A1383
Title: Skip-zeros variational inference with factored experts for ultra-huge single-cell RNA data
Authors: Teppei Shimamura - Institute of Science Tokyo (Japan) [presenting]
Abstract: Single-cell RNA sequencing now produces ultra-large and highly sparse count matrices in which biological variation, batch effects, and other covariates are strongly entangled. Existing deep generative models provide powerful latent representations, but their black-box decoders often limit interpretability and make it difficult to add new covariates without redesigning the architecture. A scalable variational framework for single-cell count data is proposed that combines skip-zeros inference with a factored product-of-experts decoder. The approach models gene expression with a negative binomial likelihood and uses the zero sufficient statistic to decompose the reconstruction term into observed nonzero entries and a sampled zero component with control-variate correction. This gives an efficient stochastic estimator compatible with automatic differentiation and large-scale mini-batch training. The decoder represents the log mean as an additive combination of experts for biological latent factors, batch effects, categorical covariates, continuous covariates, and interactions, enabling structural separation of variation sources. In PBMC and pancreas benchmarks, the framework achieves clustering accuracy comparable to or better than scVI while using fewer decoder parameters and yielding interpretable gene programs and covariate effects.
