B1466
Title: Scalable count-based models for unsupervised detection of spatially variable genes
Authors: Boyi Guo - Johns Hopkins University (United States) [presenting]
Lukas Weber - Johns Hopkins University (United States)
Stephanie Hicks - Johns Hopkins University (United States)
Abstract: Unsupervised feature selection methods are well sought in analysing high-dimensional genomics data. The recent development of spatially resolved technologies poses novel computational challenges, including identifying and ranking genes that vary in a non-random way across a 2D space, commonly referred to as spatially variable genes (SVG). While many SVG methods have been proposed to model continuous normalized gene expression data, they are susceptible to any bias attributed to normalization strategies and vulnerable to the violation of isotropic assumption, leading to erroneous findings. Available count-based SVG methods are theoretically sound but practically infeasible due to their computationally prohibitive model fitting. To address these challenges, a scalable approach is proposed that extends the generalized geo-additive framework to the analysis of spatially resolved transcriptomics data. The method identifies genes whose expression exhibits spatial patterns and accounts for effect differences across pre-defined spatial domains when applicable. In addition, the method provides flexibility in modelling raw gene expression data, accommodating multiple count-based distributions, including Poisson, Negative Binomial and Tweedie. In simulation studies and real-world applications, it is demonstrated that the proposed count-based models outperform the state-of-the-art SVG methods.