This chapter investigates the spread and suppression of COVID-19 in Japan during the fifth to seventh waves, focusing on the roles of vaccination, viral variants, and social behavior. We estimated population-level immunity by integrating real-world vaccination effectiveness, waning immunity, and asymptomatic infection. To forecast daily positive cases in three major prefectures, Tokyo, Osaka, and Aichi, we developed a deep learning model using a multipath architecture that combines long short-term memory (LSTM) modules for time-series forecasting with fully connected layers to capture feature correlations. The input data include mobility patterns, vaccination records, and social media activities. Results indicate that at least 40% population-level immunity was required to achieve temporary suppression of outbreaks, with earlier vaccination or reduced high-risk behavior significantly lowering peak case counts. The model achieved a mean absolute percentage error of less than 30% during major outbreaks. Simulations of various hypothetical scenarios have demonstrated the model potential utility in pandemic planning. These findings suggest that combining mechanistic immunity modeling with LSTM-based deep learning provides a robust framework for forecasting and evaluating policies in dynamic pandemic setting.