Using machine learning to see inside live human cells

Looking at human cells isn’t as simple as it sounds, especially when scientists want to examine clearly, in 3D, how multiple structures change over a time while cells are alive. Now, the Allen Institute for Cell Science has created a new tool, enabled by machine learning, that ticks all those boxes.

Meet the Allen Integrated Cell, “a new way to see inside living human cells,” as Rick Horwitz, Ph.D., executive director of the Allen Institute for Cell Science, put it.

Funded by Microsoft co-founder and billionaire philanthropist Paul Allen, the Allen Institute previously created a large collection of human-induced pluripotent stem cells (hiPSCs), gene-edited with fluorescent tags that illuminate major structures. Scientists took pictures of tens of thousands of these glowing cells and trained two artificial intelligence algorithms to understand them.

First, one algorithm studied the shape of the plasma membrane, the nucleus and other structures to learn their spatial relationships. A powerful probabilistic model emerged from this training. With only a fluorescent image of cell membrane and nucleus, it can accurately predict the most probable shape and location of other structures within single cells.

Second, a model uses what it learned from those fluorescent cells to find cellular structures in cells, but without those labels. It only needs relatively easy-to-collect brightfield microscope images to predict, in real time, where fluorescent labels are, Molly Maleckar, Ph.D., director of modeling at the Allen Institute for Cell Science, told FierceBiotechResearch.

The result is a real-time 3D image of cells on whole fields—instead of single cells—as if they had been fluorescently labeled. Viewed side by side, the images generated by the label-free method look nearly identical to the fluorescently labeled photographs of cells, according to the institution.

Why is the label-free analysis important? The fluorescence microscope system is expensive, for one thing. For another, researchers could only see limited numbers of cell structures simultaneously—and over a relatively short period of time, because the cells die as they're blasted with lasers, an integral part of the fluorescence microscope, explained Maleckar.

The cheaper brightfield method, on the other hand, isn’t powerful enough by itself to provide clear identification of cell structures.

Allen Institute’s new tools combine the advantages of both methods. They should help scientists save money and time, and allow them to observe cell structures over long periods of time without having to worry about the toxic effect of lasers, said Maleckar.

“[T]he Allen Integrated Cell is like the ultimate free lunch. We get to sample a ‘buffet’ of many different proteins and organelles, without having to label anything at all. This opens up a totally new and much more powerful way of doing cell biology. It’s a total game changer,” said Michael Elowitz, Ph.D., professor of biology, bioengineering and applied physics at California Institute of Technology, in a statement. 

RELATED: Altering frog embryos' bioelectric patterns 'rescued' them from birth defects

Although the models are based on hiPSCs, Maleckar said they could feasibly be applied to any human cells if training data are available. “We’re current exploring how well these models apply to images of different cell types, but the real power is using these models to determine what changes in cell organization from one cell type to another,” she said.

The tool also could be applied in drug discovery and disease research. The probabilistic model allows researchers to determine spatial relationships in single cells between structures and how they change in the presence of drugs. And the label-free model makes it possible to measure the dynamics of cells over long time periods, while drastically simplifying sample preparation.

Like all Allen Institute tools and resources, the Allen Integrated Cell is publicly and freely available on the Allen Cell Explorer at All code is open source and downloadable from GitHub.