Osteoporosis, a bone disease affecting over 200 million people worldwide, presents a significant therapeutic challenge, with Cathepsin K (CatK) being a primary target for inhibitor development due to its role in bone resorption. While conventional drug discovery methods are often slow and costly, machine learning offers a promising alternative. This study addresses the need for more accurate predictive models by developing a robust framework for assessing CatK inhibitor bioactivity. A Long Short-Term Memory (LSTM) network, chosen for its proficiency in handling complex sequential data typical of molecular structures, was optimized using a Simulated Annealing (SA) metaheuristic. The model was trained on a dataset of 1568 molecules from the ChEMBL database, with bioactivity classified based on ? values into four categories: Potent ( ), Active , Intermediate , and Inactive . The SA-optimized LSTM model significantly outperformed three baseline LSTM models, which achieved a peak average accuracy of 0.77. The optimal SA-tuned configuration (the col_rate95 scheme) attained an average accuracy and F1-score of 0.81. Notably, the model demonstrated exceptional performance in identifying Potent inhibitors, achieving an F1-score of 0.92. However, a key limitation was the difficulty in distinguishing between the Active and Intermediate classes, where misclassifications were more frequent. This research highlights the effectiveness of the SA-LSTM approach in accelerating the discovery of high-bioactivity compounds for osteoporosis treatment. Future work could focus on enhancing model robustness by integrating additional molecular descriptors or exploring alternative deep learning architectures to improve classification accuracy.