A key objective in all small molecule drug discovery programs is to identify molecular entities that can sufficiently modulate the activity of the biological target during the dosing interval. This requires small molecules to  be presented at the biological target in a concentration consistent with the target’s PKPD requirements. Learning in potency SAR and the PKPD relationships are only marginally transferable from target to target. In contrast the underlying ADME properties defining the PK of the compounds are common across projects, and offer unique opportunities for predictive modeling due to the corpus of knowledge and data progressively developed.<br/>Machine learning models are utilized to predict individual (or closely related) in vitro ADME properties starting from chemical structure. Recent advancements in the area of deep learning are opening new doors by allowing some degree of transfer learning between different endpoints. For a given compound, many different endpoints can be modeled under a common mechanistic physiologically based pharmacokinetic (PBPK) framework to simulate PK profiles across different species. Although large datasets may be available for pre-clinical PK data, datasets describing human PK are scarce and their size does not allow developing reliable machine learning models; PBPK is used instead, typically for a few selected compounds during later discovery stages. This presentation describes Genentech’s internally developed machine learning and PBPK models, and their use at scale to assist medicinal chemists in the ADME multi-parameter optimization process.

Driving ADME multi-parameter optimization by utilizing hybrid machine learning and physiologically based PK models at scale

ACS Fall 2021

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