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

Abstract

Sequence-to-function models predict genomic tracks from DNA sequence. AlphaGenome has been shown to outperform previous models, including Borzoi, on most track-prediction and variant-effect benchmarks. However, like its predecessors, it predicts tracks only for predefined, mostly bulk-derived cell types and tissues. While Scooby has extended Borzoi to single-cell resolution, a single-cell adaptation of AlphaGenome is still lacking. We present Scooby-2, which combines AlphaGenome with a cell-specific decoder to predict single-cell RNA-seq coverage and chromatin-accessibility genomic tracks from DNA sequence. The package provides the implementation, the parameter-efficient fine-tuning setup, the training pipeline and inference scripts. When evaluated on the NeurIPS bone marrow single-cell dataset, Scooby-2 achieved a Pearson correlation of 0.56 when predicting cell-type-specific gene expression differences, compared with 0.53 for Scooby. When using the distilled AlphaGenome backbone, variant effect prediction concordance of fine-mapped eQTLs from the OneK1K cohort improved from 0.37 to 0.47 Spearman's rank correlation across all associations, an improvement consistent with that of AlphaGenome over Borzoi itself.