There is increasing interest in examining the simultaneous exposures to chemical mixtures and rising public health concerns on the effects of interacting chemicals. The challenge lies in appropriate analysis of these complex mixtures to account for the unique characteristics of environmental exposure assessments such as multicollinearity, interactions, non-linear dose response relationships, and low signal to noise ratio. Existing statistical approaches suitable for high dimensional data have not been extensively validated or compared when applying to environmental mixture exposure data, which need to account for these unique characteristics. We aim to build predictive models incorporating interaction effects that will have more predictive power to inform individualized risk classification based on environmental exposure profiles. We evaluate combinations of different approaches: machine learning approaches, penalized regression approaches and Bayesian variable selection methods, and compare their performances through extensive simulation studies and application to a real dataset that have measured participants’ exposures to toxic metals and adverse health outcomes.