#16 · ML / AI methods

Pretrained sequence models adapted to single-cell and individual-level genomic readouts

The dominant thread is adapting pretrained sequence-based models to new readouts. Scooby-2 fine-tunes AlphaGenome with a cell-specific decoder to predict single-cell RNA-seq and chromatin accessibility tracks from DNA sequence. A 3'UTR reporter screen of ~2M inserts pairs mapped regulatory elements with a context-aware deep model that prioritizes ClinVar variants. Leakage-controlled benchmarking of variant effect predictors, including a gradient-boosting consensus model, addresses how such predictors are evaluated. Representation learning over somatic missense mutations from ~50,000 tumors, along with somatic and structural variant tools (SomatiX, Verix), forms the surrounding infrastructure, alongside a xenograft reporting workflow and a multi-agent cross-cancer analysis framework. What is new is the explicit move of pretrained sequence models toward single-cell outputs.