The corrdeconv preprint examines how disease can cause algorithms to mistake changes in cell state for changes in cell numbers
The corrdeconv preprint examines how disease can cause algorithms to mistake changes in cell state for changes in cell numbers
On 24 October, the authors of the corrdeconv preprint compared estimates of sample composition with direct cell counts from the same people. Their analysis shows that accurate estimates for individual samples can coexist with errors in the estimated difference between groups.
Deconvolution uses a tissue's combined RNA signal, a molecular readout of gene activity, to estimate the proportions of different cell types, such as immune T-cells and B-cells. Researchers use it to compare blood or tissue from healthy people and people with a disease. In these comparisons, they want to know whether one group actually has more or fewer cells of a particular type. Standard validation usually asks a different question: whether the estimate agrees with cell counts for each individual sample.
Disease can change both cell proportions and gene activity. Because the algorithm compares the RNA signal with a fixed reference of cell profiles, it can interpret altered gene activity as an increase or decrease in a cell type's proportion. In the authors' model, the estimates preserved the ranking of samples within a group and remained highly correlated with cell counts, even as the estimated difference between groups became biased.
In simulated mixtures, the authors kept cell proportions unchanged but replaced the profile of CD4 T-cells, a type of T-cell, with a profile from systemic lupus erythematosus. MuSiC, a deconvolution method, reduced its mean estimate of this proportion from 41,1% to 17,6%; the true proportion remained around 43,9%. The estimates preserved the sample ranking almost entirely: their correlation with the known composition reached 0,93.
“Claims about cell numbers should rest on cell counts.”
Across 21 participant cohorts, the authors compared 28 methods with direct cell counts from 4 622 people. The counts contradicted 1 119 of 2 660 estimated differences between groups, with discrepancies in the direction of change, statistical significance, or effect size. Of these errors, 96% passed standard validation based on correlation or agreement with another method.
This bias also changed the results of a reanalysis of R4RA, a randomized trial of two drugs for rheumatoid arthritis. RNA-based estimates from joint tissue identified a subgroup with low B-cell numbers and high numbers of myeloid cells, which are cells of the innate immune system. In this subgroup, tocilizumab produced a better response than rituximab. When the corrdeconv authors used histology and immunohistochemistry, examining tissue under a microscope with protein markers, 13 of the 20 patients left this subgroup and 13 others entered it. The statistical association between treatment effect and B-cell subgroup disappeared.
corrdeconv selects a correction according to where the estimates fail in the intended analysis: in differences between groups or in gene lists. The authors selected the correction in cohorts with both RNA data and direct cell counts, then fixed it before testing a new cohort. They report that corrdeconv ranked first on predefined task sets under every tested scenario for the availability of cell counts. The method calls for validating the specific group comparison that motivated the analysis against measured data.