Python Encoders for Archetypal Convex Hulls (PEACH): PyTorch-Based Archetypal Analysis

Abstract

Samples of closely related cells often contain substantial cell state heterogeneity, which traditional clustering-based analyses struggle to de-convolve. Archetype analysis is an alternative analysis approach that identifies a minimal convex hull enclosing all data points in a dimensionally-reduced space, such as PCA space. The points of this convex manifold, or hull, each represent theoretical extremal states specialized in a subset of functions that are mutually exclusive to the extremal states found at each other vertex. The transition between two points represents a Pareto front of tradeoffs between potential cell states. Cell distances from each archetype and direction between archetypes then carry interpretable biological information revealing features of resource tradeoffs in cell state regulation. The archetypal points themselves are theoretical pure extremal states rarely accessed by real cells, which instead display a combination of archetype weights/mixture coefficients representing a mixture of specializations. By characterizing the relationships between gene expression and archetype distance, archetypal analysis can identify the phenotypic features of theoretically pure specialist states, the phenotypic tradeoffs of real cells, and the boundaries of accessible cell states in the analyzed sample. Here we present Python Encoders for Archetypal Convex Hulls (PEACH), a new PyTorch-based archetype analysis package compatible with the scVerse ecosystem that recapitulates and extends features available in previous archetype analysis packages. This includes a hyperparameter search function to identify the best k archetypes that fit an input dataset. PEACH uses a deliberately constrained autoencoder architecture to directly learn an archetypal latent space, replacing previous alternating least squares methods with a highly performant, GPU-accelerated archetype analysis method that is scalable to modern scRNAseq datasets of hundreds of thousands of cells. Crucially, PEACH includes automated hyperparameter search with cross-validation to identify optimal archetype and initialization configurations, enabling users to implement archetype analysis without extensive manual hyperparameter optimization and testing.