A method, system, device, medium and application for designing protein-ligand binding pocket sequences based on a heterogeneous graph neural network, and a method for evaluating protein mutations

Abstract

The application discloses a protein-ligand binding pocket sequence design method based on a heterogeneous graph neural network, a protein mutation evaluation method, a system, a device, a medium and application, and belongs to the technical field of computational biology and artificial intelligence assisted protein design. The method constructs a heterogeneous graph structure containing protein residue nodes and ligand atom nodes, fuses protein sequence evolution information, protein spatial geometry information and protein-ligand interaction information, and realizes joint modeling of the local microenvironment of the protein binding pocket. The method can improve the protein sequence recovery rate and the binding pocket design accuracy, and can be applied to the fields of protein engineering, enzyme activity optimization, protein-small molecule binding interface optimization and artificial intelligence assisted drug design.

Description

Technical Field

[0001] This invention belongs to the field of computational biology and artificial intelligence-assisted protein design technology. Specifically, it relates to a design method for protein-ligand binding pocket sequences based on heterogeneous graph neural networks, a protein mutation assessment method, system, device, medium, and application. Background Technology

[0002] Inverse folding is a core computational task in protein design, aiming to deduce the amino acid sequence that can fold to form the target protein's three-dimensional scaffold structure. Essentially, it solves the problem of reverse mapping from structure to sequence. Its scientific significance lies in overcoming the evolutionary limitations of natural proteins, enabling "on-demand customization"—for example, designing high-temperature stable industrial enzymes, high-affinity targeted drug proteins, or artificial catalytic centers for biodegradable plastics, thus promoting the development of synthetic biology, precision medicine, and green chemistry. Simultaneously, inverse folding models can reveal deep sequence-structure-function relationships and predict the properties of mutant proteins with "zero-sample" results, thereby bridging the gap between fundamental understanding of protein physics and applications in life science engineering.

[0003] Existing backsheet models typically aim for high sequence recovery rates, but higher sequence recovery rates do not necessarily indicate stronger protein function. Therefore, they perform poorly in zero-sample prediction tasks such as catalytic activity optimization, and the designed sequences may lose valuable evolutionary advantages. Furthermore, these models rely heavily on main chain information, making it difficult to accurately describe the physicochemical microenvironment of the side chains directly involved in catalysis.

[0004] Relevant patent documents retrieved: This document, published in China (CN122090950A) on May 26, 2026, discloses a training method, design method, apparatus, and equipment for a protein reverse folding model. The training method includes: acquiring a training sample set, which includes structural information of the target protein and the binding agent protein; inputting the training samples into the protein reverse folding model to obtain multiple candidate amino acid sequences of the binding agent protein; calculating the binding performance indices of the target protein and each candidate amino acid sequence of the binding agent protein, including complex conformation quantification indices and / or binding affinity quantification indices; constructing positive and negative sample pairs based on the binding performance indices; and training the protein reverse folding model using a reinforcement learning algorithm and the positive and negative sample pairs to obtain the trained protein reverse folding model.

[0005] This document, published in China (CN120032720A) on May 23, 2025, discloses a method, system, and apparatus for predicting protein mutation effects based on a Boltzmann-aligned backfolding model. By using a backfolding model, the document selects the protein backbone structures of mutant and wild-type protein complexes before and after binding, preprocesses them to obtain the structural and sequence features of the protein backbone, and then captures local and global features of the protein backbone structure to predict the conditional probability of the protein backbone sequence. Based on the conditional probability, it combines the Boltzmann distribution and thermodynamic cycle principles to predict the change in binding free energy. The parameters are then fine-tuned using data from wet-scale experiments to solve for the final change in binding free energy. The method provided by this invention predicts the conditional probability of the protein backbone sequence based on a backfolding model, predicts the change in binding free energy, and introduces wet-scale experimental data labels for fine-tuning, enhancing prediction accuracy and generalization ability, and improving the success rate of protein compound design.

[0006] The prior art represented by the aforementioned documents has at least the following unresolved technical problems or defects: 1. In CN122090950A, the confidence scores of prediction models for unseen mutants (especially non-natural amino acid combinations) are often inflated, time-consuming, and costly. The relevant evidence is that pLDDT, PAE, and ipTM predicted by models such as HelixFold3 are used as the "gold standard" for conformation quality, but tend to generate sequences that make the prediction model "feel good" but are actually folded incorrectly. Reinforcement learning (especially DPO / PPO) requires tens of thousands or even hundreds of thousands of labeled samples to converge, while the throughput of real SPR / ITC wet experiments is extremely low.

[0007] Unlike traditional global backfolding methods, this application focuses on the key region determining catalytic performance—the substrate binding pocket and its surrounding structures—thus directly achieving "function-oriented" structure-to-sequence mapping. We have developed a protein backfolding model centered on the active pocket, focusing on protein redesign in the molecular binding region to achieve high binding strength and high catalytic activity. Compared to existing methods, our model integrates both evolutionary and local interaction information, resulting in superior performance not only in structure-to-sequence mapping prediction but also in zero-sample prediction of protein fitness, thermostability, and pocket-structural disruptive mutations. This improvement is expected to provide a more reliable computational tool for function-oriented enzyme design and targeted drug development. Summary of the Invention

[0008] The purpose of this invention is to provide: A design method, evaluation method, system, device, medium, and application of protein-ligand binding pocket sequences based on heterogeneous graph neural networks, and related technologies, to solve technical problems such as low prediction probability, dependence on main chain information, inaccurate description of the physicochemical microenvironment of the side chains directly involved in catalysis, low aggregation strength, and low catalytic activity, or combinations thereof.

[0009] Terminology Explanation: Unless otherwise defined, all technical terms in this document have the same meanings as commonly understood by one of ordinary skill in the art to which the subject matter of the claims pertains. Unless otherwise stated, all patents, patent inventions, and publications cited in this document are incorporated herein by reference in their entirety. If multiple definitions exist for terms in this document, the definitions in this chapter shall prevail.

[0010] It should be understood that the above brief description and the following detailed description are exemplary and for illustrative purposes only, and do not limit the subject matter of the invention in any way. In this invention, the singular is used in conjunction with the plural unless otherwise specifically stated. It should also be noted that, unless otherwise stated, the use of “or” or “or” means “and / or”. Furthermore, the use of the term “comprising” and other forms such as “including,” “containing,” and “contains” are not limiting.

[0011] The definition of the standard term can be found in the reference "Biocomputational Simulation and Protein Design (Chemical Industry Press, 14th Five-Year Plan Textbook for General Higher Education), Editor: Duan Mojie, Editor-in-Chief: Fu Zhousi, ISBN: 9787122497024, Publication Date: 2026-03, Publisher: Chemical Industry Press".

[0012] Unless specifically defined herein, the use of all commercially available products herein employs standard techniques. For example, it may be carried out using the manufacturer's instructions for use with the kit, or in accordance with methods known in the art or the description of this invention. The techniques and methods described herein can generally be implemented according to conventional methods well known in the art, based on the descriptions in the various summary and more specific documents cited and discussed in this specification.

[0013] The terms “optional / arbitrary” or “optionally / arbitrarily” mean that the event or situation described below may or may not occur, including both the occurrence and non-occurrence of the event or situation.

[0014] The term "heterogeneous graph neural network" used in this paper refers to a graph neural network architecture capable of processing graph data containing multiple types of nodes and edges. In protein-ligand interaction prediction, heterogeneous graph neural networks encode protein-ligand complexes as heterogeneous graphs containing two types of nodes (e.g., protein residue nodes and ligand atom nodes) and multiple edge types. By designing dedicated information propagation and aggregation strategies for different node and edge types, it captures high-order complex dependencies and rich semantic information in biological heterogeneous graphs. For example, the GIF-PLA method represents protein-ligand binding complexes as heterogeneous graphs with metapaths to capture topological interactions and high-order nonlinear relationships.

[0015] The term "ligand heavy atom" used in this article refers to heavy atoms other than hydrogen atoms in the ligand molecule (such as non-hydrogen atoms like carbon, nitrogen, oxygen, sulfur, and phosphorus). In graph-based protein-ligand interaction modeling, ligand heavy atoms are typically encoded as "ligand atom nodes" in isomorphic graphs, with node features including one-hot encoding of the atomic element type (such as C, N, O, S, P, etc.). Heavy atoms carry richer structural information and chemical characteristics than hydrogen atoms and are the core building blocks of ligand molecule graph representations.

[0016] The term "protein residue node" used in this paper refers to the amino acid residues in a protein as the basic unit of a node in a graph neural network. In protein structure graph representation, each amino acid residue corresponds to a node, carrying features such as residue type (amino acid species), residue index, and position encoding. The edges between protein residue nodes represent spatial or sequence relationships between residues (such as distance, contact, hydrogen bonds, etc.). In protein-ligand isomer graph representation, protein residue nodes and ligand atom nodes constitute two core node types.

[0017] The term "ligand atom node" used in this paper refers to using atoms from ligands (small molecules) as the basic unit of nodes in a graph neural network. In protein-ligand isomer graph representations, ligand atom nodes and protein residue nodes are considered two core node types. The characteristics of ligand atom nodes typically include one-hot encoding of the atomic element type (e.g., C, N, O, S, P, Zn, Fe, Cu, etc.), the coordinates of the atom in three-dimensional space, and the chemical group to which the atom belongs. Ligand atom nodes are connected by covalent bonds to form a molecular graph.

[0018] The term "embedding encoding" used in this paper refers to the process of converting discrete, high-dimensional raw input data (such as atom type, residue type, molecular structure, etc.) into a continuous low-dimensional vector representation. In protein design, embedding encoding is typically implemented through learnable neural network layers, mapping the raw features of each node (residue or atom) to a unified latent space dimension, generating node embedding vectors. Embedding encoding enables graph neural networks to process different types of node information in a unified vector space, providing a foundation for subsequent message passing and feature fusion.

[0019] The term "residue node encoding module" used in this paper refers to a neural network module specifically designed to encode the raw features of protein residues into node embedding vectors. This module receives input features for each residue (such as amino acid type, main chain atom coordinates, residue index, etc.) and converts them into a fixed-dimensional latent space representation through a learnable encoding layer. In the protein-ligand heteromorphism graph model, the residue node encoding module and the ligand node encoding module work in parallel, encoding features for the two types of nodes respectively to generate node embeddings of a unified dimension.

[0020] The term "ligand node encoding module" used in this paper refers to a neural network module specifically designed to encode the raw features of ligand atoms into node embedding vectors. This module receives input features for each ligand atom (such as atom element type, 3D coordinates, chemical bond information, etc.) and converts them into a fixed-dimensional latent space representation through a learnable encoding layer. In the protein-ligand heterogeneity graph model, the ligand node encoding module and the residue node encoding module share the same architecture but accept different input feature dimensions; both generate node embeddings of a unified dimension.

[0021] The term "intra-residue message passing module" used in this paper refers to a module in a protein-ligand isomorphic graph neural network specifically responsible for information propagation and aggregation between protein residue nodes. This module, through the message passing mechanism of the graph neural network, enables each residue node to collect information from its neighboring residue nodes (such as residues that are close in three-dimensional space or residues connected by a contact graph) and update its own node representation. The intra-residue message passing module aims to capture long-range interactions and co-evolutionary information within proteins.

[0022] The term "ligand internal message passing module" used in this paper refers to a module in a protein-ligand isomeric graph neural network specifically responsible for information propagation and aggregation between ligand atomic nodes. This module utilizes the message passing mechanism of the graph neural network to enable each ligand atomic node to collect information from its covalently bonded neighboring atomic nodes and update its own node representation. The ligand internal message passing module aims to capture information about the chemical structure and electronic effects within the ligand molecule.

[0023] The term "protein-ligand cross-node type message passing module" used in this paper refers to a module in a protein-ligand heterogeneous graph neural network specifically responsible for cross-type information propagation between protein residue nodes and ligand atom nodes. This module enables protein nodes to obtain information from ligand nodes, and vice versa, thereby achieving bidirectional information interaction between proteins and ligands. This module is a core component for capturing protein-ligand intermolecular interactions (such as hydrogen bonds, hydrophobic interactions, and van der Waals forces).

[0024] The term "Graph Attention Convolutional Network GATConv" used in this paper refers to: a graph neural network convolutional layer based on an attention mechanism, developed by Veli. kovi GATConv is the core convolutional layer in Graph Attention Networks proposed by [Authors' Name] at the 2018 ICLR conference. Unlike traditional graph convolutional networks, GATConv assigns different weights to different neighbor nodes through an attention mechanism when aggregating neighbor node information, enabling the model to automatically focus on the neighbor information most important to the current node. In protein-ligand interaction modeling, GATConv is often used for the transformation and updating of node features.

[0025] The term "attention-based cross-message module" used in this paper refers to a cross-type information transfer module based on an attention mechanism in protein-ligand heterogeneous graph neural networks. This module borrows the cross-attention mechanism from Transformers, using features from one type of node as the query and features from another type of node as the key and value. Selective information exchange between the two types of nodes is achieved by calculating attention weights. The attention-based cross-message module can dynamically focus on the most important features and relationships in protein-ligand interactions.

[0026] The term "Query Vector Generation Module" used in this paper refers to the submodule responsible for generating Query vectors within the Attention Cross-Message Passing Module. This module projects the embedding vector of the source node (the party initiating the information query) onto the query space through a linear transformation to generate the Query vector. The Query vector is used to calculate an attention score with the Key vector, thereby determining the degree of attention the source node pays to each feature of the target node.

[0027] The term "Key Vector Generation Module" used in this article refers to the submodule responsible for generating the Key vector within the Attention Cross-Message Passing Module. This module projects the embedding vector of the target node (the queried party) onto the key space through a linear transformation to generate the Key vector. The Key vector and the Query vector are used together to calculate the attention score, which determines how the source node should aggregate information from the target node.

[0028] The term "Value vector generation module" used in this paper refers to the submodule responsible for generating Value vectors within the attention cross-message passing module. This module projects the embedding vector of the target node onto the value space through a linear transformation to generate the Value vector. The Value vector carries the actual feature information of the target node. After calculating the attention weights, the Value vectors are aggregated through a weighted summation to obtain the updated node representation.

[0029] The term "attention normalization module" used in this paper refers to the submodule within the attention cross-messaging module responsible for normalizing the calculated attention scores. This module typically uses the Softmax function to convert the original attention scores (the dot product or similarity between the query and key) into normalized attention weights, ensuring that the sum of the attention weights for all target nodes is 1. The attention normalization module ensures that the model can focus on the most important features in the form of a probability distribution.

[0030] The term "mutation confidence" used in this paper refers to the degree of confidence a model has in predicting an amino acid mutation at a specific position in protein sequence design or mutation effect prediction. In deep learning-based protein design, models typically output the probability distribution of each amino acid type at each position. Mutation confidence reflects the model's certainty about the prediction at that position and is usually positively correlated with the predicted probability value. Mutation sites with high confidence indicate that the model has a high degree of confidence in their predictions and can be considered as priority candidates for experimental validation.

[0031] The term "negative log-probability fitness index" used in this article refers to an indicator used in protein sequence design and mutation assessment to measure the degree of fit between the amino acid sequence predicted by the model and the target structure. This index is quantified by calculating the negative value of the log probability assigned by the model to a given sequence—the lower the negative log probability value, the higher the model's assessment of the sequence's fit with the target structure. In protein design tasks, this index is commonly used to screen and rank candidate mutants.

[0032] The term "mutation probability score" used in this article refers to the probability value assigned by a model to a specific amino acid type at a given position in protein sequence design or mutation effect prediction. This score is typically generated by the model's output layer (such as a Softmax layer) and represents the likelihood of a particular amino acid appearing at that position given a protein structure and environmental conditions. A higher mutation probability score indicates that the model favors that mutation in terms of structural compatibility and functional adaptability, and it can serve as an important basis for mutation screening and sequence design.

[0033] In a first aspect, the present invention provides: a method for designing protein-ligand binding pocket sequences based on heterogeneous graph neural networks, comprising the following steps: (1) Read and parse the information; (2) Construct a protein-ligand isomer map; (3) Calculate the structural geometry and non-covalent interaction features between protein residues; (4) Calculate the characteristics of non-covalent interactions between protein residues and ligand atoms; (5) Mask the residues to be designed; (6) Construct a heterogeneous graph neural network and perform message passing; (7) Output the predicted probabilities of the 20 natural amino acids corresponding to the residues to be designed, and determine the target amino acid type based on the predicted probabilities; (8) Output the model after performing iterative sequence design.

[0034] Preferably, the information read in step (1) includes a protein receptor structure file and a ligand small molecule structure file, wherein the protein receptor structure file is a PDB format file and the ligand small molecule structure file is a MOL2 format file.

[0035] Preferably, the content analyzed in step (1) includes protein residue number, chain number, atomic coordinates, ligand atom type, atomic charge, atomic bonding information, aromatic ring information, and hydrogen bond donor and acceptor information.

[0036] Preferably, the protein-ligand isomer diagram described in step (2) consists of protein residue nodes, ligand atom nodes, and edge structures.

[0037] Preferably, the protein residue node features are residue embedding features, residue main chain dihedral features, and amino acid type one-hot encoding features generated by the protein language model.

[0038] Preferably, the protein language model is the ESM2 protein language model.

[0039] Preferably, the dimension of the residue embedding feature is 640.

[0040] Preferably, the ligand atom node is characterized by a one-hot encoded feature of the ligand atom type SYBYL.

[0041] Preferably, the edge structure is a protein residue-residue peptide bond edge, a protein residue-residue geometric edge, a protein residue-residue interaction edge, a protein residue-ligand atom geometric edge, a protein residue-ligand atom interaction edge, and a ligand atom-ligand atom chemical bond edge.

[0042] Preferably, the features of the geometric edges are encoded by expanding the distance matrix using radial basis functions and combining it with local directional projection features.

[0043] Preferably, the characteristics of the interaction edges are encoded using a non-covalent interaction contact counting method.

[0044] Preferably, the protein residue-residue peptide bond edges are used to characterize the topological connectivity of the protein backbone; The protein residue-residue geometric edges are used to characterize the spatial geometric relationships between residues; The protein residue-residue interaction edges are used to characterize the non-covalent interaction relationships between residues; The protein residue-ligand atom geometric edge is used to characterize the spatial geometric relationship between the residue and the ligand atom; The protein residue-ligand atom interaction edge is used to characterize the non-covalent interaction relationship between the residue and the ligand atom; The ligand atom-ligand atom chemical bond edge is used to characterize the covalent bond connection relationship within the ligand.

[0045] Preferably, the geometric features described in step (3) are the features of the interatomic distance matrix after radial basis function expansion, the local coordinate system directional projection features, and the normalized distance scalar features.

[0046] Preferably, the geometric features described in step (3) include: geometric edge features of protein residue-residue and geometric edge features of protein residue-ligand atom.

[0047] Preferably, the non-covalent interactions between protein residues and between protein residues and ligand atoms in step (4) include: One or more of the following: hydrophobic interaction, hydrogen bond, weak hydrogen bond, π-π stacking, salt bridge, amide-π stacking, cation-π interaction, halogen bond, disulfide bond and halogen multipolar interaction.

[0048] Preferably, the interaction is encoded using a contact counting method and encoded as an edge feature vector.

[0049] Preferably, in step (5), the site of the residue to be designed is replaced with a marker to replace the original amino acid type in order to avoid leakage of target residue information.

[0050] Preferably, the edge features of the related interactions of the residues to be designed in step (5) are all set to zero.

[0051] Preferably, the heterogeneous graph neural network described in step (6) is a residue-internal message passing module, a ligand-internal message passing module, and a protein-ligand cross-node type message passing module.

[0052] Preferably, the residue intra-residue message passing module aggregates neighborhood node information along peptide bond edges, residue geometric edges, and residue interaction edges, respectively.

[0053] Preferably, the protein-ligand cross-node type message passing module realizes information interaction between protein residue nodes and ligand atom nodes through residue-ligand geometric edges and residue-ligand interaction edges.

[0054] Preferably, the message aggregation of the peptide bond edges is implemented using a graph attention convolutional network GATConv.

[0055] Preferably, the residue geometry edges, residue interaction edges, residue-ligand geometry edges, and residue-ligand interaction edges are implemented using an attention-crossing message passing module.

[0056] Preferably, the attention cross-message passing module comprises a query vector generation module, a key vector generation module, a value vector generation module, and an attention normalization module.

[0057] Preferably, the Value vector is generated by concatenating source node features, target node features, and edge features.

[0058] Preferably, the message passing in step (6) specifically involves: after each round of message passing, synchronously updating the edge features based on the updated node features.

[0059] Preferably, the heterogeneous graph neural network described in step (6) includes a 4-round message passing process.

[0060] Preferably, the predicted probability in step (7) is obtained through the output layer of a fully connected neural network.

[0061] Preferably, the output layer of the fully connected neural network is a LayerNorm layer, a ReLU activation layer, a Dropout layer, and a fully connected layer.

[0062] Preferably, the iterative sequence design method described in step (8) is as follows: (8-1) Calculate the predicted confidence of the residues to be designed based on the predicted probability, and lock residue A after sorting; (8-2) Write the predicted amino acids corresponding to the locked residues back into the protein sequence, recalculate the protein language model embedding features, and reconstruct the heterogeneity graph; (8-3) Continue to predict the remaining unlocked residues; (8-4) Repeat steps (8-1) to (8-2) until all designed residues are predicted.

[0063] Preferably, the prediction confidence level in step (8-1) is calculated using the softmax probability value.

[0064] Preferably, the locking method in step (8-1) is as follows: if the prediction is iterated M times, the residues are ranked according to the confidence level of the current round in each iteration, and the residues from the first 1 / M, 2 / M to (M-1) / M are locked in turn. In the last round, the remaining 1 / M unknown residues are predicted, where M≥1.

[0065] Preferably, the iterative sequence design method described in step (8) further includes calculation steps for protein sequence recovery rate, prediction confidence, and Fitness score.

[0066] This invention uses heterogeneous graph structures to uniformly model the spatial, chemical, and sequence evolutionary relationships between protein residues and ligand atoms, and utilizes attention message passing mechanisms on multiple types of edges to achieve fine modeling of the local microenvironment of the protein binding pocket, thereby improving the accuracy of protein sequence design.

[0067] Secondly, the present invention provides a method for protein mutation assessment based on heterogeneous graph neural networks, comprising the following steps: S1. Read and parse the information; S2. Construct mutant protein sequences and generate mutant embedding features; S3. Construct a protein-ligand isomer map using the same isomer map construction method as step (2) in the design method described in claim 1; S4. Use the same heterogeneous graph neural network construction method as step (6) in the design method of claim 1 and perform message passing; S5. Use the same method as step (7) in the design method described in claim 1 to output the predicted probability and evaluate the rationality of the mutation; S6. Output the mutation score results.

[0068] Preferably, the protein mutation assessment method uses a single forward propagation approach to score mutations, eliminating the need for iterative sequence optimization.

[0069] Preferably, the information read in step S1 includes protein receptor structure files, ligand small molecule structure files, and mutation information to be evaluated.

[0070] Preferably, the mutation information to be evaluated includes chain number, original amino acid type, residue number, and target mutant amino acid type.

[0071] Preferably, the method for constructing the mutant protein sequence in step S2 is as follows: based on the input mutation information, the original amino acid at the corresponding position in the protein sequence is directly replaced with the target mutant amino acid.

[0072] Preferably, the indicators for evaluating the rationality of mutations in step S5 are mutation confidence, negative logarithmic probability fitness index, and target mutation amino acid probability value.

[0073] Preferably, the mutation confidence level is calculated by the predicted probability corresponding to the target mutated amino acid.

[0074] Preferably, the higher the predicted probability of the target mutated amino acid, the more reasonable the model considers the mutation to be.

[0075] Thirdly, the present invention provides the application of the above-described design method or the above-described protein mutation assessment method in protein engineering.

[0076] Preferably, the applications include one or more of the following: protein binding pocket optimization, enzyme activity optimization, protein stability optimization, protein-small molecule binding interface optimization, and drug design.

[0077] Fourthly, the present invention provides: a system for designing protein-ligand binding pocket sequences based on heterogeneous graph neural networks, comprising: The structure parsing module is used to read protein structure files and ligand structure files and parse the structural information. Pocket screening module for screening protein residues within the binding pocket region; The isomer map building module is used to construct protein-ligand isomer maps; The sequence embedding module is used to generate residue embedding features using a protein language model; The interaction calculation module is used to calculate the interaction characteristics between protein residues and between protein residues and ligand atoms. The heterogeneous graph neural network module is used for multi-round message passing and node feature updates; The prediction module is used to output the amino acid prediction results of the residues to be designed; The iterative design module is used for iterative sequence optimization based on prediction confidence.

[0078] Preferably, the heterogeneous graph neural network module includes: a residue-internal message passing unit, a ligand-internal message passing unit, and a cross-node type message passing unit.

[0079] Preferably, the cross-node type message passing unit realizes information exchange between ligand atom nodes and protein residue nodes through residue-ligand geometric edges and residue-ligand interaction edges.

[0080] Fifthly, the present invention provides: a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the above-described design method or the above-described protein mutation assessment method.

[0081] In a sixth aspect, the present invention provides: a computer device including a memory and a processor, wherein the memory stores a computer program capable of running on the processor, and the processor executes the computer program to implement the steps of the above-described design method or the above-described protein mutation assessment method.

[0082] Based on further solutions to the technical problems of the present invention, or simultaneous solutions to multiple technical problems, the preferred solution in the technical solution provided in the first aspect of the present invention includes: The first priority solution is a method for designing protein-ligand binding pocket sequences based on heterogeneous graph neural networks. This solution addresses the technical problem that "existing models mainly rely on main chain information, making it difficult to accurately describe the physicochemical microenvironment of the side chains directly involved in catalysis," and further solves the technical problems of "low prediction probability, aggregation strength, and catalytic activity."

[0083] The beneficial effects of this invention are as follows: Based on the input protein structure, after multiple rounds of iteration, a higher sequence recovery rate can be achieved compared to existing reverse-folding models. Zero-shot scoring of the structure-sequence pair is supported; the zero-shot scoring results are highly correlated with the adaptive function of the protein, effectively assessing functional changes caused by mutations. Attached Figure Description

[0084] Figure 1 This is a flowchart of a protein reverse folding design method with zero-sample prediction capability.

[0085] Figure 2This section compares the sequence design performance of different protein reverse folding methods. A represents the sequence recovery rate comparison between zPolaron and four baseline models; B represents the sequence recovery rate of zPolaron in confidence-guided design iterations; C represents the residue recovery rate as a function of the native residue PSSM score; D represents the pLDDT score distribution of the AlphaFold 3 predicted structure reconstructed from the designed sequence; E represents the structural similarity (TMscore) between the AlphaFold 3 predicted structure of the designed sequence and the input experimental structure (left half of each violin plot) and the corresponding native sequence's AlphaFold 3 predicted structure (right half); F represents the residue recovery rate within different B-factor intervals; and GK represents the distribution of ΔPSSM for the predicted sequences of the five models, defined as the difference between the PSSM score of the predicted residue and the PSSM score of the corresponding native residue.

[0086] Figure 3 This section compares the performance of different protein refolding methods for fitness scoring. In this section, A shows the Pearson correlation between zero-sample scores and experimentally measured fitness values ​​on the protein fitness benchmark dataset; B shows the Spearman correlation between zero-sample scores and experimentally measured fitness values ​​on the protein fitness benchmark dataset; C shows the t-SNE visualization of the final zPolaron embedding space for the TIM_SULSO target; and D shows the t-SNE visualization of the final zPolaron embedding space for the DAOx target with D-Met as the substrate.

[0087] Figure 4 This section compares the performance of different protein refolding methods in terms of thermal stability and mutation scoring. A represents the Pearson correlation between zero-sample score and ΔΔG value on the S2648 protein stability dataset; B represents the Spearman correlation between zero-sample score and ΔΔG value on the S2648 protein stability dataset; C represents the ROC curves of the five methods on the PETase dataset; D represents the precision-recall curves of the five methods on the PETase dataset; E represents the ROC curves of the five methods on the ADK dataset; F represents the precision-recall curves of the five methods on the ADK dataset; G represents the AlphaFold 3 predicted structure of a representative wild-type enzyme in the PETase dataset; H represents the AlphaFold 3 predicted structure of a representative pocket-property-reversed mutant enzyme in the PETase dataset; I represents the AlphaFold 3 predicted structure of a representative wild-type enzyme in the ADK dataset; and J represents the AlphaFold 3 predicted structure of a representative pocket-property-reversed mutant enzyme in the ADK dataset. Detailed Implementation

[0088] The following non-limiting embodiments are intended to enable those skilled in the art to gain a more comprehensive understanding of the present invention, but do not limit the invention in any way. The following content is merely an exemplary description of the scope of protection claimed by the present invention, and those skilled in the art can make various changes and modifications to the present invention based on the disclosed content, and such changes should also fall within the scope of protection claimed by the present invention.

[0089] The present invention will be further described below by way of specific embodiments. Unless otherwise specified, all instruments, devices, equipment, reagents, products, etc., used in the embodiments of the present invention are obtained through conventional commercial means.

[0090] Example 1: A method for designing protein-ligand binding pocket sequences based on heterogeneous graph neural networks (1) Read and parse the information: The program was used to parse the protein receptor PDB file and the ligand MOL2 file to extract the protein residue number, chain number, atomic coordinates, atom type and ligand chemical bond information; Simultaneously identify aromatic ring structures, hydrogen bond donors and acceptors, charged groups, and halogen atom information in ligands; The geometric center of all heavy atoms of the ligand is used as the binding pocket center, and protein residues within a 40 Å range of the geometric center are selected as pocket residues. (2) Constructing a protein-ligand isomer diagram: The isomorphism diagram consists of protein residue nodes, ligand atom nodes, and edge structures. Among them, protein residue node features include: The ESM2 protein language model generates 640-dimensional residue embedding vectors, sin / cos coding features of residue main chain dihedral angles φ, ψ, and ω, and one-hot coding features of amino acid types. Ligand atom node features include: One-heat coding features of SYBYL type ligand atoms; The edge structure of the heterogeneous graph is as follows: Protein residue-residue peptide bond edges are used to describe the topological connections of the protein backbone; Protein residue-residue geometry edges are used to describe the spatial geometric relationships between residues; Protein residue-residue interaction edges are used to describe non-covalent interactions between residues; Protein residue-ligand atom geometric edges are used to describe the spatial geometric relationship between residues and ligand atoms; Protein residue-ligand atom interaction edges are used to describe the non-covalent interaction relationships between residues and ligand atoms; The ligand atom-ligand atom chemical bond edge is used to describe the covalent bond connection relationship within the ligand; (3) Calculate the geometric features of protein structure: For spatially adjacent white protein residue pairs, the distance matrix between main chain atoms is calculated and encoded using radial basis function (RBF) expansion. At the same time, the protein residue-residue geometric edge features are generated by combining local coordinate system orientation projection features and normalized distance scalar features. The dimension of the protein residue-residue geometric edge features is 260. To determine the spatial relationship between protein residues and ligand atoms, the distances from the main chain atoms of the residues and virtual CB atoms to the ligand atoms are calculated and encoded using radial basis function expansion. At the same time, local directional projection features are added to generate geometric edge features of protein residues and ligand atoms. The geometric edge features of protein residues and ligand atoms have a dimension of 43.

[0091] (4) Calculate the characteristics of non-covalent interactions between protein residues and between protein residues and ligand atoms: The non-covalent interactions include: Hydrophobic interactions, hydrogen bonds, weak hydrogen bonds, π-π stacking, salt bridges, amide-π stacking, cation-π interactions, halogen bonds, disulfide bonds, and halogen multipolar interactions; For each interaction, the corresponding number of contacts is counted and encoded as an edge feature vector; Among them, the dimension of the protein residue-residue interaction edge feature is 13-dimensional; the dimension of the protein residue-ligand atom interaction edge feature is 19-dimensional. (5) Masking the residues to be designed: For the site to be designed, the corresponding amino acid in the protein sequence is replaced with a label and input into the ESM2 protein language model to generate embedding features, thereby avoiding the leakage of the original amino acid information; All interaction edge features connected to the residues to be designed are set to zero to avoid the model directly obtaining the original interaction information of the target residues; (6) Construct a heterogeneous graph neural network and perform message passing: First, the node features and edge features are uniformly projected into a 128-dimensional vector space through a linear mapping layer and a LayerNorm normalization layer; Subsequently, a heterogeneous graph neural network is used for message passing and node feature updating, where: Peptide bond edges within protein residues are aggregated using a graph attention convolutional network (GATConv). Geometric edges and interaction edges between protein residues are feature-aggregated using an attention-crossing message-passing module; The chemical bond edges inside the ligands are updated using GATConv to update the ligand atom node features; Protein residue nodes and ligand atom nodes communicate across node types via residue-ligand geometric edges and residue-ligand interaction edges; After each round of message passing is completed, the edge features are updated synchronously based on the updated node features. This heterogeneous graph neural network performs a total of 4 rounds of message passing. (7) Output the predicted probabilities of the 20 natural amino acids corresponding to the residues to be designed, and determine the target amino acid type based on the predicted probabilities: After computation by the heterogeneous graph neural network, the predicted probabilities of the 20 natural amino acids corresponding to the residues to be designed are output through the fully connected output layer, and the protein design sequence is generated based on the amino acid type with the highest predicted probability. The predicted probability is obtained through the output layer of a fully connected neural network: the output layer of the fully connected neural network consists of a LayerNorm layer, a ReLU activation layer, a Dropout layer, and a fully connected layer; (8) Output the model after performing iterative sequence design: (8-1) First, calculate the predicted confidence of the residues to be designed based on the predicted probability. Then, sort them from high to low according to the predicted confidence and select the top 1 / M number of residues with the highest confidence to lock. (8-2) Write the predicted amino acids corresponding to the locked residues back into the protein sequence, recalculate the sequence embedding features using the ESM2 protein language model, and reconstruct the heterogeneity map; (8-3) Continue to predict the unlocked residues; (8-4) Repeat steps (8-1) to (8-2) until all residues to be designed are predicted and the model is output.

[0092] The evaluation metrics output by this model include: Recovery is used to measure the recovery rate between the predicted sequence and the natural sequence. Confidence measures the level of confidence a model has in its predictions. Fitness is used to measure the plausibility of predicted amino acids in their current structural context.

[0093] Example 2: A protein mutation assessment method based on heterogeneous graph neural networks This embodiment provides a protein mutation assessment method based on heterogeneous graph neural networks. The method is used to assess the rationality of a given protein mutation in the corresponding protein structure and ligand binding environment, and includes the following steps: S1 reads and parses the information: Read the protein receptor structure file, ligand small molecule structure file, and mutation information to be evaluated. The mutation information to be evaluated includes chain number, original amino acid type, residue number, and target mutant amino acid type. Mutation information is represented in the format of "chain number_original amino acid type + residue number + target amino acid type"; S2. Construct mutant protein sequences and generate mutant embedding features: Based on the input mutation information, the original amino acids at the corresponding positions in the protein sequence are directly replaced with the target mutant amino acids, thereby generating a mutant sequence. The mutant sequence is input into the ESM2 protein language model to obtain the embedding vector features corresponding to each residue. The ESM2 embedding feature has a dimension of 640 and carries the semantic information of the mutated sequence, which is used to characterize the evolutionary rationality of the mutated protein sequence. S3. Construct a protein-ligand isomer map using the same isomer map construction method as step (2) in the above design method: Heterogeneity maps are constructed using protein structure information and ligand structure information.

[0094] The heterogeneous graph includes protein residue nodes, ligand atom nodes, and different types of edge structures, wherein the edge structures include: Protein residue-residue peptide bond edge; Protein residue-residue geometry; Protein residue-residue interaction edges; Protein residue-ligand atom geometry; Protein residue-ligand atom interaction edges; Ligand atom-ligand atom chemical bond edge.

[0095] Geometric edge features are encoded by expanding the distance matrix using radial basis functions and combining it with local directional projection features; Interaction edge features are encoded using a non-covalent interaction contact counting method; S4. Use the same heterogeneous graph neural network construction method as step (6) in the above design method and perform message passing: First, the node features and edge features are uniformly projected into a 128-dimensional vector space through a linear mapping layer and a LayerNorm normalization layer; Subsequently, a heterogeneous graph neural network is used for message passing and node feature updating, where: Peptide bond edges within protein residues are aggregated using a graph attention convolutional network (GATConv). Geometric edges and interaction edges between protein residues are feature-aggregated using an attention-crossing message-passing module; The chemical bond edges inside the ligands are updated using GATConv to update the ligand atom node features; Protein residue nodes and ligand atom nodes communicate across node types via residue-ligand geometric edges and residue-ligand interaction edges; After each round of message passing is completed, the edge features are updated synchronously based on the updated node features. This heterogeneous graph neural network performs a total of 4 rounds of message passing. S5. Use the same output prediction probability method as step (7) in the above design method and evaluate the rationality of the mutation: After computation by the heterogeneous graph neural network, the predicted probabilities of the 20 natural amino acids corresponding to the residues to be designed are output through the fully connected output layer. Based on the predicted probabilities, the mutation confidence and fitness index are calculated. The predicted probability is obtained through the output layer of a fully connected neural network: the output layer of the fully connected neural network consists of a LayerNorm layer, a ReLU activation layer, a Dropout layer, and a fully connected layer; The higher the predicted probability of the target mutated amino acid, the more reasonable the model considers the mutation to be in the current protein structure and ligand binding environment; The fitness index is calculated using the negative logarithm of the probability of the target mutated amino acid. The lower the fitness value, the better the target mutation matches the current structural environment. S6. Output the mutation score results.

[0096] In this embodiment, the model uses a single forward propagation to complete the mutation scoring, without the need for iterative sequence optimization.

[0097] The final output includes: the probability value of the target mutated amino acid, the confidence index, and the fitness index.

[0098] Comparative Example 1 The design method of Example 1 is modified as follows: In step (2), the ESM2 evolutionary embedding is removed: all ESM2 pre-trained language model embeddings in the residue node features are set to zero, and only the dihedral geometric structure features and amino acid identity one-hot encoding features are retained. This experiment is used to test whether evolutionary prior information provides incremental contributions beyond structural geometric encoding. The remaining steps are the same as in Example 1.

[0099] Comparative Example 2 The design method of Example 1 is modified as follows: The main chain structure features in step (2) are removed: the main chain dihedral geometry and amino acid identity one-hot encoding features in the residue node features are all set to zero, and only the ESM2 embedding is retained. This experiment is used to test the necessity of explicit structural geometry features, that is, whether the implicit structure perception of the language model can replace explicit geometric encoding. The remaining steps are the same as in Example 1.

[0100] Comparative Example 3 The design method of Example 1 is modified as follows: In step (2), the interaction edge features are removed: the edge attributes of the interaction edges between residues and between residues and ligands are all set to zero, and the edge prediction auxiliary training task is removed. The model relies only on covalent bond edges (peptide bonds, ligand chemical bonds) and geometric edges for message passing. This experiment is used to test whether explicit non-covalent interaction encoding provides chemical constraints that cannot be replaced by geometric distance information. The remaining steps are the same as in Example 1.

[0101] Detection Example 1 1. Experimental Methods Using the design methods of Example 1 and Comparative Examples 1-3, sequence recovery tasks were performed on newly released PDB structures from 2023 to 2025 (with sequence similarity of less than 0.4 to PDB structures released before 2023).

[0102] 2. Experimental Results The method (zPolaron) of Embodiment 1 of this invention achieved an average recovery rate of 81.68% after 10 iterations. The ablation model removing ESM2 evolutionary embeddings achieved an average recovery rate of 72.20% after 10 iterations, the ablation model removing main chain structural features achieved an average recovery rate of 55.08% after 10 iterations, and the ablation model removing interaction edge features achieved an average recovery rate of 42.49% after 10 iterations. This comparison demonstrates that the ESM2 embedding information, protein main chain geometry information, and interaction edge information in the method (zPolaron) of Embodiment 1 of this invention all significantly contribute to the sequence recovery task, indicating that only by adopting the technical solution of this application can the technical effect of this application be achieved.

[0103] Detection Example 2 1. Experimental Methods The design method of Example 1 is used to perform sequence recovery tasks on PDB structures newly released between 2023 and 2025 (with sequence similarity of less than 0.4 with PDB structures released before 2023).

[0104] 2. Experimental Results The results are as follows Figure 2The diagram illustrates the method (zPolaron) of Embodiment 1 of this invention and four mainstream reverse folding methods: ABACUS-T, CARBonAra, LigandMPNN, and ProteinMPNN. These four methods generally only use protein backbone structure features as input and do not include ESM2 embedding information or interaction feature information.

[0105] The results show that the method of Example 1 of the present invention (zPolaron) has a higher sequence recovery rate than other reverse folding methods, and the recovery rate increases significantly with the number of iterations, with 4 iterations.

[0106] This test evaluated the recovery of each residue. The method (zPolaron) in Example 1 of this invention showed a clear evolutionary pattern. As the PSSM of the initial residue increased, the success rate gradually increased, and the average PSSM after mutation was the largest, which also conformed to evolutionary constraints. Other methods are more sensitive to the uncertainty of the input structure, while the method of this invention still maintains a higher recovery success rate with the help of evolutionary information.

[0107] This test also evaluated the structure of the designed sequence. The structure predicted by AlphaFold3 had a pLDDT confidence level of nearly 90% and an average structural similarity of more than 0.95 with the input structure, indicating that the output sequence can maintain the same structure as the input protein backbone.

[0108] Detection Example 3 1. Experimental Methods Based on the protein mutation assessment method in Example 2, the scoring performance of various backfolding models was evaluated.

[0109] 2. Experimental Results Figure 3 This indicates that the protein mutation assessment method (zPolaron) of Example 2 of the present invention exhibits the best performance in protein fitness testing, showing optimal correlation at almost every target site. Furthermore, the embedding vectors output by the model show a clear pattern for mutants with different fitness levels after dimensionality reduction, with high-fitness and low-fitness mutations clustering at different locations.

[0110] Figure 4 This demonstrates that the protein mutation assessment method (zPolaron) of Example 2 of the present invention exhibits stable performance in protein thermostability testing, with at least a Pearson correlation of 0.331 and a Spearman correlation of 0.226 at all target sites. Furthermore, the protein mutation assessment method (zPolaron) of Example 2 of the present invention can identify mutants that cause fatal damage to the catalytic pocket with the highest accuracy.

[0111] In summary, this invention provides a method for designing protein-ligand binding pocket sequences and evaluating protein mutations based on heterogeneous graph neural networks. This method integrates protein sequence evolutionary information, spatial geometric information, and protein-ligand interaction information. By constructing a protein-ligand heterogeneous graph and combining it with an attention message passing mechanism, it achieves joint modeling of the binding pocket microenvironment, thereby improving the accuracy of protein sequence design and mutation evaluation. This invention not only enables intelligent design of key residues in protein binding pockets but also allows for quantitative evaluation of the rationality of a given mutation in a specific structural environment, demonstrating high application value in protein engineering, enzyme modification, and AI-assisted drug design.

[0112] Through sequence recovery task testing, the method of this invention (zPolaron) achieved an average recovery rate of 81.68% over 10 iterations on newly released PDB structures from 2023 to 2025, significantly outperforming mainstream methods such as ProteinMPNN and LigandMPNN. Ablation experiments further demonstrated that ESM2 evolutionary embedding, main chain structure features, and interaction edge features contributed approximately 9.5%, 26.6%, and 39.2% to the recovery rate improvement, respectively. The explicit encoding of interaction edges is the core innovation of this invention, effectively solving the problem of traditional methods relying on main chain information and struggling to accurately describe the side chain microenvironment. On the protein fitness zero-shot scoring benchmark, the Pearson and Spearman correlations of the method of this invention were superior to the comparative methods for almost all target sites, and t-SNE visualization of the embedding space showed natural separation between high-fitness and low-fitness mutants, proving that the model possesses reliable zero-shot prediction capabilities. In terms of thermal stability prediction, this invention achieved a Pearson correlation of no less than 0.331 on the S2648 dataset and could identify mutants that cause fatal damage to the catalytic pocket with the highest accuracy, verifying its practical value in function-oriented design. Furthermore, the designed sequence achieved a pLDDT score close to 90 after AlphaFold3 prediction and an average TM-score exceeding 0.95 compared to the input structure, indicating good foldability and structural fidelity. In summary, this invention effectively improves sequence recovery rate, zero-sample prediction accuracy, and function-oriented design performance by fusing sequence evolution information, spatial geometry information, and protein-ligand interaction information, solving problems such as low prediction probability and insufficient side-chain microenvironment modeling in existing technologies.

[0113] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention do not depart from the essence and scope of the technical solution of the present invention.

Claims

1. A method for designing protein-ligand binding pocket sequences based on heterogeneous graph neural networks, characterized in that, Includes the following steps: (1) Read and parse the information; (2) Construct a protein-ligand isomer map; (3) Calculate the structural geometry and non-covalent interaction characteristics between protein residues; (4) Calculate the characteristics of non-covalent interactions between protein residues and ligand atoms; (5) Mask the residues to be designed; (6) Construct a heterogeneous graph neural network and perform message passing; (7) Output the predicted probabilities of the 20 natural amino acids corresponding to the residues to be designed, and determine the target amino acid type based on the predicted probabilities; (8) Output the model after performing iterative sequence design.

2. The design method according to claim 1, characterized in that, The information read in step (1) includes a protein receptor structure file and a ligand small molecule structure file. The protein receptor structure file is a PDB format file, and the ligand small molecule structure file is a MOL2 format file.

3. The design method according to claim 1, characterized in that, The protein-ligand isomer diagram described in step (2) consists of protein residue nodes, ligand atom nodes, and edge structures; The protein residue node features mentioned are residue embedding features, residue main chain dihedral features, and amino acid type one-hot encoding features generated by the protein language model; The ligand atom nodes are characterized by one-hot encoded features of the SYBYL type of ligand atoms; The edge structures described are protein residue-residue peptide bond edges, protein residue-residue geometric edges, protein residue-residue interaction edges, protein residue-ligand atom geometric edges, protein residue-ligand atom interaction edges, and ligand atom-ligand atom chemical bond edges.

4. The design method according to claim 1, characterized in that, The geometric features mentioned in step (3) are the features of the interatomic distance matrix after radial basis function expansion, the local coordinate system direction projection features, and the normalized distance scalar features; This includes: geometric edge features of protein residue-residue and geometric edge features of protein residue-ligand atom.

5. The design method according to claim 1, characterized in that, The non-covalent interactions between protein residues and between protein residues and ligand atoms described in step (4) include: One or more of the following: hydrophobic interaction, hydrogen bond, weak hydrogen bond, π-π stacking, salt bridge, amide-π stacking, cation-π interaction, halogen bond, disulfide bond and halogen multipolar interaction.

6. The design method according to claim 1, characterized in that, In step (5), the site of the residue to be designed is replaced with a marker to replace the original amino acid type in order to avoid leakage of target residue information.

7. The protein-ligand binding pocket sequence design method according to claim 1, characterized in that, The heterogeneous graph neural network described in step (6) consists of a residue-internal message passing module, a ligand-internal message passing module, and a protein-ligand cross-node type message passing module.

8. The design method according to claim 7, characterized in that, The residue intra-residue message passing module aggregates neighborhood node information along peptide bond edges, residue geometry edges, and residue interaction edges, respectively. The protein-ligand cross-node type message passing module enables information exchange between protein residue nodes and ligand atom nodes through residue-ligand geometric edges and residue-ligand interaction edges.

9. The design method according to claim 8, characterized in that, The message aggregation of the peptide bond edges is implemented using the graph attention convolutional network GATConv. The residue geometry edges, residue interaction edges, residue-ligand geometry edges, and residue-ligand interaction edges are implemented using an attention-crossing message passing module.

10. The design method according to claim 1, characterized in that, The predicted probability mentioned in step (7) is obtained through the output layer of a fully connected neural network; The output layer of the fully connected neural network consists of a LayerNorm layer, a ReLU activation layer, a Dropout layer, and a fully connected layer.

11. The design method according to claim 1, characterized in that, The iterative sequence design method described in step (8) is as follows: (8-1) Calculate the predicted confidence of the residues to be designed based on the predicted probability, and lock residue A after sorting; (8-2) Write the predicted amino acids corresponding to the locked residues back into the protein sequence, recalculate the protein language model embedding features, and reconstruct the heterogeneity graph; (8-3) Continue to predict the remaining unlocked residues; (8-4) Repeat steps (8-1) to (8-2) until all design residues are predicted.

12. The design method according to claim 11, characterized in that, The locking method described in step (8-1) is as follows: if the prediction is iterated M times, the residues are ranked according to the confidence level of the current round in each iteration, and the residues from the first 1 / M, 2 / M to (M-1) / M are locked in turn. In the last round, the remaining 1 / M unknown residues are predicted, where M≥1.

13. A protein mutation assessment method based on heterogeneous graph neural networks, characterized in that, Includes the following steps: S1. Read and parse the information; S2. Construct mutant protein sequences and generate mutant embedding features; S3. Construct a protein-ligand isomer map using the same isomer map construction method as step (2) in the design method described in claim 1; S4. Use the same heterogeneous graph neural network construction method as step (6) in the design method of claim 1 and perform message passing; S5. Use the same method as step (7) in the design method described in claim 1 to output the predicted probability and evaluate the rationality of the mutation; S6. Output the mutation score results.

14. The protein mutation assessment method according to claim 13, characterized in that, The information read in step S1 includes protein receptor structure files, ligand small molecule structure files, and mutation information to be evaluated. The mutation information to be evaluated includes chain number, original amino acid type, residue number, and target mutant amino acid type.

15. The protein mutation assessment method according to claim 13, characterized in that, The method for constructing the mutant protein sequence described in step S2 is as follows: based on the input mutation information, the original amino acid at the corresponding position in the protein sequence is directly replaced with the target mutant amino acid.

16. The protein mutation assessment method according to claim 13, characterized in that, The indicators for evaluating the rationality of mutations mentioned in step S5 are mutation confidence, negative logarithmic probability fitness index, and target mutation amino acid probability value.

17. The application of the design method according to any one of claims 1-12 or the protein mutation assessment method according to any one of claims 13-16 in protein engineering.

18. The application according to claim 17, characterized in that, include: One or more of the following: protein binding pocket optimization, enzyme activity optimization, protein stability optimization, protein-small molecule binding interface optimization, and drug design.

19. A system for designing protein-ligand binding pocket sequences based on heterogeneous graph neural networks, characterized in that, include: The structure parsing module is used to read protein structure files and ligand structure files and parse the structural information. The pocket screening module is used to screen for protein residues within the binding pocket region; The isomer map building module is used to construct protein-ligand isomer maps; The sequence embedding module is used to generate residue embedding features using a protein language model; The interaction calculation module is used to calculate the interaction characteristics between protein residues and between protein residues and ligand atoms. The heterogeneous graph neural network module is used for multi-round message passing and node feature updates; The prediction module is used to output the amino acid prediction results of the residues to be designed; The iterative design module is used for iterative sequence optimization based on prediction confidence.

20. The system according to claim 19, characterized in that, The heterogeneous graph neural network module includes: residue-internal message passing units, ligand-internal message passing units, and cross-node type message passing units.

21. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, wherein when the computer program is executed by a processor, it implements the steps of the design method according to any one of claims 1-12 or the protein mutation assessment method according to any one of claims 13-16.

22. A computer device comprising a memory and a processor, wherein the memory stores a computer program capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the design method according to any one of claims 1-12 or the protein mutation assessment method according to any one of claims 13-16.

Images