C-PLANK

A fragment that binds a purified protein in a test tube and one that binds the same protein inside a cell are not the same event. The affinity datasets most drug–target interaction models train on — BindingDB, Davis, BioSNAP — record isolated proteins under controlled conditions, so a model fit to them learns binding chemistry while staying blind to protein complexes, post-translational state, and the competition a compound faces in a cell. C-PLANK (Chemi-Proteome Language Attention NetworK) changes the training substrate rather than the objective: it learns from fragment–protein interactions measured in living cells by fully functionalized fragment (FFF) chemoproteomics, where a probe carrying a variable fragment, a diazirine photocrosslinker and an alkyne handle is UV-crosslinked to whatever it engages and read out proteome-wide by mass spectrometry.

C-PLANK encodes protein sequence with frozen last-layer residue embeddings from ESM-2 concatenated with per-residue physicochemical descriptors, encodes ligand SMILES atom-by-atom through a learnable embedding matrix augmented with atom-level properties, and fuses the two through a bilinear attention network scoring every residue–atom pair. Because the chemoproteomic readout is proteome-wide, each pair carries context the assay supplies for free, distilled into the Cellular Interaction State Index (CISI): a metric, not a separate model, measuring how enriched a pair is against the global promiscuity of that ligand and protein. CISI stratifies pairs into high-, medium- and low-CISI groups that feed back as a knowledge loss on the evidential layer, so a prediction the wider interaction landscape does not support is penalised into lower confidence.

The framework comes from Shanghai Jiao Tong University and City University of Hong Kong, released as a preprint in August 2026.

#Key Features

• Trained on in-cell measurements: The corpus comes from FFF chemoproteomics in living cells with inactive fragment controls rather than in vitro affinity assays, so the learned interaction states carry cellular context.
• Dual protein and ligand representation: Protein language embeddings are concatenated with per-residue molecular weight, polarity, hydropathicity, pKa and topology; ligand atoms carry LogP, charge, topological polar surface area and hybridisation.
• Interpretable residue–atom fingerprints: The bilinear attention map yields per-residue and per-atom scores that project onto a structure or chemical drawing, showing which moiety and pocket drive a prediction.
• Evidence-aware confidence: An evidential Dirichlet layer reports uncertainty alongside probability, and CISI ties that uncertainty to independent chemoproteomic evidence rather than model confidence alone.
• Fragment-scale input window: Sequences are truncated or padded to 1,200 residues and ligands to 60 atoms, a window tuned to the low molecular weight of FFF probes.

#Technical Details

From 431 ligand interactomes curated across eight independent chemoproteomic studies, enrichment filtering yielded 45,247 positive fragment–protein pairs spanning 332 ligands and 2,629 proteins; promiscuity-matched negatives brought the balanced set to 90,494 pairs over 407 ligands, split 8:1:1. Both encoders are 3-layer 1D-CNNs with channel dimensions [32, 64, 128] and kernel sizes [3, 7, 11]; embeddings are 64-dimensional and the bilinear attention network uses 4 heads with an output dimension of 256. Optuna-tuned training used AdamW at 1e-4 with cosine annealing over up to 100 epochs, selecting the checkpoint on best validation AUROC. Evaluation ran under a random split and a stringent cold-protein split, in which proteins are clustered by 3-mer frequency into mutually exclusive partitions, both with 10-fold cross-validation. Against seven baselines including HyperAttentionDTI, TransformerCPI, MolTrans, DrugBAN and GraphBAN, C-PLANK led on AUROC and accuracy in both settings, by roughly 1–2% on the harder cold-protein split.

#Applications

The single validation-selected checkpoint transfers without retraining: applied to independent chemoproteomic datasets containing 24 previously unseen ligands, it reconstructed the interaction landscapes of 13 of them at AUROC and AUPRC ≥ 0.6, reaching 0.8497 on one. Attention maps projected onto crystal and predicted structures overlapped Fpocket-predicted druggable pockets and fell mostly within 10 Å of experimentally labelled binding sites. Prospectively, the team screened 5,669 commercial photoreactive probes against SIRT3, filtered the top 1% by predicted uncertainty, promiscuity and LogP, and advanced the hit SH00 — shown to enrich SIRT3 in cells and raise its deacetylase activity — to the optimised agonist SH10, with an EC50 of 9.23 μM and an SPR-measured Kd of 10.9 μM.

#Impact

C-PLANK shows that chemoproteomic data, expensive and scattered across labs, can be pooled into one trainable resource that generalises beyond its source studies and can be run forward into a campaign ending in a validated chemical probe. Its limits are stated plainly by the authors: 407 ligands is a narrow slice of fragment space, only diazirine photoreactive chemistry was included, inputs are 1D sequence and SMILES with no 3D structure, and pairs are scored in isolation even though FFF probes often pull down interconnected complexes. Code, the curated split dataset and a trained checkpoint ship together in one repository under GPL-3.0; the work is a preprint and has not yet been peer reviewed.