Background: Even when the same treatment is employed, some patients are cured, while others are not. The patients that are cured may have beneficial microbes in their body that can boost treatment effects, but it is vice versa for the patients that are not cured. That is, treatment effects can vary depending on the patient's microbiome. If the effects of candidate treatments are well-predicted based on the patient’s microbiome, we can select a treatment that is suited to the patient’s microbiome or can alter the patient’s microbiome to improve treatment effects.
Methods: Here, I introduce a streamlined analytic method, named microbiome virtual twins (MiVT), to evaluate the interplay between microbiome and treatment. MiVT is based on the subgroup identification framework, called virtual twins, that involves a two-step algorithm, 1) treatment effect prediction through machine learning and 2) subgroup identification using a decision tree. MiVT, however, employs a new prediction method, named distance-based machine learning (dML), to improve prediction accuracy in microbiome studies and a new significance test, named bootstrap-based test for regression tree (BoRT), to test if each subgroup's treatment effect is the same with the overall treatment effect.
Results: I demonstrate in silico that dML robustly reaches a high prediction accuracy and BoRT is a valid significance test with correctly controlled type I error rates. I also demonstrate the use of MiVT in praxis through the gut microbiome study on the effects of cancer immunotherapies on melanoma patients.
Conclusions: The results from MiVT can serve as a useful guideline in microbiome-based personalized medicine to select the therapy that is most suited to the patient’s microbiome or to use dietary supplements or therapeutics to tune the patient’s microbiome to be suited to the treatment. MiVT can be implemented using an R package, MiVT, freely available at https://github.com/hk1785/MiVT.