Absorption, Distribution, Metabolism and Excretion (ADME), which collectively defines the concentration profile of a drug at the site of action, is of critical significance to the success of a drug candidate. Recent improvements in machine learning (ML) algorithms and Biogen in-house in vitro ADME data sets allow us to investigate the performance of different ML models over time. In this study, we collected 120 Biogen prospective datasets across 6 ADME endpoints (including HLM CLint, RLM CLint, MDR1-MDCK efflux ratio, solubility, human, and rat plasma protein binding) over 20 months. A variety of ML regression models including SVM, Random Forest, XGBoost, LightGBM, deep fully-connected, and message passing neural network were evaluated by the predictivities on 120 prospective data sets, where physicochemical descriptors, 2D fingerprints, and molecular graph were used to featurize the chemical structures. Our results indicated that 1) The gradient boosting decision tree model, LightGBM, showed consistent improvement over the current random forest model in terms of predictivity and efficiency. 2) Message passing neural network model can achieve slightly better performance than LightGBM, but it is far more expensive. 3) Hyperparameter tuning can improve the prospective predictivity consistently but marginally. Given the cost, the recommended sets of hyperparameters for each model are provided for future use. 4) We developed an internal unsupervised computational platform to automate data curation, prospective prediction, and model upgrade every month. We integrated high-throughput ADME prediction via a desktop application enabled by RESTFul API.

Gradient boosting decision tree models for better temporal ADME prediction from an industrial perspective

ACS Fall 2020 Virtual Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.