Fig. 4
From: SimVec: predicting polypharmacy side effects for new drugs
The ROC AUC curves on the validation dataset during the training of SimVec modifications with different (a) similarity weighting strategies and (b) chemical initialization options. a Three different weighting strategies were examined: proportional weighting with similarity functions based on morgan fingerprints (weight_morgan) and hyperbolic embeddings (weight_hvae), and window weighting based on molecular descriptors (weight_mol_descriptors). Albeit all the strategies show similar performance, molecular descriptors with window weighting work best. b Among three available chemical initialization options, namely, morgan fingerprints (chem_morgan), RDKit-based molecular descriptors (chem_mol_descriptors) and hyperbolic embeddings (chem_hvae), morgan fingerprint work best