Deep generative models, particularly Variational Autoencoders (VAEs), have emerged as powerful tools for modeling the high-dimensional complexities of single-cell RNA-seq data. However, canonical deep VAEs operate as opaque black boxes, severely obscuring the precise biological mechanisms and gene regulatory networks driving cellular heterogeneity. To bridge the critical gap between high predictive power and biological interpretability, we introduce a novel, biologically grounded VAE architecture that replaces the standard dense neural network decoder with a Generalized Additive Model (GAM).

Guided by a domain-specific binary prior matrix mapping individual genes to known biological pathways or gene programs, our model enforces a structurally sparse decoder. In this framework, each latent space dimension explicitly represents a distinct, interpretable gene program. We model the expected expression of each gene as the sum of independent, non-linear contributions from its specifically associated programs. To achieve computational efficiency suitable for large-scale atlases, we parameterize these univariate GAM functions using uniform B-spline basis expansions. A structural sparsity mask mathematically guarantees that unassociated latent programs exert strictly zero influence on a gene, thereby preserving exact additive interpretability.

By enforcing structural independence through this additive decoder and a scaled Kullback-Leibler (KL) divergence loss, our framework learns a highly disentangled latent representation. This architecture empowers researchers to uncover complex, non-linear gene regulatory dynamics without sacrificing the transparency of linear models, providing a highly scalable, interpretable solution for advanced mechanistic single-cell analysis.

Co-authors: Giulio Caravagna