Donor-aware integrative single-cell analysis distinguishes consistent memory B-cell programmes from donor-sensitive state boundaries in untreated rheumatoid arthritis

Abstract

Single-cell transcriptomics resolves immune-cell heterogeneity, but whether cluster boundaries or coordinated programmes are the more transferable unit across donors remains uncertain. We reanalyzed 14,414 circulating memory B cells from four untreated rheumatoid arthritis donors. Harmony increased mean local donor entropy from 0.506 to 0.912, while original labels showed only moderate agreement with uncorrected and Harmony reclustering (adjusted Rand indices 0.508 and 0.473). Conditional label permutations and analyses based on pan-B-centred residuals or state-marker genes supported donor-level consistency for most major states, but identified C3-PECAM1 and the rare C10-IFN state as weakly specific or quality-control-sensitive. In training-only donor-held-out transfer, accuracy was 66.0–70.5% and adjusted Rand index was 0.432–0.525, indicating partial transfer rather than independent state recovery. Removal of DoubletFinder-predicted doublets preserved 15–20 of each state’s top 20 pseudobulk markers and did not change programme rank order across states. Rank-based UCell scores agreed with AddModuleScore in both datasets. In an independent cohort of 5,725 memory B cells from 34 donors, adapted programme scores could be quantified, but no disease-activity or serological association remained significant after correction; treatment and unavailable disease-duration data limited clinical interpretation. These results distinguish recurrent donor-level transcriptional structure from analysis-sensitive state boundaries and support programme-level interpretation without claiming independent cell-state validation or biomarker utility.