Speaker
Description
Mechanistic ordinary differential equation (ODE) models are a powerful tool for studying biological systems. However, their predictive power is constrained by gaps, biases, and inconsistencies in the literature. They typically also require quantitative time-lapse data for training, which is time-consuming to collect. While training could benefit from integrating other modalities such as omics and patient metadata, doing so remains an open challenge. Conversely, machine-learning models lack interpretability and require large datasets. Hybrid scientific machine learning (SciML) models aim to address these shortcomings by combining mechanistic and data-driven modules.
Despite this promise, adoption of SciML modeling in biology remains limited by insufficient infrastructure. Dedicated software packages are needed because implementing end-to-end differentiable SciML workflows for state-of-the-art gradient-based training (parameter estimation) is technically challenging. In addition, model exchange is hindered by the absence of a standardized, reproducible format for specifying SciML training problems, analogous to the PEtab standard for ODE models. To address these gaps, we developed the PEtab-SciML extension to the PEtab format and implemented support in PEtab.jl and AMICI. We here present the PEtab-SciML format, show how it enables efficient training strategies such as curriculum learning and multiple shooting, and report benchmark results comparing training approaches.
