Wastewater treatment plants (WWTPs) serve as critical barriers against the dissemination of antibiotic resistance genes (ARGs) from urban water environments to nature, yet the molecular mechanisms governing their biological removal remain poorly understood. By combining experimental metagenomic data from 19 Chinese WWTPs with additional data from 31 global WWTPs (50 WWTPs in total), an explainable machine learning (ML) framework was developed. The RFE-SHAP (Recursive Feature Elimination-SHapley Additive exPlanations) based on feature importance was applied to identify key biological features driving ARG removal. The study revealed that low-abundance microbial functional genes particularly those involved in DNA repair, energy metabolism, and quorum sensing exhibit threshold-dependent control over ARG attenuation. ML models (BFGs-GBDT) incorporating the RFE-SHAP-selected functional genes achieved exceptional predictive accuracy (R²_test = 0.967), outperforming taxonomy-based models (average R²_test = 0.805). Strikingly, these functionally critical genes, despite their low abundances (0.04 – 0.15%), exerted disproportionate influence on ARG removal efficiency, challenging the prevailing high-abundance-centric paradigm in WWTPs design. The findings not only elucidated the molecular mechanisms of ARG mitigation but also provided a predictive framework for precision engineering of microbial communities to enhance ARG elimination. This study advances wastewater treatment strategies from empirical ARG removal to mechanism-driven environmental risk control.