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

Single-cell RNA sequencing (scRNA-seq) can simultaneously capture gene expression and genetic variants in individual cells, offering a way to connect cellular phenotypes with somatic evolutionary history. However, extracting genetic types (GTs) from such data is challenging because most variant positions are missing in any given cell, and observed base calls contain significant error rates. The authors evaluated existing phylogenetic and imputation methods on one simulated and two empirical tumor datasets, finding that extreme sparsity prevented reliable GT recovery when multiple GTs were present.

To address this, they adapted the STICI transformer architecture to train a separate model de novo on each sparse cell-variant (CV) matrix. These data-specific models imputed millions of missing bases, greatly reducing matrix sparsity. Phylogenetic analyses of the imputed CV matrices showed substantially improved recovery of GTs in both simulated and empirical datasets, and in the empirical data revealed finer-scale genetic structure within previously reported GTs that had been missed by existing methods.

The results suggest that highly sparse scRNA-seq datasets hold substantially more recoverable lineage information than previously recognized, and that de novo transformer modeling is an effective approach for accessing it. However, sequencing errors persist and still limit reconstruction of cellular lineage structure, indicating that methodological improvements are needed before expression phenotypes can be reliably examined in the context of cellular evolutionary relationships.