Researchers team up to teach computers to predict cell fate.
To answer his questions about cell differentiation, Michel Cayouette (left) needed just one impossible thing: the ability to predict the future. He and his colleagues knew that the population of cells that eventually form a rat's retina comprised highly different subpopulations. Some retinal progenitor cells (RPCs) would undergo self-renewing divisions to produce another RPC and a differentiating cell; other cells would divide terminally to produce various differentiated cell types: light-sensing photoreceptors, signal-transmitting bipolar cells or signal-coordinating amacrine cells.
Cayouette and his postdoc Francisco Gomes knew that gene expression patterns differed among RPCs before these various types of divisions but not how (or whether) distinct patterns matched daughter cells' fate. It was a Catch-22. Analyzing cells before division destroyed the cells, leaving no way to know what cell division would have yielded. And waiting until after the cell divided was not useful either because Cayouette and Gomes wanted to know the conditions leading up to a cell's decision.
They tried making movies of cells, hoping that the way a cell migrated or vibrated during the days before cell division held clues as to what the division would yield. But after watching videos of hundreds of individual cells, the scientists could come to only one firm conclusion: their eyes alone made lousy soothsayers.
Then a colleague, Sally Temple, told Cayouette to call Badrinath Roysam, a computer scientist at the Rensselaer Polytechnic Institute. He told Cayouette that his team had developed ways to very accurately quantify how cell features change over time. The Montreal researchers shipped several hard drives of cell images to Troy, New York, USA and waited to learn whether a computer algorithm could be trained to do what their eyes could not.
“As long as you can track the cells, you can do this.”— Michel Cayouette
The conditions for collaboration were ideal, says Cayouette: “We had a hypothesis, and they were interested in developing methods of image analysis. It was a perfect match of our interest, and everybody wanted to contribute.” That does not mean it was always easy for the biologists and computer scientists to communicate. When Roysam spoke at length about 'dynamic features', for example, it took a while for the biologists to realize that he was simply referring to how cells behaved in a dish.
The single most exciting day was getting the results of the first prediction. Gomes had sent a hard drive of cell-imaging movies to Roysam's graduate student Andrew Cohen without including the outcomes of cell divisions. Cohen sent an analysis back to Montreal, saying that the cells clearly fell into two distinct groups and giving predictions for each. “I picked up the phone,” Cayouette recalls, “I said, 'This is amazing. Let's see if we can do this again'.” Ultimately, the method they named algorithmic information theoretic prediction proved itself jaw-droppingly accurate: it could predict whether a cell would undergo a terminal or self-renewing division with 99% accuracy, and the combination of offspring (photoreceptor, amacrine or bipolar cell) was correct 87% of the time.
Now that a particular cell's division can be tied to a very likely outcome, Cayouette can move on to some big questions. What genes are responsible for making certain types of neurons? Could manipulating these genes let researchers make desired cells in sufficient quantity and quality for therapy or drug discovery or better basic research?
These questions should be accessible not just for RPCs but for any cell that generates new cell types. Roysam is making the relevant software available in the hope that others will use and improve it. For biologists, Cayouette says, the main difference is the requirement to grow the cells on a microscope stage in such a way that a computer can be taught to recognize and follow individual cells. Predictive algorithms will be more difficult in cells growing in dense cell cultures or in three-dimensional matrices, but Cayouette is confident that can be overcome: “As long as you can track the cells, you can do this.”
Cohen, A.R., Gomes, F.L.A.F., Roysam, B. & Cayouette, M. Computational prediction of neural progenitor cell fates. Nat. Methods 7, 213–218 (2010).
About this article
Cite this article
Baker, M. Michel Cayouette and Badrinath Roysam. Nat Methods 7, 165 (2010). https://doi.org/10.1038/nmeth0310-165