@article{ART003397420},
author={Sungmin Yang and Hong-Gee Kim},
title={External Transferability of TCGA-Trained Pan-Cancer Transcriptomic Classifiers Across ICGC and GEO Cohorts},
journal={Journal of The Korea Society of Computer and Information},
issn={1598-849X},
year={2026},
volume={31},
number={10},
pages={129-145}
TY - JOUR
AU - Sungmin Yang
AU - Hong-Gee Kim
TI - External Transferability of TCGA-Trained Pan-Cancer Transcriptomic Classifiers Across ICGC and GEO Cohorts
JO - Journal of The Korea Society of Computer and Information
PY - 2026
VL - 31
IS - 10
PB - The Korean Society Of Computer And Information
SP - 129
EP - 145
SN - 1598-849X
AB - In this paper, we propose a controlled external transferability evaluation framework for pan-cancer transcriptomic classifiers across independent cohorts and measurement platforms. The key idea of this framework is to distinguish internal predictive accuracy from feature stability, external generalization, and preservation of cancer-discriminative transcriptomic structure. We trained and selected models using only TCGA RNA-seq data for five cancer types and directly applied the locked models, without retraining or feature reselection, to independent ICGC RNA-seq and GSE2109 microarray cohorts. Also, we evaluated 48 model configurations and performed diagnostic analyses of calibration, cancer-centroid geometry, measurement resolution, perturbation sensitivity, and biological interpretation. Experimental results showed that models with nearly identical TCGA performance exhibited substantially different external transferability. In particular, the primary model retained high performance in ICGC but declined in GSE2109, whereas ANOVA500+XGBoost(XGB) showed more stable cross-platform performance. Post hoc analyses showed that reduced external performance was accompanied by weaker preservation of cancer-type–specific transcriptomic structure, a pattern consistent with the possibility that shifts in cross-cohort representation contributed to the loss of external generalizability. These findings demonstrate that independent cross-cohort and cross-platform validation is necessary for robust evaluation of transcriptomic cancer classifiers.
KW - pan-cancer classification;transcriptomic transferability;external validation;domain shift;cross-platform generalization
DO -
UR -
ER -
Sungmin Yang and Hong-Gee Kim. (2026). External Transferability of TCGA-Trained Pan-Cancer Transcriptomic Classifiers Across ICGC and GEO Cohorts. Journal of The Korea Society of Computer and Information, 31(10), 129-145.
Sungmin Yang and Hong-Gee Kim. 2026, "External Transferability of TCGA-Trained Pan-Cancer Transcriptomic Classifiers Across ICGC and GEO Cohorts", Journal of The Korea Society of Computer and Information, vol.31, no.10 pp.129-145.
Sungmin Yang, Hong-Gee Kim "External Transferability of TCGA-Trained Pan-Cancer Transcriptomic Classifiers Across ICGC and GEO Cohorts" Journal of The Korea Society of Computer and Information 31.10 pp.129-145 (2026) : 129.
Sungmin Yang, Hong-Gee Kim. External Transferability of TCGA-Trained Pan-Cancer Transcriptomic Classifiers Across ICGC and GEO Cohorts. 2026; 31(10), 129-145.
Sungmin Yang and Hong-Gee Kim. "External Transferability of TCGA-Trained Pan-Cancer Transcriptomic Classifiers Across ICGC and GEO Cohorts" Journal of The Korea Society of Computer and Information 31, no.10 (2026) : 129-145.
Sungmin Yang; Hong-Gee Kim. External Transferability of TCGA-Trained Pan-Cancer Transcriptomic Classifiers Across ICGC and GEO Cohorts. Journal of The Korea Society of Computer and Information, 31(10), 129-145.
Sungmin Yang; Hong-Gee Kim. External Transferability of TCGA-Trained Pan-Cancer Transcriptomic Classifiers Across ICGC and GEO Cohorts. Journal of The Korea Society of Computer and Information. 2026; 31(10) 129-145.
Sungmin Yang, Hong-Gee Kim. External Transferability of TCGA-Trained Pan-Cancer Transcriptomic Classifiers Across ICGC and GEO Cohorts. 2026; 31(10), 129-145.
Sungmin Yang and Hong-Gee Kim. "External Transferability of TCGA-Trained Pan-Cancer Transcriptomic Classifiers Across ICGC and GEO Cohorts" Journal of The Korea Society of Computer and Information 31, no.10 (2026) : 129-145.
