trRosettaRNA

trRosettaRNA predicts RNA three-dimensional (3D) structures from sequence by using a transformer network to infer 1D and 2D geometries and then refines models via energy minimization.

Key Features:

• Transformer network utilization: Employs a transformer-based deep learning model to predict 1D and 2D geometries of RNA molecules.
• 1D and 2D geometry prediction: Infers linear sequence features and base-pairing interactions that provide foundational geometrical data for 3D modeling.
• Energy minimization for 3D folding: Applies energy minimization algorithms to fold predicted geometries into refined 3D structures.
• Benchmark performance: Demonstrated competitive results in benchmark evaluations, with superior Z-score performance in Root-Mean-Square Deviation (RMSD) comparisons versus other deep learning methods.

Scientific Applications:

• CASP15 evaluation: Evaluated during the 15th Critical Assessment of Structure Prediction (CASP15) and achieved competitive results compared to top human predictions for natural RNAs.
• RNA-Puzzles benchmarking: Participated in RNA-Puzzles assessments and showed competitive performance on blind prediction targets.
• Natural RNA structure prediction: Used to predict 3D structures of natural RNAs with competitive accuracy in RMSD-based comparisons.
• Synthetic RNA limitation assessment: Automated predictions are less accurate for synthetic RNAs, indicating limitations in current automated modeling for designed sequences.

Methodology:

trRosettaRNA predicts 1D and 2D geometries (including base-pairing interactions) using a transformer network and then applies energy minimization algorithms to fold and refine the 3D structure.