How the Technology that Unlocks Your Phone is Accelerating Disease Research

The Problem: If you’ve set up your phone to unlock when it sees your face, then you’re already familiar with image recognition technology. Image recognition software and artificial intelligence have a wide range of applications, especially for disease research. As scientists embark on studies using cells from hundreds of different patients, the need for programs that assess cell quality and survival based on images will be imperative for scaling up experiments.

The Study: A new AI-driven software called Monoqlo developed by NYSCF Research Institute scientists led by Brodie Fischbacher and Daniel Paull, PhD, uses image recognition technology to streamline the cell production process on the NYSCF Global Stem Cell Array®, our automated platform for producing high quality stem cells. The study appears in Nature Machine Intelligence.

The Importance: Monoqlo helps optimize the Array’s workflow and reduce human error, leading to more accurate research and freeing up time for scientists.

“Having automated technologies like the Array, which can produce large numbers of stem cells at a time, is very exciting. However, as we refine workflows and expand new capabilities  the more arduous process development and quality control becomes,” noted Dr. Paull, SVP of Discovery and Platform Development. “We need software that can integrate with our robotic technology to streamline the entire process.”

The Monoclonalization Problem

To conduct accurate disease research, scientists often carry out a process called ‘monoclonalization’ in which they isolate a single cell and allow it to multiply. 

“Monoclonalization is where you take a population of thousands or millions of cells and isolate a single cell,” explained Mr. Fischbacher, the lead author of the study. “The reason you do this can vary from needing monoclonal lines following genome editing, such as with CRISPR-Cas9, or because cell mixtures can be quite varied following the processes to create stem cells. You want to make sure that all the cells that you’re using in a study are as standardized as possible – that they’ve all descended from the same, ancestral cell.”

However, sometimes this process can go awry: a colony of cells descended from the parent cell can mature into the wrong kind of cell or die off. More importantly, there are a number of scenarios in which a polyclonal well will result, meaning that the cell population is descended from more than just one cell. Researchers must triage out groups of cells that haven’t successfully monoclonalized by manually examining images – a time consuming and not very scalable endeavor.

“Historically, the way we can tell if a dish as successfully monoclonalized is by having technicians look at images of the cells that are taken once a day,” remarked Mr. Fischbacher. “That manual image review process is time consuming, and arguably a potential source of technical variability. Every technician is different, and each one may treat the data differently.”

Monoqlo to the Rescue

“Having amassed hundreds of thousands of images from these monoclonalization workflows, it became evident that artificial intelligence had the potential to significantly improve both the accuracy of detecting monoclonal colonies as well as decreasing the amount of time our researchers need to spend looking at images,” said Dr. Paull. “We were fortunate to onboard Zhongwei Wang as a summer intern in 2019 who kickstarted our adventures into image-based machine learning. At the end of a fantastic summer we had the proof of concept that a tool like Monoqlo was achievable.”

“Monoqlo is based on a subfield of machine learning called deep learning,” said Mr. Fischbacher. “By being shown labelled images of cells, it can learn characteristics of cells that are monoclonalized versus ones that have died or matured into the wrong cell type. The idea is that the technology will be able to remove human assessment from the equation so scientists don’t have to spend time making these judgments every day and there won’t be as much variability.”

“What’s also exciting about this type of technology is that it can theoretically be used to optimize many different processes in the lab: things like gene editing or even monoclonal antibody development, which we’ve seen as a potential treatment for COVID-19,” he added. 

Next, the team will continue to optimize Monoqlo’s performance and diversify its capabilities.

“There’s always room for improvement, so we’ll continue iterating on Monoqlo’s performance,” said Mr. Fischbacher. “It could also be useful as a data querying tool: we have hundreds of thousands of images, and if we want to go back and ask any questions about that data, then it’s hugely time consuming. Monoqlo could allow us to automate that process of data extraction, often referred to as “data mining,” which would give us new insights into the biology underlying our stem cell derivation processes.”

“We are thrilled to build upon the work of our talented intern and leverage our large cellular-image-based datasets to drive this project, ” said Dr. Paull. “This work is also a good example of our integrated approach to disease research: not only do we produce fantastic biological results, but through the intersection of biology, hardware engineering, and software engineering (including data science) we are able to develop cutting edge technology that will hopefully benefit workflows for ourselves and for others in the field.”

 Journal Article:

Modular deep learning enables automated identification of monoclonal cell lines
Brodie Fischbacher, Sarita Hedaya, Brigham J. Hartley, Zhongwei Wang, Gregory Lallos, Dillion Hutson, Matthew Zimmer, Jacob Brammer, The NYSCF Global Stem Cell Array Team, & Daniel Paull. 2021. Nature Machine Intelligence. DOI: https://doi.org/10.1038/s42256-021-00354-7

Cover image: Daily scans of stem cells growing in a dish. Image credit: Brodie Fischbacher