Fig. 9
From: Improving VAE based molecular representations for compound property prediction
On the left side, 2D-PCA visualization of the logS dataset is displayed on the latent space of a PVAE model trained jointly with a MolLogP predictor. The colors indicate the clusters found by the K-means algorithm. Few molecules from selected clusters are visualized. On the right side, the relation between average RMSE error and KL divergence of the clusters and the prior is shown. Each point corresponds to one cluster (using the same color map)