Source-linked AI summary
Structure-aware Interactive Graph Neural Networks for the Prediction of Protein-Ligand Binding Affinity
Shuangli Li, Jingbo Zhou, Tong Xu, Liang Huang, Fan Wang, Haoyi Xiong, Weili Huang, Dejing Dou, Hui Xiong
TL;DR
Protein-ligand affinity prediction is important for drug discovery, but existing GNNs underuse 3D structure and long-range interactions. SIGN addresses this with PGAL for distance- and angle-aware graph learning and PiPool for global interaction modeling. Across two benchmarks, SIGN achieved the best performance, including 6.5% and 3.9% RMSE improvements over the best baselines on PDBbind and CSAR-HiQ.
Problem
Existing GNNs for protein-ligand affinity prediction inadequately use biomolecular 3D structure and essential long-range interactions.
Method
SIGN combines polar-inspired graph attention layers for distance- and angle-aware spatial modeling with pairwise interactive pooling and reconstruction learning for global interactions.
Results
SIGN achieved the best performance on two benchmark datasets, improving RMSE over the best baselines by 6.5% on PDBbind and 3.9% on CSAR-HiQ.
Takeaways & Limitations
The experiments support SIGN's effectiveness and generalizability for learning fine-grained structural and interaction information in protein-ligand complexes.
Abstract
from arXiv · showhide
Drug discovery often relies on the successful prediction of protein-ligand binding affinity. Recent advances have shown great promise in applying graph neural networks (GNNs) for better affinity prediction by learning the representations of protein-ligand complexes. However, existing solutions usually treat protein-ligand complexes as topological graph data, thus the biomolecular structural information is not fully utilized. The essential long-range interactions among atoms are also neglected in GNN models. To this end, we propose a structure-aware interactive graph neural network (SIGN) which consists of two components: polar-inspired graph attention layers (PGAL) and pairwise interactive pooling (PiPool). Specifically, PGAL iteratively performs the node-edge aggregation process to update embeddings of nodes and edges while preserving the distance and angle information among atoms. Then, PiPool is adopted to gather interactive edges with a subsequent reconstruction loss to reflect the global interactions. Exhaustive experimental study on two benchmarks verifies the superiority of SIGN.
1 INTRODUCTION
Protein-ligand binding-affinity prediction is important for computational drug discovery but remains challenging because existing models underuse 3D structure and long-range interactions. SIGN addresses these challenges with structure-aware attention and interactive pooling.
- Binding affinity measures protein-ligand interaction strength, and computational prediction can reduce reliance on laborious, time-consuming biological tests during early drug discovery.
- Existing deep-learning approaches represent complexes as sequences, 3D grids, or graphs, but modeling 3D spatial structure remains a key challenge.
- Many existing GNNs lose effectiveness on large protein-ligand complexes and inadequately capture their spatial structure and interactions.
- SIGN uses polar-inspired graph attention layers to preserve distance and angle information through iterative node-edge aggregation.
- PiPool incorporates global protein-ligand interactions by pooling edges according to atomic types and reconstructing the atomic interaction matrix with an auxiliary learning task.
- SIGN captures 3D spatial information and global long-range interactions in a semi-supervised framework, with experiments on two benchmark datasets supporting its effectiveness and generalizability.
2 RELATED WORK
Related work spans empirical, machine-learning, 3D grid, and graph-based approaches to protein-ligand affinity prediction. These approaches motivate structure-aware graph modeling for complexes with abundant spatial information.
- Protein-ligand binding-affinity prediction is important for efficient and accurate drug screening.
- Earlier empirical methods used docking and scoring functions that required expert knowledge to encode biochemical interactions.
- Structure-based approaches learn from 3D protein-ligand complexes, including 3D grid representations processed with 3D convolutions.
- Table 1 summarizes the mathematical notations used in the paper.
- The paper argues that existing approaches do not adequately model 3D-structure interactions and therefore focuses on incorporating abundant spatial information.
- GNNs aggregate structural information from molecular graphs for drug-discovery tasks such as molecular-property prediction and protein-interface prediction.
3 PRELIMINARIES
The paper formulates protein-ligand affinity prediction over a spatially constructed complex interaction graph. Its preliminaries define the graph, edge-oriented neighborhoods, and the structure-preserving prediction objective.
- The preliminaries introduce definitions and notation for the model and structure-based protein-ligand binding-affinity prediction problem.
- A complex interaction graph represents protein and ligand atoms as nodes with edges constructed from their spatial positions and 3D coordinates.
- Edge-oriented neighbors of an atom or directed edge are directed edges pointing to that atom or target edge.
- For edge e13, edges e21 and e41 are edge-oriented neighbors through their common node a1; analogous incoming-edge sets are defined for atom nodes.
- The prediction objective maps the complex interaction graph and 3D position matrix M to binding affinity y while preserving spatial structure.
4 MODEL FRAMEWORK
SIGN addresses the complex’s spatial structure and long-range protein–ligand interactions through alternating node→edge and edge→node updates, followed by interactive pooling. PGAL preserves angle and distance information, while PiPool aggregates interaction evidence globally.
- Overall framework: SIGN combines PGAL for 3D spatial modeling with PiPool for long-range interaction modeling in protein–ligand affinity prediction.The framework takes a complex interaction graph as input and uses the two components to address distinct structural challenges.
- Overall framework: PGAL alternates node→edge and edge→node interaction layers to update node and edge embeddings while learning spatial angle and distance information.The two interaction directions work together to model the complex’s spatial structure.
- Pairwise Interactive Pooling: PiPool operates on PGAL edge representations to form an atomic type-based interaction matrix that approximates global protein–ligand interactions.This pooling mechanism is designed for interaction information omitted when the complex graph contains only a partial protein structure.
- Angle-oriented Node→Edge Interaction Layer: Angle-oriented node→edge aggregation partitions neighboring edges into angle domains and applies independent propagation layers to capture directional structure and long-range dependencies.Neighbors are reorganized using a polar-coordinate angle divider before domain-specific attention and aggregation.
- Distance-aware Edge→Node Interaction Layer: Distance-aware edge→node aggregation maps atom-pair distances into spatial-relation buckets and incorporates them into attention over neighboring edges.The layer extends graph attention to distinguish multiple spatial relations among atoms before aggregating edge information into node representations.
5 EXPERIMENTS
Experiments on two benchmark datasets evaluate SIGN against diverse baselines, generalization to larger lower-quality data, spatial and interaction factors, and parameter settings. SIGN achieves the best reported performance, generalizes to the PDBbind general set, benefits from synergistic spatial and interactive modeling, and is sensitive to cutoff distance and angle divisions.
- Datasets: The evaluation uses PDBbind and CSAR-HiQ, with refined PDBbind complexes for primary training and held-out core and CSAR-HiQ sets for testing.A supplemental experiment trains on the PDBbind general set and evaluates on the core set because of dataset overlap.
- Overall Comparison (RQ1): SIGN achieves the best performance on PDBbind and CSAR-HiQ, improving RMSE over the best baselines by 6.5% and 3.9%, respectively.Results are averaged across five random runs using four regression indicators.
- Generalizability (RQ2): SIGN obtains the lowest prediction error under both training settings and improves performance by around 8% when trained on the larger general set.The general set contains more complexes but has lower quality than the refined set.
- Impact of Spatial and Interactive Factors (RQ3): SIGN outperforms variants that remove distance, angle, complete spatial information, or interaction loss, supporting joint modeling of spatial and interactive factors.Removing both distance and angle produces especially high prediction error, while removing long-range interactions also reduces performance.
- Parameters Analysis (RQ4): Increasing the cutoff distance improves performance for θ_d≤5 Å, but longer distances introduce redundancy and degrade performance.The cutoff distance is varied from 3 to 6 Å during analysis.
- Parameters Analysis (RQ4): 5 or 6 angle domains produce the best performance, whereas finer or coarser divisions degrade spatial-information learning.Fine-grained divisions may be indistinguishable, while coarse divisions contain sparse atomic neighbors.
6 CONCLUSION
The paper proposes SIGN to improve protein-ligand binding-affinity prediction by combining fine-grained structural information with atom-level interaction information. Its PGAL and pooling-based interaction reconstruction design is reported effective and generalizable on two benchmarks.
- Conclusion: SIGN learns protein-ligand complex representations using fine-grained structural and interaction information among atoms.The model is designed for binding-affinity prediction.
- Conclusion: PGAL integrates distance and angle information for modeling the complexes’ 3D spatial structure.This is the structure-modeling component of SIGN.
- Conclusion: A pooling process with reconstruction learning for the interaction matrix further improves prediction performance.The pooling design targets interaction information.
- Conclusion: Experiments on two benchmarks show the effectiveness and generalizability of SIGN.This is the paper’s concluding experimental claim.
A APPENDIX
The appendix documents graph construction, experimental settings, baseline descriptions, additional parameter-analysis results, training pseudocode, implementation resources, and computational hardware.
- Appendix: The appendix introduces complex interaction graph construction, experimental settings, baseline descriptions, additional parameter results, and the SIGN training pseudocode.The pseudocode is provided as Algorithm 2.
- Appendix: The implementation uses PaddlePaddle and trains models on 24 Intel CPUs and a Tesla V100 GPU with 32 GB memory.The code is made available in the cited PaddleHelix repository.
A.1 Complex Interaction Graph Construction
The complex interaction graph is constructed to represent non-local protein-ligand correlations while avoiding the redundancy of including the complete protein structure. The construction takes ligand positions, ligand nodes, and a cutoff distance as inputs.
- Complex Interaction Graph Construction: The graph construction adds spatial-based complex interactions because ligand-protein intermolecular connections are unavailable and local covalent bonds lack adequate 3D information.The method targets non-local correlations in the complex.
- Complex Interaction Graph Construction: Because proteins are much larger than ligands, the construction avoids including the complete protein structure while retaining relevant interaction information.This design addresses graph-size redundancy and information loss together.
- Complex Interaction Graph Construction: Algorithm 1 constructs the complex interaction graph G_I=<V,E> from the ligand position matrix, ligand node set, and cutoff distance r_θ.These are the listed algorithm inputs and output.
A.2 Instruction of the Binding Affinity
Experimental binding affinity is represented using negative logarithms of measured constants, while SIGN training is described algorithmically for batches.
- The training procedure is identified as Algorithm 2 for SIGN.
- Training processes each batch from the training samples.
- Binding affinity is expressed as the negative logarithm of K_d, K_i, or IC_50 on PDBbind and CSAR-HiQ.
A.3 Experiment Details
The experiments evaluate prediction performance using four regression metrics, including error, correlation, and standard-deviation measures.
- The experiments use four metrics to evaluate regression performance.The passage introduces the metric set but does not enumerate all four names.
A.3.1 Evaluation Metrics.
Evaluation metrics are defined using predicted and experimental binding affinities, with regression-line parameters used for standard-deviation calculation.
- Predicted and experimental binding affinities are denoted by ŷ_i and y_i for the i-th complex in dataset D.
- RMSE, MAE, and Pearson correlation coefficient R quantify prediction error and correlation in regression.
- The intercept a and slope b define the regression line used in the standard-deviation metric.
A.3.2 Input Graph and Features.
The experiments compare SIGN with machine-learning, convolutional, and graph-based baselines using constructed complex graphs and atom-level features, while analyzing PGAL depth and loss balancing.
- Input Graph and Features: SIGN and GNN-based baselines use the same constructed protein-ligand complex graph as input.
- Input Graph and Features: Atom features are extended to 36 dimensions, separating ligand and protein atom features, while edge features combine atom features with encoded distances.
- Baseline Descriptions: The comparison includes ML regressors, 3D CNN models, GraphDTA variants, and structure-aware or edge-based GNNs.
- Additional Parameters Analysis: Two PGAL interaction layers are sufficient to capture spatial information, whereas more than three layers gradually degrade performance through over-fitting.
- Additional Parameters Analysis: Performance improves and then slightly declines as λ increases, while SIGN remains stable across coefficients and outperforms all baseline methods.