<p>Bioaccumulation, toxicity, and persistence of heavy metals including Cd, Pb, Cu, Zn, As, and Ni ranked as the critical environmental challenges of freshwater ecosystem. Adsorbents made of biomass are a viable solution in the removal of these metals due to the intricate interaction of compositional, structural and operational variables, which tend to limit the predictive power of conventional isotherm and kinetic equations. The machine learning (ML) models comprised of linear regression, Random Forest (RF), and Extreme Gradient Boosting (XGBoost) models were created to forecast adsorption capacity (q, mmol g^− 1) and adsorption efficiency (S, %). The best predictive model evaluated was XGBoost with coefficients of determination (R²) of 0.76 and 0.88 on a separate test dataset of q (mmol g^− 1) and S (%) of biomass adsorbents for removal of the above-mentioned metals, respectively. SHapley Additive exPlanations (SHAP) and partial dependence plots (PDPs) were used to investigate the idea of model interpretability, where elemental ratios ((O + N)/Ca and O/Ca (Here, Ca represents Calcium)), total carbon content, and cation exchange capacity (CEC) were found to be the major descriptors to govern adsorption performance and demonstrated strong nonlinear threshold effects. Such results indicate that interpretable machine learning is not only predictive but also provides an explanatory meaning, which is one of the efficient tools of rational designing and optimization of biomass-based adsorbents to be used as heavy metal removal agents in water treatment.</p>