Protein-ligand affinity prediction method, system, electronic device and storage medium for virtual screening

Abstract

The present application belongs to the technical field of biological information combined with deep learning, and more particularly relates to a protein ligand affinity prediction method and system for virtual screening, an electronic device and a storage medium. The method comprises: preprocessing protein-ligand complex samples in a PDBbind dataset to construct a training sample set; constructing an affinity prediction model and training the affinity prediction model using the training sample set; obtaining structure files of the protein and the ligand in pdb format or pdbqt format respectively, performing protein-ligand docking through AutoDock molecular docking to obtain a target protein-ligand complex, and preprocessing the target protein-ligand complex; inputting the preprocessed target protein-ligand complex into the trained affinity prediction model to generate a protein-ligand affinity prediction result. The present application improves the accuracy and efficiency of protein-ligand affinity prediction.

Description

Technical Field

[0001] This invention belongs to the technical field of bioinformatics combined with deep learning, and more specifically, relates to a method, system, electronic device and storage medium for predicting protein ligand affinity for virtual screening. Background Technology

[0002] Protein-ligand affinity refers to the strength of the interaction between a protein and its ligand. Protein function typically depends on interactions arising from binding with other molecules, and the strength of this affinity determines the stability and specificity of protein-ligand binding. Therefore, predicting protein-ligand affinity is of significant guiding importance for drug development and protein molecule design. Using deep learning methods to predict protein-ligand affinity can help achieve high-throughput screening, explain complex biological processes, and guide the molecular-level modification of enzymes.

[0003] The interaction between proteins and ligands is the foundation and core of protein function, and the affinity between proteins and ligands is an important indicator of the strength of protein function. In the fields of chemistry or biology, the method for quantitatively studying protein ligands is to assess the tightness of binding between protein ligand molecules based on affinity. The corresponding affinity indicators include the dissociation binding constant Kd, the inhibition constant Ki, and the half maximal inhibitory concentration IC50. The measurement of these indicators requires not only setting up experimental environments under complex and restrictive conditions, but also a large number of positive and repeated experiments. This places high demands on cost, the effort and technical skills of the experimental personnel, and prevents large-scale implementation.

[0004] Molecular dynamics simulations use computational methods, based on physics, to estimate the affinity between protein ligand molecules using three-dimensional structures. By simulating different energy terms, physics-based methods are considered relatively accurate. However, molecular dynamics simulations require a large amount of computational resources, making them too expensive.

[0005] In recent years, with the development of artificial intelligence technology, especially machine learning, which has been applied to enzyme engineering, machine learning methods are data-centric compared to traditional physics- and theory-based methods. They build models to fit data and learn from the data to achieve predictions. Currently popular deep learning methods utilize both atomic three-dimensional structure information and one-dimensional amino acid sequence information for prediction. However, using only one type of information cannot fully utilize all the information contained in the protein and ligand complex. Combining atomic three-dimensional structure information and one-dimensional amino acid sequence information to build a model for prediction can effectively improve prediction accuracy.

[0006] For example, Chinese patent document CN115148279A discloses a method, apparatus, device, and computer-readable storage medium for predicting the affinity between proteins and ligand molecules. The method is based on node features, edge features, and geometric features extracted from a three-dimensional structural diagram of the protein-ligand molecule binding. It obtains the affinity between the protein and ligand molecules through a pre-trained prediction model and generates an interaction graph indicating the interactions between the atoms of the protein and ligand molecules. This improves the performance of affinity prediction and allows for the determination of the correctness of the predicted atomic interactions between the protein and ligand molecules, making the prediction results interpretable. The prediction model is obtained through error correction of affinity prediction and interaction graph prediction; therefore, the method of the embodiments of this disclosure can learn more accurate affinity prediction and interaction graph prediction based on error correction.

[0007] Chinese patent document CN116312758A discloses a method for predicting the affinity between proteins and ligands based on decoupling using graph neural networks. The method includes: S1, obtaining a training dataset of protein-ligand complexes; wherein the training dataset includes atomic information, bond information, and spatial structure information of the protein-ligand complexes; S2, preprocessing the obtained training dataset of protein-ligand complexes; S3, constructing a decoupled graph neural network model for predicting the affinity between drugs and proteins; S4, training the constructed decoupled graph neural network model using the preprocessed training dataset to obtain a trained graph neural network model; S5, using the trained decoupled graph neural network model to predict the affinity of the protein-ligand complex to be predicted to obtain the affinity prediction result.

[0008] The aforementioned predictive model is highly technical and requires a certain level of computer programming background to apply proficiently. To overcome this technical barrier and make the model more user-friendly, this invention proposes a protein ligand affinity prediction method and system for virtual screening, encapsulating this technology in a web system to address the aforementioned issues. Summary of the Invention

[0009] The present invention aims to overcome at least one of the defects of the prior art and provide a protein-ligand affinity prediction method for virtual screening, so as to improve the accuracy and efficiency of protein-ligand affinity prediction.

[0010] The present invention also discloses a protein ligand affinity prediction system for virtual screening.

[0011] The detailed technical solution of this invention is as follows:

[0012] A method for predicting protein ligand affinity for virtual screening, the method comprising:

[0013] S1. Obtain the PDBbind dataset and preprocess the protein-ligand complex samples in the PDBbind dataset to construct a training sample set;

[0014] S2. Construct an affinity prediction model and train the affinity prediction model using the training sample set;

[0015] S3. Obtain the structure files of the protein and ligand in pdb or pdbqt format respectively, perform protein-ligand docking through AutoDock molecular docking to obtain the target protein-ligand complex, and preprocess the target protein-ligand complex.

[0016] S4. Input the preprocessed target protein-ligand complex into the trained affinity prediction model to generate protein-ligand affinity prediction results.

[0017] According to a preferred embodiment of the present invention, the protein-ligand complex samples in the PDBbind dataset or the target protein-ligand complex are pretreated, wherein the atoms constituting the protein and ligand mainly include C, H, O, N, P, S, F, Cl, Br, I, K, Na, Fe, Mg, Cu, etc., and different atoms are connected by chemical bonds.

[0018] Preprocessing includes: encoding individual protein atoms in the protein-ligand complex sample or target protein-ligand complex using a 59-dimensional vector; that is, predicting the secondary structure of each protein sequence using the SSPro tool and encoding it using an 8-dimensional one-hot vector. The secondary structure of each protein sequence includes α-helices, isolated β-bridge residues, residues involved in β-steps, hydrogen bond turns, and 3D helices. 10 Spirals, π-spirals, bends, and coils;

[0019] The physicochemical properties of the protein are encoded using an 11-dimensional one-hot vector, which includes polarity, nonpolarity, acidity, basicity, and seven functional groups. The 20 standard amino acid residues and one type of non-standard residues of the protein are encoded using a 21-dimensional one-hot vector.

[0020] Nine-dimensional one-hot vectors are used to encode nine types of atoms contained in the protein, including boron, carbon, nitrogen, oxygen, phosphorus, sulfur, selenium, halogen, and metal atoms.

[0021] One-dimensional one-hot vectors are used to encode the atomic hybridization type of the protein, the number of bonds between the atom and other heavy atoms (excluding hydrogen atoms), and the number of bonds between the atom and other heteroatoms (excluding carbon and hydrogen atoms).

[0022] The atomic properties of proteins are encoded using 5-dimensional one-hot vectors, including hydrophobicity, aromaticity, acceptor, donor, and ring properties.

[0023] One-dimensional one-hot vectors are used to encode the atomic charge and atom affiliation of proteins, where the atom affiliation is protein.

[0024] According to a preferred embodiment of the present invention, preprocessing the protein-ligand complex samples in the PDBbind dataset or the target protein-ligand complex further includes: encoding individual ligand atoms of the ligands in the protein-ligand complex samples or target protein-ligand complexes using a 19-dimensional vector, that is, encoding the nine types of atoms contained in the ligands using a 9-dimensional one-hot vector, the nine types of atoms including boron atoms, carbon atoms, nitrogen atoms, oxygen atoms, phosphorus atoms, sulfur atoms, selenium atoms, halogen atoms and metal atoms;

[0025] One-dimensional one-hot vectors are used to encode the atomic hybridization type of the ligand, the number of bonds between the atom and other heavy atoms (excluding hydrogen atoms), and the number of bonds between the atom and other heteroatoms (excluding carbon and hydrogen atoms).

[0026] The atomic properties of the ligands are encoded using a 5-dimensional one-hot vector, and the atomic properties of the ligands include hydrophobicity, aromaticity, acceptor, donor and ring.

[0027] One-dimensional one-hot vectors are used to encode the atomic charge and atom affiliation of the ligands, respectively, whereby the atom affiliation is a ligand.

[0028] According to a preferred embodiment of the present invention, preprocessing the protein-ligand complex samples in the PDBbind dataset or the target protein-ligand complex further includes: trimming individual protein atoms and individual ligand atoms in the encoded protein-ligand complex samples or target protein-ligand complexes according to the input grid size of the affinity prediction model, retaining individual protein atoms and individual ligand atoms located within the corresponding grids; the trimming process is as follows:

[0029] The position coordinates of individual protein atoms and individual ligand atoms were obtained using the tf.bio library and mapped to their positions in the affinity prediction model.

[0030] Calculate the average values ​​of the ligands on the X, Y, and Z axes of the spatial coordinate system, and use these average values ​​as the center of the protein-ligand complex sample or the target protein-ligand complex. Then, readjust the coordinate positions of the protein-ligand complex sample or the target protein-ligand complex.

[0031] The individual protein atoms and individual ligand atoms are trimmed based on the coordinates of the adjusted protein-ligand complex sample or the target protein-ligand complex.

[0032] According to a preferred embodiment of the present invention, in step S2, an affinity prediction model is constructed based on a convolutional neural network. The affinity prediction model includes a two-layer Block module for extracting protein features, a two-layer Block module for extracting ligand features, a three-layer RES Block module, and a fully connected layer. The output of each RES Block module is averaged and then input into the next layer structure.

[0033] According to a preferred embodiment of the present invention, the Block module includes depthwise convolution and pointwise convolution, and uses 3D convolution with a filter size of 3×3×3 for feature extraction; the RES Block module consists of two Block modules and three RES residual blocks.

[0034] According to a preferred embodiment of the present invention, step S3 further includes: converting the pdb format protein or ligand structure file into pdbqt format, that is, firstly removing solvent molecules, metal ions and hydrogenation operations from the protein or ligand, then using AutoDock Tools to adjust the charge of the protein or ligand, determining the root node of the ligand and selecting the twistable bond of the ligand, and finally saving it as a pdbqt format file.

[0035] In another aspect of the invention, a protein ligand affinity prediction system for virtual screening is provided, the system comprising:

[0036] The training sample acquisition module is used to acquire the PDBbind dataset and preprocess the protein-ligand complex samples in the PDBbind dataset to construct a training sample set.

[0037] A model building module is used to build an affinity prediction model and train the affinity prediction model using the training sample set.

[0038] The target data acquisition module is used to acquire the structure files of proteins and ligands in pdb or pdbqt format, perform protein-ligand docking through AutoDock molecular docking to obtain the target protein-ligand complex, and preprocess the target protein-ligand complex.

[0039] The affinity prediction module is used to input the preprocessed target protein-ligand complex into the trained affinity prediction model to generate protein-ligand affinity prediction results.

[0040] In another aspect of the invention, an electronic device is also provided, comprising:

[0041] At least one processor; and

[0042] A memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the protein ligand affinity prediction method for virtual screening as described above.

[0043] In another aspect of the invention, a machine-readable storage medium is also provided, which stores executable instructions that, when executed, cause the machine to perform the protein ligand affinity prediction method for virtual screening as described above.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] (1) The present invention provides a protein ligand affinity prediction method for virtual screening. By designing an affinity prediction model based on a neural network, the target protein-ligand complex that has been molecularly docked and pretreated is input into the affinity prediction model for affinity prediction, and finally outputs a one-dimensional affinity value. The present invention improves the accuracy and efficiency of protein-ligand affinity prediction and provides a reference for experimenters to perform virtual screening of protein-ligand complexes.

[0046] (2) Based on the system requirements and functions, the present invention designs a Web system for predicting the affinity of protein ligands for virtual screening. Each functional module contained in the system is applied to each step of implementing the method of the present invention. By running each functional module on the Web server, the experimenter is provided with visualized protein-ligand affinity prediction results, and the experimenter can perform interactive operations. Attached Figure Description

[0047] Figure 1 This is a flowchart of the protein ligand affinity prediction method for virtual screening described in this invention.

[0048] Figure 2 This is a schematic diagram of the affinity prediction model based on neural networks in Embodiment 1 of the present invention.

[0049] Figure 3 This is an interface diagram of the protein-ligand molecular docking generated by the system in Embodiment 2 of the present invention.

[0050] Figure 4is an interface diagram for protein-ligand affinity prediction generated by the system in Embodiment 2 of the present invention. Detailed Description of the Embodiments

[0051] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0052] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure pertains.

[0053] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only, and is not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly dictates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they specify the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0054] The embodiments of the present disclosure and the features in the embodiments may be combined with each other without conflict.

[0055] Aiming at the deficiencies existing in the prior art, the present invention provides a protein-ligand affinity prediction method for virtual screening. By designing an affinity prediction model based on a neural network, and inputting the preprocessed target protein-ligand complex that has completed molecular docking into the affinity prediction model for affinity prediction, the method finally outputs the protein-ligand affinity prediction result, which improves the accuracy and efficiency of protein-ligand affinity prediction, and provides a reference for experimenters to virtually screen protein-ligand complexes.

[0056] The protein-ligand affinity prediction method and system for virtual screening of the present invention will be further described below with reference to specific embodiments.

[0057] Embodiment 1.

[0058] Refe Figure 1 , this embodiment provides a protein-ligand affinity prediction method for virtual screening, and the method includes:

[0059] S1. Obtain the PDBbind dataset, preprocess the protein-ligand complex samples in the PDBbind dataset, and construct a training sample set.

[0060] Specifically, this embodiment can use the PDBbind2016 dataset as the dataset for training the neural network model. The PDBbind2016 dataset contains 19,443 protein-ligand complex samples. By collecting the PDBbind2016 dataset and cleaning it to remove missing and redundant data, the dataset is divided into a training set and a validation set according to a certain ratio (e.g., 8:2 ratio) for model training and validation, respectively. At the same time, the PDBbind2016 core set dataset can be used as a test set for model testing.

[0061] In constructing the training sample set for training the neural network model, protein-ligand information requires specific encoding and transformation before it can be utilized by the neural network, and the input to the neural network needs to be fixed to a certain size. Therefore, this embodiment requires corresponding preprocessing of the protein-ligand complex samples in the PDBbind dataset. It is known that the atoms composing proteins and ligands mainly include C, H, O, N, P, S, F, Cl, Br, I, K, Na, Fe, Mg, Cu, etc., and different atoms are connected by chemical bonds. The specific preprocessing process is as follows.

[0062] (1) Define each atom of the protein in the protein-ligand complex sample (i.e., a single protein atom) using a 59-dimensional vector, and define each atom of the ligand corresponding to the protein in the protein-ligand complex sample (i.e., a single ligand atom) using a 19-dimensional vector.

[0063] That is, extract the structural characteristics of the protein pocket (small molecule ligands do not possess this structural characteristic); use the SSPro tool of ScratchProtein Predictor to predict the secondary structure of each protein sequence, including α-helix (H), isolated β-bridge residues (B), β-stepping residues (E), hydrogen bond turns (T), 3 10 Helices (G), π-helices (I), bends (S), and coils (C) are encoded using 8-dimensional one-hot vectors for these 8 structural characteristics; physicochemical properties of proteins are encoded using 11-dimensional one-hot vectors, including polarity, nonpolarity, acidity, basicity, and 7 functional groups; and standard amino acid residues (20 types) and non-standard residues (classified into 1 type) are encoded using 21-dimensional one-hot vectors.

[0064] Furthermore, the Chimera tool was used to convert the pdb format structure files of proteins and ligands into mol2 format files, respectively. The openbabel tool of the Python library was used to read the mol2 files, and the tf.bio library was used to obtain the characteristics of each atom of the protein and ligands. Then, the characteristics of each atom of the protein and ligands were encoded, including: using a 9-dimensional one-hot vector to encode 9 types of atoms (B, C, N, O, P, S, Se, halogen atoms, and metal atoms), where halogen and metal atoms refer to the types of atoms contained in the protein; using a 1-dimensional vector to encode the hybridization type of the atom; using a 1-dimensional vector to encode the number of bonds between the atom and other heavy atoms (atoms other than hydrogen atoms); using a 1-dimensional vector to encode the number of bonds between the atom and other heteroatoms (atoms other than carbon and hydrogen atoms); using a 5-dimensional vector to encode the atomic properties defined by the SMART pattern (including hydrophobicity, aromaticity, acceptor, donor, and ring); using a 1-dimensional vector to encode the charge of the atom; and using a 1-dimensional vector to encode the atom's affiliation (i.e., belonging to a protein pocket or ligand).

[0065] Based on the above encoding process, the individual protein atoms in the protein-ligand complex sample ultimately have 59-dimensional features, while the individual ligand atoms in the protein-ligand complex sample have 19-dimensional features.

[0066] (2) The individual protein atoms and individual ligand atoms in the protein-ligand complex sample are fixed to a certain size. That is,

[0067] The position coordinates of each atom in the protein and ligand are obtained using the tf.bio library, mapped to their positions in the convolutional neural network, and the average values ​​of the ligand molecule along the X, Y, and Z axes of the spatial coordinate system are calculated, respectively denoted as X. mean Y mean Z mean The average value was used as the center of the protein-ligand complex sample. Experiments showed that fixing the neural network grid size (i.e., the input size) to a 24×24×24 box yielded optimal results. The coordinates of the protein-ligand complex sample were then readjusted, denoted as X. new =XX mean +12, Y new =YY mean +12, Z new =ZZ mean +12, where X new Y new Z new This refers to the coordinates of the updated protein or ligand atoms on the X, Y, and Z axes, respectively. Based on these coordinates, the atoms of the protein and ligand are trimmed, retaining only the atoms within the corresponding grid, and finally the atoms are mapped to the input grid.

[0068] Furthermore, to enhance the generalization ability of the neural network model, this embodiment employs a data augmentation method based on rotation along a specified axis to expand the PDBbind dataset to 24 times its original size. Specifically, the atomic coordinates of proteins and ligands are rotated by 90, 180, or 270 degrees around the X-axis, and similarly by 90, 180, or 270 degrees around the Y and Z axes, respectively, expanding the data by 9 times. Additionally, rotations of 180 degrees around the diagonals of the XY, XZ, and YZ planes expand the data by 6 times, and rotations of 120 and 240 degrees around each of the four body diagonals of the grid box expand the data by 8 times. Finally, the 59-dimensional vector of a single protein atom and the 19-dimensional vector of a single ligand atom are respectively filled into individual grid cells. Ultimately, a single protein is represented as a 24×24×24×59 vector, and a single ligand is represented as a 24×24×24×19 vector, both stored in the HDF file.

[0069] Finally, all the data samples stored in the hdf file are used to construct the training sample set.

[0070] S2. Construct an affinity prediction model and train the affinity prediction model using the training sample set.

[0071] In this embodiment, an affinity prediction model is constructed based on a convolutional neural network model. Specific parameters... Figure 2 As shown:

[0072] The processed protein-ligand file (i.e. the constructed training sample set) is input into the convolutional layer of the constructed affinity prediction model for convolutional processing to extract features. First, the protein features (Pocket_input, 59-dimensional vector) and ligand features (ligand_input, 19-dimensional vector) are processed through two layers of depthwise separable blocks. During this process, the feature maps generated retain their shape and size except for the number of channels. Convolution operations within the depthwise separable blocks are then used to adjust the number of channels for the protein and ligand to the same size (64 dimensions), and these are then added together to fuse the small molecule features of the protein and ligand, forming a shallow feature map of the processed protein-ligand complex sample. Next, three layers of depthwise separable residual blocks (RES Block) are used to further extract deeper features from the protein-ligand complex sample. Each layer doubles the number of channels in the protein-ligand complex sample (i.e., 128 in the first layer, 256 in the second, and 512 in the third), increasing the model depth while further reducing overfitting. Each layer of the depthwise separable residual block is followed by an average pooling layer with a 2×2×2 kernel size. The feature map of the last layer is flattened into a one-dimensional vector and used as the input of the fully connected layer. Three fully connected layers (Liners) with 1000, 500, and 200 neurons are used, and the prediction result (out) is finally output.

[0073] The Block module refers to a depthwise separable convolution module, which includes depthwise convolution and pointwise convolution, and uses 3D convolution with a filter size of 3×3×3 for feature extraction.

[0074] The depthwise separable residual block (RES Block) refers to a combination of two depthwise separable convolutional modules (Block) and three residual blocks (Res), such as... Figure 2 As shown in the RES Block, each Block module is connected to a Residual Block (Res), and two Block modules are connected by another Residual Block Res.

[0075] The affinity prediction model was built using Leaky ReLU as the activation function and trained using the Adam optimizer with a learning rate of 10. -5 The loss function used is the mean squared error loss function (MSE), and an L2 norm of λ = 0.0001 is used to reduce the risk of overfitting. After 150 training epochs, the model begins to fit, and the results are saved as the optimal model for protein-ligand affinity prediction.

[0076] Testing was performed on the PDBbind2016 core set dataset, which contains 290 proteins and ligands with known affinities. Evaluation was performed using the root mean square error (RMSE). Where n represents the size of the dataset, truth i ,prediction i represents the true and predicted values ​​of the affinity of the i-th protein-ligand complex, and represents the model's prediction result. On the above test set, the model's RMSE reaches 1.20. The smaller the RMSE value, the better the predictive performance of the affinity prediction model.

[0077] The existing models were evaluated and compared with the models mentioned above using the PDBbind2016 core set dataset. The evaluation metrics of the models mentioned above are better than the Pafnucy model (RMSE = 1.418) and DeepAtom model (RMSE = 1.318) which only use atoms, and the DeepDTAF model (RMSE = 1.355) which only use sequence information.

[0078] S3. Obtain the structure files of the protein and ligand in pdb or pdbqt format respectively, perform protein-ligand docking through AutoDock molecular docking to obtain the target protein-ligand complex, and preprocess the target protein-ligand complex.

[0079] The purpose of molecular docking is to obtain the conformation of the ligand small molecule when it interacts with the protein, so as to provide the protein conformation and small molecule conformation for the affinity prediction module.

[0080] If the obtained protein and / or ligand structure files are in PDB format, the PDB files must first undergo default processing. This involves using a Python script in the molecular visualization software PyMOL to remove solvent molecules, metal ions, and add hydrogen to the protein and ligands. Then, the AutoDock Tools program is used to adjust the charges, determine the root node of the ligand, and select the twistable bonds of the ligand, saving the file as a PDBQT format file. Finally, the AutoDock Vina program is used to perform molecular docking, using the protein as the acceptor and the small molecule ligand as the ligand, executing a semi-flexible docking task.

[0081] After the calculation is complete, you will get 10 docking results. You can view the docking results and scores on the system's web interface. Figure 3 As shown, the results can be exported for further observation and processing in other software, and provide data preparation for the next step of affinity prediction.

[0082] After performing the above operations, the target protein-ligand complex is obtained. Then, the target protein-ligand complex is preprocessed and input into the trained affinity prediction model for affinity prediction.

[0083] In this embodiment, the pretreatment operation of the target protein-ligand complex is basically the same as the pretreatment operation of the protein-ligand complex sample mentioned above, specifically including: (1) defining each atom (i.e., a single protein atom) of the protein in the target protein-ligand complex using a 59-dimensional vector, and defining each atom (i.e., a single ligand atom) of the ligand corresponding to the protein in the target protein-ligand complex using a 19-dimensional vector; (2) fixing both the single protein atom and the single ligand atom in the target protein-ligand complex to a certain size. The specific process has been described in detail above and will not be repeated here.

[0084] S4. Input the preprocessed target protein-ligand complex into the trained affinity prediction model to generate protein-ligand affinity prediction results.

[0085] That is, the preprocessed target protein-ligand complex can be used as target data and input into the trained affinity prediction model. The model then undergoes feature extraction through convolutional layers and feature integration through fully connected layers, ultimately outputting the protein-ligand affinity prediction result. This result can be viewed on the system's web interface, such as... Figure 4 As shown.

[0086] The method described in this embodiment improves the accuracy and efficiency of protein-ligand affinity prediction, providing a reference for researchers to perform virtual screening of protein-ligand complexes.

[0087] Example 2

[0088] This embodiment provides a protein ligand affinity prediction system for virtual screening, the system comprising:

[0089] The training sample acquisition module is used to acquire the PDBbind dataset and preprocess the protein-ligand complex samples in the PDBbind dataset to construct a training sample set.

[0090] A model building module is used to build an affinity prediction model and train the affinity prediction model using the training sample set.

[0091] The target data acquisition module is used to acquire the structure files of the protein and ligand in pdb or pdbqt format, perform protein-ligand docking through AutoDock molecular docking to obtain the target protein-ligand complex, and preprocess the target protein-ligand complex.

[0092] The affinity prediction module is used to input the preprocessed target protein-ligand complex into the trained affinity prediction model to generate protein-ligand affinity prediction results.

[0093] Furthermore, this embodiment can design a web system for predicting protein ligand affinity in virtual screening, based on system requirements. This system uses the Flask framework to integrate the aforementioned functional modules into an API interface and ultimately encapsulates them into a web system. By running the functional modules on a web server and combining the front-end and back-end functional applications of the system, interactive operations such as file uploading, downloading, and viewing can be achieved on the web page.

[0094] Example 3

[0095] This embodiment also provides an electronic device, including:

[0096] At least one processor; and

[0097] A memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the protein ligand affinity prediction method for virtual screening as described above.

[0098] In this embodiment, the electronic device may include, but is not limited to: personal computer, server computer, workstation, desktop computer, laptop computer, notebook computer, mobile computing device, smartphone, tablet computer, cellular phone, personal digital assistant (PDA), handheld device, messaging device, wearable computing device, consumer electronic device, etc.

[0099] Example 4

[0100] This embodiment also provides a machine-readable storage medium storing executable instructions that, when executed, cause the machine to perform the protein ligand affinity prediction method for virtual screening as described above.

[0101] Specifically, a system or apparatus equipped with a readable storage medium may be provided, on which software program code implementing the functions of any of the embodiments described above is stored, and the computer or processor of the system or apparatus can read and execute the instructions stored in the readable storage medium.

[0102] In this case, the program code read from the readable medium itself can perform the functions of any of the above embodiments, and therefore the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of this specification.

[0103] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer or the cloud via a communication network.

[0104] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0105] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0106] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0107] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A method for predicting protein ligand affinity for virtual screening, characterized in that, The method includes: S1. Obtain the PDBbind dataset and preprocess the protein-ligand complex samples in the PDBbind dataset to construct a training sample set; S2. Construct an affinity prediction model and train the affinity prediction model using the training sample set; S3. Obtain the structure files of the protein and ligand in pdb or pdbqt format respectively, perform protein-ligand docking through AutoDock molecular docking to obtain the target protein-ligand complex, and preprocess the target protein-ligand complex. S4. Input the pretreated target protein-ligand complex into the trained affinity prediction model to generate protein-ligand affinity prediction results; In step S2, an affinity prediction model is constructed based on a convolutional neural network. The affinity prediction model includes a two-layer deep separable module Block for extracting protein features, a two-layer deep separable module Block for extracting ligand features, a three-layer deep separable residual module RES Block, and a fully connected layer. The output of each deep separable residual module RES Block is averaged and then input into the next layer structure. The depth-separable module Block includes depthwise convolution and pointwise convolution, and uses a filter size of [missing value]. Feature extraction is performed using 3D convolution; the depth-separable residual module RES Block consists of two depth-separable modules Block and three RES residual blocks.

2. The protein ligand affinity prediction method for virtual screening according to claim 1, characterized in that, Preprocessing of protein-ligand complex samples in the PDBbind dataset or of target protein-ligand complexes includes: encoding individual protein atoms in the protein-ligand complex samples or target protein-ligand complexes using 59-dimensional vectors, i.e., predicting the secondary structure of each protein sequence using the SSPro tool and encoding it using an 8-dimensional one-hot vector, wherein the secondary structure of each protein sequence includes α-helices, isolated β-bridge residues, β-steps, hydrogen bond turns, and 3D protein-ligand complexes. 10 Spirals, π-spirals, bends, and coils; The physicochemical properties of the protein are encoded using an 11-dimensional one-hot vector, which includes polarity, nonpolarity, acidity, basicity, and seven functional groups. The 20 standard amino acid residues and one type of non-standard residues of the protein are encoded using a 21-dimensional one-hot vector. Nine-dimensional one-hot vectors are used to encode nine types of atoms contained in the protein, including boron, carbon, nitrogen, oxygen, phosphorus, sulfur, selenium, halogen, and metal atoms. One-dimensional one-hot vectors are used to encode the atomic hybridization type of the protein, the number of bonds between the atom and heavy atoms other than hydrogen atoms, and the number of bonds between the atom and heteroatoms other than carbon and hydrogen atoms. The atomic properties of proteins are encoded using 5-dimensional one-hot vectors, including hydrophobicity, aromaticity, acceptor, donor, and ring properties. One-dimensional one-hot vectors are used to encode the atomic charge and atom affiliation of proteins, where the atom affiliation is protein.

3. The protein ligand affinity prediction method for virtual screening according to claim 2, characterized in that, Preprocessing the protein-ligand complex samples in the PDBbind dataset or the target protein-ligand complex further includes: encoding individual ligand atoms of the ligands in the protein-ligand complex samples or target protein-ligand complexes using a 19-dimensional vector, that is, encoding the nine types of atoms contained in the ligands using a 9-dimensional one-hot vector, the nine types of atoms including boron atoms, carbon atoms, nitrogen atoms, oxygen atoms, phosphorus atoms, sulfur atoms, selenium atoms, halogen atoms and metal atoms; One-dimensional one-hot vectors are used to encode the atomic hybridization type of the ligand, the number of bonds between the atom and heavy atoms other than hydrogen atoms, and the number of bonds between the atom and heteroatoms other than carbon and hydrogen atoms. The atomic properties of the ligands are encoded using a 5-dimensional one-hot vector, and the atomic properties of the ligands include hydrophobicity, aromaticity, acceptor, donor and ring. One-dimensional one-hot vectors are used to encode the atomic charge and atom affiliation of the ligands, respectively, whereby the atom affiliation is a ligand.

4. The protein ligand affinity prediction method for virtual screening according to claim 3, characterized in that, Preprocessing the protein-ligand complex samples in the PDBbind dataset or the target protein-ligand complex further includes: pruning individual protein atoms and individual ligand atoms in the encoded protein-ligand complex samples or target protein-ligand complexes according to the input grid size of the affinity prediction model, retaining individual protein atoms and individual ligand atoms located within the corresponding grid; the pruning process is as follows: The position coordinates of individual protein atoms and individual ligand atoms were obtained using the tf.bio library and mapped to their positions in the affinity prediction model. Calculate the average values ​​of the ligands on the X, Y, and Z axes of the spatial coordinate system, and use these average values ​​as the center of the protein-ligand complex sample or the target protein-ligand complex. Then, readjust the coordinate positions of the protein-ligand complex sample or the target protein-ligand complex. The individual protein atoms and individual ligand atoms are trimmed based on the coordinates of the adjusted protein-ligand complex sample or the target protein-ligand complex.

5. The protein ligand affinity prediction method for virtual screening according to claim 1, characterized in that, Step S3 further includes: converting the pdb format protein or ligand structure file into pdbqt format, that is, firstly removing solvent molecules, metal ions and hydrogenation operations from the protein or ligand, then using AutoDock Tools to adjust the charge of the protein or ligand, determining the root node of the ligand and selecting the twistable bond of the ligand, and finally saving it as a pdbqt format file.

6. A protein ligand affinity prediction system for virtual screening, characterized in that, The system includes: The training sample acquisition module is used to acquire the PDBbind dataset and preprocess the protein-ligand complex samples in the PDBbind dataset to construct a training sample set. A model building module is used to build an affinity prediction model and train the affinity prediction model using the training sample set. The target data acquisition module is used to acquire the structure files of proteins and ligands in pdb or pdbqt format, perform protein-ligand docking through AutoDock molecular docking to obtain the target protein-ligand complex, and preprocess the target protein-ligand complex. The affinity prediction module is used to input the preprocessed target protein-ligand complex into the trained affinity prediction model to generate protein-ligand affinity prediction results. The protein ligand affinity prediction system for virtual screening performs the protein ligand affinity prediction method for virtual screening as described in any one of claims 1 to 5.

7. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the protein ligand affinity prediction method for virtual screening as described in any one of claims 1 to 5.

8. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores executable instructions that, when executed, cause the machine to perform the protein ligand affinity prediction method for virtual screening as described in any one of claims 1 to 5.

Images