Beyond Sequence: Impact of Geometric Context for RNA Property Prediction

Junjie Xu, Artem Moskalev, Tommaso Mansi, Mangal Prakash, Rui Liao

Abstract

Traditional RNA property prediction models often treat RNA as a 1D sequence, overlooking the critical role of 2D and 3D structural information. This study systematically evaluates the impact of geometric context across various deep learning architectures. We find that while 2D and 3D models offer superior performance and data efficiency—reducing prediction error by approximately 12%—they exhibit increased sensitivity to sequencing noise. Our findings highlight a trade-off between geometric precision and model robustness in real-world genomic applications.

Introduction

RNA function is fundamentally linked to its complex folding patterns, yet the industry standard remains 1D sequence-based modeling. This approach ignores the physical reality of 2D base-pairing and 3D atomic arrangements. The primary barriers to adopting geometry-aware models are the scarcity of experimental 3D data and the potential for noise propagation from predicted structures. This study addresses these challenges by conducting a comprehensive "stress test" of geometric models under varying conditions of data availability and noise.

Overview of the study's framework, comparing 1D, 2D, and 3D RNA representations and analysis metrics.

Methodology

The research compares three distinct model families:

  1. 1D Models: Transformers and RNA-specific foundation models (e.g., RNA-FM).
  2. 2D Models: Graph Neural Networks (GNNs) such as GCN, GAT, and ChebNet, which represent RNA as a graph of base pairs.
  3. 3D Models: Equivariant GNNs (e.g., EGNN, SchNet) processing atomic XYZ coordinates.

A key innovation is "Nucleotide Pooling," which groups atoms into nucleotide-level representations to help 3D models handle longer sequences. The models were evaluated on four datasets: Tc-Riboswitches, Open Vaccine COVID-19, Ribonanza-2k, and Fungal, with structural annotations inferred via EternaFold and RhoFold.

Key Insights and Results

The Advantage of Geometry

Incorporating 2D and 3D geometry resulted in an average 12% reduction in Root Mean Square Error (RMSE) across all tasks. Spectral GNNs, specifically ChebNet, emerged as the top performers in the 2D category. By approximating global graph features, these models effectively capture long-range interactions that are physically proximal but distant in the 1D sequence.

Data Efficiency and Inductive Bias

Geometry-aware models demonstrate significant data efficiency. The study reveals that 1D sequence models require 2 to 5 times more training data to achieve the same accuracy as geometric models. This makes 2D and 3D approaches particularly valuable for small-scale datasets where experimental labels are limited.

Performance (MCRMSE) of various models across different fractions of training data for COVID, Ribonanza, and Fungal datasets.

The Robustness Paradox

Despite their accuracy, geometry-aware models are more fragile. While 1D models are robust to sequencing errors, 2D and 3D models are highly sensitive; a single mutation altering the predicted fold can increase prediction error by up to 82%. This suggests that the "stiffness" of 1D models provides a protective layer against noise that current geometric models lack.

Conclusion

This systematic evaluation proves that geometric context is a powerful tool for RNA property prediction, especially in low-data regimes. However, the sensitivity of these models to structural noise remains a significant hurdle. Future research should focus on hybrid ensemble architectures that combine the robustness of 1D sequences with the precision of 3D geometric graphs.

Summary updated

BibTeX
@inproceedings{xu2025beyond,
  author = {Junjie Xu and Artem Moskalev and Tommaso Mansi and Mangal Prakash and Rui Liao},
  title = {{Beyond Sequence: Impact of Geometric Context for RNA Property Prediction}},
  booktitle = {International Conference on Learning Representations},
  year = {2025},
  url = {https://openreview.net/forum?id=9htTvHkUhh},
}
Figure