DRBP

Whether a protein binds DNA or RNA is settled experimentally one protein at a time, by ChIP-seq or by crosslinking immunoprecipitation, each run needing a working antibody or a tagged construct. Coverage is therefore thin: across the 18,221 protein-coding genes of the human reference proteome, only 1,447 proteins carry experimentally validated DNA-binding evidence and 351 carry eCLIP-validated RNA-binding evidence. Predictors built to fill that gap have mostly been trained on labels that were themselves inferred — Gene Ontology terms, homology transfer, structure-derived annotations — so they reproduce the biases of earlier tools rather than the underlying biology.

DRBP takes the opposite trade. Built at Chonnam National University with collaborators at Chungnam National University and Seoul Women's University, it trains only on experimentally supported annotations and accepts the small positive sets that follow. The release is a pair of binary classifiers under one name — one for DNA-binding proteins, one for RNA-binding — each a low-rank adaptation (LoRA) of ESM-2 that consumes amino acid sequence and nothing else.

A binary call says a protein binds but not which part of it does; residue-level predictors answer the reverse and skip the prior question of whether it binds at all. DRBP recovers region-level attribution from the classifier itself using value-aware attention (VAT) scores, which weight each attention value by the L2 norm of its value vector and average across all 33 layers and heads. Those scores concentrate on annotated binding domains: C2H2 zinc fingers and homeodomains for DNA, RNA recognition motifs and KH domains for RNA.

#Key Features

• Experimentally grounded labels only: DNA-binding positives come from hPDI microarray data and hTFtarget's curation of 7,190 ChIP-seq samples, RNA-binding positives from ENCODE eCLIP. Computationally inferred annotations were excluded outright.
• Parameter-efficient adaptation: LoRA is applied at rank 8 to the query, key, and value projections, leaving the pretrained backbone untouched and training a few million parameters plus the classification head.
• Region-level attribution: VAT profiles are baseline-corrected with a cubic smoothing spline to remove the positional bias near the [CLS] token, then z-scored per sequence so residue importance is comparable within and across proteins.
• Two tasks, two calibrated thresholds: the DNA and RNA heads are thresholded independently, at MCC-maximizing cutoffs of 0.7 and 0.9, reflecting how differently imbalanced the two label sets are.
• Proteome-scale inference: trained adapters for every cross-validation fold ship in the repository, so the classifiers run on new human sequences without retraining.

#Technical Details

The backbone is esm2_t33_650M_UR50D, an encoder-only transformer of 33 layers, hidden dimension 1,280 and roughly 650 million parameters, with a feedforward classification head on top. Inputs are truncated or zero-padded at the C-terminus to 1,024 residues, keeping 91% of DNA-binding and 88% of RNA-binding sequences intact. Training used binary cross-entropy with logits, AdamW at a learning rate of 1e-5, batch size 16 and five epochs with early stopping on a single RTX 4090, positives oversampled 12-fold for DNA and 50-fold for RNA.

Evaluation is 20-fold cross-validation over the same proteome, 95% training and 5% held out per fold at the original class balance. The DNA head reached an AUROC of 0.84 with an MCC of 0.40; the RNA head an AUROC of 0.92 with an MCC of 0.46. Re-run on this dataset, ESM-DBP scored an MCC of 0.343 on the DNA task, while DRBP-EDP and HydRA scored MCCs of 0.361 and 0.251 on the RNA task, DRBP-EDP posting a marginally higher AUROC of 0.933. These are within-dataset figures, not an external cohort, and the negative class is defined by absence of experimental evidence rather than demonstrated non-binding — some false positives were later found annotated as binding in UniProt.

#Applications

Applied across all 18,221 human proteins, the models nominate 2,412 candidate DNA-binding and 529 candidate RNA-binding proteins above their thresholds, each paired with the attention profile explaining the call. For a lab studying an uncharacterized protein, that turns a score into a testable proposition: VAT peaks over a compact stretch of sequence suggest where to place site-directed mutagenesis, an electrophoretic mobility shift assay, or a CLIP experiment. The same readout offers a route to asking how a disease-associated variant shifts predicted binding.

#Impact

DRBP's contribution is less a leaderboard position than a demonstration that a narrow, experimentally honest training set plus a frozen protein language model recovers domain-level structure the model was never shown. VAT profiles from the fine-tuned adapters and from the untuned backbone put peaks in near-identical positions over annotated binding domains, placing the signal in ESM-2's pretrained representations and casting LoRA as calibration rather than teaching. The limits are stated plainly by the authors: human proteins only, 1,024-residue truncation dropping 9%–13% of sequences, a 351-protein RNA positive set that cannot span the diversity of RNA-binding mechanisms, and attention that reflects correlation rather than causation.