Group Ligands Docking to Protein Pockets

TL;DR

GroupBind addresses molecular docking by coordinating multiple ligands binding the same protein pocket. It introduces group ligand message passing and a triangle attention mechanism within a diffusion-based docking framework to exploit pose similarity across ligands. The approach achieves state-of-the-art performance on the PDBBind benchmark, with augmented ligand sets further boosting accuracy and robustness across scenarios. Limitations include computational cost and dependence on available group ligand data, suggesting future work to leverage non-binding ligands and selective aggregation strategies. Overall, GroupBind provides a principled mechanism to transfer information across related docking tasks, enhancing pose prediction in real-world drug discovery.

Abstract

Molecular docking is a key task in computational biology that has attracted increasing interest from the machine learning community. While existing methods have achieved success, they generally treat each protein-ligand pair in isolation. Inspired by the biochemical observation that ligands binding to the same target protein tend to adopt similar poses, we propose \textsc{GroupBind}, a novel molecular docking framework that simultaneously considers multiple ligands docking to a protein. This is achieved by introducing an interaction layer for the group of ligands and a triangle attention module for embedding protein-ligand and group-ligand pairs. By integrating our approach with diffusion-based docking model, we set a new S performance on the PDBBind blind docking benchmark, demonstrating the effectiveness of our proposed molecular docking paradigm.

Figures (8)

• Figure 1: Multiple ligands that can bind to the same protein pocket exhibit similar docking poses.
• Figure 2: Overview of the proposed GroupBind paradigm. (a) GroupBind leverages other ligands ${\mathcal{G}}_{L_{1:K}}$ known to bind the same target protein ${\mathcal{G}}_{P}$ to facilitate the docking of ${\mathcal{G}}_{L_0}$. (b) GroupBind introduces the message passing across group ligands and the triangle attention module between group ligands and proteins. (c) GroupBind is not limited to any specific docking framework and can be integrated into existing models (dashed gray box) with the aforementioned modules (blue ones).
• Figure 3: Visualization of two examples (UniProt ID: P03211, Q9H7Z6) with reference ligand docking poses, GroupBind's best predicted docking poses and DiffDock's best predicted docking poses. There are two and three different ligands that can bind to P03211 and Q9H7Z6, respectively, highlighted in distinct colors. Reference ligands tend to bind the target protein with similar poses, which is well captured with GroupBind and results in lower average RMSD compared to DiffDock.
• Figure 4: Analysis of the effect of augmented ligands. (a) The docking success rate (ligand RMSD $<$ 2 Å) increases with the number of selection increases. (b) Effect of the group size. "SG" and "AG" indicate group ligands within test set and further utilize ligands from the training set respectively. (c) Effect of Group Similarity. X-axis denotes the Tanimoto similarity within the ligand group and Y-axis denotes the average best docking RMSD. (d) Effect on Different Subsets. "w/SG, w/AG and w/o G" indicates the subsets where augmented ligands from the test set, the training set, and no augmented ligands can be used respectively. "SG, AG, NG" indicates the corresponding model performance under different augmented ligands settings: test set, training set and no augmented ligands. For (b)(c)(d), the best docking pose among 40 candidates is applied for the success rate.
• Figure 5: PDB entries 6g2c and 6g29 contain distinct ligands (Tanimoto similarity = 0.26) that bind to the same pocket but share similar binding poses.