Drug metabolism liabilities are a major bottleneck in drug development, necessitating accurate metabolite prediction tools. While AI-driven methods have advanced, challenges remain in achieving high-precision, robustness, broad reaction coverage and seamless integration into the drug development workflow. To address this, we first constructed the most comprehensive human-specific drug metabolism database to date, encompassing 11,665 human metabolic reactions and 2,497 high-quality rules. Leveraging this resource, we developed ConMeter (Contrastive learning and feature interaction based Metabolic Reaction prediction), a novel model integrating chemical feature interaction and contrastive learning. ConMeter demonstrates significant improvements, achieving ~80% accuracy in identifying at least one metabolite within the top 5 predictions and showing an average 25% relative enhancement in ranking true positives compared to state-of-the-art methods. The comprehensive database also expands the range of predictable metabolic reactions. To demonstrate the utility of ConMeter, we integrated it into a metabolite prediction-guided drug design strategy aimed at addressing the low oral bioavailability of a metabolically vulnerable RSK4 lead compound-a pervasive issue among known RSK inhibitors. This strategy successfully yielded R636, a novel RSK4 inhibitor exhibiting remarkably improved bioavailability (a 63-fold increase, reaching 63% absolute bioavailability) while maintaining its potent activity. Notably, R636 showed favorable safety profile and significant anti-tumor efficacy against esophageal squamous cell carcinoma (ESCC) in vitro and in two distinct patient-derived xenograft (PDX) mouse models. This work introduces ConMeter as a powerful and practical tool for high-accuracy metabolite prediction, demonstrating its potential to accelerate lead optimization, exemplified by the discovery of R636 as a promising orally bioavailable candidate for ESCC.