Fig. 1: Schematic representation of TS-DAR for transition state identification. | Nature Communications

Fig. 1: Schematic representation of TS-DAR for transition state identification.

From: Exploring transition states of protein conformational changes via out-of-distribution detection in the hyperspherical latent space

Fig. 1

a Overview of the TS-DAR framework. Similar to VAMPnets37, TS-DAR takes transition pairs \({{{\rm{x}}}}_{t}\), \({{{\rm{x}}}}_{t+\tau }\) from simulation trajectories as input, generates the Softmax outputs, and estimates the VAMP-2 loss. Meanwhile, TS-DAR introduces an L2-norm/scale layer to create the hyperspherical embeddings at the penultimate layer. These embeddings, combined with the pseudo state assignments obtained from the Softmax outputs, are used to estimate the dispersion loss. The framework then optimizes the neural networks using a combined loss function of the VAMP-2 loss and the dispersion loss, weighted by a constant \(\beta\). b Utilization of the L2-norm/scale layer to confine original feature embeddings \(\widetilde{{{\rm{z}}}}\) at the penultimate layer within a hypersphere of radius \(\gamma\), producing the hyperspherical embeddings \({{\rm{z}}}\). c Identification of the transition states in the hyperspherical latent space. The VAMP-2 loss enhances data compactness within each metastable state, while the dispersion loss encourages the centers of different metastable states to be far apart across the hypersphere.

Back to article page