SFU physics student uses machine learning technique to speed up image analysis

August 05, 2021

A team of SFU researchers, led by then-undergraduate student Mathew Schneider in the Department of Physics, have developed a new computational tool capable of analysing images of individual biological molecules up to 10 times faster than before.

Schneider says that the project started with a student award he received from the National Sciences and Engineering Research Council of Canada (NSERC) that allowed him to image collagen proteins using SFU’s Atomic Force Microscope (AFM).

AFM uses a probe that touches the surface of a specimen to provide resolution that is 1,000 times higher than images from a standard microscope.

He began the project using SmarTrace, a semi-manual program developed by former SFU Physics PhD student Naghmeh Rezaei. It traces the thread-like backbones of proteins such as DNA or collagen to determine molecular flexibility important for understanding physical dynamics.

Collagen is a long, chain-like molecule that forms the foundation of the physical structure of our bodies, such as muscle fibres, skin and bones. It is hoped that a better understanding of collagen will explain how DNA can be deformed by proteins and how collagen affects the ageing process.

To measure the backbone, the team used SmarTrace to manually click on points near the protein to provide an initial guess as to where the backbone might be.

It wasn’t until after Schneider had analyzed thousands of protein chains that a colleague suggested that the tedious and time-consuming task might be better handled by machine learning.

Schneider developed specialized software he named AutoSmarTrace to handle this task, training it with 7,600 AFM images he generated. This allowed him to automatically classify pixels within the images as belonging to a part of the chain, background, or other features of the protein.  

Schneider explains that AFM imaging can assign each pixel in an image to a set of categories enabling the team to dig deeper into their category of interest to determine which pixels contained a centered line that goes from one end to the other.

“That line represents an initial guess for the backbone of the chain,” he says.

They were then able to turn the guess into a good trace of the chain’s backbone using the SmarTrace pattern-matching algorithm.

Schneider says, “The AutoSmarTrace algorithm demonstrated a remarkable ability to extract the correct input backbone flexibilities, easily outperforming the manual process that we used previously.” He adds, “We were also able to broaden the applicability of the tool to include electron microscopy and fluorescence imaging.”

The team’s supervisor, physics professor Nancy Forde says it is unusual for an undergraduate to take on a project of this magnitude. “Mathew came up with the idea, led this project, and provided a rigorous quantification of the success of the algorithm. It is an impressive feat, particularly for an undergraduate student.”

She adds that the tool will be a valuable resource to her research group as well as many others’.

Schneider graduated from SFU last spring and is now a master’s degree candidate studying theoretical physics at McMaster University. AutoSmarTrace is freely available on github, and the work is published as a Computational Tool in Biophysical Journal. Click here to read the paper.