Oxide supported transition metals are common heterogeneous catalysts in chemical industry. Single atom catalysts (SACs) are the most efficient way to utilize all the metal atoms. To synthesize the SACs, the interaction strength between transition metals and the oxide supports is critical since weak interaction cannot resist sintering and thus form metal nanoparticles on the oxide substrates. Beyond the impact on metal particle size distribution, electronic metal-support interactions (EMSI) offer a path to tune the oxidation state of the transition metals and the catalytic reactivity or selectivity. Here, we use density functional theory (DFT) together with statistical learning (SL) to construct physical descriptors that can predict the metal binding energy on the oxide supports. We showed that the derived descriptors can capture the extent in the change of metal binding energy on modified MgO(100) in response to substituent dopants or surface adsorbates. Along with developing the understanding of EMSI, a novel SL algorithm, named iterative Bayesian additive regression trees (iBART), was proposed to construct the physical descriptors more efficiently than state-of-the-art methods. In summary, this work yields a systematic understanding in EMSI and an original SL method to build up physical descriptors.