Understanding protein structure and dynamics is essential for understanding their function. This is a challenging task due to the high complexity of the conformational landscapes of proteins and their rugged energy levels. In particular, it is important to detect highly populated regions which could correspond to intermediate structures or local minima. In this work we train a neural network model on the MD simulations data to create a low-dimensional latent space. The latent space is further used to explore the protein’s conformational space. It can be used to be interpolated or extrapolated to produce new intermediate protein conformations that might not have been previously seen. The latent space visualization also assists in visualizing the conformational path of the respective proteins. We compare the performance of a linear autoencoder and a variational autoencoder and discuss their advantages and shortcomings for exploring the pathways of the Cdc42 protein.