Predicting biomolecular conformations is essential for understanding their functions, stability, and dynamics. Theoretical modeling approaches, such as all-atom molecular dynamics (AAMD) simulations, provide atomistic insight into protein conformational landscapes. However, the substantial computational cost of AAMD limits its applicability to large biomolecular complexes and long time-scale processes such as protein folding. To overcome these limitations, enhanced sampling methods have been developed to accelerate conformational transitions and improve sampling efficiency. Our research focuses on a family of enhanced sampling approaches that explore free energy surfaces defined by collective variables. We have developed unique methods that integrate enhanced sampling with artificial intelligence to further improve efficiency and generalizability. Beyond enhanced sampling, recent advances in deep learning have enabled the direct generation of protein conformations consistent with the Boltzmann distribution. A major challenge of this approach lies in accurately modeling conformational ensembles under diverse physical or environmental constraints—scenarios we refer to as “proteins in constrained ensembles.” To address this, we have developed several AI-based frameworks capable of generating constrained conformational ensembles. These include methods that leverage the embedding space of protein language models and diffusion models guided by physics-based interactions. Together, these tools enable the generation of protein conformations conditional on experimental observations, interactions with material surfaces, or structural constraints imposed by organic linkers.