Pharmaceuticals and personal care products (PPCPs) are continuously introduced into aquatic environments, raising concerns about potential health risks from long-term exposure. Metal-organic frameworks (MOFs), owing to their abundant adsorption sites and tunable pore structures, show considerable potential for PPCP removal. However, experimental screening across diverse adsorbents and pollutants is time-consuming. This study developed a machine-learning framework to predict the adsorption capacity of PPCPs by MOFs at contact time and identify the key factors influencing model predictions. In total, 7324 adsorption records were collected; 7075 records covering 100 MOFs and 39 PPCPs were used for model development and internal evaluation, and 249 additional records were reserved for external validation. MACCS keys, binary Morgan fingerprints (BMF), and count-based Morgan fingerprints (CMF) were applied to encode PPCP structures, and eight machine learning models were systematically compared. Among all model–fingerprint combinations, BMF–CatBoost achieved the best predictive performance, with a test-set R² of 0.9723, and was selected as the optimal model. SHAP and accumulated local effects (ALE) analyses identified key factors and favorable response ranges associated with enhanced adsorption capacity. Overall, this study provides a data-driven methodology for rapid adsorption capacity prediction, key factor identification, and experimental condition optimization in MOF-based PPCP adsorption.