Research map / Vision-language and multimodal / Cross-modal learning
CrossBind: Collaborative Cross-Modal Identification of Protein Nucleic-Acid-Binding Residues
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 CrossBind, a cross-modal model fusing an atom-level point cloud structure encoder (sparse U-Net with SSL and atom-wise attention) with ESM-2 sequence embeddings to predict protein nucleic-acid-binding residues.
Problem
Identifying protein residues that bind DNA or RNA from single-modality (sequence-only or structure-only) representations, which miss either semantic context or fine-grained 3D geometry.
Approach
Combines a self-supervised sparse-convolution point cloud encoder over atoms, an atom-wise attention module that aggregates atoms into residue representations, and ESM-2 sequence embeddings fused by a learnable weighted sum with co-evolutionary features, plus a residue propensity filter.
Key results
- On DNA 129, CrossBind reaches F1 0.602, MCC 0.581, AUPR 0.628, versus GraphSite's F1 0.543 and MCC 0.519.
- On RNA 117, CrossBind improves F1 by 17.3% and MCC by 24.8% over GraphBind (F1 0.420 vs 0.358).
- Ablation shows removing atom-wise attention drops DNA 129 F1 from 0.602 to 0.582 and AUPR from 0.628 to 0.581.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine protein structure and sequence features with CrossBind modules | CrossBind · DNA 129 | F1: 0.543 -> 0.602 | Helped |
| combine protein structure and sequence features with CrossBind | CrossBind · DNA 181 | F1: 0.420 -> 0.475 | Helped |
| combine protein structure and sequence features with CrossBind | CrossBind · RNA 117 | F1: 0.358 -> 0.420 | Helped |
| add the protein structure encoder to the language model (ablation) | CrossBind · DNA 129 | F1: 0.511 -> 0.574 | Helped |
| add the sequence language model to atom point cloud segmentation (ablation) | CrossBind · DNA 129 | F1: 0.429 -> 0.574 | Helped |
| add atom-wise attention (ablation) | CrossBind · DNA 129 | F1: 0.582 -> 0.602 | Helped |
| add self-supervised learning (ablation) | CrossBind · DNA 129 | F1: 0.588 -> 0.602 | Helped |
| add the residue propensity filter (ablation) | CrossBind · DNA 129 | F1: 0.595 -> 0.602 | Helped |
| add co-evolutionary features (ablation) | CrossBind · DNA 129 | F1: 0.599 -> 0.602 | Helped |
| increase the ESM-2 encoder from 6 to 12 transformer layers (ablation) | CrossBind · DNA 129 | MCC: 0.432 -> 0.457 | Helped |
| increase the ESM-2 encoder from 12 to 30 transformer layers (ablation) | CrossBind · DNA 129 | MCC: 0.457 -> 0.525 | Helped |
| increase the ESM-2 encoder from 30 to 33 transformer layers (ablation) | CrossBind · DNA 129 | MCC: 0.525 -> 0.559 | Helped |
Surprising findings
- Adding co-evolutionary features led to an insignificant performance drop, as the protein language model already captured most of that information.
Limitations
- Relies on native protein structures; not evaluated on predicted structures except DNA 181 with AlphaFold2 predictions.
- Structure encoder limited to residue-level binding prediction; general 3D structure pre-training left to future work.
Experimental setup
Datasets: DNA 573 DNA 129 RNA 117 DNA 181
Benchmarks: DNA 129 DNA 181 RNA 117
Metrics: precision recall F1 MCC AUC AUPR
Baselines: GraphBind GraphSite DNABind DNAPred NucBind SVMnuc NCBRPred COACH-D NucleicNet aaRNA RNABindR Plus
Builds on: SimCLR GraphBind GraphSite ESM-2 MinkowskiNet
Closest papers by meaning
- Accurate Nucleic Acid-Binding Residue Identification Based Domain-Adaptive Protein Language Model and Explainable Geometric Deep Learning (AAAI 2025 · similarity 0.87)
- CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening (NeurIPS 2026 · similarity 0.86)
- UniSite: The First Cross-Structure Dataset and Learning Framework for End-to-End Ligand Binding Site Detection (NeurIPS 2025 · similarity 0.85)
- Pre-Training Protein Bi-level Representation Through Span Mask Strategy On 3D Protein Chains (ICML 2024 · similarity 0.85)
- DualEqui: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules (NeurIPS 2025 · similarity 0.85)
- CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.84)
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.