Preclinical Development – Chemical
2019 PharmSci 360
LigandNet is a machine learning toolbox that combines different models into an open source platform that can predict if a ligand may have activity to a specific protein. Here, we have applied advanced Machine Learning approaches such as Random Forest, XGBoost, Support Vector Machine, and Artificial Neural Networks to classify the ligands as actives/binders and not-actives/nonbinders. We obtained the known active ligands for each of ~800 proteins from Pharos (pharos.nih.gov) database and generated decoys using Decoy finder. Machine learning models were developed for each of the protein-ligand sets by using known actives and generated decoys. ECFP6 fingerprints and Topological Pharmacophore Atomic Triplet Fingerprint were employed as features. Models were validated using highest positive predictive value, sensitivity, and area under the curve of receiver operating characteristic plots (ROC-AUC) to determine which model works best with the each given datasets of the proteins, yielding the accuracy and precision of each ML approach.