During early development, tissues and organs begin to flourish by shifting, dividing, and growing many thousands of cells.
A team of MIT engineers has now developed a way to predict, minute by minute, how individual cells will assemble, divide and rearrange themselves during the earliest stages of fruit fly development. The new method could one day be used to predict the development of more complex tissues, organs and organisms. It may also help scientists identify cellular patterns consistent with early-onset diseases such as asthma and cancer.
In the study will appear in the magazine today Nature's methodsthe team presents a new deep learning model that learns and then predicts how certain geometric properties of individual cells will change as the fruit fly develops. The model records and tracks properties such as the location of a cell and whether it is currently touching a neighboring cell.
The team applied this model to videos of the development of fruit fly embryos, each of which begins with a cluster of about 5,000 cells. They found that the model could predict with 90 percent accuracy how each of the 5,000 cells would fold, shift and rearrange, minute by minute, during the first hour of development as the embryo transformed from a smooth, uniform shape to more defined structures and features.
“This initial phase is called gastrulation and lasts about an hour as individual cells rearrange themselves on a scale of minutes,” says study author Ming Guo, an associate professor of mechanical engineering at MIT. “By carefully modeling this early period, we can begin to discover how local cellular interactions give rise to global tissues and organisms.”
The researchers hope to use this model to predict cell-by-cell development in other species, such as zebrafish and mice. They can then begin to identify patterns common to all species. The team also predicts that this method could be used to recognize early patterns of diseases such as asthma. The lung tissue of people with asthma looks significantly different than healthy lung tissue. The process of initial development of asthma-prone tissue is unknown, and the team's new method could potentially reveal it.
“Asthmatic tissues show different cell dynamics during live imaging,” says co-author and MIT graduate student Haiqian Yang. “We anticipate that our model can capture these subtle dynamic differences and provide a more comprehensive representation of tissue behavior, potentially improving diagnostics or drug screening tests.”
Co-authors of the study include Markus Buehler, McAfee Professor of Engineering in MIT's Department of Civil and Environmental Engineering; George Roy and Tomer Stern of the University of Michigan; and Anh Nguyen and Dapeng Bi of Northeastern University.
Points and foams
Scientists typically model embryo development in one of two ways: as a point cloud, with each point representing a single cell as a point moving through time; or as “foam”, which represents individual cells in the form of bubbles that slide and glide over each other, much like bubbles in shaving foam.
Instead of choosing between two approaches, Guo and Yang adopted both.
“There is an ongoing debate about whether to model as a point cloud or as a foam,” Yang says. “But they both fundamentally differ in the way they model the same graph, which is an elegant way to represent living tissues. By combining them into one graph, we can highlight more structural information, such as how cells are connected to each other as they rearrange over time.”
At the heart of the new model is a “dual graph” structure that represents the developing embryo as both moving points and bubbles. With this dual representation, the researchers hoped to capture more detailed geometric properties of individual cells, such as the location of the cell nucleus, whether a cell is in contact with a neighboring cell, and whether it is folding or dividing at a given moment.
As a proof of principle, the team trained a new model to “learn” how individual cells change over time during gastrulation in fruit flies.
“The overall shape of the fruit fly at this stage is more or less an ellipsoid, but during gastrulation there are enormous dynamics at the surface,” Guo says. “It goes from being completely smooth to forming multiple folds at different angles. We want to predict all of these dynamics, moment by moment, cell by cell.”
Where and when
In the new study, researchers applied the new model to high-quality videos of fruit fly gastrulation recorded by collaborators at the University of Michigan. The videos are hour-long recordings of developing fruit flies, recorded at single-cell resolution. Moreover, the videos contain labels of the edges and nuclei of individual cells – data that is extremely detailed and difficult to obtain.
“These videos are extremely high quality,” Yang says. “This data is very sparse and allows for submicron resolution of the entire 3D volume at quite high frame rates.”
The team trained the new model on data from three of four videos of fruit fly embryos, so the model could “learn” how individual cells interact and change as the embryo develops. They then tested the model on a brand new video of fruit flies and found that it was able to predict with high accuracy how most of the embryo's 5,000 cells were changing from minute to minute.
In particular, the model can predict the properties of individual cells, such as whether they will fold, divide, or continue to share an edge with a neighboring cell, with approximately 90% accuracy.
“At the end of the day, we predict not only whether these things will happen, but also when,” Guo says. “For example, will this cell disconnect from this cell in seven minutes or eight minutes? We can tell when that will happen.”
The team believes that, in principle, the new model and dual-graph approach should be able to predict the cell-by-cell development of other multicellular systems, such as more complex species and even some human tissues and organs. The limiting factor is the availability of high-quality video data.
“From a model standpoint, I think it's ready,” Guo says. “The real bottleneck is data. If we have good quality tissue-specific data, the model can be directly applied to predict the development of many more structures.”
This work is supported in part by the United States National Institutes of Health.
