scFLAME: a unified generative model for interpretable clustering, hierarchical structure discovery and marker-gene identification in single-cell RNA-seq data

Abstract

Identifying cell types from single-cell RNA sequencing (scRNA-seq) data typically requires several separate and often uninterpretable steps: dimensionality reduction, batch-correction, clustering, marker-gene identification and the discovery of finer-grained structure. Here we introduce scFLAME (single-cell Factor Latent Analysis with Mixture Embeddings), a probabilistic generative model that unifies these tasks: a negative binomial factor analysis of the raw counts - which can be adjusted for batch - is coupled to a Gaussian mixture prior over the latent space, learning the embedding and clustering jointly, while a shared linear decoder provides cluster-specific marker genes directly from the fitted model, and a merging procedure recovers a probabilistic hierarchy of finer-grained partitions. On simulated and real data, scFLAME matches or exceeds state-of-the-art clustering accuracy, is robust across sequencing platforms, and scales near-linearly to hundreds of thousands of cells. scFLAME thus replaces a chain of separate tools with a single, interpretable model for single-cell analysis.