External Transportability and Tissue Context of Parsimonious Breast Tumour-Normal Transcriptomic Score
Year: 2026
Type: preprint
Abstract: Background Bulk tumour-versus-normal breast transcriptomic classifiers commonly show high internal discrimination, but a high area under the curve may reflect the biology of tissue contrast, cohort construction or technical separation as much as a portable malignant-cell program. We asked whether a pre-specified, interpretable score trained in one public cohort could retain rank discrimination in a locked cohort while its underlying biology was examined in parallel. Methods Processed Affymetrix GPL570 data from GSE42568 were used for discovery and nested model development. GSE7904 was held as an external cohort. Within-study normalization, probe-to-gene collapse and gene intersection produced 22,737 shared genes. Differential expression and enrichment were conducted. The elastic-net logistic score was the primary model; feature selection and tuning were contained in discovery-cohort nested cross-validation. Results The discovery cohort had 1,213 upregulated and 1,369 downregulated genes at false-discovery rate < 0.05 and log2 fold change ≥ 1. Of 2,582 discovery differentially expressed genes, 1,729 (67.0%) had a concordant external direction; 563 met the stringent criterion with the same direction in both cohorts. Cell-cycle/mitotic programmes dominated the upregulated signal, while PPAR, lipid-metabolism and adipocyte-lipolysis terms dominated the downregulated signal. The 20-feature elastic-net score produced a nested discovery AUC of 0.933 (95% CI 0.830–0.997) and a locked external AUC of 0.970 (95% CI 0.900-1.000). External basal-like and non-BLC AUCs were 0.968 and 0.986, respectively; the BRCA1-associated subgroup was too small for credible estimation. Conclusion The score transports as a rank discriminator, but its uniformly negative coefficients and lipid/adipose-associated features indicate that it likely captures loss of normal breast-tissue programmes in addition to tumour biology. The evidence supports a contrast-aware transportability signal. Prospective cohorts with representative benign controls, unified preprocessing and formal calibration assessment are required in future works.
Source: Research Square
Authors: Subhajit Ghosh, Abhirup Saha, Vetrivel Baskar
Institution: National Center for Disease Control
Language: English
Subfield: Cancer Research
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7214321663