Inass Soukarieh
Séminaire à 13h30.
Résumé : Physics-Informed Neural Networks (PINNs) have proven effective for solving partial differential equations (PDEs) across scientific domains. However, standard PINNs deal with a single PDE parameterization, limiting their ability to evaluate solutions across multiple parameter sets. This work introduces HyperSBINN, a meta-learning based approach that integrates Hypernetworks with System Biology-Informed neural network (SBINN) to address this challenge in the context of drug cardiosafety assessment. Traditionally, cardiosafety assessments rely on early-stage in silico/in vitro assays to measure a compound’s effect on cardiac ion channels. Here, we focus on predicting changes in action potential duration (APD90) at increasing drug concentrations. HyperSBINN tackles the issue of parametrization by efficiently learning the relationship between input parameters (representing drug properties) and the solution of the governing mathematical model for cardiac electrophysiology. This allows for the characterization of multiple compounds simultaneously within the cardiosafety assessment pipeline, making the process more efficient and cost-effective. Furthermore, HyperSBINN’s performance will be compared to standard machine learning techniques focused solely on predicting the quantity of APD90.