ProSiteHunter

Protein function is often mediated at a handful of residues—the binding sites where a protein contacts DNA, RNA, another protein, or an antigen. Predicting these interface residues from sequence alone is valuable because experimental structures are unavailable for most proteins, yet most existing predictors are specialized for a single interaction type and do not transfer across them. ProSiteHunter, developed at Zhejiang University of Technology and released as a preprint in October 2025, is a unified sequence-based framework that predicts protein–DNA, protein–RNA, protein–protein, and antibody–antigen binding sites within a single approach.

The method pairs a fine-tuned protein language model with a multi-source feature-fusion network. Its language-model component, SiteT5, is produced by fine-tuning ProtT5 so that its embeddings emphasize the evolutionarily conserved features characteristic of functional sites; these are then combined with complementary evolutionary, geometric, and statistical descriptors.

By handling multiple interaction categories in one framework, ProSiteHunter aims to replace a patchwork of task-specific tools with a consistent, sequence-only predictor for interface residues.

#Key Features

• Unified multi-task prediction: A single framework predicts protein–DNA, protein–RNA, protein–protein, and antibody–antigen binding sites, rather than requiring a separate model per interaction type.
• SiteT5 language model: A fine-tuned ProtT5 variant tailored to capture the conserved sequence signatures of functional residues.
• Multi-source feature fusion: Language-model embeddings are fused with evolutionary, geometric, and statistical features to strengthen residue-level predictions.
• Sequence-only input: Predictions require only amino-acid sequence, making the method applicable to proteins that lack an experimental or predicted structure.

#Technical Details

ProSiteHunter's core representation comes from SiteT5, a fine-tuned version of ProtT5-XL-UniRef50, whose per-residue embeddings are combined with evolutionary, geometric, and statistical features in a fusion network that outputs residue-level binding-site probabilities. On comprehensive benchmarks the authors report a substantial average improvement—on the order of 38%—in area under the precision-recall curve (PRAUC) across the protein–DNA, protein–RNA, and protein–protein tasks relative to prior state-of-the-art methods, along with a further gain on the especially challenging antibody–antigen task, where positive interface residues are sparse and heterogeneous.

#Applications

ProSiteHunter serves structural biologists, immunologists, and drug-discovery teams who need to localize functional residues without an experimental structure. Typical uses include annotating nucleic-acid-binding residues in transcription factors and RNA-binding proteins, mapping protein–protein interfaces for interaction studies, and predicting antibody paratope or antigen epitope residues to guide antibody engineering. Because it operates from sequence alone within one framework, it streamlines pipelines that would otherwise stitch together several interaction-specific predictors.

#Impact

ProSiteHunter demonstrates that a single fine-tuned protein language model, coupled with complementary hand-crafted features, can outperform specialized methods across a range of binding-site prediction tasks—including the difficult antibody–antigen case. By unifying protein–nucleic acid, protein–protein, and antibody–antigen prediction, it reduces the tooling burden for interface analysis and provides a strong sequence-only baseline. As a preprint awaiting peer review, its reported gains are established through in-silico benchmark comparisons.