Learning Deep Koopman Operators for Data-driven Model Discovery of Synthetic Biological Systems
Abstract
This effort developed a deep learning framework to discover Koopman network models that map causal dependencies between all measured biocircuit outputs, experimental input parameters, and contextual design parameters. The research discovered dynamic network models of how measured quantities (e.g. multiple fluorescent reporters) and monitored experimental parameters (e.g. OD, temperature, inducer concentration, age of media) are causally related, thus generalizing the concept of dynamical structure functions (and transfer functions) for arbitrary nonlinear systems. These models were used to identify categories and relationships of biological parts or biocircuits with analogous behavior, infer the state of latent variables to generate hypotheses for experimental surprise, predict and evaluate operating envelopes for stable system behavior, and quantitatively predict biocircuit dynamic response as a function of experimental parameters. The methods in this project combined the expressive power, scalability of deep learning algorithms and rigor of Koopman operator theory to discover data-driven dynamical system models for hypothesis generation and biocircuit characterization.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 2021
- Accession Number
- AD1170384
Entities
People
- Enoch Yeung
Organizations
- University of California