Scooby-2: fine-tuning AlphaGenome for single-cell genomic track prediction

Scooby-2 fine-tunes AlphaGenome with a cell-specific decoder to predict single-cell RNA-seq coverage and chromatin-accessibility tracks, reaching Pearson 0.56 on NeurIPS bone marrow.

• Why it matters: Existing AlphaGenome-style models mainly predict tracks for predefined, mostly bulk-derived cell types, leaving single-cell genomic track prediction underdeveloped.
• What they did: Scooby-2 adds a cell-specific decoder and uses parameter-efficient fine-tuning to predict single-cell RNA-seq coverage and chromatin-accessibility from DNA sequence, evaluated on the NeurIPS bone marrow single-cell dataset.
• The result: On cell-type-specific gene expression differences, Scooby-2 achieves Pearson correlation 0.56 versus 0.53 for Scooby, and improves fine-mapped eQTL concordance from 0.37 to 0.47 Spearman’s rank correlation in OneK1K.