A single-cell reference can flip the verdict: CD8 T-cell age link reverses in all 80 six-type references

When a single-cell reference is used to infer cell proportions from bulk transcriptomes, the reference carries its own error into downstream association analyses, enough to reverse a sign or make a null trend significant. Earlier benchmarks compared proportion accuracy across references, sizes, platforms and state resolution, but did not ask whether the reference changes the association finally drawn, nor how a reference's cells are spread over its donors. In 981 OneK1K donors the labelled CD8 T-cell age slope was −0.072 points per year (t = −2.6), while all 80 six-type references gave positive slopes of +0.04 to +0.24 per year, 96% with |t| > 1.96; the CD4 slope was inflated 5- to 8-fold, turning a non-significant t = −1.6 trend significant in every reference.

The error is reference-borne rather than fitter-borne, yet the direction of the islet error is set by the estimator class, so one reference yields different conclusions under different estimators. The CD8 reversal and CD4 inflation recurred across weighted MuSiC, MuSiC NNLS, the CIBERSORT-type ν-SVR and the InstaPrism algorithm. Across 117 islet reference sets, three regression-type estimators inflated the alpha-cell slope 1.11-, 1.43- and 2.06-fold and called it significant in 117, 95 and 46 sets, whereas InstaPrism attenuated it to 0.59-fold and called it significant in 6.

Reference error splits into a systematic part that moves with the covariate, removable only by a reference resolving the shifting states, and a random part measurable by a within-reference donor bootstrap and foldable into the interval. Confidence intervals computed from bulk samples ignore the uncertainty contributed by which reference donors were drawn, which in small references rivals the bulk-side standard error. With three reference donors the reference-side spread of the CD8 slope was 1.9 times the bulk-side standard error and 1.3 times for NK cells, falling to 0.5–0.8 times at 15–30 donors; the donor bootstrap SD was 0.77–1.49 times the SD across eight independent donor sets (median 1.07).

Donor sampling is the removable part: equal-donor averaging and cell pooling are two limits of one precision-weighted estimator whose threshold n* = 1/ρ is estimable from the reference itself. Public references are often thin, with 60% of 226 CELLxGENE datasets containing a cell type whose median donor contributes fewer than 10 cells. In 336 three-donor PBMC references n* was 9–17 cells by type; inserting the precision weight into unchanged MuSiC 1.0.0 lowered weighted MuSiC's 10:1:1 allocation penalty at 60 cells per type from +0.591 to +0.322 points, a 45% reduction against a predicted 48%.

The conclusions rest on two pseudobulk cohorts that share platform and annotation with their references, 981 OneK1K PBMC donors and 48 islet donors, together with a census of 226 public references and refits on frozen grids; in the real bulk cohort GSE50244 the reference and bulk profiles did not match, leaving beta cells unidentifiable before donor design could matter.

Both cohorts are pseudobulks sharing platform and annotation with their references, which understates systematic bias, and the OneK1K truth inherits Azimuth label noise. The model did not hold when every type had two-cell donors and does not explain the over-estimation of rare states or the source of the islet residual. The SVR class used EpiDISH's implementation of the CIBERSORT procedure rather than CIBERSORTx, and the InstaPrism algorithm was transcribed from released source without a check against the compiled package.