Detecting a change in cell-type composition between conditions is harder than it looks. Relative abundances are coupled, sample-to-sample variability is high, and the naive comparison produces misleading answers.

LODopt models per-sample, per-cluster cell counts with a generalized linear mixed-effects model, and pairs it with sample-specific normalization factors chosen to minimize total log-odds variance. The result is inference about absolute compositional change rather than an artifact of the denominator.

What it does

• Estimates the difference in log-odds of cluster or cell-type membership between conditions
• Optimal, data-driven normalization
• Maintains correct Type I error and minimizes parameter bias against existing methods, at comparable power, validated on simulated data
• Works with Seurat and SummarizedExperiment objects out of the box

Thomas, Agrawal, Gill, Traglia (2026). LODopt v1.1.0. Zenodo. doi:10.5281/zenodo.19239199. The methodology was previously applied in Koutsodendris et al., Nature Aging (2023).

Repository · Documentation