Research map / Probabilistic and scientific modeling / Molecules and proteins
BAnG: Bidirectional Anchored Generation for Conditional RNA Design
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 Bidirectional Anchored Generation (BAnG) and the RNA-BAnG transformer to design RNA sequences binding a given protein from its sequence and AlphaFold2 structure, without target-specific binding data.
Problem
Designing RNA sequences that bind specific proteins usually requires many known binders for each target or accurate RNA structure knowledge, which limits practical use.
Approach
Generate sequences outward from an anchor placed inside a binding motif, alternating right and left token prediction with a special bidirectional attention mask and two anchor tokens so that a single training pass learns all conditionals; protein conditioning uses sequence plus AlphaFold2-style geometric attention.
Key results
- On the synthetic SingleBind/DoubleBind tasks BAnG reached 0.98/0.97 motif containment versus 0.94/0.53 for autoregressive generation.
- RNA-BAnG produced more high-affinity sequences than random for 56 of 71 test samples; threshold-curve area 0.57 vs 0.39 for random.
- Against GenerRNA on four SRSF1/ELAVL1 samples, RNA-BAnG had higher high-affinity proportions (e.g., 0.91 vs 0.64 for ELAVL1 sample 121).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BAnG instead of autoregressive generation | synthetic sequence generation · SingleBind | correct synthetic motif proportion: 0.94 -> 0.98 | Helped |
| Use BAnG instead of autoregressive generation | synthetic sequence generation · DoubleBind | correct synthetic motif proportion: 0.53 -> 0.97 | Helped |
| Use BAnG instead of random sequence generation | RNA sequence design · RNAcompete test set | area under threshold-dependent performance curve: 0.39 -> 0.57 | Helped |
| Use RNA-BAnG instead of GenerRNA | conditional RNA generation · GenerRNA comparison set | area under threshold-dependent performance curve: 0.66 -> 0.81 | Helped |
| Use RNA-BAnG instead of RNAFLOW | conditional RNA generation · RNAFLOW comparison set | area under threshold-dependent performance curve: 0.13 -> 0.53 | Helped |
| Remove geometric attention (ablation) | RNA-BAnG · RNAcompete comparison set | proportion above DeepCLIP threshold | Hurt |
| Add geometric attention (ablation) | RNA-BAnG · RNAcompete comparison set | proportion above DeepCLIP threshold | Helped |
| Remove relative position encodings in cross-attention (ablation) | RNA-BAnG · RNAcompete comparison set | proportion above DeepCLIP threshold | Hurt |
| Add relative position encodings in cross-attention (ablation) | RNA-BAnG · RNAcompete comparison set | proportion above DeepCLIP threshold | Helped |
| Remove intrinsically disordered regions from the protein structure | RNA-BAnG · RNCMPT00133 | number of generated high-affinity RNAs: tenfold increase | Helped |
| Increase sequence length from 50 to 100 | BAnG · SingleBind | correct synthetic motif proportion: 0.98 -> 0.98 | No clear effect |
| Increase sequence length from 50 to 200 | BAnG · SingleBind | correct synthetic motif proportion: 0.98 -> 0.95 | Hurt |
Surprising findings
- RNAFLOW generated sequences with unusually high G and C content (62% G, 32% C) regardless of protein, yielding high-affinity sequences for only a few proteins.
- Standard iterative masked generation (ESM3-style) performed very poorly on the synthetic motif task (0.06 and 0.04 on SingleBind).
Limitations
- Evaluation relies on a predictive DeepCLIP scoring model rather than experimental validation of binding.
- Performance depends on protein preprocessing (e.g., domain splitting or IDR removal) and on AlphaFold2 structure quality.
- Synthetic task uses a very small 17k-parameter toy model and simplified motifs.
Experimental setup
Datasets: PDB PPI3D RNAcentral RNAcompete RNA Compendium
Benchmarks: RNA Compendium (RNAcompete-derived test set, 71 samples) synthetic SingleBind/DoubleBind motif tasks
Metrics: proportion of high-affinity sequences (DeepCLIP > 0.75) area under threshold-dependent performance curve diversity (0.9 identity clusters) novelty fraction of sequences containing motif
Baselines: GenerRNA RNAFLOW autoregressive iterative entropy iterative logit max random sequences
Builds on: autoregressive transformer AlphaFold2 invariant point attention ESM3 iterative masked generation RoPE
Closest papers by meaning
- Bridging Sequence-Structure Alignment in RNA Foundation Models (AAAI 2025 · similarity 0.86)
- RiboFlow: Conditional De Novo RNA Co-Design via Synergistic Flow Matching (NeurIPS 2025 · similarity 0.86)
- gRNAde: Geometric Deep Learning for 3D RNA inverse design (ICLR 2025 · similarity 0.85)
- RNAFlow: RNA Structure & Sequence Design via Inverse Folding-Based Flow Matching (ICML 2024 · similarity 0.85)
- CoPRA: Bridging Cross-domain Pretrained Sequence Models with Complex Structures for Protein-RNA Binding Affinity Prediction (AAAI 2025 · similarity 0.85)
- Enforcing Constraints in RNA Secondary Structure Predictions: A Post-Processing Framework Based on the Assignment Problem (ICML 2024 · similarity 0.85)
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.