Fig. 2: HetSIREN landscape and exploration for Empiar 10028 dataset.
(a) shows the UMAP19 representation of the landscape obtained from the latent space encoded by HetSIREN from the particle images after training. The landscape without pose and CTF decoupling is presented on the left, while decoupling is implemented on the right landscape. Each dot in the landscape corresponds to the centroid of the cluster representative obtained from a KMeans clustering of the decoupled HetSIREN latent space. (b) shows the decoded HetSIREN maps obtained from the decoupled latent space coefficients assigned to every representative, as shown in (a) (right). The maps provide a sensible exploration of the different conformational states identified by HetSIREN and a comparison of the structural features learned by the network.
