Overview

Part of the Image-Based Scientific Machine Learning for Theories of Biological Dynamics Across Scales program

​

Living systems grow, learn, sense, and heal through the collective interactions of a multitude of distinct molecule types. Modern imaging allows us to observe live cell dynamics of molecules at single-particle or population-level scales, but typically only one or a few types of molecules at a time. Alternative “snapshot” methods support imaging diverse molecules in space, but not time. For computational methods that provide descriptive statistics from a single data type, this limits predictive power under perturbations. Our workshop focuses on existing efforts and challenges for learning dynamical, mechanistic models that couple together multiple variables via inference over complementary datasets. By training models that obey physical rules, we aim for predictive dynamics across diverse contexts. 

​

In the higher-dimensional space of multi-component models, we identify several key challenges to emphasize in the workshop as we seek to infer useful (predictive and/or interpretable) and transferrable models from imaging data. i) Live-cell imaging data of single components must be augmented by other datasets, such as from other imaging modalities, structural databases, and ‘omics. Loss functions must be designed to deal with multi-modal experimental data. ii) Tools from structural identifiability of parameters, model selection, and uncertainty can be exploited for improved model development and iteration with experimental design. Iii) The inference pipeline must be efficient enough to consider multiple models; techniques for accelerating forward solvers of spatio-temporal dynamics and force evaluations may be essential for successful optimization. iv) Measurement noise in imaging data can be incorporated during model inference. 

​

The workshop will be designed for interactive hands-on sessions, including test cases and research problems with shared datasets for on-site analysis, training, and modeling. We highly encourage PhD students and postdocs to apply.