mRNABench: A curated benchmark for mature mRNA property and function prediction

Area of Science:

• Computational Biology
• Molecular Biology
• Machine Learning

Background:

• Messenger RNA (mRNA) plays a critical role in gene expression, influencing cellular phenotypes through its half-life, localization, and translation efficiency.
• Self-supervised foundation models offer advanced capabilities for studying the mRNA regulatory code, surpassing limitations of traditional supervised learning.
• The lack of standardized benchmarks hinders the evaluation and comparison of different self-supervised models in mRNA biology.

Purpose of the Study:

• To introduce mRNABench, a comprehensive benchmarking suite for assessing the performance of self-supervised nucleotide foundation models in mature mRNA biology.
• To evaluate the representational quality of mRNA embeddings generated by these models across diverse biological tasks.
• To facilitate the identification of superior models and training methodologies for mRNA research.

Main Methods:

• Curated ten diverse datasets and 59 prediction tasks covering key properties of mature mRNA.
• Assessed the performance of 18 families of nucleotide foundation models, conducting a total of 135,000 experiments.
• Investigated parameter scaling, compositional generalization, and the relationship between sequence compressibility and model performance.

Main Results:

• Identified synergistic effects between specific self-supervised learning objectives.
• Developed a novel Mamba-based model that achieves state-of-the-art performance with significantly reduced parameters (700x fewer).
• Demonstrated the effectiveness of mRNABench in evaluating and comparing foundation models for mRNA biology.

Conclusions:

• mRNABench provides a robust framework for benchmarking self-supervised models in mRNA biology.
• The study highlights the potential of efficient, large-scale models like the new Mamba-based architecture.
• This resource will accelerate the development and application of AI in understanding mRNA regulation and function.