Neutrophil annotation is not just a detection count: one fixed pipeline, two QC rules, and nine patient pairs that stop agreeing
Synopsis
How many neutrophils a single-cell pipeline labels is not the same as how well it estimates sample composition. Using protein measurements and physical sorting labels from four public cohorts as independent references, a fixed annotation pipeline and two preset QC rules were tested: under frozen QC, SingleR recovered 99.10% of paired whole-blood protein-positive cells and estimated a neutrophil-fraction decrease in all nine patient pairs; under the alternative QC rule the same pipeline estimated a decrease in only three pairs, with a mean paired change of +0.12 percentage points. Most of the shift came from which cells were filtered out, not from the annotator.
Interpretation
Whether an annotation pipeline is useful depends not only on how many target cells it labels, but also on how many negative controls it mislabels and how many cells defining the denominator it retains. Higher neutrophil detection has usually been read as an analytical improvement, but a detection gain alone does not show that the labels agree with independent measurements. Four public cohorts supplied 629,185 source-retained cells across 99 computational partitions, and the protein references were built from measured antibody counts without RNA values, RNA-predicted labels or patient outcomes.
On paired whole-blood samples, the direction of patient-level change from one annotation pipeline depends on the QC rule applied. Under frozen QC, pruned SingleR estimated a neutrophil-fraction decrease in all nine acute-to-convalescent pairs, with a mean change of −17.20 percentage points and a descriptive patient-bootstrap 95% interval of −20.30 to −14.35. Applying the unchanged alternative QC rule to the same RNA data left SingleR estimating a decrease in only three of nine pairs, with a mean paired change of +0.12 percentage points (interval −0.45 to +0.66); holding source-population predictions fixed and swapping only the QC mask shifted the mean paired estimate by +18.77 percentage points, while recalculating query-dependent pruning added +0.25.
Recovery has to be checked against several independent references, because each reference answers a different question. Beyond the whole-blood protein panel, physical sorting and tissue protein measurements supplied separate positive and negative controls. Among 103 physically sorted mature neutrophils, pruned SingleR recovered 99 (96.12%) and M4 recovered 50 (48.54%); among 68 plaque protein-positive cells, SingleR recovered 42 (61.76%) and M4 recovered 16 (23.53%), neither CellTypist model recovered a positive, and all four procedures made zero neutrophil calls among 6,267 retained negative controls.
Perspective
The conclusions rest on a reanalysis of four public human single-cell cohorts; the protein references define operational high-confidence compartments, the whole-blood panel lacks CD66b and isotype controls, and the plaque positive compartment holds four to 29 cells per patient. The setting is neutrophil composition work and other rare or technically sensitive populations whose abundance is inferred from a selected single-cell sample.
The protein references do not adjudicate every cell, and letting all unadjudicated cells take either identity leaves all nine whole-sample paired bounds spanning zero; these are reference-coverage bounds, not sampling confidence intervals. Six of the nine clinical pairs changed processing batch and only three are same-batch; physical-sorting aliases cannot be assumed to identify the same people across lineages; and the ABC negative-control discordance sits mainly in 230 NK-sorted cells, which computational doublet restriction generally did not resolve.