·       How to extract value when mining large scale Clinical Trial Data

·       Assessing ML approaches to mine date more effectively

·       Assessing how to use NPL to address challenges in extracting data value

·       Assessing challenges and problems faced in data mining e.g., data silos

Shameer Khader

Senior Director, AI/ML, Data Science, Digital Health and Bioinformatics
AstraZeneca

See Bio

Predictive computational methods are key enablers to support the investigation and development of new pharmaceutical molecules in all phases of the R&D process. The accurate prediction of ADME-Tox properties, for example, is an efficient task in early discovery helping to identify molecules with an appropriate profile of interest. At development or LCM stages, support via computational methods is also provided, e.g., towards identification and classification of reactive impurities. At all different stages, deep neural nets (DNN) emerged as transformative technology to analyze large datasets and to predict ADME-Tox properties and even drug adverse effects. Nonclinical or clinical ADME-Tox datasets are sometimes sparsely populated, but specifically DNNs allow to overcome some limitation of classical machine learning; multitask (MT)-DNN can ultimately combine different endpoints in a predictive multitask network. Model application however can be put outside the hands of experts through a portal solution that combines data acquisition, data normalization, prediction, and storage with analysis guidance.

In this talk, I will describe a fully industrialized approach to generate, apply and visually interpretation of different types of DNNs to model ADME-Tox data. Quantitative regression models (>50k data) targeting microsomal metabolic lability (ML), passive permeability, and lipophilicity will be described. Focusing on compound safety classification endpoints, e.g., phototoxicity, it can be shown that DNNs based on comparably smaller data sets (< 5k) can also exhibit strong predictive power when using quantum chemical descriptors. Finally, I will introduce a computational quantum chemistry approach to assess the reactivity of genotoxic Nitrosamine impurities demonstrating the usage of computational methods for marketing products.

• ADME-Tox prediction using deep neural networks (DNN)
• Comparison of different DNN methods
• Applicability in a pharmaceutical environment
• Computational quantum chemistry approach to assess the reactivity of N-Nitrosamine impurities

Jan Wenzel

Preclinical Safety - Digital Toxicology
Sanofi R&D

See Bio