Please note that the account you create here is different than your Keystone Symposia account at www.keystonesymposia.org used to register for our multi-day conferences and is uniquely for viewing our virtual content.
Drug discovery and development is a costly and lengthy process that involves balancing multiple attributes of molecules to achieve desired clinical efficacy and safety. Among those attributes, absorption, distribution, metabolism, and excretion (ADME) properties, nonclinical safety profiles, and pharmacokinetics/pharmacodynamics (PK/PD) characteristics are key components to the molecules’ success. Recent advancement in Artificial Intelligence/Machine Learning (AI/ML) techniques has the potential to increase probability of success and speed up decision-making in drug discovery and development. The AI/ML model works by learning from large datasets that contain chemical structure information and their corresponding experimental readouts and then elucidates the underlying relationships between chemical structure and biological responses. The AI/ML models can be subsequently used to predict the likely biological responses of a molecule that has never been tested, or sometimes not even synthesized, entirely based on its chemical structure. In the ADME area, discovery scientists apply such AI/ML models to identify small molecules (out of hundreds of thousands of theoretical options) with the most promising ADME profiles to synthesize, test, and advance into the pipeline. With the advent of advanced AI/ML techniques, such as the “Graph Convolutional Neural Network”, the ADME field is enjoying a renaissance of sorts. Recently, Lilly scientists have executed state-of-the-art multi-task neural network modeling and have observed consistently improved predictions over traditional ML methods, thereby decreasing cycle time and traditional costs of drug discovery. Albeit incremental, consistent improvement of the new neural network method over traditional ML methods illustrates the potential of emerging AI/ML techniques to enhance the accuracy of ADME property predictions. To speed up innovation, we have initiated an interdisciplinary effort to tackle ADME, nonclinical safety, and PK/PD problems that can likely benefit from applying the state-of-the-art AI/ML techniques. Currently, our toxicology team is actively working on extracting morphological features from historical in-house rat primary hepatocyte cell images using AI/ML techniques to improve understanding of translation between in vitro phenotypic changes and in vivo toxicity observations. Similarly, our PK/PD modelers are experimenting with AI/ML methodologies to build models that predict patient response time course and simulate the effects of untested dosing regimens without requiring additional clinical trials. Collectively, these efforts will integrate AI/ML into various aspects of early drug discovery to enhance the probability of success and speed up decision-making to move potential medicines into development.