Methods questions on `metadata/pseudobulk_differential_expression`
Hi Tahoe team,
Thanks for releasing the table metadata/pseudobulk_differential_expression - it is exactly the layer we were about to compute ourselves. Before we build on it we would like to understand how it was produced, since we could not find it described in the v1 preprint.
The schema reads unambiguously as a DESeq2 results table, so these 4 questions assume that and ask for the specifics:
1. Design formula, and which samples enter each fit. plate is part of the contrast key, so each contrast appears to be computed within a plate and cell line. Does the model include a plate or batch term, and is a separate fit run per (cell line, plate), or one larger fit from which contrasts are extracted?
2. Unit of observation, and how dispersion was estimated. Grouping the released cell-level metadata by (compound, dose, cell line) gives 56,879 conditions, of which 49,165 (86.4%) appear on a single plate; 7,464 appear on two, 150 on three, and 100 on all fourteen. So for the large majority there is no second well to serve as a replicate, yet every contrast carries an lfcSE and a padj. What is a column of the count matrix that goes into the fit - individual cells, random pools of cells within a condition, or one pseudobulk profile per (sample, cell line)? And in the pseudobulk case, where does the
dispersion come from when the treated side has a single sample: a trend fitted across conditions, the replicated vehicle samples, shrinkage, or something else?
3. Composition of the control pool. n_cells_ctrl is consistently larger than n_cells_trt (e.g. 1,378 vs 4,862), so controls are pooled. Since each plate holds 96 samples, is the pool every vehicle sample on that plate for that cell line, one designated vehicle sample, or something wider? And are cells pooled across vehicle samples before fitting, or kept as separate columns in the count matrix?
4. Filters applied to this table. The cell-level filters are documented (full,pass_filter, the 50-cell floor per cell line–drug condition, etc. ). Which of these does this table inherit, and are there additional filters applied at the contrast level? We ask because counting (condition, plate) pairs in the released metadata gives 65,943, while the table contains 65,218 contrasts (4,089,820,780 rows / 62,710 genes) - a 1.1% difference we would rather account for than guess at.
Again, thank you for the great effort! Happy to contribute a short methods note back to the dataset card once we understand it.