An AI that can predict cell structures
New 3D models of living human cells generated by machine-learning algorithms are allowing scientists to understand the structure and organization of a cell's components from simple microscope images.
Why it matters: The tool developed by the Allen Institute for Cell Science could be used to better understand how cancer and other diseases affect cells or how a cell develops and its structure changes — important information for regenerative medicine.
"Each cells has billions of molecules that, fortunately for us, are organized into dozens of structures and compartments that serve specialized functions that help cells operate," says Allen Institute's Graham Johnson, who helped develop the new model.
What they did: The researchers used gene editing to label the nucleus, mitochondria and other structures inside live human induced pluripotent stem cells (iPSC) with fluorescent tags and took tens of thousands of images of the cells.
- They then used those images to train a type of neural network known as Generative Adversarial Networks (GANs). That yielded a model that can predict the most likely shape of the structures and where they are in cells based on just the cell's plasma membrane and nucleus.
- Using a different algorithm, they created a model that can take an image of a cell that hasn't been fluorescent-labeled — in which it's difficult to distinguish the cell's components ("it looks like static on an old TV set," Graham Johnson says) — and find the structures.
What they found: When they compare the predicted image to actual labeled ones, the Allen Institute researchers said they are nearly indistinguishable.
The advance: Gene editing and fluorescent dyes often used to study cells only allow a few components to be visualized at once and can be toxic, limiting how long researchers can observe a cell.
Plus, "knowledge gained from more expensive techniques or ones that take a while to do and do well can be inexpensively applied to everyone’s data," says the Allen Institute's Greg Johnson, who also worked on the tool. "This provides an opportunity to democratize science."