A method and system for predicting protein-ligand affinity

Abstract

This invention relates to a method and system for predicting protein-ligand affinity. The method includes the following steps: obtaining a target protein-ligand complex; modeling the ligand conformation in the target protein-ligand complex as a stochastic dynamic process under pocket geometry constraints using a pre-trained conditional molecular dynamics refinement module, performing conditional denoising to obtain a refined complex; constructing a dual-view atomic map structure based on the refined complex; extracting features from the dual-view atomic map structure using a first encoder to obtain local geometric features; extracting residue-level semantic features of the protein pocket and molecular-level semantic features of the ligand using a second encoder to obtain global semantic features; fusing the local geometric features and global semantic features, and inputting the fused features into a pre-trained decoder for predicting the affinity of the target protein-ligand complex.

Description

Technical Field

[0001] This invention relates to a method and system for predicting protein-ligand affinity, belonging to the field of bioinformatics prediction technology. Background Technology

[0002] Predicting protein-ligand binding affinity is a core task in drug discovery, involving lead compound screening and structure-activity relationship optimization. Current computational prediction methods in this field are based on whether or not the three-dimensional structure of the protein-ligand complex is utilized, and are mainly divided into two categories: complex-free methods and complex-based methods. Meanwhile, technologies related to conformational dynamic modeling are also gradually developing, but all methods have significant technical limitations. Traditional regularized computational methods also suffer from inherent defects such as conformational sampling and energy approximation, while experimental determination methods, although providing reliable results, suffer from high cost and low throughput.

[0003] Complex-free methods do not explicitly construct the 3D structure of protein-ligand complexes. Instead, they learn the fusion characterization of proteins and ligands separately to predict binding affinity. These methods typically take protein sequences / graphs or ligand strings / graphs as input, learning affinity-related association patterns from the sequence or topological structure. Early representative works, such as DeepDTA, used convolutional neural networks to extract local patterns from smiles and amino acid sequences, respectively, and performed feature fusion and regression prediction at higher levels. Subsequently, GraphDTA expanded the ligand representation from linear sequences to molecular graph structures, introducing graph neural networks to more effectively model chemical topological relationships, thereby enhancing the ability to characterize structural changes. Building on this, methods such as MONN further attempted to explicitly model the correspondence between ligand substructures and protein residues, approximating potential interactions without relying on 3D complexes through attention or alignment mechanisms. These methods are computationally efficient, have a simple workflow, and are weakly dependent on structural data, making them suitable for large-scale virtual screening. However, because they mainly rely on sequence or 2D semantic associations and lack explicit characterization of real-space contact and geometric constraints, their performance degrades in scenarios with significant induced fit or numerous conformational variations.

[0004] To compensate for the lack of explicit 3D interaction evidence in complex-free methods, complex-based methods directly use the 3D structure of the protein-ligand complex as input, learning to explicitly characterize the interaction patterns at the binding interface through geometric representation. These methods typically model intramolecular and intermolecular geometric constraints around atomic-level spatial adjacency relationships, thus more directly capturing the structural factors determining binding affinity. In the area of ​​geometric graph learning, HybridGeo models both intramolecular and intermolecular interactions separately, aggregating spatial information through a hybrid strategy to characterize local geometry and interfacial interactions within a unified framework. EHIGN further explicitly introduces interaction-driven inductive bias into model design, constructing an interpretable heterogeneous interaction graph neural network. Regarding multi-scale modeling, MHAN-DTA proposes a multi-scale hybrid attention network that models the target protein through a hierarchical strategy and incorporates a pocket feature aggregation module based on a self-attention mechanism to introduce global perception capabilities to alleviate the problem of insufficient local feature mining. Furthermore, Geo-PLA constructs a geometric equivariance graph representation learning framework with local-global structure awareness capabilities, which captures geometric information in protein-ligand complexes to achieve high-precision prediction of binding affinity.

[0005] However, existing protein-ligand affinity prediction methods still have significant limitations in conformational modeling, dynamic coupling, and feature representation. In conformational modeling, most models rely on a single static protein-ligand complex conformation, failing to systematically characterize conformational uncertainties and dynamic changes. They are highly sensitive to initial posture quality and conformational noise, limiting their generalization ability across datasets and targets, resulting in poor prediction stability. In dynamic modeling, existing dynamic or generative conformational refinement methods typically follow a molecular-type decoupling modeling paradigm, treating proteins and ligands as independent update objects. This makes it difficult to characterize the coupling dynamics between proteins and ligands in the complex system, weakening the constraint of the protein binding pocket on ligand conformational evolution and easily introducing additional inference overhead and geometric instability. In structural representation, existing methods mainly focus on atomic-level local geometric features, lacking a global representation of the overall structural semantics of the protein pocket and ligand, resulting in insufficient generalization ability in conformational perturbations, new backbone migration, and out-of-distribution chemical spaces.

[0006] Therefore, how to achieve conformational dynamic modeling of protein-ligand coupling under protein binding pocket constraints, and construct a molecular characterization method that takes into account both local geometry and global semantic information, thereby improving the accuracy and robustness of affinity prediction, has become an urgent technical problem to be solved. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention proposes a method and system for predicting protein-ligand affinity.

[0008] The technical solution of the present invention is as follows: On the one hand, this invention proposes a method for predicting protein-ligand affinity, comprising the following steps: Obtain the target protein-ligand complex; The pre-trained conditional molecular dynamics refinement module models the ligand conformation in the target protein-ligand complex as a stochastic dynamic process under pocket geometry constraints, performs conditional denoising, and obtains the refined complex. Based on the refined complex, a dual-view atomic diagram structure containing molecular inter diagrams and molecular intra diagrams was constructed. The first encoder extracts features from the dual-view atomic graph structure to obtain local geometric features; The global semantic features are obtained by extracting residue-level semantic features of protein pockets and molecular-level semantic features of ligands from the refined complex using a second encoder. Local geometric features and global semantic features are fused, and the fused features are input into a pre-trained decoder to predict the affinity of the target protein-ligand complex. In a preferred embodiment, the step of modeling the ligand conformation in the target protein-ligand complex as a stochastic dynamic process under pocket geometry constraints through a pre-trained conditional molecular dynamics refinement module, and performing conditional denoising, specifically includes: The refinement of ligand conformation is defined as a stochastic dynamic process from the initial state to the target state, and intermediate perturbation dynamics are constructed through a stochastic interpolation framework. And set pocket geometry constraints, including: Pocket conditional vector field constraint, with pocket conformation as hard constraint, parameterizes conditional vector field through message passing network to learn the expected evolution direction of ligand in intermediate perturbation dynamic conformation space; Geometric constraints are used to penalize unreasonable spatial contacts and constrain conformational drift by defining the protein-ligand geometric potential function.

[0009] In a preferred embodiment, the stochastic dynamics process is numerically integrated by discrete time steps, and a stochastic perturbation term is introduced to model the inherent uncertainty of the ligand conformation.

[0010] In a preferred embodiment, during the pre-training process of the conditional molecular dynamics refinement module, a loss function is constructed with the regression loss of the pocket conditional vector field as the training objective.

[0011] As a preferred embodiment, the construction of a dual-view atomic diagram structure comprising a molecular inter diagram and a molecular intra diagram based on the refined complex specifically involves: For the refined complex, an undirected graph based on spatial contact between the ligand and the protein pocket is constructed as a molecular inter graph. The edges in the graph are determined by the ligand atomic coordinates, the protein pocket atomic coordinates, and the non-covalent contact distance threshold.

[0012] For the ligand and protein pockets in the refined complex, undirected graphs based on covalent bonds were constructed as molecular intragraphs, including ligand intragraphs and protein pocket intragraphs, with the edges in the graphs determined by covalent bonds.

[0013] In a preferred embodiment, the first encoder includes: A message passing network is used to perform independent message passing and aggregation on molecular inter-graphs and molecular intra-graphs, respectively, to obtain inter-graph features and intra-graph features. The feature fusion module is used to perform independent nonlinear projections on inter-graph features and intra-graph features, and then obtain preliminary fused features by feature summation; The radial distance-weighted feature aggregation module transforms the Euclidean distance between atoms into a gated weight vector by introducing a radial basis function, and performs distance-weighted message aggregation on the preliminary fused features to obtain local geometric features.

[0014] In a preferred embodiment, the second encoder includes: The residue-level semantic feature extraction module is used to construct a residue-level contact map with protein pocket residues as geometric anchors; then, a pre-trained protein language model is used to extract the semantic representation of the protein pocket residue sequence, and mean pooling is performed on the residue-level contact map to obtain the residue-level semantic features of the protein pocket. The molecular-level semantic feature extraction module is used to decompose ligand molecules into substructure token sequences, obtain token embeddings through a pre-trained molecular representation learning model, and obtain molecular-level semantic features of the ligands through mean pooling.

[0015] On the other hand, the present invention also proposes a protein-ligand affinity prediction system, comprising: The refinement module is used to obtain the target protein-ligand complex; and through the pre-trained conditional molecular dynamics refinement module, the ligand conformation in the target protein-ligand complex is modeled as a stochastic dynamic process under pocket geometry constraints, and conditional denoising is performed to obtain the refined complex. The graph structure construction module is used to construct a dual-view atomic graph structure containing molecular inter-graphs and molecular intra-graphs based on the refined complex. The feature extraction module is used to extract features from the dual-view atomic diagram structure through the first encoder to obtain local geometric features; and to extract residue-level semantic features of protein pockets and molecular-level semantic features of ligands from the refined complex through the second encoder to obtain global semantic features. The affinity prediction module is used to fuse local geometric features and global semantic features, and input the fused features into a pre-trained decoder to predict the affinity of the target protein-ligand complex.

[0016] In another aspect, the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the protein-ligand affinity prediction method as described in any embodiment of the present invention.

[0017] In another aspect, the present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the protein-ligand affinity prediction method as described in any embodiment of the present invention.

[0018] The beneficial effects of this invention are as follows: The conditional molecular dynamics refinement module with pocket constraints performs conformational refinement, correcting initial conformational noise and attitude bias, and enhancing the model's ability to capture real protein-ligand binding interactions. Dual-view decoupled modeling avoids confusion between different physical semantic features, accurately characterizing the core structural determinants of the binding interface. By fusing local geometric features and global semantic features, the bias of single local geometric modeling is compensated for, significantly improving the model's physical consistency and generalization performance in conformational perturbation and scaffold transition scenarios.

[0019] Additional aspects and advantages of the invention will be set forth in the following description, and some of them will be obvious from the description, or may be learned by practicing the invention. Furthermore, various aspects and advantages of the invention may be realized and obtained by means of method steps and combinations particularly pointed out in the appended claims. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a distribution diagram of ablation experimental data provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0023] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0024] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0025] The term “and / or” refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes these combinations.

[0026] See Figure 1 This invention proposes a protein-ligand affinity prediction method that achieves unified modeling of complex dynamic conformation information, intramolecular-intermolecular interaction information, and local-global feature information. Compared with existing models that rely on a single static conformation, molecular type decoupling modeling, or only focus on local geometric features, this patent establishes a multi-level protein-ligand complex representation system that coordinates dynamics, local geometry, and global semantics through a pocket-condition molecular dynamics refinement mechanism, a dual-view atomic characterization system, and a local-global feature fusion strategy. While maintaining the physical consistency of conformational evolution and the geometric equivariance of feature modeling, it achieves efficient fusion and information complementarity of dynamic conformation, local interactions, and global semantic information. Through pocket-condition constraint conformation refinement, initial conformational noise and attitude bias are corrected, enhancing the model's ability to capture real protein-ligand binding interactions; through dual-view decoupling modeling, confusion between different physical and semantic features is avoided, accurately characterizing the core structural determinants of the binding interface; and through local-global feature fusion, the bias of single local geometric modeling is compensated for, significantly improving the model's physical consistency and generalization performance under conformational perturbation and scaffold transition scenarios.

[0027] The method specifically includes the following steps: S100, Obtain the target protein-ligand complex ,in, Indicates ligand conformation. Indicates protein conformation.

[0028] S200, using the pre-trained Conditional Molecular Dynamics Refinement (CMD-Refine) module, refines the ligand conformation in the target protein-ligand complex. Modeled as a stochastic dynamic process under pocket geometry constraints, the complex is subjected to conditional denoising to obtain a refined composite. .

[0029] In one embodiment, the conditional molecular dynamics refinement module models ligand conformational evolution as a stochastic dynamic process under pocket geometry constraints, rather than independent monomer updates. This corrects initial conformational noise, generating physically consistent protein-ligand complex conformations and providing a stable geometric basis for subsequent feature modeling. Specifically: For the input protein-ligand complex Ligand conformation refinement is defined as a time interval. stochastic dynamics ,in For the initial state, To refine the target state, intermediate perturbation states are constructed using a random interpolation framework: ; in It is a linear interpolation function. This is the noise amplitude control function. To achieve a stochastic characterization of ligand conformational evolution by including noise terms that follow a standard normal distribution, using pocket conformations... Under hard constraints, CMD-Refine learns a conditional vector field to characterize the desired evolutionary trend of ligands in the perturbed conformational space. Ensure the equivariance of ligand conformation updates to rigid body transformations, thus ensuring the physical consistency of geometric modeling.

[0030] In one embodiment, a messaging network The SE(3)-equivariant network is used for parameterization. In other embodiments, other network structures with spatial rotation and translation equivariance can also be used. The advantage of the SE(3)-equivariant network is that it can directly manipulate atomic coordinates in three-dimensional space. The design of its network layers strictly follows the transformation rules of the SE(3) group, making the output conditional vector field... When ligands undergo rigid body transformations such as translation and rotation, they can maintain the corresponding changes in direction and size, thereby accurately capturing the spatial interaction patterns between ligands and pockets and avoiding feature representation distortion caused by coordinate system transformation or conformational rotation.

[0031] In one embodiment, during the inference phase of the stochastic dynamics process, numerical integration of the stochastic dynamics process is performed through discrete time steps, assuming... The iterative update rule for ligand atom coordinates is as follows: ; in, Let be the random disturbance intensity control function. This parameter, representing standard normal noise, is introduced to simulate conformational uncertainty and prevent the finishing process from degenerating into a single deterministic correction. This design ensures that the ligand is explicitly aware of pocket geometric constraints throughout the finishing process, thereby avoiding conformational drift inconsistent with the binding environment under unconstrained dynamics.

[0032] In one embodiment, to avoid atomic collisions and non-physical displacements, CMD-Refine introduces geometrical feasibility guidance during the sampling process. This involves defining the protein-ligand geometrical potential function. To achieve the penalty for unreasonable spatial contact and the constraint of conformational drift: .

[0033] in, This is a smoothing barrier function used to penalize unreasonable spatial contacts. It is a soft anchoring coefficient used to balance the constraint strength between local geometric repair and overall conformational drift. , These are the sets of ligand and protein pocket atoms, respectively. After each coordinate update, a gradient correction is applied to the ligand coordinates: ; in, To guide the strength coefficient, For time scheduling functions, The analytical gradient of the geometric potential function with respect to the coordinates of the ligand atoms is used to enhance geometric correction in the early stage and ensure convergence stability in the later stage, thereby improving the geometric stability of the refined trajectory without compromising the expressive power of the stochastic system.

[0034] In one embodiment, during the pre-training process of the conditional molecular dynamics refinement module, a loss function is constructed with the regression loss of the pocket conditional vector field as the training objective, guiding the model to learn reasonable conformational evolution trends: ; in, The first-order time derivative of the random interpolation function represents the instantaneous evolution direction of the ligand conformation from the initial state to the target state under noise-free conditions. The final output is the refined complex. .

[0035] S300. In protein-ligand complexes, interatomic relationships are simultaneously influenced by both intramolecular covalent structural constraints and transmolecular non-covalent interactions. Covalent bonds define a stable, conformationally invariant topological framework within the molecule, while non-covalent interactions between ligand-pockets are determined by spatial contact relationships and are highly sensitive to conformational changes. Modeling these two types of relationships as a single graph structure can easily confuse structural priors with geometrically induced signals during message passing, thus weakening the model's ability to characterize real interaction patterns. Therefore, based on the refined complex... We construct a dual-view atomic graph structure containing molecular inter-graphs and molecular intra-graphs to achieve decoupled modeling of intramolecular covalent structural constraints and intermolecular non-covalent interactions, thus avoiding the confusion of feature information with different physical semantics during modeling.

[0036] In one embodiment, the dual-view atomic diagram structure is specifically as follows: For the ligand and protein pockets of the refined complex, undirected graphs based on covalent bonds were constructed as molecular intragraphs (internal views), including ligand intragraphs. and protein pocket intra diagram , and The edge sets are the sets of edges in the ligand intragraph and the protein pocket intragraph, respectively, and together they form the edge set. , Determined by chemical covalent bonds, it is used to characterize the covalent skeleton, local chemical environment and bond connection patterns within molecules, and to capture topological features that remain unchanged during conformational changes.

[0037] For the refined complex, an undirected graph based on spatial contact between the ligand and protein pocket is constructed as a molecular intergraph. The edge set of a molecular inter graph is defined as: ; in, These are the refined coordinates of the ligand atoms. Pocket atom coordinates, The non-covalent contact distance threshold is preferably 5 Å, where Å is a unit used to measure minute lengths, 1 Å = ), used to explicitly model the spatial contact distribution and non-covalent interactions between ligands and pockets.

[0038] S410. The first encoder extracts features from the dual-view atomic graph structure to obtain local geometric features.

[0039] In one implementation, the first encoder is a Local Dual-View Encoder (LDVE), which performs independent message passing and feature fusion on the dual-view graph structure to extract stable features that reflect the local spatial interactions between proteins and ligands. At the same time, it ensures the invariance of features to rigid body transformations, specifically including: A message passing network is used to perform independent message passing and aggregation on molecular inter-graphs and molecular intra-graphs, respectively, to obtain inter-graph features and intra-graph features. In one embodiment, the message passing network adopts an SE(3)-equivariant network for any view. Its nodes The rules for message aggregation, feature update, and coordinate update are as follows: ; ; ; in, , , For learnable mapping functions, This is a feature splicing operation that enables hierarchical updates of node features and adaptive adjustments to geometric coordinates.

[0040] The feature fusion module is used to process the intra graph features obtained after S-layer iterations. and inter-graph features Independent nonlinear projections are performed to eliminate distribution and scale offsets between views. Preliminary fusion features are then obtained through feature summation, using the following formula: ; in, , These are dedicated nonlinear projection functions, adapted to the characteristic distributions of covalent structures and non-covalent interactions, respectively.

[0041] The radial distance-weighted feature aggregation module transforms the Euclidean distance between atoms into a gated weight vector by introducing a radial basis function (RBF). Distance-weighted message aggregation is performed on the preliminary fusion features to selectively amplify the contribution of short-range interactions and suppress sparse contact noise. ; in, For view-specific mapping functions, This is an element-wise multiplication operation. Finally, global pooling is used to obtain the local geometric interaction features of the complex: .

[0042] S420. The refined complex is processed using a second encoder to extract residue-level semantic features of the protein pockets and molecular-level semantic features of the ligands, resulting in global semantic features. In one embodiment, the second encoder is a designed Global Semantic Prior Encoder (GSPE) composed of ESM-2 and Mol2Vec. Based on a pre-trained model, it extracts residue-level semantic features of the protein pockets and molecular-level semantic features of the ligands, compensating for modeling biases of local geometric features, providing the model with global structural and functional semantic information, and improving the model's generalization ability in out-of-distribution scenarios. Specifically, this includes: The residue-level semantic feature extraction module, for the pocket, denoted as . C of protein pocket residues α (α-carbon atom, referring to the carbon atom adjacent to the functional group) Coordinates are geometric anchor points Construct residue-level contact maps , ; in, This represents the residue contact threshold. Furthermore, a pre-trained ESM2 model is used to extract the semantic representation of the pocket residue sequence. and in the residue contact diagram Mean pooling is performed on the protein pocket to obtain the residue-level semantic features. .

[0043] The molecular-level semantic feature extraction module is used to decompose ligand molecules into substructure token sequences. Token embeddings are obtained through a pre-trained Mol2Vec model, and molecular-level global semantic features of protein ligands are obtained through mean pooling. This reflects the semantic features of the overall chemical skeleton and substructure of the ligand.

[0044] S500: Local geometric features and global semantic features are fused together, and the fused features are input into a pre-trained decoder to predict the affinity of the target protein-ligand complex. In one embodiment, local geometric interaction features are deeply fused with global semantic prior features to construct a prediction decoder based on a multilayer perceptron (MLP), enabling quantitative prediction of protein-ligand binding affinity. An end-to-end training loss function is designed to optimize the overall model parameters. Specific steps include: Local geometric interaction features Global semantic features of protein pockets Global semantic features of ligands Feature concatenation is performed to form a complex feature that combines local geometric interactions and global semantic information, giving full play to the complementary advantages of local and global features: ; in, This is a feature splicing operation.

[0045] The fused integrated features are input into the MLP-based prediction decoder. Output predicted values ​​of protein-ligand binding affinity. : ; Using mean squared error (MSE) as the supervised loss, the deviation between predicted and experimental values ​​is minimized to achieve end-to-end supervised regression training of the model. ; in, For the sample size, For the first The predicted value for each sample, These are the corresponding experimentally measured values.

[0046] In this invention, a pre-training of the conformation refinement module is first performed using large-scale small molecule conformation datasets, including PCQM4Mv2 and ANI-1x. Then, the overall model training and performance validation are completed based on a protein-ligand complex benchmark dataset. PCQM4Mv2 is a quantum chemistry dataset containing approximately three million small molecules, whose equilibrium conformations are obtained through density functional theory (DFT) optimization. ANI-1x consists of approximately five million DFT calculations, covering the energy and geometric changes of organic small molecules in different conformations. Both datasets model single-molecule systems, covering a rich variety of chemical backbone types and continuous conformational perturbation modes, making them suitable for learning the geometric evolution and recovery of molecules under random noise conditions. Subsequently, the performance of CMD-PLA is evaluated on three publicly available benchmark datasets: PDBbind (v2020), CASF-2013, and CASF-2016. PDBbind-2020 includes 19,443 protein-ligand samples with experimentally determined 3D complex structures, along with corresponding binding affinity annotations; it serves as the primary training and internal evaluation data source. Considering potential sample overlap between different datasets and the possibility that some complexes might not be processed into the structured input required by the model, the data was first cleaned to remove overlapping and unprocessable samples. The cleaned PDBbind-2020 data was then randomly divided into 14,557 training sets, 1,796 validation sets, and 1,828 test sets in an 8:1:1 ratio, totaling 18,181 samples. To evaluate the model's generalization ability across data distributions, we further conducted independent tests using two standard external benchmarks, CASF-2013 and CASF-2016. Both are derived from the PDBbind core set and are commonly used for cross-sectional comparisons of structure-based affinity prediction methods. After undergoing the same data cleaning process as described above, CASF-2013 and CASF-2016 retained 188 and 284 independent test samples, respectively, and used only in the external testing phase to verify the model's generalization performance on unseen complexes.

[0047] To verify the effectiveness of the above embodiments, the present invention provides a specific verification example: The trained model was validated on the PDBbind (v2020) internal test set, CASF-2013, and CASF-2016 independent test sets. The same evaluation metrics and experimental settings were used as existing mainstream methods such as DeepDTA, GraphDTA, SchNet, IGN, and HybridGeo. The experiment was repeated 3 times and the average value was taken. The core evaluation metrics were root mean square error (RMSE), mean absolute error (MAE), and Pearson correlation coefficient (R).

[0048] The experimental results are shown in Table 1. The method provided in this invention outperforms existing methods overall on the three datasets. On the PDBbind-2020 test set, it achieves the best results on all metrics. The RMSE decreases from 1.208 to 1.160 compared to the suboptimal HybridGeo method, and the R-squared increases to 0.792, indicating a stable improvement in both prediction accuracy and ranking consistency. On CASF-2016, the method provided in this invention maintains the same advantage, demonstrating its good generalization ability for heterogeneous complexes. Although the method provided in this invention is slightly inferior to IGN on CASF-2013, it still achieves suboptimal results. This is presumably related to the data distribution and quality of CASF-2013, as all models experience the most severe performance drop on this dataset, and its scale is relatively small. In this case, models relying on strong geometric priors may be more advantageous. The performance advantage of the method provided in this invention on the PDBbind-2020 test set and CASF-2016 mainly stems from a more stable synergy between conformational processing and geometric modeling. By introducing conditional conformational refinement, the method provided in this invention mitigates the impact of initial pose noise on the ligand-pocket interaction neighborhood to some extent, allowing atomic-level representations to be built upon more consistent spatial relationships. Simultaneously, separating the modeling of intramolecular covalent structures from transmolecular non-covalent contacts helps reduce interference from different physical semantics during message passing. At CASF-2013, due to the small data size and significant conformational and experimental noise, the benefits of conformational refinement were relatively limited.

[0049] Table 1. Performance comparison of the model on three major datasets: PDBbind (v2020), CASF-2013, and CASF-2016 (RMSE, MAE, SD, and R are evaluation metrics, and the values ​​in parentheses are standard deviations).

[0050] To evaluate the actual contribution of each key design element in the above method to affinity prediction, this embodiment conducted an ablation study on the PDBbind-2020 dataset. Specifically, while keeping all other settings unchanged, six model variants were constructed: removing the inner view to test the importance of intramolecular covalent structure modeling (w / o Intra); removing the inter-view to evaluate the role of intermolecular spatial contact modeling (w / o Inter); removing CMD-Refine and directly using the original complex conformation (w / o CMD); retaining dynamic updates but without imposing pocket geometry constraints and conformational drift regulation (w / o LM); and removing the global representations of ligands and pockets, respectively (w / o LG and w / o PG). All variants were trained three times under the same settings and evaluated on the internal test set. The results are as follows: Figure 2As shown, removing CMD-Refine or relaxing its geometric constraints (w / o CMD and w / o LM) leads to the most significant performance degradation across all metrics, indicating that conformational stability plays a fundamental role in downstream affinity prediction. Simultaneously, removing Intra and Inter both significantly increase prediction errors, demonstrating that information on intramolecular structural relationships and intermolecular interactions is complementary in complex modeling. Furthermore, removing the global representation of ligands or pockets also impairs performance, suggesting that global semantics primarily functions as a complement and stabilizing agent.

[0051] This invention also proposes a protein-ligand affinity prediction system, comprising: The refinement module is used to obtain the target protein-ligand complex; and through the pre-trained conditional molecular dynamics refinement module, the ligand conformation in the target protein-ligand complex is modeled as a stochastic dynamic process under pocket geometry constraints, and conditional denoising is performed to obtain the refined complex; this module is used to implement the functions of steps S100 and S200 in the above method, which will not be described in detail here. The graph structure construction module is used to construct a dual-view atomic graph structure containing molecular inter-graphs and molecular intra-graphs based on the refined complex; this module is used to implement the function of step S300 in the above method, and will not be described in detail here. The feature extraction module is used to extract features from the dual-view atomic diagram structure using the first encoder to obtain local geometric features; and to extract residue-level semantic features of protein pockets and molecular-level semantic features of ligands from the refined complex using the second encoder to obtain global semantic features. This module is used to implement the functions of steps S410 and S420 in the above method, which will not be described in detail here. The affinity prediction module is used to fuse local geometric features and global semantic features, and input the fused features into a pre-trained decoder to predict the affinity of the target protein-ligand complex. This module is used to implement the function of step S500 in the above method, and will not be described in detail here.

[0052] This invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the protein-ligand affinity prediction method as described in any embodiment of this invention.

[0053] This invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the protein-ligand affinity prediction method as described in any embodiment of this invention.

[0054] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, A and B simultaneously, or B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0055] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0056] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0057] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0058] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for predicting protein-ligand affinity, characterized in that, Includes the following steps: Obtain the target protein-ligand complex; The pre-trained conditional molecular dynamics refinement module models the ligand conformation in the target protein-ligand complex as a stochastic dynamic process under pocket geometry constraints, performs conditional denoising, and obtains the refined complex. Based on the refined complex, a dual-view atomic diagram structure containing molecular inter diagrams and molecular intra diagrams was constructed. The first encoder extracts features from the dual-view atomic graph structure to obtain local geometric features; The global semantic features are obtained by extracting residue-level semantic features of protein pockets and molecular-level semantic features of ligands from the refined complex using a second encoder. Local geometric features and global semantic features are fused, and the fused features are input into a pre-trained decoder to predict the affinity of the target protein-ligand complex.

2. The protein-ligand affinity prediction method according to claim 1, characterized in that, The step of modeling the ligand conformation in the target protein-ligand complex as a stochastic dynamic process under pocket geometry constraints through the pre-trained conditional molecular dynamics refinement module, and performing conditional denoising specifically includes: The refinement of ligand conformation is defined as a stochastic dynamic process from the initial state to the target state, and intermediate perturbation dynamics are constructed through a stochastic interpolation framework. And set pocket geometry constraints, including: Pocket conditional vector field constraint, with pocket conformation as hard constraint, parameterizes conditional vector field through message passing network to learn the expected evolution direction of ligand in intermediate perturbation dynamic conformation space; Geometric constraints are used to penalize unreasonable spatial contacts and constrain conformational drift by defining the protein-ligand geometric potential function.

3. The method for predicting protein-ligand affinity according to claim 2, characterized in that, In the stochastic dynamic process, the stochastic dynamic process is numerically integrated by discrete time steps, and a stochastic perturbation term is introduced to model the inherent uncertainty of the ligand conformation.

4. The protein-ligand affinity prediction method according to claim 2, characterized in that, During the pre-training process of the conditional molecular dynamics refinement module, a loss function is constructed with the regression loss of the pocket conditional vector field as the training objective.

5. The method for predicting protein-ligand affinity according to claim 1, characterized in that, The construction of a dual-view atomic diagram structure containing both molecular inter diagrams and molecular intra diagrams based on the refined complex is specifically as follows: For the refined complex, an undirected graph based on spatial contact between the ligand and the protein pocket is constructed as a molecular inter graph. The edges in the graph are determined by the ligand atomic coordinates, the protein pocket atomic coordinates, and the non-covalent contact distance threshold. For the ligand and protein pockets in the refined complex, undirected graphs based on covalent bonds were constructed as molecular intragraphs, including ligand intragraphs and protein pocket intragraphs, with the edges in the graphs determined by covalent bonds.

6. The protein-ligand affinity prediction method according to claim 1, characterized in that, The first encoder includes: A message passing network is used to perform independent message passing and aggregation on molecular inter-graphs and molecular intra-graphs, respectively, to obtain inter-graph features and intra-graph features. The feature fusion module is used to perform independent nonlinear projections on inter-graph features and intra-graph features, and then obtain preliminary fused features by feature summation; The radial distance-weighted feature aggregation module transforms the Euclidean distance between atoms into a gated weight vector by introducing a radial basis function, and performs distance-weighted message aggregation on the preliminary fused features to obtain local geometric features.

7. The method for predicting protein-ligand affinity according to claim 1, characterized in that, The second encoder includes: The residue-level semantic feature extraction module is used to construct a residue-level contact map with protein pocket residues as geometric anchors; then, a pre-trained protein language model is used to extract the semantic representation of the protein pocket residue sequence, and mean pooling is performed on the residue-level contact map to obtain the residue-level semantic features of the protein pocket. The molecular-level semantic feature extraction module is used to decompose ligand molecules into substructure token sequences, obtain token embeddings through a pre-trained molecular representation learning model, and obtain molecular-level semantic features of the ligands through mean pooling.

8. A protein-ligand affinity prediction system, characterized in that, include: The refinement module is used to obtain the target protein-ligand complex; The ligand conformation in the target protein-ligand complex is modeled as a stochastic dynamic process under pocket geometry constraints by a pre-trained conditional molecular dynamics refinement module, and conditional denoising is performed to obtain the refined complex. The graph structure construction module is used to construct a dual-view atomic graph structure containing molecular inter-graphs and molecular intra-graphs based on the refined complex. The feature extraction module is used to extract features from the dual-view atomic graph structure through the first encoder to obtain local geometric features; The refined complex was then processed by a second encoder to extract residue-level semantic features of protein pockets and molecular-level semantic features of ligands, resulting in global semantic features. The affinity prediction module is used to fuse local geometric features and global semantic features, and input the fused features into a pre-trained decoder to predict the affinity of the target protein-ligand complex.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the protein-ligand affinity prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the protein-ligand affinity prediction method as described in any one of claims 1 to 7.

Images