Figure 8.

From featurized data, an analysis of molecular kinetics can proceed under the assumption of microscopic reversibility (i.e., equilibrium) or not. In the former case, TICA can be applied to reduce dimensionality and may serve as a final or intermediate model. When TICA serves as a step in the MSM construction pipeline, clustering is performed in TICA space and the MSM is estimated from the cluster assignments. Both TICA and MSMs adhere to the VAC, which acts on the eigenvalues of the propagator approximation. A macrostate MSM can be created if further interpretability is desired. In the latter case (no reversibility assumption), TCCA is used instead of TICA to reduce the dimensionality of the data set. From there, clustering can be performed and a reversible MSM can be constructed, or the TCCA results can be regarded as the final model. The relevant variational principle is VAMP, which acts on the singular values of the propagator approximation. A VAMPnet bypasses the majority of the construction pipeline by creating an interpretable model directly from featurized data. VAMPnets employ the VAMP criterion as a loss function in a neural network scheme.