Figure 1.

Structure-based drug classification pipelines. Chemicals from 12 medical subheadings (MeSH) therapeutic use classifications were converted to either two-dimensional color molecule images or Morgan molecular fingerprints. Molecule images were used to train a convolutional neural network (IMG + CNN) classifier, and fingerprints were used to train a random forest (MFP + RF) classifier. The models were used separately to predict classes of drugs using stratified cross-validation.