New deep learning approach greatly improves RNA structure prediction

Researchers present ARES (Atomic Rotationally Equivariant Scorer) – a machine learning method that significantly improves the computational prediction of RNA structures over previous approaches. Like proteins, RNA molecules twist and fold into intricate three-dimensional shapes that are crucial to their function. Understanding these structures could help uncover the biological functions of RNA, including non-coding RNA, and pave the way for discovering novel drugs for diseases that remain untreatable. However, experimentally solving RNA structure remains a challenge despite decades of effort, and only a few RNA structures are currently known. Additionally, using machine learning to predict RNA structure has proven far more difficult – and less successful – than it has for protein structure prediction. To address these challenges, Raphael Townshend and colleagues developed ARES, a deep neural network that can consistently produce accurate RNA structural models, despite being trained using data for only 18 recently experimentally determined RNA structures. According to the authors, ARES significantly outperformed other computational approaches in the community-wide RNA-Puzzles structure prediction challenge. Townshend et al. note that ARES’s performance is particularly notable as it learned to make its predictions based solely on atomic structure and does not incorporate prior assumptions about what RNA-specific structural features might be important, such as base pairs, nucleotides or hydrogen bonds. As well, it was able to accurately predict the structures of RNAs larger and more complex than those it was trained on. “ARES is still short of the level consistent with atomic resolution or sufficient to guide identification of key functional sites or drug discovery efforts, but Townshend et al. have achieved notable progress in a field that has proven recalcitrant to transformative advances,” writes Kevin Weeks in a related Perspective.