Predicting CYP2D6 phenotype from resting brain perfusion images by gradient boosting

The cytochrome P450 enzyme 2D6 is involved in the metabolism of 20% of all commonly used drugs, including many psychotropic drugs and CNS-active substances. CYP2D6 is among the CYP enzymes with the highest expression levels in the brain, suggesting a role in the local brain metabolism of psychotropic drugs and the existence of endogenous substrates. The genetic polymorphism of CYP2D6, which causes individual differences in activity levels of the enzyme, has also been characterized functionally in human brain imaging studies. Here we explore the feasibility of predicting CYP2D6 phenotype using component-wise gradient boosting on fMRI resting brain perfusion images. The images belonged to subjects showing a range of genetic CYP2D6 variants. We achieved sensitivity and specificity values between 85% and 87% for the classification of ultrarapid metabolisers, and between 71% and 79% for poor metabolisers. An extension of the boosting algorithm, developed to improve the clinical plausibility of the inherently sparse models, produced enhanced models in agreement with the results of previous studies, showing some brain regions as positively associated with genotypic variation, most prominently in the prefrontal white matter and the corpus callosum. With further development, such a probabilistic method might constitute a valuable, non-invasive alternative to actual genotyping.