GeSite

Protein–nucleic acid recognition is usually mediated by a small, highly conserved binding domain, but identifying which residues make contact remains slow experimentally: as of late 2024 UniParc held over 833 million protein sequences against roughly 226,000 protein–DNA and protein–RNA complex structures in the PDB. Computational predictors split into sequence-driven methods, which scale but extract limited signal, and structure-driven methods, which are more accurate but bounded by structural coverage.

GeSite, from Shaoliang Peng's group at Hunan University, argues that the binding-site problem is bottlenecked less by the network than by the representation feeding it. General-purpose protein language models characterize properties common to all protein families and are correspondingly under-specialized for the families that dominate nucleic-acid binding. GeSite therefore builds on domain-adapted language models — ESM-DBP for DNA-binding proteins, from the group's earlier work, and a new ESM-RBP for RNA-binding proteins — and feeds their embeddings into an E(3)-equivariant graph convolutional network over the query structure. The paper's own conclusion is that the gain comes from the adapted language model rather than the geometric head, and the ablation supports it.

The method is released as two independent single-task predictors, one for DNA-binding residues and one for RNA-binding residues, both driven from a single prediction.py entry point.

#Key Features

• Domain-adaptive language models: ESM-RBP is produced by continuing masked-language pretraining of the last five transformer blocks of ESM-2 650M on UniRBP40, a 459,656-sequence non-redundant set of RNA-binding proteins clustered at 40% identity from 9.7 million UniProtKB entries.
• Equivariant geometric head: The structure becomes a graph with residues as nodes and edges between Cα atoms closer than 14 Å; three equivariant graph convolutional layers update node embeddings jointly with coordinates, giving rotation- and translation-invariant predictions.
• Dual evolutionary inputs: Each residue's 2,048-dimensional feature vector concatenates the 1,280-dimensional domain-adapted language-model embedding with a 768-dimensional MSA Transformer embedding from an HHblits search against Uniclust30.
• Explanation by subgraph: GNNExplainer extracts, for each predicted binding residue, the subgraph driving that call, recovering annotated helix-turn-helix, homeodomain, OB-fold and Pumilio-repeat regions — including contributions from domains distant in sequence.
• Generalization testing beyond held-out splits: Evaluation covers subsets filtered to TM-score below 0.5 against training structures, and RNA-285_Test, assembled from protein–RNA complexes released after the training data was collected.

#Technical Details

Training uses DNA-573_Train (573 proteins, 14,479 binding residues) and RNA-495_Train (495 proteins, 14,609 binding residues), with CD-HIT at a 0.3 cluster threshold removing test proteins of high sequence identity to training. ESM-RBP pretraining ran roughly 35,000 steps over about three days on two Tesla A40 GPUs at a 512-token context; the prediction model was trained for 50 epochs on a single Tesla V100 with cross-entropy weighted 0.7/0.3 for positives and negatives, AdamW at a 1e-4 learning rate, and 1e-4 weight decay.

On the three independent test sets GeSite reaches MCC values of 0.522 (DNA-129_Test), 0.389 (DNA-181_Test) and 0.326 (RNA-117_Test), with AUCs of 0.941, 0.919 and 0.861 against predictors including GraphBind, GraphSite, ESM-NBR, CLAPE, ULDNA, iDRNA-ITF, EquiPNAS and hybridDBRpred. Swapping the domain-adapted embeddings for stock ESM-2 features and retraining drops MCC to 0.480, 0.362 and 0.285 — the 8.75%, 7.45% and 14.38% margins the paper attributes to domain adaptation. Across five graph architectures and three sequence architectures in cross-validation, the equivariant network leads on AUC and MCC for both tasks. Inference requires a PDB file per query named by PDB ID plus chain, an HHblits installation and the Uniclust30 database. Weights are distributed through Hugging Face with no license stated, across two repositories: the GeSite repository holds the DNA and RNA heads plus ESM-RBP and UniRBP40, while ESM-DBP and UniDBP40 sit in a separate repository from the group's earlier work.

#Applications

The predictor is aimed at structural and molecular biologists who have a structure — experimental or predicted — and want a per-residue shortlist of likely DNA or RNA contacts before committing to mutagenesis or crystallography. Because the two heads are separate, a protein can be scored independently for DNA and RNA binding, useful for the many nucleic-acid-binding proteins whose specificity is ambiguous from sequence alone. The GNNExplainer output extends this from prediction to hypothesis: the case studies recover the helix-turn-helix domain of a Thermoplasma acidophilum riboflavin kinase and the multi-domain contributions to binding in the African swine fever virus DNA ligase — the kind of evidence needed to justify a mutant panel.

#Impact

GeSite is a compact demonstration of a broadly applicable pattern: for a narrow protein family, continued pretraining of a general language model on that family's sequences buys more than architectural work on the downstream head. Publishing ESM-RBP as a separately loadable checkpoint alongside the task models makes that claim reusable rather than internal. The practical limitations are real — the preprint is not peer-reviewed and bioRxiv classifies it as confirmatory results, the RNA task remains harder than the DNA one at every reported metric, and setting up HHblits with Uniclust30 plus per-query PDB files is a heavier lift than a sequence-only predictor. No license accompanies the released weights.