activity of another molecule when the activity of one molecule is known; The multi-view feature extraction module includes a local view module and a global view module, wherein the local view module is used to extract local features of intramolecular interactions based on the initial embedding representation of each molecule, and the global view module is used to extract global features of interactions between molecules and proteins based on the initial embedding representation of each molecule and the initial embedding representation of the protein target; Utilizing the activity data of each pair of molecule pairs, the parameters of the pre-training task and the fine-tuning task of the multi-view molecule pair activity prediction model are optimized, wherein the pre-training task is to predict the activity difference between two molecules, and the fine-tuning task includes predicting the activity of another molecule based on the known activity of one molecule; The molecule pair activity prediction was performed using the multi-view molecule pair activity prediction model after parameter optimization.

2. The multi-view molecular pair activity prediction method based on twin neural network according to claim 1, characterized in that: The local view module includes multiple local feature extraction layers, each of which includes a self-attention layer and a first nonlinear layer. The initial embedding representation of each molecule sequentially passes through the self-attention layer and the first nonlinear layer to extract the local features of internal interactions. The global view module includes global feature extraction layers with the same number of layers as the local feature extraction layers. Each global feature extraction layer includes a cross-attention layer and a second nonlinear layer, and also shares the self-attention layer in the local feature extraction layer. The initial embedding representation of each molecule is cross-attended with the initial embedding representation of the protein target in the cross-attention layer after the features are extracted by the self-attention layer to extract the global contextual information of the interaction between the molecular features and the protein, and then mapped by the second nonlinear layer to obtain the global features.

3. The multi-view molecular pair activity prediction method based on twin neural network according to claim 1 or 2, characterized in that: The method further comprises: performing normalized weighted processing on the local features and global features extracted by the multi-view feature extraction module and then inputting the processed features into the activity prediction module for activity prediction.

4. The multi-view molecular pair activity prediction method based on twin neural network according to claim 1, characterized in that: In the activity prediction module, the activity difference between two molecules is predicted based on the local features and global features of the two molecules, including: The difference between the local features of the two molecules is input into the multi-layer perceptron, and the difference between the global features of the two molecules is also input into the multi-layer perceptron, which calculates and outputs the activity difference between the two molecules.

5. The method for predicting activity of multi-view molecular pairs based on twin neural networks according to claim 4, characterized in that: In the activity prediction module, the activity of another molecule is predicted when the activity of one molecule is known, including: Based on the difference in the activities of the two molecules obtained, if the activity of one molecule is known, the activity of the other molecule is the sum of the known activity of the one molecule and the difference in the activities of the two molecules.

6. The multi-view molecular pair activity prediction method based on twin neural network according to claim 1, characterized in that: Supervised learning is used to optimize the parameters of the multi-view molecular pair activity prediction model. For the pre-training task, the loss function is constructed based on the difference between the predicted activity difference and the actual activity difference for parameter optimization. For the fine-tuning task, the loss function is constructed based on the difference between the predicted activity of another molecule and the actual activity of another molecule for parameter optimization.

7. A multi-view molecular pair activity prediction device based on twin neural networks, characterized in that: include: A data construction unit, which is used to construct a molecule pair activity data set containing a protein target, each pair of molecule pair activity data includes first molecule activity data, second molecule activity data, and the protein target; An embedding representation unit, which is used to extract initial embedding representations of two molecules and initial embedding representations of protein targets respectively; A model building unit, which is used to build a multi-view molecule pair activity prediction model with a twin neural network architecture, including a multi-view feature enhancement extraction module and an activity prediction module, wherein the multi-view feature enhancement extraction module is used to perform feature enhancement extraction of local views and global views based on an initial embedding representation of each molecule and an initial embedding representation of a protein target to obtain local features and global features corresponding to each molecule, and the activity prediction module is used to predict the activity difference between two molecules based on the local features and global features corresponding to the two molecules, or to predict the activity of another molecule when the activity of one molecule is known; The multi-view feature extraction module includes a local view module and a global view module, wherein the local view module is used to extract local features of intramolecular interactions based on the initial embedding representation of each molecule, and the global view module is used to extract global features of interactions between molecules and proteins based on the initial embedding representation of each molecule and the initial embedding representation of the protein target; A parameter optimization unit, which is used to optimize the parameters of the pre-training task and the fine-tuning task of the multi-view molecule pair activity prediction model using the activity data of each pair of molecule pairs, wherein the pre-training task is to predict the activity difference between two molecules, and the fine-tuning task includes predicting the activity of another molecule based on the known activity of one molecule; The reasoning application unit is used to perform molecule pair activity prediction using the multi-view molecule pair activity prediction model after parameter optimization.

8. A computing device comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that: When the one or more processors execute the executable code, they are used to implement the multi-view molecular pair activity prediction method based on a twin neural network as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the multi-view molecular pair activity prediction method based on a twin neural network as described in any one of claims 1 to 6 is implemented.

Images