Research map / Probabilistic and scientific modeling / Molecules and proteins
Beyond Sequence: Impact of Geometric Context for RNA Property Prediction
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.
Systematically benchmarks 1D sequence, 2D graph, and 3D all-atom models for RNA property prediction on four curated datasets, under low-data, partial-label, and sequencing-noise conditions.
Problem
Whether explicit 2D and 3D RNA geometry improves RNA property prediction over sequence-only models, and how this holds under limited data, partial labels, sequencing noise, and OOD noise.
Approach
Unified evaluation of 1D, 2D (EternaFold-derived base-pairing graphs) and 3D (RhoFold-predicted all-atom graphs) models across realistic data regimes, plus a nucleotide-pooling strategy for all-atom 3D models that enlarges their receptive field.
Key results
- 2D models (e.g., ChebNet) outperform the 1D Transformer, with ChebNet achieving the best MCRMSE on COVID (0.279), Ribonanza (0.468), Tc-Riboswitches (0.621) and Fungal (0.973).
- Nucleotide-pooled 3D variants reduce MCRMSE by about 10% versus the original all-atom 3D models and outperform 1D models by about 5% on Ribonanza and Tc-Riboswitches.
- Under sequencing noise, the 1D Transformer degrades about 14% (robustness) and 27% (generalization), while 2D models degrade 30-82% and 3D models 29-56%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use 2D structural information with ChebNet instead of sequence-only Transformer1D | RNA property prediction · Open Vaccine COVID-19 | MCRMSE: 0.361$\pm$0.017 -> 0.279$\pm$0.007 | Helped |
| Use 2D structural information with ChebNet instead of sequence-only Transformer1D | RNA property prediction · Ribonanza-2k | MCRMSE: 0.705$\pm$0.015 -> 0.468$\pm$0.002 | Helped |
| Use 2D structural information with ChebNet instead of sequence-only Transformer1D | RNA property prediction · Tc-Riboswitches | MCRMSE: 0.705$\pm$0.079 -> 0.621$\pm$0.022 | Helped |
| Use 2D structural information with ChebNet instead of sequence-only Transformer1D | RNA property prediction · Fungal | MCRMSE: 1.417$\pm$0.005 -> 0.973$\pm$0.003 | Helped |
| Add nucleotide pooling to EGNN (ablation) | EGNN · Open Vaccine COVID-19 | MCRMSE: 0.480$\pm$0.025 -> 0.364$\pm$0.003 | Helped |
| Add nucleotide pooling to SchNet (ablation) | SchNet · Ribonanza-2k | MCRMSE: 0.843$\pm$0.004 -> 0.685$\pm$0.006 | Helped |
| Add nucleotide pooling to SchNet (ablation) | SchNet · Tc-Riboswitches | MCRMSE: 0.696$\pm$0.008 -> 0.655$\pm$0.038 | Helped |
| Add graph adjacency features to the attention matrix | Transformer1D · COVID | cosine similarity to adjacency: 0.107 -> 0.448 | Helped |
| Add graph adjacency features to the attention matrix | Transformer1D · COVID | cosine similarity to BPP: 0.090 -> 0.672 | Helped |
| Use sequence-only modeling rather than geometry-aware models as sequencing noise increases | RNA property prediction · Sequencing noise robustness and OOD generalization | MCRMSE increase relative to clean data: 14\% and 27\%; 2D models 30\% to 82\%; 3D models 29\% and 56\% | Helped |
Surprising findings
- Foundation models RNA-FM and SpliceBERT underperform a simple supervised Transformer1D on COVID and Ribonanza.
- Pairwise RMSD between 3D structure prediction tools reaches 11-45 Å, indicating large structural noise in predicted 3D inputs.
Limitations
- 3D all-atom models have limited receptive fields and do not scale to the Fungal dataset (OOM).
- Geometry-aware models are highly sensitive to sequencing-induced structural noise, while the sequence-only model is more robust.
Experimental setup
Datasets: Tc-Riboswitches Open Vaccine COVID-19 Ribonanza-2k Fungal
Benchmarks: Tc-Riboswitches OpenVaccine COVID-19 Ribonanza-2k Fungal
Metrics: MCRMSE
Baselines: Transformer1D RNA-FM SpliceBERT GCN GAT GraphGPS
Builds on: Transformer ChebNet EGNN FastEGNN EternaFold
Closest papers by meaning
- gRNAde: Geometric Deep Learning for 3D RNA inverse design (ICLR 2025 · similarity 0.89)
- DualEqui: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules (NeurIPS 2025 · similarity 0.87)
- RDesign: Hierarchical Data-efficient Representation Learning for Tertiary Structure-based RNA Design (ICLR 2024 · similarity 0.87)
- TANGO: RNA Topology and Geometry Co-Design (NeurIPS 2026 · similarity 0.87)
- Enforcing Constraints in RNA Secondary Structure Predictions: A Post-Processing Framework Based on the Assignment Problem (ICML 2024 · similarity 0.86)
- Geometric Hyena Networks for Large-scale Equivariant Learning (ICML 2025 · 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.