DNAreader

A residue that grips a phosphate backbone and a residue that grips ATP look much alike to a classifier trained only to separate "binds DNA" from everything else. Most predictors of DNA-binding residues are correspondingly ligand-agnostic, flagging residues that contact proteins, RNA or small molecules almost as readily as true DNA contacts. A second split compounds it: predictors train either on structured protein–DNA complexes from the PDB or on disorder annotations from DisProt and MobiDB, and each family underperforms on the other's territory — structure-trained methods fall to AUROC below 0.70 on disorder-annotated proteins, disorder-trained methods below 0.63 on structured ones.

DNAreader, from Jian Zhang's group at Xinyang Normal University with Lukasz Kurgan's lab at Virginia Commonwealth University, changes what the training signal is asked to separate. Rather than a binary binding/non-binding objective, it defines six residue populations — structured and disordered variants of DNA-binding, other-ligand-binding, and non-binding residues — and trains a multi-center contrastive loss that pulls each toward its own anchor in latent space while pushing the anchors apart. Because "binds a non-DNA ligand" is its own anchor rather than an undifferentiated negative, cross-predictions are penalized directly instead of incidentally.

Published in Nucleic Acids Research in August 2026, it is the first predictor to annotate DNA-binding residues at per-residue resolution in structured and disordered regions alike. A companion module, DNAreaderDBIDR, aggregates those calls into whole DNA-binding disordered regions, matching the output of older disorder-trained tools.

#Key Features

• Anchored contrastive objective: A multi-center loss organizes the latent space around six residue types rather than two labels. Accuracy rises and cross-prediction error falls from three anchors to six; eight adds nothing significant.
• Population-balanced batch training: Each batch of 500 residues carries all six populations, with the two rare DNA-binding classes held at a tuned 4% each so the minority signal survives gradient descent.
• Cross-prediction treated as a metric: Evaluation reports cross-prediction ratio and area under the cross-prediction curve alongside AUROC, separating false positives that bind other ligands from those that bind nothing.
• Two compatible output modes: DNAreaderDBIDR feeds the residue scores plus IUPred3 disorder predictions into a logistic regressor over 33 derived features to call whole disordered binding regions.
• Sequence-only deployment: A public web server takes a FASTA sequence, needs no structure, and returns color-coded propensities and a parsable CSV.

#Technical Details

Each residue is encoded as a 90-dimensional vector: 25 dimensions from a frozen esm2_t6_8M_UR50D ESM-2 checkpoint — its 320-dimensional embedding compressed to five dimensions by a small feed-forward network, then read across a five-residue window — plus 65 sequence-derived features covering binding propensity, ASAquick solvent accessibility, AAindex physicochemical properties, IUPred3 disorder, and HHblits-derived evolutionary couplings. Fifteen consecutive residue vectors enter three stacked transformer encoder units (8 heads, key and value width 32, feed-forward width 128); the middle position's output passes through a five-layer feed-forward head to one sigmoid propensity. The trained network holds roughly 363,000 parameters, the frozen language model contributing none. Removing both contrastive learning and batch training drops AUROC from 0.863 to 0.803 and raises the area under the cross-prediction curve from 0.121 to 0.172.

Two blind test sets are separated from training data by BLASTclust at 25% identity: test set 1 (430 proteins) and test set 2 (231 proteins), compared against 14 published predictors. On test set 1 DNAreader reaches AUROC 0.86 and AULCratio 5.7 overall — 0.90 and 6.3 on structure-annotated proteins, 0.85 and 5.4 on disorder-annotated ones — against 0.79 and about 4.3 for the best meta-predictor, HybridDBRpred, and 0.75 and 4.1 for the best individual tool, iDRNA-ITF. Cross-prediction ratio exceeds 4.5 with AUCPC below 0.13, where the meta-predictors sit near 3 and 0.20. Test set 2 gives AUROC 0.88 and AUCPC 0.10, with all margins significant at P < .01.

#Applications

The immediate use is annotating DNA-binding residues in proteins with no solved complex — the PDB holds only about 9,200 protein–DNA structures. Because folded and disordered regions are scored on one scale, it suits transcription factors and chromatin proteins whose recognition elements sit in low-complexity tails. The low cross-prediction rate matters for prioritizing mutagenesis: a flagged residue is less likely to be a protein- or small-molecule site misread as a nucleic-acid contact. Local runs additionally require ASAquick, HHblits, IUPred3 and CLIP with hand-configured install paths.

#Impact

DNAreader erases the structured/disordered split rather than bridging it with a meta-predictor, the previous best approach. Its more transferable contribution is methodological: anchoring a contrastive loss on ligand-type subpopulations is a general recipe for binding-site predictors troubled by cross-predictions, and it applies equally to structure-based models such as GeSite. The trade-offs are real — a benchmark study with no experimental validation, a small single-task network with no exposed embeddings, and a language model supplying only 25 of the 90 input dimensions, with engineered features carrying the rest.