Representation-dependent transportability of single-cell-derived myeloid- and fibroblast/stromal-associated transcriptional programs in bladder cancer

Abstract Single-cell-derived transcriptional programs are often transported to bulk and external single-cell datasets. It remains unclear, however, at which representation level these programs remain informative. We evaluated four eligible myeloid–stromal program pairs derived from exploratory single-cell analyses of bladder cancer across multiple external cohorts and data types. We tested locked 30-gene programs in TCGA-BLCA, GSE13507 and GSE31684, in the external scRNA-seq datasets GSE222315 and GSE301651, and in 22 spatial sections from GSE319536. Bulk associations were assessed by Spearman correlation. In the external scRNA-seq datasets, we examined lineage representation, fixed clustering and state matching, Harmony sensitivity, patient-wise leave-one-gene-out (LOO) coherence, and expression-matched null controls. Fixed-marker tissue-composition adjustment, spatial depth sensitivity and clinical associations were examined as secondary analyses. All 12 bulk pair-by-cohort correlations were positive (Spearman rho 0.585–0.851). However, results across the full 2 × 2 set of eligible pairs did not support a unique interpretation for any one pair. In the external scRNA-seq datasets, lineage-level patient coverage was complete, but direction agreement and top-20 marker overlap were weak for discrete-state matching. Patient-wise LOO coherence was positive at the dataset level and exceeded expression-matched null expectations. Rescoring with common genes had little effect on bulk correlations (absolute change ≤ 0.020). Fixed-marker composition adjustment substantially reduced these correlations. Spatial within-section co-variation was weak after depth adjustment and was sensitive to filtering. Section-level spatial autocorrelation was also low. Clinical associations varied across analyses, and none of the overall-survival associations was significant. Transportability depended on the representation level. Fixed programs retained detectable within-lineage coherence across external datasets, but this did not show preservation of discrete states, a specific myeloid–stromal axis, or a cellular interaction.