AiPP

Fewer than 900 human proteins have been targeted by an FDA-approved drug, and most of the remaining proteome lacks any chemical probe. Activity-based protein profiling (ABPP) is the workhorse for finding covalently ligandable cysteines at proteome scale, but its coverage is bounded by probe chemistry, protein abundance, and cell type, and the same cysteine is frequently called liganded in one study and unliganded in another. Earlier machine-learning ligandability predictors extrapolate beyond the measured sites but require a structure — a hard constraint across the structurally unresolved proteome.

AiPP (artificial intelligence protein profiling) attacks the problem from sequence alone: eight task heads and two auxiliary modules layered on frozen representations from a protein language model, developed by the Jana Shen lab at the University of Maryland School of Pharmacy with collaborators at Purdue University and posted to bioRxiv in September 2025. Its primary head, LigCys, scores every cysteine in a protein for covalent ligandability; the remaining heads annotate reversible ligand-binding residues and cysteine functional context, so a predicted site arrives with the evidence needed to judge it.

The trained artifacts are lightweight heads rather than a new backbone — AiPP builds on ESM Cambrian (ESMC) and leaves it frozen. The distinctive contributions are LatentLift, a procedure for reconciling contradictory chemoproteomic labels before training, and the resulting proteome-wide ligandability atlas.

#Key Features

• Covalent cysteine ligandability from sequence: LigCys ranks cysteines within a protein by intrinsic covalent ligandability, requiring no structure and remaining usable for disordered and structurally unresolved regions.
• Reversible pocket context: A companion head, LigBind, predicts residues that contact reversible small-molecule ligands, rescuing overlooked cysteines and filtering spurious ones by testing whether a predicted site sits near a real pocket.
• Cysteine functional annotation: Six additional heads flag cysteines engaged in disulfide bonds or coordination of Zn, Cu, Fe, Fe–S clusters, or heme, separating druggable sites from structurally or catalytically committed ones.
• LatentLift label harmonization: Cysteines are clustered by similarity of their language model embeddings and labeled by cluster consensus, converting discordant ABPP evidence into trainable labels while enforcing leakage-controlled splits.
• Human proteome atlas and web servers: Precomputed predictions for 19,486 cysteine-containing human proteins, on-demand inference, and a curated ABPP database are served publicly.

#Technical Details

LigCys is a three-layer perceptron trained on frozen 2,560-dimensional per-token embeddings from layer 76 of the 6-billion-parameter ESMC model; LigBind and the six context heads are single-layer perceptrons over the same representations. Training labels come from LigCysABPP, a database the authors assembled from 15 cysteine-directed ABPP studies published between 2016 and 2025, comprising 703,135 site-level records for 140,459 cysteines across 10,649 proteins. LatentLift assigns a cluster label only when at least four positive records from four distinct sources agree; iterative data expansion grew the training set to 1,099 and 1,744 proteins for two variants blended into the production model.

On LC3Dts — a time-stamped external test set of 68 protein chains with 77 covalently modified cysteines from PDB depositions between January 2024 and November 2025 — LigCys reached 73% Top-1 recovery and 81% AUPRC, against 49%/69% for TopCySPAL and 16%/58% for CovCysPredictor. LigBind, trained on 687,712 residue-level annotations across 1,998 proteins derived from BioLiP2, achieved 93.9% AUROC and 78.1% Top-10 recovery. Swapping the backbone for Ankh or Dayhoff Atlas lowered Top-1 recovery by 8% and 13%; ESM-2 15B matched ESMC, and fine-tuning it changed little. Adding explicit structural features slightly reduced accuracy.

#Applications

AiPP targets covalent ligand discovery against proteins that chemoproteomics has missed. Applied across the human proteome, it mapped 17,413 Top-1 ligandable cysteines in 17,379 proteins and flagged 567 ABPP-unliganded or undetected drug targets — including 78 GPCRs, 177 transporters, and 207 enzymes — carrying high-confidence sites. It ranked FOXA1 C258 first, matching a stereoselective acrylamide probe reported independently, and prioritized C480 of the phosphatase PTPN6, a cryptic allosteric site undetected by ABPP that guided discovery of an orally active first-in-class covalent inhibitor. Medicinal chemists can query the atlas or run the command-line tool, which needs an EvolutionaryScale Forge API token to compute embeddings.

#Impact

By showing that a frozen protein language model encodes enough of the local biochemical environment to rank ligandable cysteines without structures, AiPP extends covalent-site prediction to the share of the proteome that structure-based models cannot reach, and LatentLift offers a reusable recipe for training on contradictory proteomics labels. The preprint has not yet been peer reviewed, and the code is released under CC BY-NC 4.0, which precludes commercial use without author permission, although the archived weights and training datasets on Zenodo carry the more permissive CC BY 4.0. LigCys models intrinsic ligandability and is deliberately agnostic to cellular state, so phosphorylation- and redox-dependent reactivity fall outside its scope; it also handles mutation-created pockets unevenly, recovering the KRAS G12C cysteine but missing the p53 Y220C site. Coverage is limited to cysteines, with lysine and tyrosine heads named as future work.