Contact: Paul Preuss, [email protected]

“Cell biologists love a good microscope image,” according to Mary Helen Barcellos-Hoff of Berkeley Lab’s Life Sciences Division, “but a radiation biologist is likely to say, ‘Well that’s pretty, but what does it mean?'”

Barcellos-Hoff, trained as a microscopist, studies the effects of low-dose radiation and other factors on cells and their environment. “Radiation biology is classically a quantitative field. It has been difficult to put new information gained from microscopy into quantitative form.”

Bahram Parvin of the National Energy Research Scientific Computing Center (NERSC) has a particular interest in feature-based representation of scientific images. “How do you represent images so that such a representation reduces the data volume, and at the same time it is information preserving? What quantifiable insights can you obtain from an image collection?”

Now Barcellos-Hoff and Parvin have given quantitative meaning to microscope images in a way that promises far-reaching consequences for the field they have dubbed “phenomics,” which deals with the proteins a genome codes for — how proteins are regulated and expressed in the cell and how they interact with each other to condition the cell’s responses to outside stimuli and a changing microenvironment.

CA BioSig user’s view of raw and processed images from an in vivo study, with corresponding annotations. The protein expression shown in green in the left image is heterogeneous for cells near the structural feature of interest, called a lumen, a dark hole surrounded by cells. The image on the right shows cellular classification as a function of proximity to the lumen.

Called BioSig, the system Parvin and his NERSC colleagues Qing Yang and Gerald Fontenay have created can process and annotate numerous images of multicellular systems. While Yang focused on developing unique algorithms to compute meaningful features from images, Fontenay concentrated on annotating experimental images with computed features, accommodating user input, and making information available over the web. BioSig has already proved its worth in a variety of image-based studies; it can be used by researchers working alone or together on a wide range of problems.

Cellular semaphore
“Now that the human genome has been sequenced, now that we’ve entered the era of functional genomics, the problem of quantitative imaging has become critical,” says Barcellos-Hoff.

BioSig’s specialty lies in keeping track of the myriad proteins that cells use to send and receive signals, necessary for understanding the forms of cells and cell structures and their organization in tissues. The first step in any program that wants to link quantitative information to images of cells is transforming them so that they can clearly and consistently show where proteins are at work, e.g., which cell in the tissue? Which compartment in the cell? And how much protein is expressed in each cell type?

Any biological system exhibits significant variations to begin with. Coupled with technical variability in sample preparations, underlying image representation becomes quite heterogeneous: staining is nonuniform, images are noisy, and subcellular compartments can overlap each other. Each compartment within the cell must be clearly separated from its neighbors in a process called segmentation.

From 2-D confocal microscope images, BioSig computed these 3-D views of a colony of cultured cells and their gap junction.

Yang and Parvin developed means of segmentation that remove noise and clearly outline each cell or nucleus, in two-dimensional images as readily as in 3-D ones. Their method is completely automatic and does not require human interaction, as other segmentation programs do. The routine recognizes curved sections of the envelopes or membranes of nuclei; by calculating individual “centroids” from this information, adjacent objects can be distinguished even when they overlap. With further calculation, the location and magnitude of expressed proteins — identified by specific antibodies — can be resolved with great precision.

Degrees of freedom
One challenge in getting a quantitative result from an information-based imaging system is keeping track of its many degrees of freedom — where a sample was obtained, how it was exposed, how it was stained, and many more — by which an image can be related to the database.

“For example, one radiation study compared the wild strain of mice with a genetically altered strain,” Parvin says. “Images were collected from mammary sections of mice after they were sacrificed, at intervals of an hour after exposure, four hours after, and eight hours after. How do you compose a picture of time-varying quantitative information, within and across species, for many experimental variables?”

The answer Parvin and Barcellos-Hoff came up with was to annotate each image with a wide range of factors, many of which come straight out of the experimenter’s lab notebook and can be displayed in a separate window on the BioSig screen. This aspect of the work required the computer science team to gain a detailed understanding of the experimental protocol, a process that involved several sessions with different scientists in the Life Sciences Division.

The BioSig user can view a coarse representation of BioSig’s data model, shown as a graph, and can click on each object to view its content in more detail.

“You want to represent query information visually in terms of plots or images,” says Parvin. “But requirements are always changing and new information needs to be added. So the questions facing us were, how do we accommodate such changes in the future? How do we visualize a query result in a way that is informative?” One helpful approach, he remarks, was that “we leveraged the latest dot-com innovations.”

“One of the best things is, you don’t have to redo an experiment to ask new questions,” says Barcellos-Hoff. “If you start off looking for one kind of feature in BioSig, but some other feature looks more interesting, you can query the database for just that. For the first time you can follow many different threads through the tapestry.”

Putting BioSig to work
Two applications illustrate BioSig’s ability to track different variables in complex situations: an in vitro study of how colonies of human breast cells are influenced by the extracellular matrix, and an in vivo study that looked at two important regulatory proteins interacting to control DNA repair mechanisms in mouse mammary cells.

In in vitro experiments, Barcellos-Hoff and her colleagues observed the formation of hollow spheres called acina (Latin for “berries”), a process starting with a single cell in a sheet of human cells of the kind that form the lining of ducts in the breast. The researchers used BioSig to correlate numerous images showing the locations of key regulatory proteins as the acina evolved over a period of 10 days.

When challenged by radiation alone, or a protein modifying factor alone, the location of one of the key proteins was disturbed. When cell cultures were subjected to both challenges, however, results were dramatic — the cells failed to express either of the key proteins, and the acina colonies lost their symmetry, seeming to collapse into their hollow centers.

Low-dose radiation and a transforming growth factor effect the organization of cells in mouse mammary tissue structures. To measure symmetry, BioSig segmented the nuclei, then fitted an ellipse to them. An untreated sample maintains symmetry along the lumen; a treated sample loses its symmetric organization.

In the in vivo experiment, the researchers examined the role of extracellular factors on the distribution of a DNA repair protein, p53 — the so-called “guardian of the genome” — when an organism is subjected to ionizing radiation. By collating and comparing thousands of images, BioSig clearly established a link between extracellular conditions and intracellular response.

When what’s fuzzy wasn’t
To come up with BioSig, Parvin and Barcellos-Hoff had to combine biology and computation (what biologists call informatics) in a new way. “In the beginning we spoke two different languages, in which the same words might have different meanings,” Parvin says.

Barcellos-Hoff agrees. “When I spoke of a fuzzy image, Bahram translated that into a particular distribution of pixels. Sometimes we had to reiterate three or four times to get our concepts lined up.”

Other kinds of learning were involved as well. “As a microscopist, I was trained to eliminate information, to ignore everything that didn’t relate to what I was looking for,” says Barcellos-Hoff. “That’s a tremendous waste of information — the opposite of what we’re doing with BioSig.”

But, says Parvin, “this is how you break new ground. One aspect of progress comes from combining multiple disciplines.”

The ability to save and access vast amounts of quantitative information in images makes them useful in ways never before practical, for example by lending statistical significance to population evaluations. It also constitutes what has been called a “hypothesis-generating data model” — the potential for testing new ideas by querying experimental information in an existing database, and even importing “legacy” data, which may have been gathered for a completely different purpose.

These are among the reasons BioSig research has long been supported by the Office of Biological and Environmental Research in the Department of Energy’s Office of Science. In addition, all of BioSig’s algorithm developments have been funded through the office of Berkeley Lab’s Director.

Results are described in detail in the article “BioSig: an imaging bioinformatic system for studying phenomics,” by B. Parvin, Q. Yang, G. Fontenay, and M.H. Barcellos-Hoff, in the July, 2002, issue of Computer, published by the Institute of Electrical and Electronic Engineers (IEEE).

Additional information: