Proteins are complex biomolecules that perform various functions in the cell, and their conformational states determine their behavior. Many proteins have multiple conformational states separated by energy barriers -- with advanced molecular simulation strategies, we can simulate the behavior of proteins, including how they exchange between these meta-stable states. However, detecting the meta-stable states by analyzing the resulting high-dimensional datasets is cumbersome by manual inspection. Consequently, unsupervised learning techniques such as clustering have been proposed to detect the meta-stable states. Here we propose to employ a mathematical framework: the variational approach for Markov processes (VAMP) guide the learning of an E3-equivariant neural network to learn how to detect meta-stables directly from simulation data. This approach (VAMPnets) is a recent development, yet so far, using neural network Ansatze which satisfy the symmetries of molecules, has not received much attention. We anticipate a more compute and data-efficient architecture to emerge from this study, allowing for robust analysis of simulation data faster and with less data. We base these expectations on the improved performance of equivariant neural networks such as convolutional neural networks (CNNs) and transformers in image and text applications and E3-equivariant neural networks in machine-learned forcefields.