The Janssen in-house generic Target Product Profile (gTPP) machine learning (ML) model has been implemented in the workflows of over 200 Discovery Scientists since 2019 [1]. Previously, we designed the model to predict 18 pharmaceutically important endpoints and have since increased this to include endpoints such as: LogD at pH 10.5, 7.4, 2.6; LogP, Pharmaron permeability, and hERG inhibition. In order to improve gTPP performance, we eliminated multi-class labels and only included binary labels. This resulted in an average increase of 30% in the number of returned predictions at a 95% confidence level (defined herein as ‘efficiency’) for the corresponding endpoints. An important metric for gauging whether gTPP is driving design is to determine whether properties are being calculated before a ‘go’ decision is made on a compound. Through our novel ability to capture usage statistics, we observed an increase from 40% to 64% of compounds with properties predicted before synthesis. We attribute this improvement to increased education and exposure to the model, greater accessibility by exposing the model on additional cheminformatics platforms, and increased efficiencies. Finally, a more streamlined model building process was achieved in which models can be created using several tools and are stored in a model repository. These are pushed to a generic model server from which they can serve multiple end user tools. This upgraded deployment procedure has decreased the time required to build the gTPP model from several weeks to several days and allowed super-users to re-train and repurpose the model for their projects. To date, gTPP has been employed on over 30 million unique chemical structures spanning discovery projects in all therapeutic areas and has demonstrated value in both triaging compounds within large chemical libraries and optimizing late-stage molecules.

Improvements in performance and usage of enterprise-wide global ADMET model

ACS Spring 2022

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