Matthew Segall, Chief Executive Officer of Optibrium, presented on the topic of predicting pharmacokinetics (PK) using limited ADME data and deep learning. Segall began by highlighting the challenges of PK prediction in drug discovery, noting the complexity and cost of PK experiments and the limitations of traditional QSAR models and physiologically based pharmacokinetic models. He introduced a novel approach called deep learning imputation, which uses sparse experimental data to fill in missing values and improve PK predictions.
Segall explained that imputation starts with the sparse experimental data available and fills in the gaps to predict complex, expensive endpoints more accurately. This method leverages early-stage, high-throughput data to predict later-stage, in vivo data, which is typically more expensive and time-consuming to obtain. The Alchemite method, developed in collaboration with Intelligens, was used to build and validate models that predict PK parameters with high accuracy.
Segall provided examples of the method's application, including a proof of concept study with AstraZeneca and an anti-infective project. In the AstraZeneca study, the method demonstrated high accuracy in predicting a broad range of in vivo PK parameters using sparse in vitro ADME data. In the anti-infective project, the method was used to optimise PK in a mouse model, successfully identifying compounds with improved PK profiles over the existing lead compound.
Segall emphasised the practical deployment of the deep learning imputation approach through the Cerella platform, which automates data cleaning, model building, and validation, and updates models as new data becomes available. This platform makes the approach accessible, robust, and scalable for drug discovery applications.
In conclusion, Segall demonstrated that deep learning imputation provides a new way to leverage sparse data for accurate PK prediction, offering confidence in targeting high-quality compounds and saving time and effort in drug discovery.
