Illuminating the Ligandable Proteome with AI Protein Profiling

Most human proteins lack chemical probes or pharmaceutical modulators, leaving much of the proteome unexplored. 1,2 Activity-based protein profiling has enabled proteome-scale discovery of protein-ligand interactions and covalent inhibitors, 3-5 but remains limited by probe chemistry, protein abundance, and discordant ligandability assignments across studies. 6-9 Machine-learning (ML) models can in principle generate proteomewide ligandability maps, but the predictive utility of current models is limited by the requirement for structures and the use of incomplete and weak training labels. 9-12 Here we developed an artificial intelligence protein profiling (AiPP), a sequence-based multitask platform built on the ESMC protein language model, 13,14 to accelerate proteomewide therapeutic discovery and target identification. While its primary task is identification of covalently ligandable cysteines, seven additional task heads and two external modules provide broader context by annotating reversible ligand-binding residues, disordered molecular recognition features, cysteine functional context, and reactivities. Central to the development is the LatentLift clustering approach, which leverages latent space similarities to reconcile conflicting experimental labels and facilitate model training. Applied to the human proteome, AiPP generated a cysteine-directed ligandability atlas that overcomes the limitations of current chemo-proteomic maps. As a proof of concept, we demonstrate that AiPP can guide the discovery of covalent allosteric inhibitors targeting the previously undruggable protein tyrosine phosphatase PTPN6. By linking sequence, ligandability and biological context, AiPP provides an open resource for therapeutic discovery, while LatentLift offers a general strategy for harmonizing proteomics data for ML applications.