UCLA researchers develop artificial intelligence to analyze cells without killing them
UCLA researchers developed a model using artificial intelligence to analyze a cell and identify its protein content without harming the cell itself (Justin Jung/Daily Bruin senior staff).
UCLA researchers developed an artificial intelligence model that allows them to analyze the appearance of a cell and determine its protein content without killing the cell.
Before cells are injected into a patient’s body, they have to go through quality control to assess if they will be effective and are safe to use, said Sara Imboden, a graduate student from Switzerland who is conducting research at UCLA. Currently, quality control is conducted by labeling cells with fluorescent tags, a time-intensive and expensive process that also kills the cells, she added.
A group of researchers from the UCLA Henry Samueli School of Engineering and Applied Science published a study in March that used images of stem cells – specifically, a type of stem cell extracted from bone marrow known as mesenchymal stromal cells – to develop the artificial intelligence model, said Neil Lin, an assistant professor of mechanical and aerospace engineering and the lead of the study.
These types of stem cells can be used for regenerative therapies and autoimmune disease treatment, Imboden said.
The AI model consists of two AI networks, said Cho-Jui Hsieh, an assistant professor of computer science who specializes in machine-learning algorithms.
The first network is fed colorful immunofluorescent images and black-and-white microscope images from the same field of view, Imboden said. It then learns by detecting the relationship between the two inputs, she added. After studying the images, the generator attempts to output an image as close to the colorful image as possible, she said.
Then, a second network compares the predicted image to the AI-generated one and guesses which one was fake, Imboden said.
The second network points out weaknesses of the output image created by the first network, which allows the model to improve, Hsieh said.
The model uses a similar technique as deepfakes, which are images or videos using AI to replace the person in the original media with a replica. Deepfake videos featuring TikTok user Chris Ume @deeptomcruse imitating actor Tom Cruise went viral in March 2020, according to The New York Times.
“It’s pretty much the same mechanism,” Lin said. “In the Tom Cruise deepfake, you train your own images as the input and the Tom Cruise real photos as the target. But in our case, the inputs are the black-and-white images of the cells … and the protein images are the targets.”
The AI model analyzes subtle differences the human eye can’t detect and predicts what proteins are present, he said.
One limitation of this research is the AI was developed for a very specific subset of cells derived from one patient, said Marie Payne, a mechanical engineering graduate student who contributed to the study. Future studies could examine how the model responds to cells taken from another patient or under different conditions, she said.
AI models developed for one type of cell can make incorrect predictions when tested on another cell, Hsieh said.
“No matter what kind of input you give, the AI will always try to predict something,” Hsieh said. “Measuring the faithfulness of the AI prediction is very important future work.”
Future studies could also improve the model’s effectiveness, Payne said.
The study’s researchers hope that machine learning models will become more prominent in the medical field.
“I can see AI being in every single subset of medicine,” Payne said. “In this paper we focused on cells, … but we’d like to use an AI approach for the whole tissue scene.”