Predicting Metabolic Pathway Dynamics with Machine Learning, w/ Zak Costello - #163

Time Series

Time series modeling plays a crucial role in predicting metabolic dynamics, as Zak Costello explains. By using a small but detailed dataset, he models the evolution of metabolite concentrations over time, akin to predicting the path of a ball rolling down a hill 1. This approach allows for the identification of potential issues in metabolic engineering, such as toxicity, by simulating experiments in silico.

If you know how a strain starts out its life, I can tell you how it's going to evolve and ultimately, how much limonene it will end up producing over time.

--- Zak Costello

Zak is optimistic about the future of data-driven predictivity, aiming to bridge the gap between data availability and scientific knowledge 2.

   

ML Techniques

In exploring machine learning techniques, Zak Costello utilizes ensemble learning to model metabolic dynamics. By employing random forests and meta learners like Teapot, he identifies relationships between metabolic states and their derivatives 3. This method allows for the creation of new time series data, enhancing the understanding of dynamic behaviors in biological systems.

We use more traditional machine learning approaches. And for us, that meant, you know, random forests.

--- Zak Costello

Zak finds parallels in other fields, noting shared challenges in understanding dynamic behavior from data, which he finds particularly intriguing 4.