Organic micropollutants (OMPs) such as pharmaceutical, personal care products, pesticides and industrial chemicals in wastewater can threaten environment and human health. Forward osmosis (FO) is a promising technique for OMPs rejection with high anti-fouling capacity and low energy demand. However, OMPs rejection in FO system is a complex multifactorial process. It is unrealistic to assess OMPs rejection by FO process via repetitive trial-and-error experimentation. Here, an interpretable machine learning (ML)-assisted strategy was presented to optimize the OMPs retention in FO process. 18 influential factors associated with membrane properties, OMPs properties and experimental conditions were used for 10 ML models development. The optimal XGBoost-18 model was determined by performance evaluation based on multiple metrics. Interpretation for the optimal model was achieved through Shapley additive explanations (SHAP). The results showed that McGowan volume of OMPs, molecular weight of OMPs, zeta potential of FO membrane surface, and osmotic pressure of draw solution significantly affected OMPs rejection in FO process. 11 representative input features were selected from the original 18 variables based on the feature importance analysis provided by SHAP. On this basis, the SHAP-XGBoost-11 model was trained and achieved the most accurate prediction (R2adj=98 %) for OMPs rejection. The current study provides a new perspective for more efficient experimental optimization of FO system, aiming to achieve the highest rejection of target OMPs in the future. In addition, the findings in this study suggested a referential logical framework the expansion of FO process to a broader application scale.