#15 · ML / AI methods

Sequence models extend regulatory and splicing prediction to personal genomes

The dominant thread is supervised sequence-to-function modeling applied to RNA regulation, with performance judged against established baselines or on individual-level genomes. m6A-FORM is a foundation model pretrained on 22.5M MeRIP-seq regions that predicts reader binding, YTHDF2-linked decay and tissue-conserved m6A sites, extending pretraining to epitranscriptomic function. SPLAIRE, a dilated CNN trained on airway epithelial RNA and genotype data, reports outperforming leading models at splice site identification and usage. SAGE-net provides a scalable framework for training and evaluating sequence-to-function models on personal genomes. What is new is the joint use of genotype and individual-level data in training and evaluation, rather than architecture alone. Transfer across contexts is addressed only partially: m6A-FORM reports tissue-conserved sites, while the other two reports do not describe cross-context transfer results.