Fig. 20. Upper panel: timescale separation (otherwise known as spectral gap) along different TIC vectors (unbiased MD simulation of apo PlmII with length of ∼600 ns). TIC1 and TIC2 projected along χ2 angle of Tyr-79 shows that χ2 is one of the slow degrees of freedom in plasmepsin-II (corresponds with the higher weight on sin  χ2 in Fig. 15. Middle panel: bottleneck layer of a variational autoencoder (VAE)⁷¹ projected along TIC1 and TIC2. TIC1–2 were used as inputs in VAE. Different values of bottleneck layer represents different points along TIC vectors. One can see that the bottleneck layer did a better job in separating different metastable states compared to χ2. Further, time-series projection of the bottleneck layer shows its fluctuation during MD simulation. The coefficients (weights) of the bottleneck layer can be used as order parameters in metadynamics simulations. This method in principal similar to combining RAVE⁷² with SGOOP.⁵³ See the Github profile corresponding this article to see step-by-step guide on how to use variational autoencoder. Bottom panel: shows apparent free energy surface projected along TIC1, TIC2 and bottleneck layer. In this case, i have used a variational autoencoder with the following hyper-parameters: number of hidden layers = 2, number of neurons in each hidden layer = 20, number of epochs = 100, batch size = 500, learning rate = 1e − 2.