The paper introduces scooby, a novel computational framework that utilizes deep learning to model multimodal genomic profiles-specifically single-cell RNA-sequencing (scRNA-seq) coverage and chromatin accessibility (scATAC-seq insertions)-directly from DNA sequence at single-cell resolution. scooby builds upon the existing bulk multiomics model Borzoi by integrating a cell-specific decoder and employing parameter-efficient fine-tuning (LoRA) to adapt it for single-cell data, effectively capturing cellular heterogeneity. The authors demonstrate that scooby accurately predicts cell-state-specific gene expression, substantially outperforming previous models, and successfully identifies lineage-specific transcription factor (TF) activity through in silico motif mutation experiments. Furthermore, the framework proves valuable in dissecting cell-type-specific variant effects (eQTLs), which are often masked in bulk-level genomic studies. Overall, scooby is presented as a scalable and accurate tool for uncovering the genetic mechanisms underlying gene regulation in individual cells.
References:
