Optimizing ODE-derived Synthetic Data for Transfer Learning in Dynamical Biological Systems

Abstract

AbstractMotivationSuccessfully predicting the development of biological systems can lead to advances in various research fields, such as cellular biology and epidemiology. While machine learning has proven its capabilities in generalizing the underlying non-linear dynamics of such systems, unlocking its predictive power is often restrained by the limited availability of large, curated datasets. To supplement real-world data, informing machine learning by transfer learning with data simulated from ordinary differential equations has emerged as a promising solution. However, the success of this approach highly depends on the designed characteristics of the synthetic data.ResultsWe optimize dataset characteristics such as size, diversity, and noise of ordinary differential equation-based synthetic time series datasets in three relevant and representative biological systems. To achieve this, we here, for the first time, present a framework to systematically evaluate the influence of such design choices on transfer learning performance in one place. We achieve a performance improvement of up to 92% in mean absolute error for our optimized simulation-based transfer learning compared to non-informed deep learning. We find a strong interdependency between dataset size and diversity effects. The optimal transfer learning setting heavily relies on real-world data characteristics as well as its coherence with the synthetic data’s dynamics, emphasizing the relevance of such a framework.Availability and ImplementationThe code is available athttps://github.com/DILiS-lab/opt-synthdata-4tl.