The design of good collective variables for analysis and the enhancement of sampling of molecular simulations is not a trivial task. It often relies on the knowledge of the system and the experience of the scientist. Machine learning and artificial neural networks can be used for this purpose. Here, we demonstrate for the first time the use of an adversarial autoencoder to design collective variables. Similar to other autoencoders, it encodes data into the latent space and decodes them back with minimal loss of information. Furthermore, it uses the "adversarial game" to control the distribution of the latent space. The coordinates of the latent space can be used as efficient collective variables for analysis and sampling enhancement. The method was applied to the alanine dipeptide trajectory and thermal unfolding trajectory of the tryptophan cage miniprotein. We demonstrate efficient visualization of the conformational space of both molecules. The latent space coordinates of the tryptophan cage were very efficient in accelerating its folding by metadynamics.
Tedeschi et al. (Tue,) studied this question.