Machine Learning workflow for discovery of dynamics in biological systems from time-series multiomics data
Project for Mathematical Structures of Complex Systems, Heidelberg University, WS 2019/2020
by Achita Prasertwaree & Kevin Siswandi
In this project, we develop an automated machine learning (ML) workflow that learns dynamics given time-series measurements. We compare the implementation of our ML approach against the traditional Michaelis-Menten kinetic modeling as a baseline. The workflow we build here consists of data augmentation, automatic model selection, model validation/evaluation, and numerical integrator to solve the initial value problem. Furthermore, we leverage the Michaelis-Menten model to simulate additional training data in order to investigate the performance scaling of our ML method. Once the ML model is learnt, it can be used to predict dynamics, simulate new training data, or extract biological insights. This method is also general: it can be readily applied to any pathway or host/systems.
Costello, Z., Martin, H.G. A machine learning approach to predict metabolic pathway dynamics from time-series multiomics data. npj Syst Biol Appl 4, 19 (2018). https://doi.org/10.1038/s41540-018-0054-3