A1201
Title: Bayesian regression using zero-inflated, high-kurtosis count data featuring data augmentation
Authors: Toby Kettlewell - University of Glasgow (United Kingdom) [presenting]
Yiyi Cheng - University of Glasgow (United Kingdom)
Thomas Otto - University of Glasgow (United Kingdom)
Vincent Macaulay - University of Glasgow (United Kingdom)
Mayetri Gupta - University of Glasgow (United Kingdom)
Abstract: Single-cell transcriptomic (scRNA) data can provide substantial insight into cellular heterogeneity, but its high dimensionality, sparsity, and complex patterns of technical noise complicate analysis. Zero-inflated negative binomial (ZINB) distributions are commonly used to model scRNA data, but the two negative binomial parameters and the proportion-of-zeroes parameter combined with the data's size make regression computationally expensive. This is typically mitigated by applying methods that reduce technical noise separately from inferential analysis, but this leads to inflated confidence as there is no propagation of error between data preprocessing and analysis. Furthermore, ZINB distributions cannot always account for the data's high kurtosis. A Bayesian regression model for count data has been developed and applied to scRNA data. This model extends the ZINB distribution by adding a geometric component to reflect the large tail values, and the Bayesian framework ensures that propagation of error occurs naturally. Data augmentation allows for model fitting using an efficient Gibbs sampler, avoiding the need for tuning parameters. Regression is performed both for rows (e.g., cells) and columns (e.g., genes), ensuring that observation-level noise can be mitigated and described explicitly. Identifiability is enforced with modified normal priors defined on restricted supports.
