DeltaDock: A Unified Framework for Accurate, Efficient, and Physically Reliable Molecular Docking

TL;DR

A novel two-stage docking framework, DeltaDock, consisting of pocket prediction and site-specific docking is proposed, innovatively reframe the pocket prediction task as a pocket-ligand alignment problem rather than direct prediction in the first stage, and follows a bi-level coarse-to-fine iterative refinement process to perform site-specific docking.

Abstract

Molecular docking, a technique for predicting ligand binding poses, is crucial in structure-based drug design for understanding protein-ligand interactions. Recent advancements in docking methods, particularly those leveraging geometric deep learning (GDL), have demonstrated significant efficiency and accuracy advantages over traditional sampling methods. Despite these advancements, current methods are often tailored for specific docking settings, and limitations such as the neglect of protein side-chain structures, difficulties in handling large binding pockets, and challenges in predicting physically valid structures exist. To accommodate various docking settings and achieve accurate, efficient, and physically reliable docking, we propose a novel two-stage docking framework, DeltaDock, consisting of pocket prediction and site-specific docking. We innovatively reframe the pocket prediction task as a pocket-ligand alignment problem rather than direct prediction in the first stage. Then we follow a bi-level coarse-to-fine iterative refinement process to perform site-specific docking. Comprehensive experiments demonstrate the superior performance of DeltaDock. Notably, in the blind docking setting, DeltaDock achieves a 31\% relative improvement over the docking success rate compared with the previous state-of-the-art GDL model. With the consideration of physical validity, this improvement increases to about 300\%.

Figures (11)

• Figure 1: The overview of DeltaDock's two modules. (a) The pocket-ligand alignment module CPLA. Contrastive learning is adopted to maximize the correspondence between target pocket and ligand embeddings for training. During inference, the pocket with the highest similarity of the ligand is selected. (b) The bi-level iterative refinement module Bi-EGMN. Initialized with a high-quality sampled pose, the module first performs a coarse-to-fine iterative refinement. This process generates progressively refined ligand poses utilizing a recycling strategy. To guarantee the physical plausibility of the predicted poses, a two-step fast structure correction is subsequently applied. This correction involves torsion angle alignment followed by energy minimization based on the SMINA.
• Figure 2: Site-specific docking performance. (a) Overall Performance of different methods on the PDBbind test set. The search space was delineated by extending the minimum and maximum of the x, y, and z coordinates of the ligand by 4 Å respectively. For TANKBind, we directly supply the protein block with a radius of 20 Å centered around the ground-truth ligand center to the model. (b) Overall performance of different methods on the PoseBusters dataset. (c) A waterfall plot for illustrating the PoseBusters tests as filters for both DeltaDock and DeltaDock-SC predictions. The evaluation results for DeltaDock are denoted above the lines, while those for DeltaDock-SC are annotated below.
• Figure 3: Further analysis on the (a) PDBbind and (b) PoseBusters dataset. Left: DCC cumulative curve of top-1 pockets. Middle: VCR cumulative curve of top-1 pockets. Right: Scatter plot of RMSD of initial and updated poses. All experiments are conducted in the blind docking setting.
• Figure 4: The main protease of SARS-CoV-2 is depicted by the white surface. The ligand structures in pink, blue, and red correspond to PDB 5RGY, 7AQJ, and 7JU7, respectively. Left: The green pocket, a protein structure truncated to within 12.0 Å of the blue structure, is insufficient to encompass the pocket structure necessary for predicting the red structure. Right: The orange pocket, truncated within a 40.0 Å box utilized by DeltaDock, is ample to cover the entire pocket.
• Figure 5: Performance of different pocket prediction methods on the PDBbind test set. The hit rate is significantly improved by ensembling the predicted pockets from various methods.