Digital pathology images, a type of data that comes from high-resolution biopsy tissue scans, are one of the most promising data sources we can use to select lung cancer treatment. They provide direct visual information from cancerous cells regarding texture and placement that has been used in similar tasks in the oncological field. This study uses digital pathology scans to introduce a novel AI-based framework for multiple prediction tasks related to the efficacy of immunotherapy in non-small cell lung cancer (NSCLC). Using hematoxylin and eosin (H&E) slides, the proposed method bypasses conventional PD-L1 immunohistochemistry markers by extracting robust feature representations through a pre-trained foundation model and aggregating tile-level information using an attention-based multiple instance learning (ABMIL) framework to predict histology, PD-L1 expression categories and treatment response in a weakly supervised setting. An ensemble strategy is employed to enhance predictive performance and mitigate overfitting. The results of testing the proposed approach on real-world data from a cancer research institution demonstrate the potential of this strategy for refined patient stratification and improved prognostic assessment in NSCLC immunotherapy.