The structural dynamics of proteins play a crucial role in their function, yet most experimental and deep learning methods produce only static models. While molecular dynamics (MD) simulations provide atomistic insight into conformational transitions, they remain computationally prohibitive, particularly for large-scale motions. Here, we introduce DeepPath, a deep-learning-based framework that rapidly generates physically realistic transition pathways between known protein states. Unlike conventional supervised learning approaches, DeepPath employs active learning to iteratively refine its predictions, leveraging molecular mechanical force fields as an oracle to guide pathway generation. We validated DeepPath on three biologically relevant test cases: SHP2 activation, CdiB H1 secretion, and the BAM complex lateral gate opening. DeepPath accurately predicted the transition pathways for all test cases, reproducing key intermediate structures and transient interactions observed in previous studies. Notably, DeepPath also predicted an intermediate between the BAM inward- and outward-open states that closely aligns with an experimentally observed hybrid-barrel structure (TMscore = 0.91). Across all cases, DeepPath achieved accurate pathway predictions within hours, showcasing an efficient alternative to MD simulations for exploring protein conformational transitions.