Decoding the limits of deep learning in molecular docking for drug discovery.

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery
• Structural biology and molecular modeling

Background:

• Structure-based molecular docking is crucial for computational drug design.
• Deep learning (DL) is revolutionizing docking but presents translation challenges.
• Evaluating DL docking methods against traditional approaches is essential.

Purpose of the Study:

• To comprehensively assess traditional and DL-driven molecular docking methods.
• To analyze performance across pose prediction, physical plausibility, interaction recovery, virtual screening (VS) efficacy, and generalization.
• To identify limitations and propose optimization strategies for DL docking frameworks.

Main Methods:

• Comparative analysis of traditional docking with DL paradigms (generative diffusion, regression, hybrid).
• Evaluation across five critical performance dimensions: accuracy, plausibility, recovery, VS efficacy, and generalization.
• Investigation of failure mechanisms and exploration of optimization strategies.

Main Results:

• Generative diffusion models show superior pose prediction accuracy; hybrid methods offer a balanced performance.
• Regression models frequently produce physically implausible poses; most DL methods exhibit high steric tolerance.
• Significant generalization challenges observed, especially with novel protein binding pockets, limiting current DL applicability.

Conclusions:

• DL significantly impacts molecular docking, with generative models leading in accuracy.
• Current DL methods face limitations in physical plausibility and generalization, hindering broad application.
• Further research is needed to develop robust, generalizable DL frameworks for reliable in silico drug design.