KEPLA: A Knowledge-Enhanced Deep Learning Framework for Accurate Protein-Ligand Binding Affinity Prediction

TL;DR

KEPLA tackles the challenge of predicting protein-ligand binding affinity by integrating biochemical knowledge from Gene Ontology and ligand properties into a deep learning framework. It combines an ESM-based protein encoder and a GCN-based ligand encoder with a knowledge-graph embedding objective and a cross-attention-based PLA predictor, enabling knowledge-grounded joint representations. The approach yields state-of-the-art in-domain and cross-domain performance, while providing structural- and knowledge-level interpretability through attention maps and KG-derived explanations, and it introduces a novel KG dataset built on PDBbind. This knowledge-enhanced, interpretable framework advances drug discovery by improving predictive accuracy and offering actionable insights into binding mechanisms under realistic data shifts.

Abstract

Accurate prediction of protein-ligand binding affinity is critical for drug discovery. While recent deep learning approaches have demonstrated promising results, they often rely solely on structural features of proteins and ligands, overlooking their valuable biochemical knowledge associated with binding affinity. To address this limitation, we propose KEPLA, a novel deep learning framework that explicitly integrates prior knowledge from Gene Ontology and ligand properties to enhance prediction performance. KEPLA takes protein sequences and ligand molecular graphs as input and optimizes two complementary objectives: (1) aligning global representations with knowledge graph relations to capture domain-specific biochemical insights, and (2) leveraging cross attention between local representations to construct fine-grained joint embeddings for prediction. Experiments on two benchmark datasets across both in-domain and cross-domain scenarios demonstrate that KEPLA consistently outperforms state-of-the-art baselines. Furthermore, interpretability analyses based on knowledge graph relations and cross attention maps provide valuable insights into the underlying predictive mechanisms.

Figures (6)

• Figure 1: Overview of the KEPLA framework.(a) Existing PLA prediction strategy. (b) KEPLA strategy, which incorporates a biochemical KG to enhance protein and ligand encoding. (c) KEPLA framework with two encoders: an ESM-based encoder for protein sequences and a GCN-based encoder for ligand SMILES, producing local and global representations. Global representations are used for KG embedding, while local representations are fed into the cross attention module for PLA prediction. (d) Cross attention module, where an interaction map derived from protein–ligand local representations enables fine-grained feature fusion, followed by an MLP decoder for affinity prediction. (e) KG embedding module, which projects global representations into a semantic space and evaluates triplet plausibility via a scoring function.
• Figure 2: Cross-domain performance comparison on the PDBbind dataset (statistics over five independent runs).(a) Performance comparison of interaction-free methods under clustering-based pair split. (b) Performance comparison of interaction-free methods under cold pair split. The box plots display the median as the center line and the mean as a square marker. The minimum and lower percentile indicate the worst and second-worst scores, while the maximum and upper percentile represent the best and second-best scores, respectively.
• Figure 3: Ablation study in terms of RMSE and R on the PDBbind and CSAR-HiQ datasets under random split (statistics over five independent runs).(a) Performance of different variants of the KG module. (b) Performance of different variants of the cross attention module. The vertical bars represent the mean values, while the black lines denote standard deviations (error bars). Dots indicate the performance scores from each individual run.
• Figure 4: Visualization of ligands and binding pockets for structural-level interpretability study.(a) Interpretability of co-crystallized ligands. The upper section of each panel displays the 2D structures of ligands, with atoms highlighted in orange to indicate those predicted to contribute to protein binding. All 2D structures are visualized using RDKit. The lower section of each panel presents protein-ligand interaction maps derived from the corresponding crystal structures. (b) Interpretability of binding pocket structures. The 3D representations show protein-ligand binding pockets, with correctly predicted binding-site residues highlighted in orange and the corresponding ligands shown in cyan. Remaining amino acid residues, secondary structure elements, and surface maps are shown in grey. All protein-ligand interaction maps and 3D visualizations of X-ray crystal structures are generated using the Molecular Operating Environment (MOE) software.
• Figure 5: Biochemical knowledge of proteins and ligands for knowledge-level interpretability study.(a) Interpretability of protein biological knowledge. The upper part of each panel displays the GO entities most closely associated with the protein, ranked in ascending order according to their triplet scores from the KG embedding model. The lower part presents the types and detailed definitions of these entities, annotated with their corresponding GO IDs. (b) Interpretability of ligand chemical knowledge. The upper part of each panel shows the LP entities most closely related to the ligand, also ranked in ascending order by their triplet scores in the KG embedding model. The lower part provides the types and specific functions of these entities, with molecular descriptors annotated by their corresponding numbers.