TriDS: AI-native molecular docking framework unified with binding site identification, conformational sampling and scoring

TL;DR

TriDS is an AI-driven molecular docking framework that unifies binding site detection, conformational sampling, and scoring, improving accuracy and efficiency for drug discovery applications.

Contribution

It introduces a novel, unified docking method using ML-based differentiable scoring, expanding previous work and enhancing user-friendliness and computational performance.

Findings

Achieves high docking accuracy, especially for large ligands.

Improves computational efficiency in speed and GPU memory.

Successfully integrates binding site prediction with conformational sampling and scoring.

Abstract

Molecular docking is a cornerstone of drug discovery to unveil the mechanism of ligand-receptor interactions. With the recent development of deep learning in the field of artificial intelligence, innovative methods were developed for molecular docking. However, the mainstream docking programs adopt a docking-then-rescoring streamline to increase the docking accuracy, which make the virtual screening process cumbersome. Moreover, there still lacks a unified framework to integrate binding site identification, conformational sampling and scoring, in a user-friendly manner. In our previous work of DSDP and its subsequent flexible version, we have demonstrated the effectiveness of guiding conformational sampling with the gradient of analytic scoring function. As the third generation of DSDP, here we expanded the similar strategy to ML-based differentiable scoring model to device a novel docking method named TriDS under the mainstream AI training framework, which unifies the sampling and scoring steps. To be user-friendly, TriDS also integrates ML-based model for binding site prediction and has compatibility with multiple input file formats. We show here that gradients of a suitable ML-based scoring function can lead to excellent docking accuracy in the benchmark datasets, especially for large ligands. Moreover, TriDS is implemented with enhanced computational efficiency in terms of both running speed and GPU memory requirement.