Abstract
Structure-based drug design (SBDD) relies on accurate knowledge of protein structure and ligand-binding conformations. However, most of the static conformations obtained by advanced methods such as structural biology and de novo protein folding algorithms often don’t meet the needs for drug design. We introduce PackDock, a flexible docking method that combines “conformation selection” and “induced fit” mechanisms in a two-stage docking pipeline. The core module of this method is PackPocket, which uses a diffusion model to sample diverse conformations within the equilibrium distribution of the ligand binding pocket or to converge towards the side-chain conformations induced by the ligand. We evaluate our method using several tests that reflect real-world application scenarios. (1) Side-chain packing and Re-docking experiments validate the ability of PackDock to predict accurate side-chain conformations and ligand conformations. (2) Cross-docking experiments with apo and non-homologous ligand-induced holo structures align with real docking scenarios, demonstrating PackDock’s practical value. (3) Docking experiments with hypothetical models show that PackPocket can potentially conduct SBDD starting from protein sequence information only. Additionally, we found that PackDock can identify key amino acid conformation changes, which may provide insights for lead compound optimization. We demonstrate PackDock can accurately predict the complex conformations in various application scenarios, by combining the conformation selection mechanism and the induced fit mechanism, and by using the ability of PackPocket to accurately predict the side chain conformations in the pocket region. We believe this method can improve the usability of existing structures, providing a new perspective for the SBDD community.
Introduction
In structure-based drug design (SBDD), predicting the binding conformation of a small molecule (ligand) to a protein is a fundamental challenge. Traditional docking tools often treat the protein as a rigid structure, but in reality, proteins are dynamic. A significant hurdle is the structural difference between apo (unbound) and holo (bound) states. Side chains in an apo structure may obstruct a ligand, causing docking failures. While "flexible docking" methods exist to address this, they are often computationally prohibitive or lack the atomic precision required for effective drug discovery. PackDock addresses these limitations by providing a fast, deep-learning-driven framework that accounts for protein flexibility through a two-stage generative process.
Methods
PackDock utilizes a novel two-stage pipeline inspired by classical biological binding theories: Conformational Selection and Induced Fit.
- Stage 1: Conformational Selection: The system uses a diffusion-based model, PackPocket, to generate a diverse ensemble of empty binding pocket conformations. This simulates the protein's natural stochastic fluctuations, identifying states predisposed to ligand binding.
- Stage 2: Induced Fit: After initial docking into these diverse pockets (using a physics-based engine like Vina-GPU), PackPocket is redeployed—this time conditioned on the ligand's presence—to refine side-chain packing. This step resolves steric clashes and optimizes protein-ligand interactions, effectively "hugging" the ligand.
The core engine, PackPocket, is a score-based diffusion model that operates in torsion space. By predicting rotation angles rather than raw Cartesian coordinates, the model ensures that physical constraints like bond lengths and angles are strictly maintained. It employs a geometric graph neural network to process spatial relationships between protein residues, atoms, and the ligand.
Results
PackDock was rigorously evaluated against several benchmarks, including PDBbind v2020, CASP13/14, and the DISCO cross-docking dataset.
- State-of-the-Art Accuracy: In side-chain recovery, PackPocket outperformed established tools like SCWRL4 and newer deep learning models like AttnPacker. On the challenging DISCO dataset, which uses non-native protein structures, PackDock achieved a 48.21% success rate (RMSD < 2.0 Å), significantly outperforming Vina (21.07%) and DiffDock (30.36%).
- Efficiency: PackDock demonstrated a massive speed advantage, running approximately 10 times faster than SMINA and 25 times faster than traditional Induced Fit Docking (IFD) workflows.
- Refining AI Models: The model proved highly effective at refining AlphaFold2-predicted structures. When docking into these hypothetical models, PackDock reduced the average ligand RMSD from 3.68 Å to 1.77 Å, showcasing its utility in sequence-to-structure drug design pipelines.
Discussion
PackDock represents a significant advancement in flexible docking by combining the strengths of generative deep learning with traditional physics-based docking. While it excels at side-chain manipulation, a current limitation is the assumption of a rigid protein backbone. Although backbone shifts are often minor, future iterations may incorporate backbone flexibility to capture larger conformational changes. Additionally, optimizing the serial components of the pipeline for parallel architectures could further enhance its speed for massive virtual screening campaigns. By accurately predicting complex conformations across various scenarios, PackDock provides a powerful new tool for the SBDD community.
Summary updated