Method for predicting relative binding free energy of geometric characteristics of protein mutation structure

Through the integration of AlphaFold's geometric characteristics and deep learning models to predict protein mutation structures, the difficulty of predicting drug binding ability is solved, the prediction accuracy and reliability are improved, and better guidance is provided for drug design.

CN119068982APending Publication Date: 2024-12-03JIANGSU UNIV OF TECH
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411119980.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The prior art is difficult to effectively predict the impact of protein mutation structure on drug binding capacity, especially when the protein structure is time-consuming and costly to resolve, it is impossible to fully understand all possible mutation states.

Method used

AlphaFold is used to predict the geometric characteristics of protein mutation structure, and combine deep learning models (such as Transformer-Encoder, GRU, GCN/SGCN) to integrate protein sequence, deformation data and small molecule map data to predict the effect of mutations on binding free energy.

Benefits of technology

Improves the accuracy and reliability of the prediction model, and allows more comprehensive capture of key features of protein-small interactions, helping to understand and optimize drug design, especially to provide better guidance in drug resistance studies.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068982A_ABST
    Figure CN119068982A_ABST
Patent Text Reader

Abstract

The invention discloses a method for predicting relative binding free energy of geometrical characteristics of a protein mutation structure. The method comprises the following steps: establishing a database of proteins and small molecules; generating a three-dimensional structure of a wild type protein and a three-dimensional structure of a mutant protein by using AlphaFold; constructing a character token dictionary of the protein sequences; calculating a CA atomic distance matrix between the mutant and the wild-type protein to obtain a contact diagram of the mutant; respectively calculating deformation data of the wild type protein and the mutant protein; respectively converting the contact graph and the small molecule of the mutant into graph data, and integrating the graph data with integer codes and deformation data of a protein sequence in an iterator to form a training set and a test set; testing and training the deep learning model; and verifying a prediction result of the deep learning model. The prediction accuracy is further improved, the method can be used for calculating the influence of the change of binding free energy caused by protein mutation on drug resistance, and more guidance and help are provided for drug design.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method for predicting protein mutation structure, and in particular to a method for introducing AlphaFold to predict relative binding free energy of geometric features of protein mutation structure. Background Art

[0002] The interaction between proteins and small molecules involves a series of complex physical and chemical processes, the core of which is the change in free energy of interaction. Understanding and predicting the change in free energy of interaction is of great significance for drug development. With the improvement of computer performance, especially the widespread use of graphics processing units (GPUs), physical computing methods and artificial intelligence-based methods have made a series of progress in the field of drug development.

[0003] Experimental-based methods bear considerable experimental time and financial costs. With the demand for high throughput, physics-based computational methods such as molecular dynamics simulation, molecular docking, free energy perturbation FEP, etc. have been proposed. These methods are challenging in some aspects, including the selection of force field parameters, sufficient sampling, and computing resources. With the development of deep learning, artificial intelligence-based methods have shown great potential in the prediction of protein-small molecule interactions. These methods can quickly predict free energy by learning a large amount of known data.

[0004] Sequence- and structure-based methods have made significant progress in predicting protein-ligand interactions, which have greatly accelerated the process of studying protein-ligand interactions. For protein representation, one-hot encoding can be used to construct a structural descriptor based on the amino acid sequence using evolutionary information, and a contact map can be constructed based on the spatial distance between amino acids. For ligand representation, simplified molecular input line input system (SMILES), one-hot encoding, chemical fingerprints or molecular graphs can be used. Different protein and ligand representation methods provide different perspectives in predicting protein-ligand interactions.

[0005] Machine learning and deep learning have played an important role in predicting protein-small molecule interactions. By training models, these methods can learn complex patterns of protein-small molecule interactions from large amounts of data, thereby improving the accuracy and efficiency of predictions. Currently popular artificial intelligence models have been applied to the field of drug research, such as Transformer, BERT, GCN, and GraphDTA. Wallach et al. constructed a structure-based deep convolutional neural network for predicting the biological activity of small molecules. The researchers extracted features from the complex 3D network of proteins (Wallach; arxivpreprint arxiv:1510.02855; 2020). et al. proposed DeepDTA, which uses CNN to extract one-dimensional sequence information of protein and ligand SMILES. The performance of this method is better than SmiBoost's DT binding affinity prediction method ( Bioinformatics; 2018; 37). Bennett et al. used atomic dynamics simulation to construct a dataset for predicting small molecule transfer free energy, and compared spatial graph neural networks, 3D convolutional neural networks, and shallow learning methods based on chemical fingerprints. The results showed that the structural information of the molecule can improve the prediction performance (Bennett; Journal of Chemical Information and Modeling; 2020; 60).

[0006] In order to have a more comprehensive understanding of the interaction between proteins and small molecules, in recent years, more and more studies have begun to focus on integrating the characteristic information of proteins and small molecules, such as DeepDTA, GraphDTA, MolTrans, and Modality-DTA. Its characteristic coding includes integer coding of proteins and small molecules, construction of protein and small molecule fingerprint coding or small molecule graph information. There may be the following problems: 1) The one-dimensional coding of protein sequences can capture the position and composition information of each amino acid in the sequence, and the SMILES and chemical fingerprints of small molecules can effectively describe their molecular structure and chemical characteristics. However, this one-dimensional coding usually cannot capture richer spatial information and has certain limitations. 2) The graph information of small molecules can provide detailed molecular structure, including the connection mode between atoms, the chemical environment in the environment, and possible conformations. Considering only the graph information of small molecules ignores the influence of proteins on interactions.

[0007] The three-dimensional structure of the combined protein can more accurately predict and analyze the interaction between proteins and small molecules. Zheng et al. used BiLSTM and DynCNN models to learn small molecule representations based on SMILES integer encoding and protein representations based on two-dimensional distance graphs, respectively. This method achieved good performance on the BindingDB benchmark dataset (Zheng; Nature Machine Intelligence; 2020, 2). Lu et al. used the two-dimensional graph of proteins and ligands and the three-dimensional spherical query graph of the complex as fusion inputs, and the performance of the model was better than previous models (Lu; Briefings in Bioinformatics; 2023; 24).

[0008] However, previous studies were all based on the crystal structure of proteins. Since protein databases only contain partial structural information, how to overcome the lack of protein three-dimensional structure and conduct structure-based drug design has become a difficult problem in the field of drug research and development. For example, when studying drug resistance, it is necessary to study the effect of the mutant structure of the protein on the drug binding ability. Generally, there are few known mutant structures of proteins. Since the structural analysis of proteins requires a lot of time and money, it is not possible to perform structural analysis on all possible mutant structures. Taking ABL1 kinase as an example, multiple mutations have been reported for this protein, and some mutations can cause drug resistance. In order to measure the impact of different mutant structures on drug resistance, researchers need to obtain more structures of proteins in mutant states. Summary of the invention

[0009] Purpose of the invention: The purpose of the present invention is to provide a method for predicting the relative binding free energy of geometric features of protein mutation structures, solving the problem of how to predict the change in binding free energy caused by protein mutation. Another purpose of the present invention is to provide a computer-readable storage medium that stores and is used by a processor to execute the above prediction method.

[0010] Technical solution: A method for predicting the relative binding free energy of geometric features of protein mutation structures according to the present invention comprises the following steps:

[0011] Establish databases of proteins and small molecules;

[0012] Generate three-dimensional structures of wild-type and mutant proteins using AlphaFold;

[0013] Construct a dictionary of character tokens for protein sequences;

[0014] The CA atomic distance matrix between the mutant and wild-type proteins was calculated to obtain the contact map of the mutant;

[0015] The deformation data of wild-type and mutant proteins were calculated separately;

[0016] The contact maps of mutants and small molecules are converted into graph data respectively, and integrated with the integer encoding of protein sequences and deformation data in the iterator to form training and test sets;

[0017] Establish a deep learning model in the server, train it using the training set, and test the trained deep learning model using the test set;

[0018] Verify the prediction results of the deep learning model.

[0019] Preferably, the establishment of a database of proteins and small molecules comprises:

[0020] The data of kinase inhibitor molecules and protein sequences for studying point mutation resistance are obtained from the database; the sequence of wild-type protein, mutation site and experimental value of binding free energy change between molecule and protein are obtained from the literature; these data are stored in the local database to obtain a database of proteins and small molecules.

[0021] Preferably, the use of AlphaFold to generate three-dimensional structures of wild-type and mutant proteins comprises:

[0022] According to the wild-type protein sequence and mutation sites in the protein and small molecule database, the letters at the specified positions of the wild-type sequence are modified to obtain the mutant protein sequence. The wild-type and mutant protein sequences are uniformly stored in a csv file, and the AlphaFold structure prediction model is used to generate a pdb file storing the three-dimensional structure of the wild-type and mutant proteins.

[0023] Preferably, the construction of a character token dictionary of a protein sequence comprises: reading a csv file storing sequence information, character tokenizing all sequences, mapping each character to a unique index, and forming a dictionary character, which is saved as a pickle file.

[0024] Preferably, the calculation of the CA atomic distance matrix between the mutant and the wild-type protein comprises:

[0025] Load the pdb files of mutant and wild-type proteins; perform structural comparison between wild-type and mutant and obtain the coordinates of CA atoms after comparison; calculate the distance between each CA atom in the mutant and other CA atoms, and record the distance if the distance is less than 10A; save the calculated distance matrix as a csv file.

[0026] Preferably, the step of calculating the deformation data of the wild-type and mutant proteins comprises:

[0027] Load the pdb files of mutant and wild-type proteins; for each variant, first set the neighbor cutting distance and minimum plddt value parameters of the protein, and then calculate the deformation between the wild type and the mutant.

[0028] Preferably, said converting the contact map of the mutant and the small molecule into graph data respectively comprises:

[0029] The contact graph of the mutant is converted into graph data, including node features, edge indexes, edge features, and coordinate matrices. The node feature dimension of the mutant graph data is 56, and each atomic feature includes the one-hot encoding of each amino acid, the physicochemical properties of the amino acid, the mutation information of the mutant, and the position-specific matrix.

[0030] Small molecules are converted into graph data, including adjacency matrix, node features and edge features. The node feature dimension of the graph data of small molecules is 75, and each atomic feature includes atom type, atomic valence, and number of heteroatoms.

[0031] Preferably, the deep learning model includes at least one of dl-original, dl-withCM-GCN and dl-withCM-SGCN.

[0032] The method for verifying the prediction results of the deep learning model includes: calculating the free energy change value predicted by the deep learning model, and comparing the root mean square error, mean absolute error and Pearson correlation coefficient with the experimental results for verification.

[0033] Preferably, the construction of the dl-original model includes an MPNN layer, two two-layer Transformer-Encoders, a two-layer GRU layer, a fully connected layer, and a dropout layer with a loss rate of 0.2;

[0034] The construction of the dl-withCM-GCN model includes an MPNN layer, two two-layer Transformer-Encoder layers, a two-layer GRU, a GCN layer, a fully connected layer, and a dropout layer with a dropout rate of 0.2;

[0035] The construction of the dl-withCM-SGCN model includes an MPNN layer, two two-layer Transformer-Encoder layers, a two-layer GRU, an SGCN layer, a fully connected layer, and a dropout layer with a dropout rate of 0.2.

[0036] The protein geometric structure features predicted by AlphaFold provide rich three-dimensional information for understanding the interaction between proteins and small molecules. The contact map generated by AlphaFold can reveal the key interaction sites inside the protein and between the protein and small molecules based on the mutual distances and interactions between atoms in the protein. This information helps to better understand its mechanism and improve the prediction accuracy and reliability of the model. Therefore, the present invention selects integer encoding of protein sequences, deformation data, contact maps of mutants, and graph data of small molecules, and uses Transformer-Encoder, GRU, GCN / SGCN and MPNN for learning, and calculates the changes in binding free energy caused by protein mutations. Explore the predictive performance of changes in binding free energy caused by protein mutations with different input features.

[0037] Another aspect of the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program implements the steps of the above method when executed by a processor.

[0038] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0039] (1) The present invention introduces the geometric structural features predicted by AlphaFold, which can obtain the protein structures under different mutation states and capture the key features of the interaction more comprehensively.

[0040] (2) The present invention uses multiple feature inputs, including protein sequence information, structural information extracted from the predicted three-dimensional structure of the protein, and graph data description for small molecules. Using multiple deep learning models for feature learning and fusion helps to explore and optimize the performance of different models and feature combinations in prediction tasks.

[0041] (3) The present invention further improves the accuracy of prediction and can be used to calculate the effect of changes in binding free energy caused by protein mutations on drug resistance, providing more guidance and assistance for drug design. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the process of the present invention;

[0043] Figure 2 This is the structural diagram of the deep learning model dl-original;

[0044] Figure 3 This is the structural diagram of the deep learning model dl-withCM-GCN;

[0045] Figure 4 This is the structural diagram of the deep learning model dl-withCM-SGCN;

[0046] Figure 5 RMSE bar charts of predicted values ​​and true values ​​for molecular docking, deep learning model dl-original, deep learning model dl-withCM-GCN, and deep learning model dl-withCM-SGCN;

[0047] Figure 6 MAE bar graphs of predicted values ​​and true values ​​for molecular docking, deep learning model dl-original, deep learning model dl-withCM-GCN, and deep learning model dl-withCM-SGCN;

[0048] Figure 7Pearson coefficient bar chart of the predicted values ​​and true values ​​of molecular docking, deep learning model dl-original, deep learning model dl-withCM-GCN, and deep learning model dl-withCM-SGCN. DETAILED DESCRIPTION

[0049] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0050] Analyzing the changes in binding free energy caused by protein mutations is a key task in the field of drug design, which helps to understand how drug molecules interact chemically and physically with target proteins. This understanding is crucial for optimizing the binding affinity, selectivity and biological activity of drugs, thereby improving the efficacy of drugs and reducing adverse reactions.

[0051] The present invention proposes a relative binding free energy prediction method that introduces AlphaFold to predict the geometric characteristics of protein mutation structures, which can improve the accuracy and reliability of the prediction model. The present invention uses integer coding of protein sequences, deformation vectors, contact maps of mutants, and graph data of small molecules as input features, and uses deep learning models Transformer-Encoder, GRU, GCN / SGCN, and MPNN with different feature extraction and processing capabilities for learning. This combination of multiple features and multiple models can make full use of various types of feature information. Among them, the CA atomic distance matrix of mutants and wild-type proteins is also called the contact map.

[0052] like Figure 1 The figure shows a flow chart of the AlphaFold method for predicting the relative binding free energy of protein mutation structure geometric features provided by the present invention. As can be seen from the figure, the prediction method includes:

[0053] S1: Building a database of proteins and small molecules

[0054] Obtain the data of 8 kinase inhibitor molecules and protein sequences for studying point mutation resistance from the PDB database; obtain the sequence of wild-type protein, mutation site, and experimental value of binding free energy change between molecule and protein from the literature. Establish a local database in the local server to store the pdb files of small molecules and related files of proteins.

[0055] S2: Three-dimensional structures of wild-type and mutant proteins were generated using AlphaFold.

[0056] According to the wild-type protein sequence and mutation site in the established database, the letters at the specified position of the wild-type sequence are modified to obtain the mutant protein sequence, and the wild-type and mutant protein sequences are uniformly stored in a csv file. The AlphaFold structure prediction model is used to generate a pdb file storing the three-dimensional structure of the wild-type and mutant proteins.

[0057] S3: Building a dictionary of character tokens for protein sequences

[0058] Read the csv file storing sequence information, tokenize all sequences, map each character to a unique index, and form a dictionary character, saving it as a pickle file.

[0059] S4: Calculate the CA atomic distance matrix between the mutant and wild-type protein to obtain the contact map of the mutant

[0060] Load the pdb files of the mutant and wild-type proteins. Compare the structures of the wild-type and mutant and obtain the coordinates of the compared CA atoms. Calculate the distance between each CA atom in the mutant and other CA atoms. If the distance is less than 10A, record the distance. Save the calculated distance matrix as a csv file.

[0061] S5: Calculate deformation data for wild-type and mutant proteins separately

[0062] Load the pdb files of mutant and wild-type proteins. For each variant, first set the neighbor cutting distance and minimum plddt value parameters of the protein, and then calculate the deformation between the wild-type and mutant. Save the calculated effective deformation data as a csv file.

[0063] S6: Convert the contact map of the mutant and the small molecule into graph data respectively, and integrate them with the integer encoding of the protein sequence and the deformation data in the iterator to form the training set and the test set

[0064] The data with the relative free energy value of nd in the experiment were screened out, and the training set and test set were divided according to different kinase inhibitors. For example, the mutant data under the axitinib inhibitor in the initial data set were used as the test set, and the mutant data under the other 7 inhibitors were used as the training set. To input the data into the deep learning model, an iterator needs to be created, which contains the integer encoding of the wild-type and mutant sequences, deformation data, and the contact map of the mutant and the map data after small molecule conversion.

[0065] In this process, the protein sequences of wild-type and mutants are converted into data of 287 length according to the character token dictionary. The contact graph of the mutant is converted into graph data, including node features, edge indexes, edge features, and coordinate matrices. The node feature dimension of the mutant graph data is 56, and each atomic feature includes the one-hot encoding of each amino acid, the physicochemical properties of the amino acid, the mutation information of the mutant, and the position-specific matrix. For small molecules, they are converted into graph data, including adjacency matrices, node features, and edge features. The node feature dimension of the small molecule graph data is 75, and each atomic feature includes atom type, atomic valence, number of heteroatoms, etc.

[0066] S7: Build a deep learning model in the server, train it using the training set, and test the trained deep learning model using the test set

[0067] The deep learning models dl-original, dl-withCM-GCN and dl-withCM-SGCN were constructed and trained using the training set. The number of training times was set to 50. The trained deep learning models were tested using the test set. Among them, the deep learning model dl-original was used as a basic model for the prediction method of binding free energy changes caused by protein mutants. The prediction performance of the model was tested without considering the complex protein graph structure information through the features constructed by protein sequence and pdb and the two-dimensional graph of small molecules generated by SMILES strings. The deep learning model dl-withCM-GCN used GCN to capture the structural information of the contact graph between amino acid residues to improve the feature representation ability. The deep learning model dl-withSGCN used SGCN to enhance the understanding of the geometric structure by using the global coordinate information in the protein structure. The design of dl-withCM-GCN and dl-withSGCN is to use different graph representation and graph learning methods to more comprehensively capture and learn the geometric features of protein mutant structures, so as to improve the accuracy of relative binding free energy prediction of protein mutants.

[0068] Figure 2It is a deep learning model dl-original. Its input features include features constructed from protein sequences and pdb and graph data constructed from SMILES strings of small molecules. The features constructed from protein sequences and pdb include integer encodings of wild-type protein sequences (wild-encoded vectors), integer encodings of mutant protein sequences (mutation-encoded vectors), and strain data of mutants (strain). The integer encoding of the sequence is required to generate its mask, and the strain data is required to generate its hidden features (hidden). The input features of small molecules include node features, edge indexes, and edge features generated by SMILES strings. The node features involve atomic symbols, atomic degrees, implicit valence, formal charges, etc.; the edge features involve the type of bond, whether the bond is conjugated, whether the bond is in a ring, and the chirality information of the bond.

[0069] Figure 3 It is a deep learning model dl-withCM-GCN, whose input features include features constructed from protein sequences and pdb and graph data constructed from SMILES strings of small molecules. The features constructed from protein sequences and pdb include integer encodings of wild-type protein sequences (wild-encoded vectors), integer encodings of mutant protein sequences (mutation-encoded vectors), deformation data of mutants (strain), and graph data suitable for GCN. The protein graph features used in this model include node features, edge indexes, and edge features generated from protein sequences, where node features involve the chemical properties, size, and charge information of amino acid residues; the values ​​of the contact graph are used as edge features.

[0070] Figure 4 It is a deep learning model dl-withCM-SGCN, whose input features include protein sequence and pdb constructed features and graph data constructed by SMILES strings of small molecules. The features of protein sequence and pdb construction include integer encoding of wild-type protein sequence (wild-encoded vectors), integer encoding of mutant protein sequence (mutation-encoded vectors), deformation data of mutant (strain) and graph data suitable for SGCN. The protein graph features used in this model include node features, edge indexes and coordinate matrices generated by protein sequences. The coordinate matrix extracts the atomic coordinates of each residue in the protein structure by parsing the pdb file.

[0071] Figure 2The construction of the deep learning dl-original model includes an MPNN layer, two two-layer Transformer-Encoders, a two-layer GRU layer, a fully connected layer, and a dropout layer with a loss rate of 0.2. Figure 3 The deep learning dl-withCM-GCN model is constructed with an MPNN layer, two two-layer Transformer-Encoder layers, a two-layer GRU, a GCN layer, a fully connected layer, and a dropout layer with a dropout rate of 0.2. Figure 4 The construction of the deep learning dl-withCM-SGCN model includes an MPNN layer, two two-layer Transformer-Encoder layers, a two-layer GRU, an SGCN layer, a fully connected layer, and a dropout layer with a dropout rate of 0.2.

[0072] S8: Calculate the free energy change value predicted by the deep learning model, compare the root mean square error, mean absolute error and Pearson correlation coefficient with the experimental results, and verify the prediction results of the deep learning model

[0073] Calculate the root mean square error RMSE between the model prediction value and the true value:

[0074]

[0075] Where n is the number of predicted values, y i is the true value, y i is the predicted value.

[0076] Figure 5 It is the RMSE value between the predicted value and the true value in the molecular docking AutoDock Vina and the three deep learning models dl-original, dl-withCM-GCN, and dl-withCM-SGCN. As can be seen from the figure, the deep learning model dl-original has lower values ​​than molecular docking on the dasatinib, imatinib, nilotinib, ponatinib, and gefitinib datasets. Compared with Vina and dl-original, the deep learning model dl-withCM-GCN has lower RMSE values ​​on the bosutinib, dasatinib, imatinib, nilotinib, ponatinib, and gefitinib datasets. The deep learning model dl-withCM-SGCN has lower RMSE values ​​on the dasatinib, imatinib, nilotinib, ponatinib, and gefitinib datasets.

[0077] Calculate the root mean square error MAE between the model prediction value and the true value:

[0078]

[0079] Where n is the number of predicted values, y i is the true value, y i is the predicted value.

[0080] Figure 6 It is the MAE value of the predicted value and the true value in molecular docking and three deep learning models dl-original, dl-withCM-GCN, and dl-withCM-SGCN. As can be seen from the figure, the deep learning model dl-original has a lower MAE value on bosutinib, dasatinib, imatinib, nilotinib, and ponatinib data, and the deep learning models dl-withCM-GCN and dl-withCM-SGCN perform better on these data sets. By comparison, it can be seen that the introduction of the structural feature Contact Map can reduce the mean absolute error of the prediction and effectively improve the prediction accuracy of the model.

[0081] Calculate the Pearson coefficient R between the model prediction value and the true value:

[0082]

[0083] Where n is the number of predicted values, x i is the i-th true value, y i is the ith predicted value, is the mean of the true values, is the mean of the predicted values.

[0084] Figure 7It is the Pearson coefficient value of the predicted value and the true value in molecular docking and three deep learning models dl-original, dl-withCM-GCN, and dl-withCM-SGCN. As can be seen from the figure, the Pearson value of dl-original in bosutinib, dasatinib, imatinib and gefitinib is higher than that of Vina. In addition, dl-withCM-GCN and dl-withCM-SGCN have the best Pearson coefficients in the bosutinib and gefitinib datasets. In general, the dl-withCM-GCN model achieved the best results, with lower RMSE and MAE values ​​than other models on multiple datasets, and better Pearson coefficients on the bosutinib and geftinib datasets. By introducing the structural feature ContactMap and using GCN for feature extraction, the key features of the binding between protein mutants and small molecules are more fully captured.

[0085] Comprehensive RMSE, MAE and Pearson coefficient analysis show that the performance of deep learning models on multiple data sets is significantly better than molecular docking, indicating that deep learning has significant advantages in predicting the relative binding free energy of protein mutants. Furthermore, we predicted the geometric features of mutated proteins through AlphaFold and extracted structural features from the geometric features. It can be seen that the introduction of structural features can further improve the accuracy of deep learning model predictions. The deep learning model that introduces GCN and SGCN further improves the prediction performance, which shows that graph convolutional networks have unique advantages in capturing structural feature information.

[0086] The present invention proposes a relative binding free energy prediction method that introduces the geometric features of protein mutation structures predicted by AlphaFold. GCN and SGCN are used to learn the geometric features predicted by AlphaFold, which helps to better capture the interaction between proteins and small molecules, thereby improving the accuracy and reliability of prediction. This method provides new ideas and tools for the research and design of drug resistance.

Claims

1. A method for predicting the relative binding free energy of geometric features of protein mutation structures, characterized in that: The steps include: Establish databases of proteins and small molecules; Generate three-dimensional structures of wild-type and mutant proteins using AlphaFold; Construct a dictionary of character tokens for protein sequences; The CA atomic distance matrix between the mutant and wild-type proteins was calculated to obtain the contact map of the mutant; The deformation data of wild-type and mutant proteins were calculated separately; The contact maps of mutants and small molecules are converted into graph data respectively, and integrated with the integer encoding of protein sequences and deformation data in the iterator to form training and test sets; Establish a deep learning model in the server, train it using the training set, and test the trained deep learning model using the test set; Verify the prediction results of the deep learning model.

2. The method for predicting the relative binding free energy of geometric features of protein mutation structures according to claim 1, characterized in that: The method of establishing a database of proteins and small molecules includes: The data of kinase inhibitor molecules and protein sequences for studying point mutation resistance are obtained from the database; the sequence of wild-type protein, mutation site and experimental value of binding free energy change between molecule and protein are obtained from the literature; these data are stored in the local database to obtain a database of proteins and small molecules.

3. The method for predicting the relative binding free energy of geometric features of protein mutation structures according to claim 1, characterized in that: The use of AlphaFold to generate three-dimensional structures of wild-type and mutant proteins includes: According to the wild-type protein sequence and mutation sites in the protein and small molecule database, the letters at the specified positions of the wild-type sequence are modified to obtain the mutant protein sequence. The wild-type and mutant protein sequences are uniformly stored in a csv file, and the AlphaFold structure prediction model is used to generate a pdb file storing the three-dimensional structure of the wild-type and mutant proteins.

4. The method for predicting the relative binding free energy of geometric features of protein mutation structures according to claim 1, characterized in that: The method of constructing a character token dictionary of a protein sequence comprises: reading a csv file storing sequence information, tokenizing all sequences, mapping each character to a unique index, and forming a dictionary character, which is saved as a pickle file.

5. The method for predicting the relative binding free energy of geometric features of protein mutation structures according to claim 1, characterized in that: The calculation of the CA atomic distance matrix between the mutant and the wild-type protein includes: Load the pdb files of mutant and wild-type proteins; perform structural comparison between wild-type and mutant and obtain the coordinates of CA atoms after comparison; calculate the distance between each CA atom in the mutant and other CA atoms, and record the distance if the distance is less than 10A; save the calculated distance matrix as a csv file.

6. The method for predicting the relative binding free energy of geometric features of protein mutation structures according to claim 1, characterized in that: The calculation of deformation data of wild-type and mutant proteins respectively includes: Load the pdb files of mutant and wild-type proteins; for each variant, first set the neighbor cutting distance and minimum plddt value parameters of the protein, and then calculate the deformation between the wild type and the mutant.

7. The method for predicting the relative binding free energy of geometric features of protein mutation structures according to claim 1, characterized in that: The converting the contact map of the mutant and the small molecule into graph data respectively comprises: The contact graph of the mutant is converted into graph data, including node features, edge indexes, edge features, and coordinate matrices. The node feature dimension of the mutant graph data is 56, and each atomic feature includes the one-hot encoding of each amino acid, the physicochemical property characteristics of the amino acid, the mutation information of the mutant, and the position-specific matrix. The conversion of small molecules into graph data includes adjacency matrix, node features and edge features. The node feature dimension of the graph data of small molecules is 75, and each atomic feature includes atom type, atomic valence, and number of heteroatoms.

8. The method for predicting the relative binding free energy of geometric features of protein mutation structures according to claim 1, characterized in that: The deep learning model includes at least one of dl-original, dl-withCM-GCN and dl-withCM-SGCN; The method for verifying the prediction results of the deep learning model includes: calculating the free energy change value predicted by the deep learning model, and comparing the root mean square error, mean absolute error and Pearson correlation coefficient with the experimental results for verification.

9. The method for predicting the relative binding free energy of geometric features of protein mutation structures according to claim 8, characterized in that: The construction of the dl-original model includes an MPNN layer, two two-layer Transformer-Encoders, a two-layer GRU layer, a fully connected layer, and a dropout layer with a loss rate of 0.2; The construction of the dl-withCM-GCN model includes an MPNN layer, two two-layer Transformer-Encoder layers, a two-layer GRU, a GCN layer, a fully connected layer, and a dropout layer with a dropout rate of 0.2; The construction of the dl-withCM-SGCN model includes an MPNN layer, two two-layer Transformer-Encoder layers, a two-layer GRU, an SGCN layer, a fully connected layer, and a dropout layer with a dropout rate of 0.

2.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

Citation Information

Cited By

  • Protein DDG prediction model optimization method, protein DDG prediction method and related device

    CN120748503A