Benchmarking pre-trained genomic language models for RNA sequence-related predictive applications

Area of Science:

• Computational biology
• Genomics
• Molecular biology

Background:

• RNA is crucial for cellular functions, necessitating advanced computational methods for sequence analysis.
• Pre-trained genomic language models (gLMs) offer versatile tools for RNA-related prediction tasks.
• Comprehensive evaluations of gLMs for RNA analysis are currently limited.

Purpose of the Study:

• To benchmark eleven gLMs against task-specific methods for RNA sequence analysis.
• To evaluate gLM performance across four key RNA processes: classification, m6A prediction, splicing, and translation efficiency.
• To identify factors influencing gLM performance and provide recommendations for model selection.

Main Methods:

• Benchmarking eleven gLMs and task-specific methods on four RNA prediction tasks.
• Systematic profiling of pre-training datasets, input context lengths, and tokenization schemes.
• Comparative analysis of performance based on data availability and balance.

Main Results:

• Outstanding gLM performance is achieved by integrating biological context with data and algorithms, not just scale.
• gLMs outperform task-specific methods in scenarios with limited or imbalanced training data.
• Task-specific methods can be computationally efficient and achieve comparable results in certain contexts.

Conclusions:

• gLMs hold significant promise for advancing RNA sequence analysis and biomedical research.
• Optimizing gLM performance requires careful consideration of data, algorithms, and biological context.
• Targeted recommendations for gLM selection can guide researchers in diverse applications.