What does Olivier–Ricci curvature add to graph-based single-cell clustering? A matched, dose-controlled re-evaluation in the aging mouse brain

Curvature-guided edge pruning has been proposed to improve graph-based clustering of single-cell RNA-seq data by removing negatively curved shortcut edges that merge distinct populations. Results: We re-evaluate this idea under matched, dose-controlled conditions on three fully annotated GSE129788 aging mouse-brain datasets (3,193/1,489/1,666 cells; one rare class at 0.97–1.44%). Five pruning criteria (Olivier–Ricci, Forman–Ricci, longest-edge, shared-neighbor, and random) are compared across the full pruning-quantile grid q = 0.01–0.10 with 10 Leiden seeds and 20 independent random-pruning draws per point (28,800 clusterings), exact Wilcoxon tests and Holm correction. No criterion improves ARI over the unweighted graph; where pruning differs significantly, the control outperforms ORC pruning in 335 of 336 corrected comparisons. Local-scale re-weighting — not pruning — stabilizes rare-class recovery, significantly only at low resolution (rare-class F1 +0.50, p = 0.011, at resolution 0.3; +0.20, p = 0.115, at 0.5). Low ORC values do mark inter-population edges (AUROC 0.87–0.94; 18–48-fold enrichment at q = 0.01) but do not outperform plain edge length (0.94–0.98). On a synthetic bridge benchmark, only the combined prune-and-re-weight pipeline recovers the rare cluster across the whole resolution sweep. Availability: code, dose–response grids and per-seed tables are publicly released.