Rethinking Druggability in the Evaluation of AI-driven Structure-based Drug Design

Keywords: AI for Drug Discovery, Druggability, Structure-based Drug Design, Evaluation

Abstract: Structure-based drug design harnesses three-dimensional structural information to guide ligand discovery and has seen rapid progress through machine learning. Yet the evaluation of AI-driven SBDD models has largely ignored **druggability**---the propensity of a binding pocket to accept a small, drug-like molecule. As a result, generative models may appear successful by creating compounds that dock well to pockets that are not feasible drug targets. We review SBDD benchmarks and druggability assessment methods, highlight pitfalls of current evaluation protocols, and propose a methodology to incorporate continuous druggability scores into the widely used CrossDocked2020 benchmark. By weighting generative scores according to pocket druggability and analysing performance across druggable and undruggable targets, our framework encourages models to focus on realistic therapeutic targets and reveals algorithmic biases.

Submission Number: 63