Efficient reference composition sensitivity analysis for single-cell annotation

Background: Single-cell labels can depend on the cell-type proportions represented in a reference atlas. Evaluating this dependence requires repeated model fits and informative composition scenarios. We developed an optimal transport workflow that accelerates these checks and directs them toward sensitive assignments. Findings: A shared baseline curvature factor supports repeated transport correction, while implicit composition derivatives guide perturbations and prioritize cells for review. With references of up to 4,800 cells, quadratic prediction with shared curvature was faster than batched Anderson-accelerated Sinkhorn in 10 of 11 complete comparisons, with a median speedup of 2.55 at the prescribed numerical precision. In 23,658 held-out cells from 40 liver-cancer and lung donors, directed search increased donor-mean sensitive-cell detection from 7.39% to 11.65% in liver myeloid cells, 4.81% to 6.75% in lung epithelium and 4.21% to 4.78% in lung mesenchyme at the same scenario budget. A 16-configuration analysis identified stronger search gains for native multiclass annotations than for merged or binary labels. Across eight native-label configurations, sensitivity ranking recovered 70.63 randomly confirmed sensitive cells per 100 reviewed, versus 58.25 using the winning-score margin and 50.25 using confidence. Selective-annotation improvement was modest and confined to lung epithelium. Conclusions: Shared curvature reduces repeated composition-checking costs, and derivative-guided search improves detection yield. Composition sensitivity also adds review-priority information beyond the baseline winning-score margin. These diagnostics help identify reference-dependent assignments for biological review.