Machine learning and language models for RNA structure prediction: Progress and perspectives

Area of Science:

• Computational Biology
• Molecular Biology
• Bioinformatics

Background:

• The function of all RNA molecules is intrinsically linked to their three-dimensional structure.
• A significant gap exists between the vast number of known RNA sequences and experimentally determined structures.
• Accurate RNA structure prediction is essential for understanding RNA function across diverse biological contexts.

Purpose of the Study:

• To provide a comprehensive review of the current state-of-the-art in computational RNA structure prediction.
• To discuss the evolution of methods, from thermodynamic models to advanced deep learning and language models.
• To highlight key datasets, benchmarks, and performance metrics in the field.

Main Methods:

• Review of computational approaches including thermodynamic free energy minimization, supervised deep learning, and self-supervised RNA language models.
• Analysis of established training datasets and community-accepted benchmarks for evaluating prediction accuracy.
• Discussion of emerging methods like generative models and integration of multimodal data.

Main Results:

• Self-supervised RNA language models represent the current state-of-the-art, significantly improving structure prediction accuracy.
• Progress has been made in developing robust benchmarks and identifying key datasets for model training and evaluation.
• The field is rapidly evolving with new data modalities and generative approaches.

Conclusions:

• Computational RNA structure prediction has seen remarkable progress, driven by advances in machine learning.
• Integrating diverse data types and exploring generative models offer promising avenues for future research.
• Addressing challenges in generalization and noncanonical interactions remains critical for advancing the field.