Research map / Probabilistic and scientific modeling / Molecules and proteins
Protein-Nucleic Acid Complex Modeling with Frame Averaging Transformer
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 FAFormer, a frame-averaging-based equivariant Transformer that embeds FA within each layer, and uses predicted protein-nucleic acid contact maps as an unsupervised binding score for aptamer screening.
Problem
Predicting protein-nucleic acid contacts and screening aptamers for protein targets when labeled binding data is scarce and experimental screening is costly.
Approach
Integrates frame averaging as a geometric module inside each Transformer block (local-frame edge module, biased MLP attention, global-frame FFN, gated coordinate updates) rather than wrapping an encoder with FA, and ranks aptamers by maximum predicted contact probability.
Key results
- FAFormer achieves over 10% relative improvement over baselines on contact map prediction across three protein complex datasets (e.g., Protein-DNA F1 0.1457 vs 0.1283 for FA).
- On zero-shot aptamer screening, FAFormer reaches Top-10 precision 0.4000 on GFP and 0.3333 on HNRNPC, generally outperforming baselines.
- Compared with RoseTTAFoldNA, FAFormer gives comparable contact prediction and better aptamer screening while being 20-30x faster in inference.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FAFormer instead of the strongest listed baseline | Contact map prediction · Protein-RNA | F1: 0.1093.004 -> 0.1284.003 | Helped |
| Use FAFormer instead of the strongest listed baseline | Contact map prediction · Protein-DNA | F1: 0.1283.001 -> 0.1457.005 | Helped |
| Use FAFormer instead of the strongest listed baseline | Contact map prediction · Protein-Protein | F1: 0.1461.001 -> 0.1596.002 | Helped |
| Use FAFormer instead of the strongest listed baseline | Contact map prediction · Protein-RNA | PRAUC: 0.1015.002 -> 0.1113.004 | Helped |
| Use FAFormer instead of the strongest listed baseline | Contact map prediction · Protein-DNA | PRAUC: 0.1195.006 -> 0.1279.006 | Helped |
| Use FAFormer instead of the strongest listed baseline | Contact map prediction · Protein-Protein | PRAUC: 0.1245.001 -> 0.1463.003 | Helped |
| Add the Edge module (ablation) | FAFormer · Protein-DNA, Protein-RNA, Protein-Protein | F1 | Helped |
| Add the Edge module (ablation) | FAFormer · Protein-DNA, Protein-RNA, Protein-Protein | PRAUC | Helped |
| Add the attention mechanism (ablation) | FAFormer · Protein-DNA, Protein-RNA, Protein-Protein | F1 | Helped |
| Add the attention mechanism (ablation) | FAFormer · Protein-DNA, Protein-RNA, Protein-Protein | PRAUC | Helped |
| Use Global Frame FFN instead of conventional FFN (ablation) | FAFormer · Protein-DNA, Protein-RNA, Protein-Protein | F1 | Helped |
| Use Global Frame FFN instead of conventional FFN (ablation) | FAFormer · Protein-DNA, Protein-RNA, Protein-Protein | PRAUC | Helped |
Surprising findings
- RoseTTAFoldNA and AlphaFold3 fail to identify positive aptamers for some targets (e.g., NELF, UBLCP1) despite strong complex structure prediction.
- Some geometric methods underperform the vanilla Transformer on certain datasets, attributed to overfitting on crystal structures.
Limitations
- Geometric features are limited to C-alpha and C3 atom coordinates; backbone and sidechain features are not used.
- Aptamer screening is evaluated on five targets with relatively small sampled candidate sets for the RoseTTAFoldNA comparison (10% of candidates).
- Validation/test structures are predicted (ESMFold, RoseTTAFoldNA) rather than experimental, and the protein-RNA test set compared to RoseTTAFoldNA is small (16 cases).
Experimental setup
Datasets: PDB NDB RNASolo DIPS GFP aptamer candidates NELF aptamer candidates HNRNPC aptamer candidates CHK2 SELEX aptamers UBLCP1 SELEX aptamers
Benchmarks: Protein-RNA contact prediction Protein-DNA contact prediction Protein-protein contact prediction Nucleic acid binding site prediction Zero-shot aptamer screening (GFP, NELF, HNRNPC, CHK2, UBLCP1)
Metrics: F1 PRAUC Top-10 precision Top-50 precision
Baselines: Transformer SE(3)Transformer Equiformer EGNN GVP-GNN FA GraphBind GraphSite RoseTTAFoldNA AlphaFold3
Builds on: frame averaging Transformer Equiformer EGNN GVP-GNN
Closest papers by meaning
- Generalist Equivariant Transformer Towards 3D Molecular Interaction Learning (ICML 2024 · similarity 0.86)
- CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening (NeurIPS 2026 · similarity 0.83)
- PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design (ICML 2025 · similarity 0.83)
- An All-Atom Generative Model for Designing Protein Complexes (ICML 2025 · similarity 0.83)
- Triplet Interaction Improves Graph Transformers: Accurate Molecular Graph Learning with Triplet Graph Transformers (ICML 2024 · similarity 0.82)
- DualEqui: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules (NeurIPS 2025 · similarity 0.82)
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.