Understanding and classifying metabolite space and metabolite likeness

This document summarizes a study on classifying molecules as metabolite-like or non-metabolite using machine learning models. The study used metabolite data from HMDB and non-metabolite data from ZINC. Multiple descriptors and fingerprints were used as features to train support vector machines, random forests, and naive Bayes classifiers. The random forest model using MDL public keys achieved the best performance with high sensitivity and specificity on an external test set. The top-performing models were also able to predict metabolite-likeness for unknown structures generated in silico. However, the study found prediction was more accurate than interpretation, and local models may be needed.