The importance of a compound’s binding properties has long been acknowledged. For example, plasma protein binding needs to be considered to understand whether a compound’s plasma concentration will be high enough for a pharmacological effect. Similarly, the amount of binding in in vitro systems is important for correlating in vitro to in vivo properties. In the case of microsomal or hepatocyte incubations used to assess metabolic stability of a compound, only the free fraction of the compound will be metabolised in the system. A low intrinsic clearance value for a highly bound compound may indicate less metabolic stability than a somewhat higher value for a less bound compound. Thus, ways to measure – and predict – fraction unbound in the incubation, fuinc, have been suggested in literature over the past 10 to 15 years. Lipophilicity and ion class have been identified as major drivers of binding and have been used to predict fuinc. More elaborated machine-learning models have been described as well, but the models based on lipophilicity are in general more accessible. However, their applicability domain has not been intensively investigated. The present work describes the collaboration of members of the in silico ADME discussion group within the International Consortium for Innovation through Quality in Pharmaceutical Development (IQ consortium) to achieve an unbiased comparison of published fuinc prediction methods utilising large proprietary datasets from six pharmaceutical companies. Internal machine-learning models, based on the proprietary datasets, were included for comparison when available. After pre-analyses at each company, an excel template was created to easily calculate fraction unbound based on the selected methodologies and to gather categorized data on compound properties and prediction quality. This template was used to run the analysis at each company individually. The resulting data summary was shared across companies and used for the final analysis, comprising information on several thousands of compounds describing a varied chemical space. Overall lipophilicity-based models were well able to predict compounds with CLogP below three within 2-fold of the experimental value. However, more lipophilic compounds were less easily predicted. Internal machine-learning models were more likely to predict also compounds from the same company with CLogP above 3 within 2-fold.

Cross-company evaluation of in silico approaches to predict microsomal or hepatocyte binding

2019 Fall National Meeting

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