A method for predicting protein-ligand binding affinity based on shell map and base model

Abstract

The application provides a protein-ligand binding affinity prediction method based on a shell graph and a basic model. The protein three-dimensional structure, ligand three-dimensional structure, protein sequence and ligand molecular representation of a protein-ligand complex are obtained; a plurality of radial shell layers are divided with the ligand as the center to construct a shell graph, and a shell feature matrix is generated according to the residue-atom contact in each shell layer; a protein semantic representation and a ligand semantic representation are encoded by using a protein basic model and a ligand basic model respectively, and a structure representation is obtained by processing the shell feature matrix by using a shell graph encoder; the three representations are used as queries and keys for information aggregation by using a cross-attention fusion module, and a binding affinity prediction value is output by an affinity prediction network. The method realizes deep interaction between structure information and semantic information, and improves the prediction accuracy and generalization ability.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, specifically to a method for predicting protein-ligand binding affinity based on shell diagrams and basic models. Background Technology

[0002] Accurate prediction of protein-ligand binding affinity is a core task in structural biology and computer-aided drug discovery. It directly impacts the efficiency of virtual screening, the optimization of lead compounds, and the final evaluation of candidate drugs. While traditional experimental methods offer high precision, they require stringent experimental conditions, are costly, and time-consuming, making them unsuitable for the ever-increasing demands of drug screening. Therefore, utilizing computational methods to predict binding affinity from structural, sequence, and molecular information is of great significance for reducing drug discovery costs and improving the efficiency of candidate molecule screening.

[0003] With the continuous accumulation of protein sequence data, protein structure data, and small molecule chemical data, affinity prediction methods based on machine learning and deep learning have gradually developed. Existing methods can be broadly categorized into non-interaction modeling methods and interaction modeling methods. Non-interaction modeling methods primarily generate a global representation based on the protein sequence and ligand molecular diagrams or molecular strings, for example, using sequence networks, graph neural networks, or pre-trained language models to extract features, and then predict binding affinity based on this global representation. However, because this method does not explicitly utilize the three-dimensional binding structure of the protein-ligand complex, it is difficult to fully characterize the spatial contact relationships within the actual binding pocket.

[0004] Interaction modeling methods explicitly utilize the three-dimensional structure of protein-ligand complexes. Chinese patent application CN120319300A, "A Structure-Aware Protein-Ligand Binding Affinity Prediction Method and System," constructs a protein-ligand isomorphism map with atoms as nodes and captures higher-order interactions by defining meta-paths. However, because this scheme treats each atom within the binding pocket as an independent node, the size of its isomorphism map and derived isomorphism map expands dramatically with protein size, resulting in significant computational and memory overhead. Furthermore, this scheme inputs features obtained from structure map encoding and sequence encoding into different prediction branches and then performs a post-fusion averaging of the predicted values ​​from each branch. This post-fusion strategy results in simple interaction between three-dimensional structural information and sequence semantic information only at the decision level, making it difficult to achieve mutual calibration and deep coupling during the feature extraction stage.

[0005] Chinese patent application CN121565237A, entitled "A Protein-Ligand Binding Affinity Prediction Method Based on Graph Convolutional Networks," constructs protein and ligand graphs with atoms as nodes. It employs a polar coordinate-inspired graph attention mechanism for cross-graph message passing and finally concatenates the graph-level representation with sequence semantic features and molecular fingerprints for prediction. However, the model's predictive stability still has room for improvement when the test ligand has a novel chemical backbone, the test protein has low sequence similarity to the training protein, or the temporal distribution of the data changes.

[0006] Therefore, how to construct an affinity prediction method that can simultaneously characterize the three-dimensional spatial information of protein-ligand complexes and the semantic information of proteins and ligands, and achieve effective fusion of multi-source information, remains a technical problem that urgently needs to be solved. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a protein-ligand binding affinity prediction method based on shell diagrams and basic models, solving the technical problem that "existing methods are unable to effectively distinguish the affinity differences between different binding conformations of the same ligand, resulting in reduced screening accuracy".

[0008] To achieve the above objectives, the present invention is implemented using the following technical solution: In a first aspect, the present invention provides a method for predicting protein-ligand binding affinity based on shell diagrams and a basic model, comprising the following steps: S1. Obtain the three-dimensional structure of the protein, the three-dimensional structure of the ligand, the amino acid sequence of the protein, and the molecular representation of the ligand for the protein-ligand pair to be predicted. Based on the three-dimensional structure of the protein and the three-dimensional structure of the ligand, divide the ligand binding pocket into multiple radial shells with the ligand as the center. Use each radial shell as a node and establish edges based on the nearest neighbor relationship between the radial shells to construct a shell graph. Generate the shell node features of the corresponding nodes based on the contact pattern between the protein structural unit and the ligand in each radial shell to obtain the shell feature matrix. S2. The protein amino acid sequence is encoded using a protein basic model to obtain a protein semantic representation, and the ligand molecule representation is encoded using a ligand basic model to obtain a ligand semantic representation; the shell feature matrix is ​​processed using a shell graph encoder to obtain a structural representation characterizing the local three-dimensional geometric information of the protein-ligand complex. S3. Input the structural representation, protein semantic representation and ligand semantic representation into the cross-attention fusion module, so that each representation is used as a query representation and the other two representations are used as key representation and value representation to aggregate information and obtain the cross-attention fusion representation. S4. The cross-attention fusion representation is processed using an affinity prediction network to obtain the predicted binding affinity value of the protein-ligand pair.

[0009] The construction of the shell diagram described in S1 includes: For each protein structural unit, the minimum Euclidean distance between the atoms contained in the protein structural unit and the ligand atoms is calculated to obtain the representative distance of the protein structural unit relative to the ligand. The spatial range corresponding to the representative distance is divided into multiple radial intervals according to the preset shell thickness, and each protein structural unit is assigned to the corresponding radial shell according to the representative distance of each protein structural unit. The nth radial shell is mapped to a node in the shell graph, such that each node is associated with a shell node feature used to characterize the protein-ligand contact pattern in the corresponding radial region, and an edge is established between each node and one or more nearest neighbor nodes in its radial order.

[0010] The shell node features described in S1 include: Multiple contact channels are established according to the protein structural unit type and ligand atom type; the protein structural unit type includes 20 standard amino acid residue types and other residue types, and the ligand atom type includes hydrogen, carbon, oxygen, nitrogen, phosphorus, sulfur, halogen and other element types; For each protein structural unit and each ligand atom, the minimum Euclidean distance from the ligand atom to the atoms contained in the protein structural unit is calculated, and the corresponding radial shell is determined based on the minimum Euclidean distance. In the determined radial shell, the number of contacts is accumulated for the contact channels corresponding to the protein structural unit type and ligand atom type to obtain the shell node characteristics of the radial shell.

[0011] The shell graph encoder described in S2 includes multiple graph coding blocks, each of which performs the following processing: Attention aggregation is performed among the nearest neighbor nodes defined by the edges of the shell graph to obtain the updated representation of the candidate nodes; The candidate node update representation is added to the input node representation of this layer for the first time to obtain the intermediate node representation; The intermediate node representation is subjected to layer normalization and node-by-node feedforward network processing, and the processing result is added to the intermediate node representation for a second residual to obtain the output node representation of the layer; The node representations output by the final layer graph encoding block are flattened or pooled according to the radial shell order, and the structure representation is obtained through learnable projection.

[0012] The process of obtaining the shell feature matrix in S1 includes: stacking the features of each shell node in a radial order from the inside out to obtain the shell feature matrix; standardizing the shell feature matrix based on the mean and standard deviation estimated from the training samples; and inputting the standardized shell feature matrix into a bidirectional long short-term memory network, which undergoes layer normalization and random deactivation to obtain the initial node representation of the shell graph encoder.

[0013] The ligand-based model in S2 is a pre-trained molecular language model, and the ligand semantic representation is the classification marker position representation output after inputting the ligand molecular representation into the ligand-based model; the protein-based model is a pre-trained protein language model, and the protein semantic representation is obtained by pooling the residue-level representation output after inputting the protein amino acid sequence into the protein-based model; during the training phase, the parameters of the ligand-based model and the protein-based model are kept frozen.

[0014] The cross-attention fusion module described in S3 performs the following processing: The ligand semantic representation, protein semantic representation, and structural representation are respectively mapped to a unified latent space through linear layers; Using the mapped ligand semantic representation as the query, the mapped protein semantic representation and structural representation are concatenated as the key and value, and multi-head cross-attention computation is performed to obtain the updated ligand branch representation; Using the mapped protein semantic representation as the query, the mapped ligand semantic representation and structural representation are concatenated as the key and value, and multi-head cross-attention calculation is performed to obtain the updated protein branch representation. Using the mapped structural representation as the query, the mapped ligand semantic representation and protein semantic representation are concatenated as the key and value, and multi-head cross-attention computation is performed to obtain the updated structural branch representation. The updated ligand branch representation, protein branch representation, and structural branch representation are fused to obtain the cross-attention fused representation.

[0015] The affinity prediction network described in S4 is a multi-layer fully connected network; the cross-attention fusion representation is input into the affinity prediction network to obtain the affinity prediction value; the regression loss is calculated based on the affinity prediction value and the true value, and the parameters of the shell graph encoder, the cross-attention fusion module and the affinity prediction network are updated.

[0016] The method is used for virtual screening of candidate drugs, specifically including: fixing the same protein target, taking multiple candidate ligands as ligands in the protein-ligand pair to be predicted, performing steps S1 to S4 as described in claim 1, and obtaining the predicted binding affinity values ​​of each candidate ligand to the protein target; sorting the multiple candidate ligands according to the predicted binding affinity values, and using the sorting results as the output of virtual screening.

[0017] In a second aspect, the present invention provides a protein-ligand binding affinity prediction system, comprising: The data acquisition module is used to acquire the three-dimensional structure of the protein, the three-dimensional structure of the ligand, the amino acid sequence of the protein, and the molecular representation of the ligand for the protein-ligand pair to be predicted. The shell graph construction module is used to divide the ligand binding pocket into multiple radial shells with the ligand as the center, use each radial shell as a node and establish edges according to the nearest neighbor relationship between the radial shells, and generate shell node features according to the contact pattern between protein structural units and ligands in each radial shell. The structure representation extraction module is used to process the shell graph and shell node features using a shell graph encoder to obtain the structure representation of the protein-ligand complex. The semantic representation extraction module is used to obtain the protein semantic representation based on the protein amino acid sequence and to obtain the ligand semantic representation based on the ligand molecule representation; The cross-attention fusion module is used to use each of the structural representation, protein semantic representation and ligand semantic representation as a query representation, and to aggregate information using the other two representations as key representation and value representation to obtain the cross-attention fusion representation. The prediction output module is used to output a predicted value of protein-ligand binding affinity based on the cross-attention fusion representation.

[0018] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the protein-ligand binding affinity prediction method as described above.

[0019] Fourthly, the present invention provides an electronic device including a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the protein-ligand binding affinity prediction method as described above.

[0020] Compared with the prior art, the beneficial effects achieved by the present invention are: (1) The present invention transforms the three-dimensional contact relationship in the protein-ligand complex into a radially ordered graph structure by using a shell diagram centered on the ligand, so that the model can simultaneously characterize close contact and the surrounding spatial environment, thus alleviating the problem of insufficient characterization of local three-dimensional geometric information when relying solely on one-dimensional sequences or two-dimensional molecular diagrams.

[0021] (2) The present invention introduces protein sequence semantics and ligand chemical semantics through protein basic model and ligand basic model respectively, so that the model can utilize the prior representation obtained by large-scale pre-training to improve the representation stability when the training samples are limited or the distribution of test samples differs from that of training samples.

[0022] (3) This invention enables the protein semantic representation, ligand semantic representation and complex local three-dimensional geometric representation to be mutually conditional before prediction through the cross-attention fusion module, avoiding the lack of information interaction caused by simple splicing, so that affinity prediction can be constrained by sequence semantics, chemical semantics and three-dimensional contact mode at the same time; adopting a data organization method adapted to the prediction task, training and evaluation are carried out in cold start or distribution change scenarios, which can be closer to the prediction needs of new proteins, new ligands or new batches of samples encountered in actual drug discovery, and has good application and promotion value. Attached Figure Description

[0023] Figure 1 This is a schematic flowchart of the protein-ligand binding affinity prediction method of the present invention.

[0024] Figure 2 This is a schematic diagram of the module connection relationship of the protein-ligand binding affinity prediction system of the present invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1; refer to Figure 1 This embodiment provides a method for predicting protein-ligand binding affinity based on shell diagrams and a basic model, including the following steps: S1. Constructing the shell diagram of the protein-ligand complex (1) Obtain the input data for each protein-ligand pair The input data includes the three-dimensional structure of the protein, the three-dimensional structure of the ligand, the protein amino acid sequence, and the ligand molecular representation. The training samples also include realistic binding affinity labels. Specifically, the three-dimensional structures of the protein and ligand are used to construct the shell diagram, the protein amino acid sequence is used to generate the protein semantic representation, the ligand molecular representation is used to generate the ligand semantic representation, and the realistic binding affinity labels are used for supervised training.

[0027] (2) Input data preprocessing The system parses protein 3D structure files, reading atomic coordinates, residue names, and residue affiliations; it also parses ligand 3D structure files, reading ligand atomic coordinates and element types; and it unifies protein and ligand coordinates to the same coordinate system. Non-hydrogen atoms within the same residue are grouped into a single protein structural unit, using the residue as the protein structural unit, and non-standard residues are categorized into other residue types. Ligand atom types are categorized into hydrogen, carbon, oxygen, nitrogen, phosphorus, sulfur, halogens, and other element types.

[0028] After preprocessing, the input data is validated. If the protein's three-dimensional structure or the ligand's three-dimensional structure is missing, the sample will not be included in the shell diagram construction process; if the protein's amino acid sequence or the ligand's molecular representation is missing, the sample will not be included in the basic model encoding process; if the training sample lacks a true binding affinity label, the sample will not be used for supervised training.

[0029] (3) Constructing the shell diagram Based on the protein structural unit set and ligand atom set determined by the above preprocessing results, the coordinate units of the protein and ligand are unified.

[0030] Calculate the representative distance between each protein structural unit and its ligand. For the j-th protein structural unit p... j Define the representative distance between it and the ligand. j The minimum Euclidean distance between its constituent atoms and ligand atoms:

[0031] in, It is a collection of ligand atoms. for The set of atoms in and These are the three-dimensional coordinates of the corresponding atom.

[0032] Based on representative distance The ligand binding pocket is divided into A concentric shell centered on a ligand. Each shell layer Represented as:

[0033] in, ; To combine the set of all structural units in the pocket, , , The thickness is the shell layer, measured in Å. In one embodiment, , Each shell layer sequentially covers a radial region of equal thickness from the inside out.

[0034] Construct a shell diagram based on the radial order of each shell layer. The first Each shell layer Mapped to nodes in the shell graph , obtain the node set Each node Corresponding to a shell And associate the shell node feature vectors According to the shell order at each node Edging is established between the nearest neighbor shell nodes to obtain the edge set. This forms an adjacent structure that reflects the radial continuity of the shell. In one embodiment, That is, each shell node is connected to no more than 3 neighboring shell nodes in its radial sequence.

[0035] Generate a shell graph adjacency matrix based on the edge set. Generate the original adjacency matrix without self-loops, the adjacency matrix with self-loops, and the normalized adjacency matrix; wherein, the original adjacency matrix is ​​used to represent the true nearest neighbor shell relationship, the adjacency matrix with self-loops is used for graph attention masking and residual aggregation, and the normalized adjacency matrix is ​​used for optional normalized neighborhood aggregation.

[0036] S2. Extract protein, ligand, and shell map features. (1) Generate shell node features Establish residue-ligand atom contact channels. Protein residue types include 20 standard amino acid residues and one other residue category, totaling 21 categories; ligand atom types include hydrogen, carbon, oxygen, nitrogen, phosphorus, sulfur, halogens and other elements, totaling 8 categories; the combination of the two forms 168 contact channels.

[0037] For each residue and each ligand atom, calculate the minimum Euclidean distance from the ligand atom to the atoms contained in the residue, determine the shell assignment of the residue-ligand atom pair according to the shell interval, and accumulate the number of contacts of the residue type and ligand atom type combination in the corresponding shell to obtain the node feature vector of the shell. When there is no corresponding contact in a certain shell, the node feature vector of the shell remains a zero vector.

[0038] Stack all shell node features in shell order to obtain the shell feature matrix. : , where N is the number of shells and d is the feature dimension of each shell node.

[0039] The shell feature matrix is ​​standardized based on the mean and standard deviation estimated from the training set, resulting in... : , where μ c and σ c These are the mean matrix and standard deviation matrix estimated position by position based on the shell features of the training set, respectively. Both have the same N×d dimension as the shell feature matrix. This indicates element-wise division.

[0040] (2) Extracting protein semantic representation and ligand semantic representation Inputting the protein amino acid sequence into the basic protein model yields the protein semantic representation h. p In this embodiment, the protein base model is a ProtT5-XL-U50 encoder, with average pooling and an output dimension of 1024. The ligand molecule representation is input into the ligand base model to obtain the ligand semantic representation hl. In this embodiment, the ligand molecule representation is the string SMILES, the ligand base model is ChemBERTa-77M-MTR, with CLS position representation pooling and an output dimension of 384. During the training phase, the parameters of both the protein and ligand base models are frozen and do not participate in gradient updates.

[0041] (3) The local three-dimensional geometric representation of the complex is obtained by encoding the shell diagram. The standardized shell feature matrix Inputting into a bidirectional long short-term memory network, and after layer normalization and random deactivation, the initial node representation is obtained. : ; in, Representation layer normalization, This indicates random deactivation. In one implementation, , The bidirectional long short-term memory network has 1 layer.

[0042] Represent the shell diagram and initial nodes obtained from the above steps. The input consists of a shell graph encoder composed of M-layer graph attention coding blocks. In the m-th layer, the nodes of the previous layer are represented... Layer normalization is performed, and multi-head attention aggregation is carried out on the nearest neighbor relationships defined by the edge set to obtain the candidate update matrix. : ,in, ; For the first The input node representation of the layer is a matrix; This is the set of edges in the shell graph; This represents multi-head attention computation; For the first The candidate node update matrix obtained by layer attention aggregation; θ g This represents the set of learnable parameters for the graph encoder.

[0043] For the edge set For each edge inside, calculate the distance from the nearest neighbor node. To the node Attention weights : ,in, , , respectively representing the first Layer output nodes and nodes of 3D input feature vector; nodes As a query node, node As a neighboring node that transmits information to it; and These are the learnable query projection matrix and the key projection matrix, respectively; The dimension of a single attention head; In all adjacent nodes Normalize the above to make the fixed nodes The sum of the corresponding attention weights is 1. Among them, Represents a node The nearest neighbor set.

[0044] The candidate update representation is obtained by weighted summation of the value vectors of nearest neighbors based on the attention weights. : ,in, Represents a node The nearest neighbor set, The projection matrix is ​​the learnable value. For nodes The candidate update representation obtained by aggregating nearest neighbor information, and satisfying .

[0045] The node representation update is accomplished using pre-normalization and a double residual structure: ; , This is the matrix representing the intermediate nodes obtained from the first residual connection. Representation layer normalization; This represents a node-by-node feedforward network that shares parameters among all shell nodes. In this embodiment, M=3, the number of attention heads is 4, and the random deactivation ratio is 0.1.

[0046] Output of the Mth layer By flattening and projecting, a local three-dimensional geometric representation of the complex is obtained: The dimensions after flattening are The projected dimension is 512.

[0047] S3, Constructing a trimodal symmetric cross-attention fusion Protein semantic representation h p Ligand semantic representation h l and the local three-dimensional geometric representation of the complex h c We obtain z by projecting each linear layer onto the unified latent space. p z l and z c All three representations are mapped to 512 dimensions. ; ; .

[0048] z is represented by ligands l As a query, protein is represented as z. p And shell diagram representation of z c After being concatenated as keys and values, multi-head cross-attention computation is performed, enabling the ligand representation to adaptively aggregate context related to the current binding relationship based on protein sequence semantics and complex three-dimensional geometry information, resulting in an updated ligand branch representation: .

[0049] Similarly, z is represented by proteins. p As a query, the updated protein branch representation is obtained: .

[0050] Similarly, z is represented by a shell diagram. c As a query, the updated structure branch representation is obtained: , where ∥ represents the splicing operation and MHA represents the multi-head attention operation.

[0051] The above three-way computation enables each of the three representations to use itself as the query and aggregate complementary information from the other two representations, achieving bidirectional information interaction among the three types of representations. In this embodiment, the number of cross-attention heads is 8, and the random deactivation ratio is 0.1. The three-way update representations are then concatenated into a cross-attention fusion representation: The dimension is 1536.

[0052] S4, Binding Affinity Prediction (1) Constructing an affinity prediction network The cross-attention fusion representation obtained in step S3 Input the affinity prediction network to obtain the affinity prediction value: In this embodiment, fusion representation The dimension is 1536. The affinity prediction network is a three-layer fully connected network with dimensions of 1536→1536→640→1 for each layer. The outputs of the first two layers are processed by ReLU activation and random deactivation with a random deactivation ratio of 0.3. The last layer outputs the affinity prediction value.

[0053] (2) Model training According to the predicted value Calculate the mean squared error loss with the true affinity value y: ,in, This refers to the batch sample size. For the first The true affinity of each sample For the first Predicted affinity of each sample.

[0054] The trainable parameters of the shell graph encoder, cross-attention fusion module, and affinity prediction network are updated based on the loss. The parameters of the protein base model and ligand base model remain frozen during training and do not participate in gradient updates.

[0055] In this embodiment, the optimizer used is AdamW, with a learning rate of 1×10⁻⁶. -4 Weight decay 1×10 -4 The training batch size is 64, and the maximum training duration is 200 epochs. Early stopping is implemented during training based on the validation set loss. Training stops when the validation set loss fails to improve within a preset number of patience epochs, and the model parameters that best perform on the validation set are retained.

[0056] (3) Model reasoning and application The test set samples are processed sequentially according to steps S1 to S3. The resulting cross-attention fusion representation is input into the trained affinity prediction network, which outputs the affinity prediction value for each test sample.

[0057] When the testing or application task is candidate ligand screening, the predicted affinity values ​​are output for the same protein and multiple candidate ligands respectively, and the candidate ligands are ranked according to the predicted values; when the testing or application task is target matching, the predicted affinity values ​​are output for the same ligand and multiple candidate proteins respectively, and the candidate proteins are ranked according to the predicted values.

[0058] Example 2; This embodiment provides a protein-ligand binding affinity prediction system, which can implement the method described in Embodiment 1.

[0059] The system includes: a data acquisition module, a task definition and data partitioning module, a shell graph construction module, a structural representation extraction module, a semantic representation extraction module, a cross-attention fusion module, a prediction output module, and a training and evaluation module.

[0060] The data acquisition module is used to acquire the three-dimensional structure of the protein, the three-dimensional structure of the ligand, the amino acid sequence of the protein, the molecular representation of the ligand, and the real binding affinity label required for the training phase in the protein-ligand pair, and to align and standardize the data of the same protein-ligand pair.

[0061] The task definition and data partitioning module is used to determine the prediction task according to the application or evaluation purpose. The prediction task includes at least one of affinity regression prediction, candidate ligand ranking, candidate protein matching, protein cold start evaluation, ligand cold start evaluation, time distribution change evaluation, and external independent test evaluation. The module also organizes training set, validation set, and test set according to the prediction task.

[0062] The shell graph construction module is used to analyze the three-dimensional structure of proteins and ligands, determine the protein structural units, ligand atom sets, and their spatial coordinate relationships; calculate the representative distance between protein structural units and ligands, and divide the ligand binding pocket into multiple radial shells according to a preset shell thickness; use each radial shell as a node in the shell graph, and establish edges based on the nearest neighbor relationships between radial shells to obtain the shell graph and its adjacency relationships; generate shell node features based on the contact patterns between protein structural units and ligand atoms in each radial shell, and arrange the shell node features in a radial order from the inside out to obtain the shell feature matrix.

[0063] The structure representation extraction module receives the shell graph, shell graph adjacency relation, and shell feature matrix output by the shell graph construction module, standardizes the shell feature matrix and initializes the node representation, and uses the shell graph encoder to aggregate information among the nearest neighbor nodes defined by the shell graph adjacency relation to obtain a structure representation characterizing the local three-dimensional geometric information of the protein-ligand complex.

[0064] The cross-attention fusion module is used to receive the structural representation output by the structural representation extraction module and the protein semantic representation and ligand semantic representation output by the semantic representation extraction module. It maps the three representations to a unified latent space, uses each representation as a query representation, and aggregates information using the other two representations as key and value representations to obtain the cross-attention fusion representation.

[0065] The prediction output module is used to receive the cross-attention fusion representation output by the cross-attention fusion module, and to process the cross-attention fusion representation using the affinity prediction network to output the protein-ligand binding affinity prediction value.

[0066] The training and evaluation module is used to receive the protein-ligand binding affinity prediction value output by the prediction output module, calculate the training loss based on the prediction value and the real label, update the trainable parameters in the system, select model parameters based on the validation set results, and output affinity prediction value or ranking results during the testing or application phase.

[0067] like Figure 2 As shown, the data connection relationships between the above modules are as follows: the output of the data acquisition module is passed to the task definition and data partitioning module; the output of the task definition and data partitioning module is passed to the shell graph construction module and the semantic representation extraction module; the output of the shell graph construction module is passed to the structure representation extraction module; the output of the structure representation extraction module and the semantic representation extraction module is passed to the cross-attention fusion module; the output of the cross-attention fusion module is passed to the prediction output module; and the output of the prediction output module is passed to the training and evaluation module.

[0068] Example 3; This embodiment provides a method for model training, evaluation, and application in scenarios involving cold starts and distribution variations. This method does not change the process described in Embodiment 1; it simply constructs training, validation, and test sets using different data organization methods, creating distributional differences between the training and test samples to evaluate the model's generalization ability in real-world drug discovery scenarios.

[0069] Specifically, the following five evaluation paradigms are supported: (1) Conventional random partitioning, where samples are randomly partitioned into training, validation, and test sets in a ratio of 8:1:1; (2) Ligand cold start, where samples are grouped according to the Bemis-Murcko backbone of ligand molecules, so that samples with the same backbone appear only in a single set, to evaluate the model's predictive ability for new chemical backbone ligands; (3) Protein cold start, where samples are grouped according to protein sequence similarity clustering, so that proteins with sequence similarity higher than 30% do not appear in different sets at the same time, to evaluate the model's predictive ability for new target proteins; (4) Temporal distribution variation, where samples are divided according to the year the complex was released, with early data used for training and late data used for validation and testing, to evaluate the model's predictive stability under temporal distribution drift; (5) External independent testing, where samples in PDBbind v2020 that do not overlap with CASF-2016 are used for training and validation, and CASF-2016 and deduplicated CSAR-HiQ are used as external test sets, and external test samples do not participate in training, validation, and standardized statistic estimation.

[0070] In any of the above scenarios, the training phase generates the shell feature matrix, protein semantic representation, and ligand semantic representation according to Example 2. After obtaining the fused representation through the shell graph encoder and cross-attention fusion module, the affinity prediction network outputs the predicted value. The only difference between different scenarios is the partitioning method of the training / validation / test sets; the model structure and training process remain unchanged. The standardized statistics of the shell features, the contact channel order, and the training configuration obtained from the training set remain consistent throughout the validation, testing, and inference phases.

[0071] This embodiment is also used for candidate ligand screening and candidate protein matching: when screening candidate ligands, the same protein target is fixed, and multiple candidate ligands are input respectively, and sorted from high to low according to the predicted value; when matching candidate proteins, the same ligand is fixed, and multiple candidate proteins are input respectively, and sorted according to the predicted value.

[0072] Example 4; This embodiment provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the method described in Embodiment 1 or Embodiment 3.

[0073] When the computer program is executed by the processor, it performs the following operations: acquiring the three-dimensional structure of the protein, the three-dimensional structure of the ligand, the amino acid sequence of the protein, and the molecular representation of the ligand; parsing and preprocessing the three-dimensional structures of the protein and the ligand; constructing a shell graph centered on the ligand and generating shell node features to obtain a shell feature matrix; calling the protein basic model and the ligand basic model to encode the protein semantic representation and the ligand semantic representation, respectively; obtaining the local three-dimensional geometric representation of the complex using the shell graph encoder; obtaining the cross-attention fusion representation using the cross-attention fusion module; and outputting the affinity prediction value using the affinity prediction network.

[0074] During the training phase, the computer program calculates the regression loss based on the true and predicted values, updates the trainable parameters of the shell graph encoder, cross-attention fusion module, and affinity prediction network, selects model parameters based on the validation set results, and saves the trained model parameters, shell feature channel order, standardized statistics, and training configuration. During the inference phase, the program loads the aforementioned model parameters and configuration and outputs affinity prediction values ​​or ranking results for the test samples.

[0075] The non-transitory computer-readable storage medium includes disk storage, read-only memory, random access memory, erasable and rewritable memory, flash memory, optical storage, read-only optical disk, or other media capable of storing computer program instructions.

[0076] Example 5; This embodiment provides an electronic device, including a processor, a memory, an input / output interface, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method described in Embodiment 1 or Embodiment 3.

[0077] The input / output interface receives protein 3D structure files, ligand 3D structure files, protein amino acid sequences, ligand molecular representations, and true binding affinity labels from the training phase. It also outputs affinity predictions, candidate ligand ranking results, candidate protein ranking results, or model evaluation results. The memory stores input data, basic model parameters, shell graph construction configuration, standardized statistics of shell features, trained model parameters, and computer program instructions. The processor executes the computer program to perform protein sequence encoding, ligand molecular representation encoding, shell graph encoding, cross-attention fusion, and affinity prediction. During the training phase, the processor updates trainable parameters based on regression loss. During the inference phase, the processor outputs affinity predictions without updating model parameters.

[0078] The electronic device is a server, workstation, personal computer, cloud computing node, or computing device with hardware acceleration unit. It can be deployed in a single machine or in a distributed manner, with different computing nodes performing basic model encoding, shell graph feature generation, model training, or batch inference tasks respectively.

[0079] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No markings in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for predicting protein-ligand binding affinity based on shell diagrams and a basic model, characterized in that, Includes the following steps: S1. Obtain the three-dimensional structure of the protein, the three-dimensional structure of the ligand, the amino acid sequence of the protein, and the molecular representation of the ligand for the protein-ligand pair to be predicted; Based on the protein three-dimensional structure and ligand three-dimensional structure, the ligand binding pocket is divided into multiple radial shells with the ligand as the center. Each radial shell is used as a node and an edge is established according to the nearest neighbor relationship between the radial shells to construct a shell graph. The shell node features of the corresponding nodes are generated according to the contact pattern between the protein structural unit and the ligand in each radial shell, and the shell feature matrix is ​​obtained. S2. The protein amino acid sequence is encoded using a protein basic model to obtain a protein semantic representation, and the ligand molecule representation is encoded using a ligand basic model to obtain a ligand semantic representation; the shell feature matrix is ​​processed using a shell graph encoder to obtain a structural representation characterizing the local three-dimensional geometric information of the protein-ligand complex. S3. Input the structural representation, protein semantic representation and ligand semantic representation into the cross-attention fusion module, so that each representation is used as a query representation and the other two representations are used as key representation and value representation to aggregate information and obtain the cross-attention fusion representation. S4. The cross-attention fusion representation is processed using an affinity prediction network to obtain the predicted binding affinity value of the protein-ligand pair.

2. The method according to claim 1, characterized in that, The construction of the shell diagram described in S1 includes: For each protein structural unit, the minimum Euclidean distance between the atoms contained in the protein structural unit and the ligand atoms is calculated to obtain the representative distance of the protein structural unit relative to the ligand. The spatial range corresponding to the representative distance is divided into multiple radial intervals according to the preset shell thickness, and each protein structural unit is assigned to the corresponding radial shell according to the representative distance of each protein structural unit. The nth radial shell is mapped to a node in the shell graph, such that each node is associated with a shell node feature used to characterize the protein-ligand contact pattern in the corresponding radial region, and an edge is established between each node and one or more nearest neighbor nodes in its radial order.

3. The method according to claim 1, characterized in that, The shell node features described in S1 include: Multiple contact channels are established according to the protein structural unit type and ligand atom type; the protein structural unit type includes 20 standard amino acid residue types and other residue types, and the ligand atom type includes hydrogen, carbon, oxygen, nitrogen, phosphorus, sulfur, halogen and other element types; For each protein structural unit and each ligand atom, the minimum Euclidean distance from the ligand atom to the atoms contained in the protein structural unit is calculated, and the corresponding radial shell is determined based on the minimum Euclidean distance. In the determined radial shell, the number of contacts is accumulated for the contact channels corresponding to the protein structural unit type and ligand atom type to obtain the shell node characteristics of the radial shell.

4. The method according to claim 1, characterized in that, The shell graph encoder described in S2 includes multiple graph coding blocks, each of which performs the following processing: Attention aggregation is performed among the nearest neighbor nodes defined by the edges of the shell graph to obtain the updated representation of the candidate nodes; The candidate node update representation is added to the input node representation of this layer for the first time to obtain the intermediate node representation; The intermediate node representation is subjected to layer normalization and node-by-node feedforward network processing, and the processing result is added to the intermediate node representation for a second residual to obtain the output node representation of the layer; The node representations output by the final layer graph encoding block are flattened or pooled according to the radial shell order, and the structure representation is obtained through learnable projection.

5. The method according to claim 1, characterized in that, The process of obtaining the shell feature matrix in S1 includes: stacking the features of each shell node in a radial order from the inside out to obtain the shell feature matrix; standardizing the shell feature matrix based on the mean and standard deviation estimated from the training samples; and inputting the standardized shell feature matrix into a bidirectional long short-term memory network, which undergoes layer normalization and random deactivation to obtain the initial node representation of the shell graph encoder.

6. The method according to claim 1, characterized in that, The ligand basic model in S2 is a pre-trained molecular language model, and the ligand semantic representation is the classification marker position representation output after inputting the ligand molecular representation into the ligand basic model; the protein basic model is a pre-trained protein language model, and the protein semantic representation is obtained by pooling the residue-level representation output after inputting the protein amino acid sequence into the protein basic model. During the training phase, the parameters of the ligand-based model and the protein-based model are kept frozen.

7. The method according to claim 1, characterized in that, The cross-attention fusion module described in S3 performs the following processing: The ligand semantic representation, protein semantic representation, and structural representation are respectively mapped to a unified latent space through linear layers; Using the mapped ligand semantic representation as the query, the mapped protein semantic representation and structural representation are concatenated as the key and value, and multi-head cross-attention computation is performed to obtain the updated ligand branch representation; Using the mapped protein semantic representation as the query, the mapped ligand semantic representation and structural representation are concatenated as the key and value, and multi-head cross-attention calculation is performed to obtain the updated protein branch representation. Using the mapped structural representation as the query, the mapped ligand semantic representation and protein semantic representation are concatenated as the key and value, and multi-head cross-attention computation is performed to obtain the updated structural branch representation. The updated ligand branch representation, protein branch representation, and structural branch representation are fused to obtain the cross-attention fused representation.

8. The method according to claim 1, characterized in that, The affinity prediction network described in S4 is a multi-layer fully connected network; the cross-attention fusion representation is input into the affinity prediction network to obtain the affinity prediction value; the regression loss is calculated based on the affinity prediction value and the true value, and the parameters of the shell graph encoder, the cross-attention fusion module and the affinity prediction network are updated.

9. The method according to claim 1, characterized in that, The method is used for virtual screening of candidate drugs, specifically including: fixing the same protein target, taking multiple candidate ligands as ligands in the protein-ligand pair to be predicted, performing steps S1 to S4 as described in claim 1, and obtaining the predicted binding affinity values ​​of each candidate ligand to the protein target; sorting the multiple candidate ligands according to the predicted binding affinity values, and using the sorting results as the output of virtual screening.

10. A protein-ligand binding affinity prediction system, characterized in that, include: The data acquisition module is used to acquire the three-dimensional structure of the protein, the three-dimensional structure of the ligand, the amino acid sequence of the protein, and the molecular representation of the ligand for the protein-ligand pair to be predicted. The shell graph construction module is used to divide the ligand binding pocket into multiple radial shells with the ligand as the center, use each radial shell as a node and establish edges according to the nearest neighbor relationship between the radial shells, and generate shell node features according to the contact pattern between protein structural units and ligands in each radial shell. The structure representation extraction module is used to process the shell graph and shell node features using a shell graph encoder to obtain the structure representation of the protein-ligand complex. The semantic representation extraction module is used to obtain the protein semantic representation based on the protein amino acid sequence and to obtain the ligand semantic representation based on the ligand molecule representation; The cross-attention fusion module is used to use each of the structural representation, protein semantic representation and ligand semantic representation as a query representation, and to aggregate information using the other two representations as key representation and value representation to obtain the cross-attention fusion representation. The prediction output module is used to output a predicted value of protein-ligand binding affinity based on the cross-attention fusion representation.

Images