A1354
Title: Scalable stochastic gradient variational inference framework for microbiome differential abundance analysis
Authors: Hanna Jankowski - York University (Canada)
Kevin McGregor - University of Manitoba (Canada)
Saurabh Panchasara - York University (Canada) [presenting]
Abstract: Differential abundance analysis (DAA) attempts to answers a core question in microbiome studies: which microbial features change reliably between conditions, or with an exposure? DAA is challenging because microbiome data are high-dimensional, compositional, zero-inflated, and skewed in additive log-ratio transformation. The proposed Bayesian zero inflated factor analysis logistic skew normal multinomial (ZIFA-LSNM) model, with skew-normal priors on latent factors, addresses these issues effectively and can be considered as an effective tool box to perform DAA. However, standard inference via coordinate ascent variational inference (CAVI) or fixed form variational Bayes (FFVB) suffers major computational and analytical challenges in high-dimensional settings. To overcome this, a scalable inference framework based on stochastic gradient variational inference (SGVI) is developed that avoids rigid closed-form analytical updates. Across available models such as ZIPPCA-LPNM and ZIFA-LSNM, SGVI achieves comparable optimized fits while substantially reducing runtime. This efficiency enables fast, scalable estimation of latent microbial structures, providing an accessible tool for large-scale DAA.
