Deep learning model for predicting mRNA half-life based on 3'UTR sequences

mRNA stability, quantified by half-life, is central to gene expression homeostasis and post-transcriptional regulation. Predicting mRNA half-life from 3' untranslated region (3'UTR) sequences alone, however, remains challenging, particularly when both accuracy and biological interpretability are required. Here, we present a sequence-driven framework that combines a pre-trained RNA language model (RNA-FM) with a Transformer backbone to predict mRNA half-lives of Saccharomyces cerevisiae transcripts from their 3'UTRs. The model is evaluated in a within-gene setting, focusing on novel isoforms of known genes. On the independent hold-out test set, the model achieves an RMSE of 8.76 min, an MAE of 5.95 min, and an R² of 0.545, while 5-fold cross-validation on the train_val pool yields consistent performance estimates. Predictions show close agreement with experimental measurements, exhibiting minimal systematic bias (∼0.70 min) and good calibration across most of the value range. Positional masking and in-silico mutagenesis identify GU-enriched short motifs (e.g., UUGUAU, AUGCA, GUGUA) as globally destabilizing elements, with motif removal increasing predicted half-lives by 0.3-0.8 min (Δ = mut-base,FDR<0.05)on average, particularly near the 3'end. In contrast, motifs such as UUUAUG, GAGGU, ACCAC, and CACCA display stabilizing effects. Despite moderate individual effect sizes, these motifs exhibit robust positional dependencies. Overall, this framework integrates predictive performance with mechanistic interpretability, providing a practical tool for studying mRNA homeostasis and guiding mRNA design.