#33 · Epigenomics

Fine-tuned genomic models detect cryptic splicing where zero-shot AlphaGenome fails

The dominant thread is extracting regulatory and splicing information from genomic sequence and sequencing-derived readouts using deep learning. Two items test genomic foundation models in ALS and TDP-43 pathology. Zero-shot AlphaGenome shows weak QTL effect recovery and wrong UNC13A splicing direction in one study, while in another frozen AlphaGenome fails to detect TDP-43 cryptic exons, full fine-tuning generalizes, and a CLDN11 biomarker candidate emerges. SpliceAI2 aims to unify splice-site, junction and transcript-usage prediction end-to-end, and Seal predicts spatiotemporal expression and variant effects in the developing human brain. Haplotype-resolved Hi-C across 354 human and 12 ape haplotypes links structural variants to recurrent chromatin contact divergence. What is new is adaptation of models to rare events and haplotype-resolved, donor-specific readouts rather than zero-shot prediction alone.