Robust RNA secondary structure prediction with a mixture of deep learning and physics-based experts

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Deep learning (DL) models struggle with out-of-distribution (OOD) generalization in RNA secondary structure prediction.
• In-distribution (ID) sequences benefit from DL models' high accuracy, but OOD sequences require different approaches for reliable predictions.

Purpose of the Study:

• To develop a novel mixture-of-experts (MoE) approach, MoEFold2D, to enhance RNA secondary structure prediction.
• To improve both in-distribution (ID) accuracy and out-of-distribution (OOD) robustness by combining DL and physics-based models.
• To implement automated ID/OOD detection without needing training data during inference.

Main Methods:

• Developed MoEFold2D, an MoE pipeline integrating DL and physics-based models.
• Trained an ensemble of DL models, each specialized on subsets of RNA structure data.
• Implemented consensus analysis of DL predictions for automated ID/OOD sequence classification.
• Utilized DL model consensus for ID predictions and physics-based models for OOD predictions.

Main Results:

• MoEFold2D effectively distinguishes between ID and OOD RNA sequences through consensus analysis.
• DL models showed consistent predictions for ID sequences, while OOD sequences yielded inconsistent predictions.
• The approach achieved accurate ID predictions by averaging consensus DL models.
• Robust OOD predictions were obtained using physics-based models.

Conclusions:

• MoEFold2D successfully combines the strengths of DL and physics-based models for RNA secondary structure prediction.
• The method circumvents generalization gaps, offering accurate ID and robust OOD predictions.
• Automated ID/OOD detection via consensus analysis is a key innovation, improving prediction reliability.