De novo transformer modeling improves recovery of genetic cell types from sparse single-cell RNA sequencing

Single-cell RNA sequencing (scRNA-seq) simultaneously provides gene-expression profiles and genetic variants from individual cells, creating an opportunity to relate cellular phenotypes to their somatic evolutionary histories. However, delineating genetic types (GTs) from scRNA-seq remains difficult because most bases remain undetected in each cellular genome and are accompanied by false-positive and false-negative base calls. We evaluated some existing phylogenetic and imputation methods for GT inference using one simulated and two empirical tumor datasets. We found that the extreme sparsity of cell-variant (CV) alignments produced by scRNA-seq data prevented reliable GT recovery, particularly when multiple GTs were present. This led us to explore a transformer-based genotype imputation system (STICI), which has been shown to reduce sparsity in germline genomes. However, STICI was trained on germline sequence variation in the human population, making it inapplicable for imputing idiosyncratic somatic variation in tumor genomes. We tested the hypothesis that a STICI model trained on observed bases in a given CV alignment can predict missing bases in that dataset. We found that sample-specific STICI (ssSTICI) models predicted many missing bases with high imputation probabilities, and the resulting CV alignments placed the constituent cells of each GT closer together in the inferred cellular phylogenies. Therefore, sample-specific deep learning models show strong potential to reduce CV alignment sparsity, paving the way for integrative studies of sequence and expression evolution using scRNA-seq and similar technologies.