Actionable\nPredictions of Human Pharmacokinetics at\nthe Drug Design Stage

We present a novel computational approach for predicting\nhuman\npharmacokinetics (PK) that addresses the challenges of early stage\ndrug design. Our study introduces and describes a large-scale data\nset of 11 clinical PK end points, encompassing over 2700 unique chemical\nstructures to train machine learning models. To that end multiple\nadvanced training strategies are compared, including the integration\nof in vitro data and a novel self-supervised pretraining task. In\naddition to the predictions, our final model provides meaningful epistemic\nuncertainties for every data point. This allows us to successfully\nidentify regions of exceptional predictive performance, with an absolute\naverage fold error (AAFE/geometric mean fold error) of less than 2.5\nacross multiple end points. Together, these advancements represent\na significant leap toward actionable PK predictions, which can be\nutilized early on in the drug design process to expedite development\nand reduce reliance on nonclinical studies.