2025/07/08 by Ruian Shi, Taykhoom Dalal, Philip Fradkin +10 · 1 voice · 3 citations
Biochemistry, Genetics and Molecular Biology · #Benchmark (surveying) #Biology #Cartography #Computational biology #Computer science #Evolutionary biology #Function (biology) #Gene #Genetics #Geography #Messenger RNA #Philosophy #Property (philosophy) #RNA Research and Splicing #RNA and protein synthesis mechanisms #RNA modifications and cancer
paper · pdf · doi:10.1101/2025.07.05.662870
published in bioRxiv (Cold Spring Harbor Laboratory) (Cold Spring Harbor Laboratory)
openalex publication_date 2025/07/08 · openalex created_date 2025/10/10 · openalex updated_date 2026/08/01
Messenger RNA (mRNA) is central in gene expression, and its half-life, localization, and translation efficiency drive phenotypic diversity in eukaryotic cells. While supervised learning has widely been used to study the mRNA regulatory code, self-supervised foundation models support a wider range of transfer learning tasks. However, the dearth and homogeneity of standardized benchmarks limit efforts to pinpoint the strengths of various models. Here, we present mRNABench, a comprehensive benchmarking suite for mature mRNA biology that evaluates the representational quality of mature mRNA embeddings from self-supervised nucleotide foundation models. We curate ten datasets and 59 prediction tasks that broadly capture salient properties of mature mRNA, and assess the performance of 18 families of nucleotide foundation models for a total of 135K experiments. Using these experiments, we study parameter scaling, compositional generalization from learned biological features, and correlations between sequence compressibility and performance. We identify synergies between two self-supervised learning objectives, and pre-train a new Mamba-based model that achieves state-of-the-art performance using 700x fewer parameters. mRNABench can be found at: https://github.com/morrislab/mRNABench.