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

mRNABench benchmarks mature mRNA property prediction on human transcripts using 13 datasets and 81 tasks, evaluating 29 nucleotide foundation model families across 750k experiments.

• Why it matters: Standardized benchmarks are lacking, making it hard to determine which self-supervised foundation models best capture mature mRNA regulatory information for transfer learning.
• What they did: The suite curates 13 datasets and 81 prediction tasks for human transcripts and runs 750k experiments across 29 families of self-supervised nucleotide foundation models, analyzing scaling, compressibility-performance links, splitting, and training objectives.
• The result: The resulting benchmark enables systematic comparison of model representational quality for mature mRNA embeddings and clarifies how objective choice and data handling affect prediction performance.

AI summary of the abstract · may contain errors · not medical advice · read the paper.