Chemical mixtures can significantly affect human health, but understanding the interactions among various chemical exposures and identifying influential ones in relation to some health outcomes are difficult. Bayesian kernel machine regression (BKMR) is a widely used model for capturing nonlinear dynamics and interactions between multiple exposures and health outcomes. However, tools for quantifying the interactions captured by this flexible model are scarce. Utilizing the inherent connection between BKMR and Gaussian process regressions, we adopt the classic variance-based sensitivity analysis tools from the uncertainty quantification community and propose a variable clustering approach to quantify interactions, discover high-order interaction terms, and rank variable importance. The performance of this method is demonstrated in a range of simulation scenarios and applied to a real dataset to examine the interactive effects of multiple per- and polyfluoroalkyl substances exposures, dietary patterns, and gestational diabetes mellitus status on thyroid function in women during their late pregnancy.