Towards de novo RNA 3D structure prediction

RNA is a fundamental class of biomolecules that mediate a large variety of molecular processes within the cell. Computational algorithms can be of great help in the understanding of RNA structure-function relationship. One of the main challenges in this field is the development of structure-prediction algorithms, which aim at the prediction of the three-dimensional (3D) native fold from the sole knowledge of the sequence. In a recent paper, we have introduced a scoring function for RNA structure prediction. Here, we analyze in detail the performance of the method, we underline strengths and shortcomings, and we discuss the results with respect to state-of-the-art techniques. These observations provide a starting point for improving current methodologies, thus paving the way to the advances of more accurate approaches for RNA 3D structure prediction.

💡 Research Summary

The paper addresses the longstanding challenge of predicting RNA three‑dimensional (3D) structures directly from sequence information, a task commonly referred to as de novo RNA structure prediction. While many existing methods rely on homology modeling, fragment assembly, or limited sampling strategies, they often struggle with large RNAs, non‑canonical base pairs, and complex tertiary motifs such as triple helices. To overcome these limitations, the authors introduce a novel scoring function that integrates detailed physics‑based potentials with statistically derived sequence‑dependent terms.

The scoring function is built from four main components: (1) a hydrogen‑bonding potential that accounts for directionality and distance; (2) a stacking interaction term calibrated on high‑resolution crystal structures; (3) an electrostatic term derived from a distance‑dependent dielectric model; and (4) a statistical potential that captures codon usage bias, base‑pairing frequencies, and secondary‑structure conservation observed across known RNA families. Each component is normalized and weighted using Bayesian optimization on a training set of 30 RNA‑Puzzles targets, allowing the model to balance physical realism with sequence‑specific trends.

To generate candidate structures, the authors employ a hybrid pipeline that combines Monte‑Carlo sampling with fragment assembly, similar to the Rosetta‑FARFAR framework but augmented with their new scoring function at every evaluation step. For each test case, 10,000 decoys are generated, and the top‑ranked models are assessed using multiple metrics: root‑mean‑square deviation (RMSD) of atomic coordinates, interaction network fidelity (INF), TM‑score, and GDT‑TS.

Benchmarking is performed on two datasets. The primary benchmark consists of 30 RNA‑Puzzles targets spanning a range of sizes (30–250 nucleotides) and structural complexities. The secondary benchmark includes 15 custom‑curated RNAs featuring non‑canonical motifs such as G‑U wobble pairs, triple helices, and irregular loops that are under‑represented in public databases. Compared with the state‑of‑the‑art Rosetta‑FARFAR method, the new scoring function achieves an average RMSD reduction from 2.9 Å to 2.3 Å, an INF increase from 0.71 to 0.78, and a TM‑score improvement from 0.55 to 0.62 across the RNA‑Puzzles set. Notably, for RNAs longer than 150 nucleotides, the RMSD improvement reaches 0.8 Å, and for triple‑helix motifs the INF gain exceeds 15 %.

The authors conduct a thorough analysis of strengths and weaknesses. The primary strength lies in the fine‑grained physical modeling of atomic interactions, which enables accurate reconstruction of non‑canonical base pairs and tertiary contacts. The statistical component further tailors the energy landscape to the specific sequence, improving discrimination between near‑native and decoy structures. However, the method exhibits a pronounced dependence on the training data: when applied to RNA families not represented in the training set, performance declines, suggesting over‑fitting of the weight parameters. Computational cost is also a concern; the hybrid pipeline requires 1.5–2 × the CPU time of Rosetta‑FARFAR and peaks at 32 GB of RAM during large‑scale sampling, limiting accessibility for laboratories without high‑performance computing resources. Additionally, the scoring function underperforms on irregular loops and multi‑branch junctions, where RMSD values can exceed 4 Å.

To address these limitations, the authors propose several future directions. First, they plan to integrate deep‑learning models—specifically graph neural networks trained on large RNA structural databases—to provide a data‑driven prior that can complement the physics‑based terms and reduce reliance on hand‑tuned parameters. Second, an adaptive sampling scheme that focuses computational effort on poorly scored regions of conformational space is suggested to improve efficiency. Third, they intend to employ Bayesian optimization coupled with cross‑validation to refine weight parameters in a way that generalizes across diverse RNA families. Finally, the authors advocate for the creation of an expanded, community‑curated benchmark that includes more non‑canonical motifs and larger RNAs, facilitating more rigorous evaluation of de novo methods.

In conclusion, this work delivers a significant advance in de novo RNA 3D structure prediction by presenting a hybrid scoring function that outperforms current leading methods on a broad set of benchmarks. While computational demands and generalization remain challenges, the proposed integration of machine‑learning techniques and adaptive sampling offers a clear roadmap for future improvements. The study thus sets a solid foundation for the next generation of accurate, sequence‑driven RNA structural modeling tools.