Electron microscopy (EM) is a method used for high-resolution images of biological and non-biological samples. It requires precise and time-consuming steps from sample preparation to image acquisition to produce the clarity and details required to visualize small cell structures with high resolution. Additionally, extracting the biological information out of EM-created images is a very laborious and time-intensive task. This is because the current EM analysis software usually requires the skilled eye to examine hundreds of pictures manually.
A team of scientists from the Max Planck Florida Institute for Neuroscience (MPFI) has applied neural networks to create a novel analysis software ‘Gold Digger,’ aimed at streamlining part of the lengthy process. Diego Jerez and Eleanor Stuart, two high school data science students, started working on this project out of curiosity. But later, it turned into a more complex and interdisciplinary project.
In collaboration with the Christie Lab and the Electron Microscopy Core Facility, this project immensely improved upon the established computer-based approach used to analyze EM images’ protein distribution. Naomi Kamasawa, the Head of the Electron Microscopy Core Facility at MPFI, describes the synergy of the collaborative work of their team as a crucial that bridged the gap between these areas of expertise.
To visualize the cells within them, EM demands proteins be labeled with gold nanoparticles, unlike traditional light microscopy that uses fluorescent labeling. The new software uses a deep learning approach to identify gold particles bound to specific proteins of interest.
The team has developed a deep learning-based algorithm that can accurately identify different sizes of gold particles. This fully automated approach speeds up the counting process and generates more precise location information of protein distributions across a membrane, facilitating breakthroughs.
Dr. Michael Smirnov, Neural Data Scientist at MPFI, explains one of the challenges faced in software development was finding a method to train the software to recognize only gold particles that appear dark on an electron micrograph, as opposed to similarly looking shadows caused by the uneven surface of a cell. He further states that the software could distinguish gold particles from these shadow artifacts with near human-level accuracy by providing a large training dataset and correcting errors in the algorithm.
The Gold Digger Software has a small and compact architecture designed to be generalizable and compatible with various EM applications such as magnification, cell type, image area, and gold particle size. It was also used primarily for freeze-fracture replica EM. The software has a user-friendly interface and will be soon open-sourced for future research scope, allowing scientists worldwide to take full advantage of and improve upon this innovative algorithm.