Current Bioinformatics Seminar
Time: 11AM Tuesdays.
Venue: Davis Auditorium and Online
8 September 2026
This is a WEHI only event.Assumptions under the microscope: interrogating differential expression methods for imaging-based spatial transcriptomics
Raiha BrowningWEHI
The methods most often used for differential expression (DE) analysis of spatial transcriptomics data were developed for bulk and single-cell RNA-sequencing, and implicitly carry assumptions about the data and its distribution, as do the normalisation methods applied before them. Imaging-based platforms can violate these assumptions: panels are targeted, measuring only hundreds to thousands of genes, and gene counts are low, typically zero, one or two per cell. For composition normalisation, for instance, the premise that most genes are not DE fails because the panel is deliberately chosen for genes of interest. In this work, we establish when conventional assumptions hold and give guidance for model and normalisation choice in pseudobulk DE of imaging-based spatial data. We probe these assumptions by simulating pseudobulk counts under negative binomial, beta-binomial and lognormal data generating models across null and DE scenarios. We compare count-based, binomial-family and log-scale testing frameworks in terms of type I error, false discovery rate and power, and consider several normalisation strategies: composition-based scaling (TMM), total counts, and cell-number normalisation. We then extend the comparison to real imaging-based datasets. The choice of framework and normalisation matters: a poor match can leave tests underpowered or inflate false discovery rates, driven by mismatch between the generating distribution and testing framework, or when transcript counts per cell vary systematically between groups. Count-based (edgeR) and log-scale (limma-voom) frameworks perform well in terms of power and false discoveries respectively; TMM proves robust despite assumption violations, while total counts scaling offers a lighter-touch, more conservative alternative.