Expanding the druggable proteome by integrating deep learning with molecular simulations to predict cryptic pockets