De Novo RNA Tertiary Structure Prediction at Atomic Resolution Using Geometric Potentials from Deep Learning

Robin Pearce; Gilbert S. Omenn; Yang Zhang

doi:10.1101/2022.05.15.491755

bioRxivMay 15, 2022BioinformaticsPreprint

De Novo RNA Tertiary Structure Prediction at Atomic Resolution Using Geometric Potentials from Deep Learning

Robin Pearce iD, Gilbert S. Omenn iD, Yang Zhang iD

PDF DOI

Structured summary available · Jump to summary ↓

Abstract

Experimental characterization of RNA structure remains difficult, especially for non-coding RNAs that are critical to many cellular activities. We developed DeepFoldRNA to predict RNA structures from sequence alone by coupling deep self-attention neural networks with gradient-based folding simulations. The method was tested on two independent benchmark datasets from Rfam families and RNA-Puzzle experiments, where DeepFoldRNA constructed models with an average RMSD=2.69 [A] and TM-score=0.743, which outperformed state-of-the-art methods and the best models submitted from the RNA-Puzzles community by a large margin. On average, DeepFoldRNA required ~1 minute to fold medium-sized RNAs, which was ~350-4000 times faster than the leading Monte Carlo simulation approaches. These results demonstrate the major advantage of advanced deep learning techniques to learn more accurate information from evolutionary profiles than knowledge-based potentials derived from simple statistics of the PDB library. The high speed and accuracy of the developed method should enable large-scale atomic-level RNA structure modeling applications.

Abstract summary

Claude Opus 5.5 · xhigh effort · 2 Nov 2026 · from the abstract · CC BY-NC-ND 4.0

DeepFoldRNA couples deep self-attention neural networks with gradient-based folding simulations to predict atomic-resolution RNA tertiary structures from sequence in ~1 minute with 2.69 A average RMSD.

Experimental RNA structure determination is challenging, and existing computational approaches rely on slow Monte Carlo simulations and knowledge-based potentials derived from Protein Data Bank (PDB) statistics. DeepFoldRNA replaces these with geometric potentials learned from evolutionary profiles via deep self-attention networks coupled to gradient-based folding simulations.

Key results

ShownDeepFoldRNA predicted RNA tertiary structures from sequence with higher accuracy than existing methods and top community submissions across Rfam and RNA-Puzzles benchmarks.
Achieved an average RMSD=2.69 [A] and TM-score=0.743 across two independent benchmark datasets from Rfam families and RNA-Puzzle experiments. · Abstract
ShownGradient-based folding in DeepFoldRNA reduced RNA structure prediction runtime by orders of magnitude compared with leading Monte Carlo simulation approaches.
Required ~1 minute on average to fold medium-sized RNAs, which is ~350-4000 times faster than leading Monte Carlo simulation methods. · Abstract

Methods, data and reproducibility

Combined deep self-attention neural networks trained on evolutionary profiles to predict geometric potentials with gradient-based folding simulations to fold RNA tertiary structures from sequence alone.

Two independent benchmark datasets from Rfam families and RNA-Puzzle experiments · n = 2 benchmark datasets

Methods & tools: DeepFoldRNA, Rfam, RNA-Puzzles, PDB
Experimental conditions: ~1 minute on average to fold medium-sized RNAs

Limitations (1)

Because the method learns geometric potentials from evolutionary profiles, performance on RNAs with shallow or no homologous sequence alignments is not addressed in the abstract. · noted in this summary

Read this if you predict tertiary RNA structures from sequence or benchmark deep learning folding potentials against RNA-Puzzles.