lociPARSE: a locality-aware invariant point attention model for scoring RNA 3D structures

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Assessing 3D RNA structural model accuracy without experimental data is crucial for model selection and conformational sampling.
• Existing knowledge-based potentials and machine learning methods struggle with high-fidelity RNA structure scoring.

Purpose of the Study:

• To introduce lociPARSE, a novel deep learning architecture for accurate scoring of RNA 3D structures.
• To develop a method that estimates local and global structural accuracy in a superposition-free manner.

Main Methods:

• Developed lociPARSE, a locality-aware invariant point attention architecture.
• Implemented superposition-free estimation of Local Distance Difference Test (lDDT) scores for nucleotide-level accuracy.
• Aggregated local accuracy information to predict global structural accuracy.

Main Results:

• lociPARSE significantly outperforms conventional statistical potentials and existing machine learning methods.
• The model demonstrated superior performance across multiple assessment metrics on various datasets, including CASP15.
• lociPARSE accurately captures local atomic environment accuracy and predicts global structural accuracy.

Conclusions:

• lociPARSE represents a significant advancement in computational RNA structure evaluation.
• The method provides a reliable and accurate way to assess 3D RNA models without experimental structures.
• lociPARSE is freely available, facilitating its use in structural biology research.