Platform Biological Divergence: quantifying gene-level differences between bulk and single-cell transcriptomics in breast cancer

Bulk and single-cell RNA-seq offer complementary views of tissue transcriptomes. Nevertheless, pseudo-bulk (PB) profiles derived from single-cell data often differ systematically from true bulk measurements. These discrepancies are rarely quantified at the gene level, despite their impact on downstream analyses. To investigate the PB-bulk disagreement, here we introduce, based on linear mixture models, a gene-centric platform biological divergence (PBD) score. Across matched breast cancer, human adipose, and healthy retinal datasets, PBD stratifies genes into low-PBD "stable proxies" that retain true biological signals and high-PBD "divergent drivers" that shape PB-bulk differences. High-PBD (15-25%) genes show platform-specific signatures: bulk-enriched drivers highlight stromal, immune, and metabolic programs (breast cancer, adipose) and housekeeping/translational pathways (retina), whereas PB-enriched drivers reflect cell-intrinsic and regulatory processes, including immune, endothelial, progenitor, ECM, developmental, and photoreceptor programs. In contrast, low-PBD genes reliably preserve biology-driven variation and improve cross-platform concordance. Overall, the PBD framework reveals systematic gene-level biases between PB and bulk profiles, enabling more accurate comparison and integration of bulk and single-cell transcriptomes across tissues and disease contexts.