PhD Project

My PhD project was conducted under the joint supervision of Dr Diego Oyarzún (computational) and Prof Peter Swain (laboratory & computational), at the University of Edinburgh.

My project examined the metabolic cycle in budding yeast. This yeast metabolic cycle is defined by levels of metabolites in the cell increasing and decreasing in regular cycles, which are linked to cell division. Changing feeding conditions for the yeast cells or removing genes from the yeast genome that are responsible for metabolic processes affect the speed of these metabolic cycles.

Not a lot of people study this cycle. Those who do study them in chemostats – essentially, large liquid cultures of many cells – so it is difficult to see whether individual cells behave differently. This is why I used microfluidics to separate the cells. Microfluidics refers to an experimental system that moves around small amounts of liquids – in my case, I used microfluidics to move cells from cultures, trap & separate them, then add or remove nutrients over time.

My project had three parts:

1. Find out ways to analyse the thousands of oscillatory time series I produced. Microfluidics is nice because I got data from many cells. But then I had to clean, process, and categorise this data. I also found out that commonly-used computational methods to compute the periodicity of data, such as the Fourier transform, did not work so well with the noisy biological data I have. My aim was then to find alternative methods, and to combine them with other data-processing steps in a data analysis pipeline. In this code repository, I investigated how noise in the data affect some analysis methods. And using this repository, I used machine learning to tell apart oscillatory and non-oscillatory signals.
2. Show that the metabolic cycle occurs in a variety of cell conditions. My hypothesis here was that the metabolic cycle is something fundamental in the yeast cell, therefore I should be able to see oscillations in metabolite levels when the cells have certain genes removed, or are starved of nutrients. So far, so good. Here is the code that I used to process the data I obtain from my microfluidics experiments to produce reports of the cells in each experiment.
3. Use mathematical models of metabolism to explain why yeast cells have these metabolic cycles. More specifically, I answered the question: does it make sense for the cell to make different building blocks (carbohydrates, proteins, fats) at different times, and do these times fit with the periodicity of the metabolic cycle that I observed in experiments? To do this, I used a genome-scale metabolic model – essentially, a list of chemical reactions that can happen in the yeast cell – and solved linear programming problems to see how fast each reaction should be so that the cell grows as fast as possible. Here is the code that I used to solve these problems.

In addition, I was part of the team that developed aliby, a Python-based pipeline that Prof Peter Swain’s research group uses to analyse time-lapses of images of yeast cells taken from the microscopes in their microfluidics systems. This pipeline finds cells in the images (segmentation), keeps track of cells as they move through time (tracking), match new cells to the ones they divided from (lineage prediction), and creates time series for further processing. I used this software extensively in part 2 of my project.

For me, this was a valuable foray in collaborative software development. Of course, I sharpened my Python skills, but also DevOps: automated testing and deployment of software, managing software versions, maintaining servers for users, and making sure things don’t crash and burn. I learnt how to talk to other developers, project managers, and users (biologist and physicists) of varying programming or technical abilities. Naturally, I learnt to navigate lots and lots and lots of conflicts – both merge conflicts and inevitable conflicts between ideas!

My PhD work eventually led to a paper that showed that single-cell metabolic oscillations are pervasive and may alleviate a proteome constraint. During my PhD, I also worked on a paper on the accuracy and data efficiency in deep learning models of protein expression.