Protein-ligand dissociation rate prediction method based on graph neural network transfer learning
By employing a graph neural network-based transfer learning method, a model is pre-trained using a large-scale protein-ligand affinity dataset. This model is then combined with a graph neural network to construct protein-ligand interaction features, solving the accuracy and cost issues of protein-ligand dissociation rate prediction in existing technologies. This enables efficient prediction across targets and virtual screening in new drug development.
Patent Information
- Application Number
- CN202511102289.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-07
- Publication Date
- 2025-11-18
AI Technical Summary
Existing computational methods struggle to efficiently and accurately predict protein-ligand dissociation rates, especially in small datasets and cross-target predictions. Furthermore, existing machine learning methods lack three-dimensional structural information and are computationally expensive.
A graph neural network-based transfer learning method is adopted. The model is pre-trained using a large-scale protein-ligand affinity dataset. The protein-ligand interaction features are constructed by combining graph neural networks. The dissociation rate is predicted on the dissociation rate dataset through transfer learning. The three-dimensional spatial information of the protein pocket residue map and ligand atom map is used to predict the dissociation rate.
It enables efficient and accurate prediction of protein-ligand dissociation rates across targets, reduces computational costs, and better reflects the actual efficacy of drugs, making it suitable for virtual screening in new drug development.
Smart Images
Figure CN120977376A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of protein-ligand dissociation rate prediction, and more particularly relates to a protein-ligand dissociation rate prediction method based on graph neural network transfer learning. BACKGROUND
[0002] It is known that new drug development is a costly and time-consuming process. The development of machine learning and artificial intelligence is currently greatly assisting the efficiency of new drug development. There are two methods of using AI to assist in the development of small molecule drugs, one is to screen from existing molecules or fine-tune successful drugs on the market through quantitative structure-activity relationship and virtual screening, and the other is to design molecules from scratch or generate new molecules through reverse quantitative structure-activity relationship. The underlying logic of these two methods cannot be separated from the prediction of the multidimensional properties of small molecules and their interactions with proteins by machine learning methods.
[0003] The most typical one is the scoring function in virtual screening and molecular docking, which calculates the thermodynamic dissociation constant of protein-ligand binding equilibrium, i.e. the binding affinity Kd. However, in fact, the binding affinity can only represent the binding strength at the reaction equilibrium state, and often cannot accurately reflect the in vivo efficacy of the drug. In recent years, researchers have begun to realize that the dissociation rate (Koff) or residence time (τ) of the drug-target complex is often a more critical determinant, and the dissociation rate has become an important information for measuring drug efficacy in the early stage of drug development. The scoring function for dissociation rate combined with the affinity scoring function can more efficiently screen small molecules with high specificity.
[0004] Currently, experimental methods (surface plasmon resonance, radiolabeled or fluorescently labeled ligand dissociation experiments, and double-point competitive binding detection methods, etc.) are the main means to obtain drug-target binding kinetics data. However, the cost of manpower and material resources required by experimental methods is high, and it is difficult to provide detailed information about the interaction between proteins and ligands.
[0005] Therefore, in the virtual screening and design of drugs, in order to achieve efficient optimization of residence time, it is necessary to develop a computational method that can accurately and efficiently predict this property. However, in existing computational methods, conventional molecular dynamics simulation is limited by time scale and cannot simulate dissociation processes on the order of milliseconds to hours, enhanced sampling methods such as τ random acceleration molecular dynamics simulation have insufficient accuracy and still have high computational cost, and cannot achieve high-throughput prediction. Most existing machine learning methods are based on a small data set, or only for a single protein target, or have defects in feature selection. The current advanced graph neural network deep learning model is mainly designed and trained to predict Kd rather than koff.
[0006] Existing similar solutions and their limitations:
[0007] (1) Using RF-Score descriptor + protein secondary structure SS descriptor as features, training a random forest model. Where RF-Score is a one-dimensional vector recording the number of protein atom-ligand atom pairs appearing within the radius of the protein-ligand complex binding site; the SS descriptor records the contact count of the ligand atom and the amino acid residues of the different secondary structures of the protein.
[0008] Disadvantages: 1.1 RF-Score molecular descriptor is an early protein-ligand complex characterization method, which only records the number of atom pairs within a certain distance, without specific atom positions or distance information, and only records the element types of the two atoms in the atom pair, without any other context information of the atom, which is a very rough characterization of protein-ligand interactions, and its extended descriptor such as SS relies on human selection, and overall it is an old and poor performance feature extraction method.
[0009] 1.2 Random forest is a classic machine learning algorithm, which lags behind the current mainstream deep learning neural network model, and this method may be limited by the small amount of data, which cannot support the training of deep learning models
[0010] (2) Using text strings that do not involve three-dimensional structures as features, where the ligand uses the Mol2vec descriptor, and the protein is embedded into 30 dimensions according to the "high-order amino acid sequence" to form a string of each amino acid residue and its two preceding and following positions. The deep learning model uses a hybrid expert MoE architecture, which goes through the MoE layer after the dimension reduction and attention layer of the features, which contains multiple fully connected layers, i.e. experts and gate layers, and finally outputs the predicted value.
[0011] Disadvantages: The features are completely based on the SMILES sequence of the ligand and the amino acid sequence of the protein, and do not contain three-dimensional structure information, which does not conform to the basic logic of predicting dynamic properties. SUMMARY
[0012] To solve these problems, the present application develops a cross-target computational method for predicting protein-ligand dissociation kinetics parameters with low computational cost based on transfer learning. Using the method of transfer learning, the model is first pre-trained on a larger protein-ligand complex affinity dataset to fully learn the interaction information between proteins and ligands as an implicit representation of protein-ligand interactions, and then a new prediction module is trained on the pre-trained affinity model to predict the dissociation rate using the three-dimensional structure of the protein-ligand complex in the dissociation rate dataset, learning to predict the dissociation rate using the implicit representation.
[0013] In order to achieve the above purpose, the present application is realized by adopting the following technical solutions:
[0014] Dissociation rate dataset construction: integrate protein-ligand kinetics experimental data in public databases and literature, construct a database containing protein-ligand structure and kinetics parameters, protein structure is derived from PDB or AlphaFold predicted structure, ligand structure is generated from SMILES using RDkit, protein structure is repaired using pdbfixer, and molecular docking is performed using Smina tool to obtain protein-ligand complex binding pose;
[0015] Pre-training protein-ligand affinity prediction model, using PDBbind v2020 database for training, extracting protein-ligand complex interaction feature information;
[0016] Based on graph neural network to construct affinity prediction model, including protein pocket range definition and BLOSUM62 encoding, Protein EGNN processing to generate protein three-dimensional structure features, ligand represented by two-dimensional molecular graph and using fatoms and fbonds features, after processing by LigandGAT module, output ligand high-dimensional features;
[0017] Input protein and ligand features into protein-ligand interaction module protein ligand communication, output fresidues_updated and fatoms_updated feature matrix containing protein-ligand interaction information;
[0018] Input the output of the interaction module into the energy contribution prediction module and the interaction probability prediction module respectively, the energy contribution prediction module generates energy contribution prediction value for protein residue-ligand atom pair, the interaction probability prediction module outputs the interaction probability prediction value of the corresponding residue-atom pair, finally the two are multiplied and added element by element to get the affinity prediction value, where the interaction probability prediction module is optimized by binary cross-entropy loss by comparing with the label of the distance between residue-atom pairs in the real structure less than 4.0 A, and the affinity prediction value is optimized by root mean square error loss with experimental affinity;
[0019] Based on the affinity prediction model, transfer learning is performed, the parameters of the above modules are frozen, the interaction features of protein and ligand are averaged and pooled, input into the multilayer perceptron regression module to output the dissociation rate prediction value, and the experimental dissociation rate is used as the label, the model parameters are optimized by root mean square error loss.
[0020] In one scheme, the protein structure is preferably selected from the structures in the PDB database, and if there is a missing part, the pdbfixer tool is used for automatic completion to ensure the integrity of the three-dimensional structure of the protein, and if no available PDB structure is found, the AlphaFold predicted structure is selected to supplement, and the protein structure preparation process includes the completion step of atoms and residues, and ensures the accuracy of the protein pocket area.
[0021] In one scheme, the ligand structure is automatically generated by the RDkit tool according to the SMILES string to generate an sdf format structure file, which ensures the consistency and availability of each ligand structure, and cooperates with the subsequent high-throughput automatic generation of protein-ligand complex conformations.
[0022] In one scheme, the molecular docking uses the Smina molecular docking tool to perform high-throughput docking operation on the protein structure and the ligand structure, to obtain the binding pose of the protein-ligand complex, and to use it as the spatial basis for subsequent feature coding and graph neural network input.
[0023] In one scheme, for the protein part, the BLOSUM62 is used to encode the protein pocket residue sequence to obtain an initial feature matrix of [N_res, 20], and the E3 equivariant graph neural network Protein EGNN module is used to embed it into a high-dimensional feature matrix of [N_res, 300] to fully express the local three-dimensional space and sequence information of the protein pocket.
[0024] In one scheme, for the ligand part, the ligand molecule is encoded in two dimensions, and the atom features fatoms include element type, number of connected hydrogen atoms, hybridization state, whether it is a directional atom, etc., after one-hot encoding, it is [N_atom, 31], the bond information fbonds includes single and double bond state, directionality, whether in the ring, etc., and is encoded as [N_keys, 6], and is input into the Ligand GAT module of the multi-head attention mechanism, and the output dimension is [N_atom, 300] of the ligand high-dimensional features.
[0025] In one scheme, the affinity model is trained, and then migration learning is performed to realize the prediction of the dissociation rate; the parameters of the trained protein-ligand graph neural network and the interaction module are frozen, and only the new dissociation rate prediction head is trained.
[0026] Specifically, the updated feature matrices of the protein pocket and the ligand are respectively subjected to average pooling dimension reduction into [1, 300] vectors, and then spliced into [1, 600], input into the multi-layer perception regression module, output the dissociation rate prediction value of the protein-ligand complex, and optimized by the root mean square error loss of the experimental dissociation rate data.
[0027] The present application has the following advantages:
[0028] (1) The application of transfer learning in protein ligand interaction is the main protection point: using the protein ligand thermodynamic affinity data with sufficient data quantity which is easier to obtain to train an advanced graph neural network model, constructing an implicit representation of protein ligand complex interaction, and then using this representation method to learn the prediction of protein ligand complex kinetic dissociation rate on a small-scale data set, transferring the knowledge learned in a larger data set.
[0029] (2) The starting point of this implicit representation is the protein pocket residue graph and the ligand atom graph, which contains complete three-dimensional spatial position related information and residue / atom type information, and can more fully depict the interaction of protein ligand complex.
[0030] (3) The prediction method of the present scheme is not limited to a specific protein target, and can be used for cross-target prediction and screening. BRIEF DESCRIPTION OF DRAWINGS
[0031] Figure 1 The flow chart of the method of the present application is shown in the figure;
[0032] Figure 2 The protein target type distribution in the dissociation rate data set is shown in the figure;
[0033] Figure 3 The protein target distribution in the dissociation rate data set is shown in the figure;
[0034] Figure 4 The dissociation rate data distribution in the dissociation rate data set is shown in the figure. DETAILED DESCRIPTION
[0035] In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the related drawings. The typical embodiments of the present application are shown in the drawings. However, the present application can be realized in many different forms and is not limited to the embodiments described in the present application. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.
[0036] Unless otherwise defined, all technical and scientific terms used in the present application have the same meaning as understood by those skilled in the art to which the present application belongs. The terms used in the present application in the specification of the present application are only for the purpose of describing the specific embodiments and are not intended to limit the present application. In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the related drawings. The typical embodiments of the present application are shown in the drawings. However, the present application can be realized in many different forms and is not limited to the embodiments described in the present application. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.
[0037] AsFigure 1 The specific implementation process of a protein-ligand dissociation rate prediction method based on graph neural network transfer learning is as follows:
[0038] S1, dissociation rate dataset construction: integrate existing public databases and recent literature protein-ligand kinetic experiment data, construct a protein-ligand structure and dynamics parameter database covering nearly a hundred target points. The database example and basic distribution information are as follows.
[0039] pdbbind v2020 database is used to pretrain the affinity prediction model
[0040] Protein and ligand structure preparation: select one PDB structure or AlphaFold predicted structure for each protein target, repair the structure from PDB using pdbfixer, and supplement the missing atoms or residues in the pdb structure according to the sequence (if missing). The ligand is generated from SMILES to sdf structure file using RDkit.
[0041] Molecular docking: using the molecular docking tool Smina, the corresponding protein and ligand are high-throughput molecularly docked to obtain the complex binding pose.
[0042] The protein-ligand affinity prediction model is pre-trained using the pdbbind v2020 database, and the protein-ligand complex interaction features are extracted from it.
[0043] Table 1 shows an example of the integrated dissociation rate dataset
[0044]
[0045] S2, construct a graph neural network model;
[0046] Initial feature generation and feature extraction module: define the protein pocket range using the ligand position after molecular docking, define the protein pocket structure as a three-dimensional graph, and use BLOSUM62 encoding for the initial feature matrix, with a size of [N_res, 20], which is input into the E3 variable graph neural network module Protein EGNN, and the output is the updated pocket feature matrix [N_res, 300]; define the ligand molecular structure as a two-dimensional graph, and the initial atom feature matrix fatoms includes the one-hot encoding of each atom's element type, the number of connected hydrogen atoms, hybridization state, and whether it is a directional atom, etc. Physicochemical properties, size [N_atom, 31], bond feature matrix fbonds [N_keys, 6] includes single and double bonds, whether it is a directional bond, whether it is in a ring, etc. Properties, input into the multi-head attention module LigandGAT, and output the updated ligand feature matrix [N_atom, 300].
[0047] Protein ligand communication module: The protein pocket residue feature matrix and ligand atom feature matrix extracted from the above two modules are input into the protein ligand cross-attention module to obtain updated two feature matrices, so that the feature matrices contain information describing the interaction between the protein pocket and the ligand, respectively fresidues_updated[N_res, 300], fatoms_updated[N_atom, 300].
[0048] Affinity prediction module: fresidues_updated and fatoms_updated output by the protein ligand interaction module are respectively transmitted into two prediction heads, one is the energy contribution prediction module, and the other is the interaction probability prediction module. The potential energy contribution value and the interaction probability are calculated for each pocket residue-ligand atom pair, respectively, that is, a three-dimensional tensor of [N_atom, N_res, 1] is output, then element-wise multiplication and addition are performed to obtain the final affinity prediction value.
[0049] Energy contribution prediction module: fresidues_updated and fatoms_updated are expanded and element-wise multiplied through the broadcast mechanism to obtain a new tensor of [N_atom, N_res, 300], and then reduced to [N_atom, N_res, 1] through a LeakyReLU layer, representing the energy contribution prediction value of each residue-atom pair.
[0050] Interaction probability prediction module: fresidues_updated and fatoms_updated are expanded and element-wise multiplied through the broadcast mechanism to obtain a new tensor of [N_atom, N_res, 300], and then reduced to [N_atom, N_res, 1] through a Sigmoid layer, representing the interaction probability of each residue-atom pair. The output value of this module is compared with whether the distance between the residue-atom pair in the real structure is less than 4.5 angstroms, and optimized through binary cross-entropy loss.
[0051] Except for the interaction probability prediction module, the other parts are optimized through the root mean square error loss of the final affinity prediction value and the real value.
[0052] S3, migration learning, build dissociation rate prediction module: complete the above affinity prediction model training, and freeze the existing module parameters, train a new dissociation rate prediction head. Using the constructed dissociation rate dataset, the model infers to the protein-ligand interaction module, and the output feature matrix fresidues_updated and fatoms_updated are respectively averaged pooled into [1, 300] vectors and spliced into [1, 600] feature vectors, which are input into the MLP regression module to output the dissociation rate prediction value, which is optimized by the root mean square error loss with the dissociation rate experimental value.
[0053] Technical effects:
[0054] (1) Multi-round random partition verification set R2=0.65; according to the target type, the data set is divided, and the performance of one protein target as a complete verification set is better than the baseline model, and the generalization ability is better. See Figures 2-4 .
[0055] Table 2 Dissociation rate prediction module in data set according to target type Test set performance
[0056]
[0057]
[0058] (2) It can be used as the second scoring function of virtual screening in new drug development, filling the gap of screening lead compounds only according to affinity, making the screening compound standard more reflect the actual drug efficacy of the drug.
[0059] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiments of each method. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.
[0060] It should be understood that the above detailed description of the technical solutions of the present application by means of preferred embodiments is illustrative rather than limiting. Those skilled in the art can modify the technical solutions recorded in each embodiment on the basis of the description of the present application, or make equivalent replacement to part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solution deviate from the spirit and scope of the technical solutions of each embodiment of the present application.
Claims
1. A protein-ligand dissociation rate prediction method based on graph neural network transfer learning, characterized in that: The method comprises: Dissociation rate dataset construction: integrate protein-ligand dynamics experimental data in public databases and literature, construct a database containing protein-ligand structure and dynamics parameters, protein structure is derived from PDB or AlphaFold predicted structure, ligand structure is generated from SMILES using RDkit, protein structure is repaired using pdbfixer, and molecular docking is performed using Smina tool to obtain protein-ligand complex binding pose; Pre-training protein-ligand affinity prediction model, training with PDBbind v2020 database, extracting protein-ligand complex interaction feature information; Based on graph neural network, an affinity prediction model is constructed, including protein pocket range definition and BLOSUM62 encoding, Protein EGNN processing to generate protein three-dimensional structure features, ligand represented by two-dimensional molecular graph and using fatoms and fbonds features, processed by LigandGAT module to output high-dimensional features of ligand; The protein and ligand features are input into the protein-ligand interaction module protein ligand communication, and the fresidues_updated and fatoms_updated feature matrices containing protein-ligand interaction information are output; The interaction module output is respectively input into the energy contribution prediction module and the interaction probability prediction module. The energy contribution prediction module generates an energy contribution prediction value for a protein residue-ligand atom pair. The interaction probability prediction module outputs an interaction probability prediction value corresponding to the residue-atom pair. Finally, the two are multiplied element by element and added to obtain an affinity prediction value. The interaction probability prediction module is optimized by a binary cross-entropy loss through a comparison with a label in a real structure, wherein the distance between a residue-atom pair is less than 4.5 A. The affinity prediction value is optimized by a root mean square error loss with an experimental affinity. Based on the affinity prediction model, transfer learning is performed, the parameters of the module are frozen, the dissociation rate prediction head is trained, the interaction features of protein and ligand are averaged and pooled, then spliced, input into the multilayer perceptron regression module to output the dissociation rate prediction value, and the experimental dissociation rate is used as the label, and the root mean square error loss is used to optimize the model parameters.
2. The protein-ligand off-rate prediction method based on graph neural network transfer learning according to claim 1, wherein: The protein structure is preferably selected from the PDB database, if there is a missing part, the pdbfixer tool is used for automatic completion to ensure the integrity of the protein three-dimensional structure, if no available PDB structure is found, the AlphaFold predicted structure is selected to supplement, and the protein structure preparation process includes the completion steps of atoms and residues, and ensures the accuracy of the protein pocket area.
3. The protein-ligand off-rate prediction method based on graph neural network transfer learning according to claim 1, wherein: The ligand structure is automatically generated by RDkit tool according to SMILES string to generate sdf format structure file, which ensures the consistency and availability of each ligand structure, and cooperates with the molecular docking step to automatically generate high-throughput protein-ligand complex conformation.
4. The protein-ligand dissociation rate prediction method based on graph neural network transfer learning according to claim 1, wherein: Molecular docking uses Smina molecular docking tool to perform high-throughput docking operation on protein structure and ligand structure to obtain the binding pose of protein-ligand complex, which is used as the spatial basis for subsequent feature coding and graph neural network input.
5. The protein-ligand off-rate prediction method based on graph neural network transfer learning according to claim 1, wherein: For the protein part, BLOSUM62 is used to encode the protein pocket residue sequence to obtain an initial feature matrix of [N_res, 20], and the E3 graph neural network Protein EGNN module is used to embed it into a high-dimensional feature matrix of [N_res, 300], which fully expresses the local three-dimensional space and sequence information of the protein pocket.
6. The protein-ligand off-rate prediction method based on graph neural network transfer learning according to claim 1, wherein: For the ligand part, the ligand molecule is encoded in a two-dimensional graph, with atom features fatoms containing element type, number of connected hydrogen atoms, hybridization state, and whether it is a directional atom, after one-hot encoding, [N_atom, 31], bond information fbonds containing single and double bond state, directionality, and whether in the ring, encoded as [N_keys, 6], input into the LigandGAT module of the multi-head attention mechanism, and output ligand high-dimensional features with a dimension of [N_atom, 300].
7. The protein-ligand off-rate prediction method based on graph neural network transfer learning according to claim 1, wherein: After the affinity prediction model is trained, transfer learning is performed to realize the prediction of the dissociation rate; the parameters of the protein-ligand graph neural network and the interaction module that have been trained are frozen, and only the new dissociation rate prediction head is trained; Specifically, the updated feature matrix of the protein pocket and the ligand is respectively reduced to a [1, 300] vector after average pooling, and then spliced into a [1, 600] vector, input into a multi-layer perception regression module, output the dissociation rate prediction value of the protein-ligand complex, and optimized through the root mean square error loss with the experimental dissociation rate data.