Reliable machine learning models in genomic medicine using conformal prediction

Abstract

Machine learning and genomic medicine are the mainstays of research in delivering personalized healthcare services for disease diagnosis, risk stratification, tailored treatment, and prediction of adverse effects. However, potential prediction errors in healthcare services can have life-threatening impact, raising reasonable skepticism about whether these applications are beneficial in real-world clinical practices. Conformal prediction is a versatile method that mitigates the risks of singleton predictions by estimating the uncertainty of a predictive model. In this study, we investigate potential applications of conformalized models in genomic medicine and discuss the challenges towards bridging genomic medicine applications with clinical practice. We also demonstrate the impact of a binary transductive model and a regression-based inductive model in predicting drug response and the performance of a multi-class inductive predictor in addressing distribution shifts in molecular subtyping. Additionally, we employed a regression-based inductive predictor to estimate the resistance of cancer cell lines to the anticancer drug afatinib. The main conclusion is that as machine learning and genomic medicine are increasingly infiltrating healthcare services, conformal prediction has the potential to overcome the safety limitations of current methods and could be effectively integrated into uncertainty-informed applications within clinical environments.