AI-Based Protein Structure Prediction in Drug Target Identification

Abstract

The deployment of AlphaFold2 and its successors--including ESMFold, RoseTTAFold, and AlphaFold3--has resolved the
protein structure prediction problem for the majority of the human proteome, generating a structural database of over 214
million protein models covering virtually all catalogued proteins across 48 organisms. This transformative capability has
fundamentally accelerated the drug target identification pipeline by enabling structure-based druggability assessment,
cryptic binding site discovery, and protein-protein interaction interface characterisation at proteome scale--without the
time and cost constraints of experimental structure determination. This study applies an integrated AI-driven pipeline
combining AlphaFold2 structural models, FoldSeek structural similarity search, DiffDock-L deep learning molecular
docking, and PLINDER protein-ligand interaction scoring to systematically assess druggability and identify novel binding
sites across 847 previously unstructured proteins in three high-priority disease proteomes: type 2 diabetes (T2D, 312
targets), Alzheimer's disease (AD, 287 targets), and colorectal cancer (CRC, 248 targets). Of 847 targets assessed, 312
(36.8%) were classified as druggable (DScore >= 0.6) using structure-predicted models compared to only 187 (22.1%)
classified as druggable from sequence-based predictions alone--a 66.8% expansion of the druggable target space.
Cryptic binding site analysis using CryptoSite and MDPocket identified 1,247 previously uncharacterised pockets, of
which 89 showed druggability scores exceeding known drug targets in the same disease class. Molecular docking of the
ChEMBL35 approved drug library (9,847 compounds) against newly identified binding sites identified 34 high-confidence
drug repurposing opportunities (DiffDock confidence score > 0.7, PLINDER similarity to known drug-target complexes >
0.6), including metformin against an AD-associated AMPK regulatory subunit conformation not previously characterised
as a drug-binding site.