Homology Modelling in Structure-Based Drug Design: Tools, Validation and Pitfalls

Abstract

Homology modelling builds a three-dimensional model of a protein from its sequence and the experimentally solved structure of a related protein, and it is often the only route to a structure for targets that crystallography or cryo-electron microscopy has not yet reached. This narrative mini review asks how far such models can be trusted when they are used for structure-based drug design, and what happens to that trust now that deep-learning predictors deliver models for almost any sequence. It draws on published work on template detection, alignment, model building, validation, docking and the treatment of flexibility, and it reports no new data. The evidence points to a conditional answer. Models built on close templates can support the analysis of a binding site and, in well-known receptor examples, the discovery of new ligands, but accuracy at the level of individual side chains and loops, which is the level that matters for docking, depends on the alignment, on the state of the template and on how the pocket moves. Global validation scores describe the whole fold and say little about the pocket. Predicted confidence from deep-learning methods is useful but does not remove the need for checks tied to the question being asked. The review ends with a short list of practices that make model-based design more reliable and more honestly reported.