Translational Risk Governance in Gene-Expression-Based Drug-Response Prediction

Abstract

Gene-expression-based drug-response prediction can prioritize translational experiments, but strong internal performance does not establish transportability or decision readiness. This study develops a management-analytics framework for governing the evidence chain from gene-expression measurement to Machine Learning (ML) prediction and action. It combines an evidence synthesis with a Monte Carlo stress test using 360 source samples, 200 external target samples, 80 genes, three levels of batch/domain shift, and 250 replications per level. Ridge regression and Partial Least Squares (PLS) regression were evaluated by five-fold internal cross-validation and independent target-domain testing. Under no shift, Ridge achieved internal R²=0.588 and external R²=0.602. Under strong shift, internal R² increased to 0.664 while external R² fell to 0.478, yielding an optimism gap of 0.185; PLS showed the same pattern. The paper therefore proposes a six-gate Translational Drug-Response Analytics Governance (T-DAG) framework covering intended use, data integrity, response validity, model validity, transportability, and decision readiness. The implication is that predictive authority should scale with external evidence and decision consequence rather than internal accuracy alone.