The microscope, one of the earliest scientific tools, allowed researchers to collect information about a cell's shape and features to begin to understand this basic unit of life. Later, by knocking out genes and observing the resulting phenotypes, researchers could identify associations between genes, moving toward the challenging goal of mapping signaling networks.

Most of these studies have been qualitative, but a group of researchers at Harvard Medical School wanted to build these genetic relationships quantitatively. In a recent article in Science, Norbert Perrimon, George Church and colleagues report a method to study the roles of individual genes in regulating cell morphology and to then use this information to describe signaling networks.

The authors performed a screen, exposing cells to one of 249 RNA interference (RNAi) or gene overexpression treatment conditions, and stochastically labeled cells in a culture with a fluorescent protein, which allowed them to accurately analyze several cells, rather than extracting information from all the cells in a field of view (Fig. 1).

Figure 1: Cell identification.
figure 1

Micrograph of cells stochastically labeled with GFP (left) and a map of the cells in the field of view identified by CellSegmenter (right). Images courtesy of Bakal and Aach.

Then they designed image-processing techniques to detect cell boundaries. “Most of the automated image-analysis programs that people use to do segmentation still weren't applicable for our needs because they approximate the boundaries and the shape,” explains Chris Bakal, a postdoc in Perrimon's lab and co-first author on the paper with John Aach from the Church lab.

Although accurate cell boundaries could be obtained by simple intensity thresholding of fluorescent cells, the complex cell shapes and variable fluorescence often required setting a different threshold for each cell. To manage this, they developed a software application called CellSegmenter, designed by Aach, which allowed them to vary and set intensity thresholds, and select cells using simple point-and-click operations. “It's semi-automated, as we call it,” says Bakal, “which is just a fancy way of saying that we do it by hand.”

For each of the over 12,000 cells analyzed in this work, the cell images captured by CellSegementer were evaluated using image-analysis algorithms to yield 145 morphological features. Using neural networks and clustering techniques, the researchers separated cells with similar signatures into 'phenoclusters', which then allowed them to describe a signal network that regulates cell protrusion, adhesion and tension. “The computational and statistical approach that we've applied after data extraction was what led us to some good insights,” notes Bakal.

One of the goals is to study more genes using this technique, so the group is trying to automate the process by using confocal microscopy to obtain images with less variability between them and thus eliminate the need for the manual correction. Bakal also envisions using this technique for small-molecule screening—generating a cell signature after treatment, and integrating it with signatures from RNAi studies to make predictions about the small molecule's targets.

This work also has potential application as a diagnostic tool. Just as scientists observe cells, pathologists examine cellular shape and features in clinical samples to determine the identity of a disease or what oncogene may be expressed. “Taking that a step further,” hypothe-sizes Bakal, “we could make this quantitative in nature—take clinical samples, for example, assign a quantitative signature to their shape, and then by integrating that information with our database, match it up with cells overexpressing oncogenes or cells in which a tumor suppressor gene has been knocked out.”

For now, CellSegmenter can be used on any image for which intensity thresholding can be used. “It's a very simple tool that allows a lot of user interface, so that simplicity allows you to do many different kinds of things with it” says Bakal.