Research map / Label-efficient and robust learning / Recognition and long-tail
Accurate Nucleic Acid-Binding Residue Identification Based Domain-Adaptive Protein Language Model and Explainable Geometric Deep Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GeSite, which combines a DNA/RNA-binding domain-adaptive ESM2 protein language model (ESM-RBP fine-tuned on 459,656 RBPs; ESM-DBP reused) with an E(3)-equivariant GNN to predict nucleic acid-binding residues and explains predictions with GNNExplainer.
Problem
Identifying DNA- and RNA-binding residues from protein sequence/structure is inaccurate and lacks interpretability, and general-purpose protein language models are not specialized for nucleic acid-binding families.
Approach
Continue-train ESM2 on nucleic acid-binding protein sequences (domain-adaptive PLM) to get sequence embeddings, concatenate them with MSA-based ESM-MSA embeddings, and feed a residue-distance graph into an EGNN; GNNExplainer is used to show which neighboring residues drive predictions.
Key results
- GeSite MCC 0.522 on DNA-129_Test and 0.326 on RNA-117_Test; AUC 0.941, 0.919 and 0.861 on DNA-129_Test, DNA-181_Test and RNA-117_Test.
- ESM-RBP vs. original ESM2 raised MCC on RNA-117_Test from 0.285 to 0.326 (about 14.38% relative improvement).
- GeSite is superior to or comparable with 9 DBS and 7 RBS predictors; the gain over the second-best RBS method is reported as 38.14% in MCC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ESM-DBP embeddings instead of ESM2 embeddings (ablation) | GeSite · DNA-129_Test | MCC: 0.480 -> 0.522; 8.75% | Helped |
| Use ESM-DBP embeddings instead of ESM2 embeddings (ablation) | GeSite · DNA-181_Test | MCC: 0.362 -> 0.389; 7.45% | Helped |
| Use ESM-RBP embeddings instead of ESM2 embeddings (ablation) | GeSite · RNA-117_Test | MCC: 0.285 -> 0.326; 14.38% | Helped |
| Use GeSite instead of GraphSite | DNA-binding residue prediction · DNA-129_Test | MCC: 0.519 -> 0.522 | Helped |
| Use GeSite instead of GraphSite | DNA-binding residue prediction · DNA-181_Test | MCC: 0.397 -> 0.389 | Hurt |
| Use GeSite instead of iDRNA-ITF | RNA-binding residue prediction · RNA-117_Test | MCC: 0.236 -> 0.326 | Helped |
Surprising findings
- RBS prediction is consistently lower than DBS prediction, which the authors attribute to fewer RBP training samples and more complex RNA binding patterns.
Limitations
- RBS prediction performance is notably lower than DBS prediction (RNA-117_Test MCC 0.326 vs DNA-129_Test 0.522).
- Interpretability is shown only on a handful of hand-picked cases, qualitatively.
- Relies on predicted MSA features (HHblits/Uniclust30) and fine-tuning only the last five ESM2 blocks.
Experimental setup
Datasets: UniRBP40 DNA-573_Train RNA-495_Train UniProtKB RBPs Uniclust30
Benchmarks: DNA-129_Test DNA-181_Test RNA-117_Test
Metrics: Spe Rec Pre F1 MCC AUC AP PCC p-value
Baselines: GraphBind GraphSite ESM-NBR CLAPE DRNAPred ULDNA iDRNA-ITF DNApred hybridDBRpred hybridRNAbind Pprint2 ESM2
Builds on: ESM2 ESM-DBP EGNN GNNExplainer GraphBind
Closest papers by meaning
- CrossBind: Collaborative Cross-Modal Identification of Protein Nucleic-Acid-Binding Residues (AAAI 2024 · similarity 0.87)
- MMSite: A Multi-modal Framework for the Identification of Active Sites in Proteins (NeurIPS 2024 · similarity 0.84)
- CoPRA: Bridging Cross-domain Pretrained Sequence Models with Complex Structures for Protein-RNA Binding Affinity Prediction (AAAI 2025 · similarity 0.84)
- DualEqui: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules (NeurIPS 2025 · similarity 0.84)
- ESM All-Atom: Multi-Scale Protein Language Model for Unified Molecular Modeling (ICML 2024 · similarity 0.83)
- LEMON-ZEST: Evolution-Informed Tokenization for Efficient Protein Language Modeling (NeurIPS 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.