[An RNA Scoring Function for Tertiary Structure Prediction Based on Multi-layer Neural Networks].

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning

Background:

• Accurate scoring functions are crucial for *ab initio* RNA tertiary structure prediction.
• Traditional scoring functions often lack flexibility and face challenges with reference state selection.

Purpose of the Study:

• To investigate the efficacy of a machine learning-based approach as a scoring function for RNA tertiary structure prediction.
• To develop and evaluate neural network models for scoring RNA structural candidates.

Main Methods:

• Constructed and trained two multi-layer neural networks: one coarse-grained and one all-atom.
• Developed an RNA database with 322 training, 70 validation, and 70 testing RNAs, each with 300 decoys from molecular dynamics simulations.
• Employed an early-stop strategy based on validation set loss for network optimization.

Main Results:

• The machine learning scoring functions demonstrated superior performance compared to traditional methods.
• The developed neural networks consistently outperformed a recent knowledge-based all-atom potential in evaluating RNA structural candidates.
• The approach offers greater flexibility in feature incorporation and avoids the reference state problem.

Conclusions:

• Machine learning-based scoring functions represent a powerful and flexible advancement for *ab initio* RNA tertiary structure prediction.
• The developed neural networks provide a more accurate and robust method for evaluating RNA structural models.