#24 · ML / AI methods

Deep learning predicts splicing and isoform-level RNA regulation from limited data

The dominant thread is deep learning for sequence-to-function prediction in RNA processing. Two items address splicing directly: a benchmark of eleven RNA splicing predictors on thousands of synthetic recoded GFP variants with long-read measurements, reporting that deep-learning models generalise, and ASPIRE, which predicts alternative splicing PSI values from low-depth RNA-seq using a minimal gene set and embedded feature selection. circMAC extends sequence labelling to nucleotide-level miRNA binding sites on full-length circRNA isoforms. Related architectures are applied to expression and UTR biology, where base-pair resolution conservation improves cell type specific expression prediction and pretrained RNA language model embeddings with contrastive learning link 5' and 3' UTR pairs. What is new is the emphasis on learning from limited or synthetic data and on pretrained language-model representations of sequence.