PRISM-GEP: Viewing Single-Cell Expression through the Lens of Topic Modeling
Yanir Buznah, Tal Ishon, Uri ShahamAbstract
Motivation
Gene-expression programs (GEPs) are the co-regulated gene modules whose coordinated activity encodes biological processes. Recovering them from single-cell RNA-seq is hard because regulation is many-to-many. A cell can run several programs at once, and a gene can take part in several. Methods that resolve programs along a trajectory break this structure by assigning each gene to a single program.
Results
PRISM-GEP discovers overlapping gene-expression programs with Latent Dirichlet Allocation, giving it a prior estimated from gene co-expression in place of the flat default it normally assumes. No reference dataset or pathway database is required. Across 15 human and mouse datasets, PRISM-GEP matches specialized factorizations such as cNMF and scHPF on Gene Ontology Biological Process metrics. Over the 43 rankable dataset-by-metric entries its mean rank is level with the best, and unlike every specialized method it never ranks last. The same co-expression geometry orders the genes within each program, recovering developmental cascades on par with GeneTrajectory.
Availability and Implementation
The implementation of our methods, including all code, datasets, and experimental workflows, is available in Python at https://github.com/shaham-lab/PRISM-GEP. The version used in this paper is archived at https://doi.org/10.5281/zenodo.22659056.
Contact
Uri Shaham, Bar-Ilan University.
Supplementary information
Supplementary data are available at Bioinformatics online.