Impact d'AlphaFold sur le drug design : Avancées, limites et perspectives

Details

Supervisors

Spillier, Quentin

Faculty

Faculté de pharmacie et des sciences biomédicales

Degree label

Master [120] en sciences pharmaceutiques, à finalité spécialisée

Abstract

enThis thesis examines the real impact of AlphaFold on drug design, beyond the improvement of protein structure prediction alone. Because the three-dimensional structure of proteins determines their biological function and their interactions with ligands, the ability to predict protein structures directly influences drug discovery and optimization. In the context of a growing gap between the number of known protein sequences and the availability of experimental structures, computational approaches, followed by artificial intelligence, have progressively become essential tools. This work reviews the evolution of structural prediction approaches: experimental methods, which provide highly accurate reference structures but are time-consuming and costly; homology modeling, which is effective when suitable templates are available but rapidly loses accuracy as sequence identity decreases; and free modeling methods, which allow template free prediction at the expense of high computational cost and limited reliability. The thesis then shows how the integration of deep learning, combined with evolutionary information, has overcome many of these limitations, leading to the emergence of AlphaFold, whose architecture and successive versions are analyzed in detail. The contribution of AlphaFold to drug design is evaluated through two case studies. The first, on the TAAR1 receptor, shows that AlphaFold-guided virtual screening provides better enrichment, higher hit rates, and enables the identification of potent, selective, and in vivo active ligands compared with homology-based approaches. The second, involving the 5-HT2A and σ2 receptors, confirms that AlphaFold2 models can rival experimental structures for exploring novel chemical space. Finally, this thesis discusses the limitations of AlphaFold and proposes strategies to improve its use in drug design. Overall, the results show that AlphaFold can, in certain contexts, substitute for experimental structures to efficiently guide drug discovery, but that model quality is not systematically guaranteed. Therefore, AlphaFold models must be interpreted critically, refined when necessary, and compared with experimental data whenever available.