Fig. 2
From: MolPROP: Molecular Property prediction with multimodal language and graph fusion
Latent Embedding Visualization of the MolPROP ESOL Regression Model. The learned neural network embeddings of the ESOL test set are projected into 2-dimensional space utilizing the UMAP algorithm for A MolPROPGATv2-ChemBERTa-2-77Â M-MLM, B GATv2 (ablated), and C ChemBERTa-2-77Â M-MLM (ablated) models. All panels display the 1st UMAP dimension as the x-axis and the 2nd UMAP dimesion as the y-axis. The 2-dimensional UMAP projection is determined with the 10 nearest neighbors, utilizing the Chebyshev distance metric, and a minimum distance of 0.25. The color scheme is displayed on the right panel as a colorbar where the scalar values range from red to blue and represent the logarithm of water solubility in mol/L. Therefore, red clusters of molecules have high water solubility and the blue clusters of molecules have low water solubility