Abstract
Molecular docking is a pivotal process in drug discovery. While traditionaltechniques rely on extensive sampling and simulation governed by physicalprinciples, these methods are often slow and costly. The advent of deeplearning-based approaches has shown significant promise, offering increases inboth accuracy and efficiency. Building upon the foundational work of FABind, amodel designed with a focus on speed and accuracy, we present FABind+, anenhanced iteration that largely boosts the performance of its predecessor. Weidentify pocket prediction as a critical bottleneck in molecular docking andpropose a novel methodology that significantly refines pocket prediction,thereby streamlining the docking process. Furthermore, we introducemodifications to the docking module to enhance its pose generationcapabilities. In an effort to bridge the gap with conventionalsampling/generative methods, we incorporate a simple yet effective samplingtechnique coupled with a confidence model, requiring only minor adjustments tothe regression framework of FABind. Experimental results and analysis revealthat FABind+ remarkably outperforms the original FABind, achieves competitivestate-of-the-art performance, and delivers insightful modeling strategies. Thisdemonstrates FABind+ represents a substantial step forward in molecular dockingand drug discovery. Our code is in https://github.com/QizhiPei/FABind.