Emerging frontiers for AI in single-cell genomics

Abstract

We are just over a decade into the single-cell revolution in biology, which has been coupled to the growing use of artificial intelligence (AI) in single-cell genomics. Most early single-cell RNA embedding models and subsequent “foundation models”—representational models pretrained on large single-cell datasets—did not gain wide traction among experimental biologists. However, as massive single-cell RNA-sequencing (scRNA-seq) atlas efforts ramp up, there is a clear rationale for exploiting AI models to interpret new datasets in the context of vast existing data resources. Meanwhile, the advent of single-cell epigenetic/multiomic assays and Perturb-seq technologies presents new opportunities for AI modeling. Here, we review AI embedding and foundation models for scRNA-seq and single-cell assay for transposase-accessible chromatin using sequencing (scATAC-seq), extensive machine-learning efforts to exploit single-cell multiome in regulatory genomics, and AI strategies to learn “virtual cell” models from Perturb-seq, ending with a description of current limitations of single-cell AI models and an outlook for future advances.