Reductionism has served cell biologists well in the past decades, but systems biologists are now promising to broaden the picture. The challenge will be to provide not just the lists and quantitation but real mechanistic insights.

We've all seen the list papers. There are lists of genes in genomes, lists of transcripts in different cell types, and lists of protein interactions in model organisms. If those eye-glazing lists are all that systems biology has to offer, some old fashioned reductionists are happy to stick to their hypothesis-driven, one-gene-at-a-time experiments.

But the department names continue to change. Methods certainly changed in the 1970s and 1980s as departments of biochemistry morphed to departments of molecular biology and cell biology; now the backers of new departments of systems biology are confident that their new approach will bring a shift in how biology is investigated.

One of those proponents is Lew Cantley, cofounder of the new systems biology department at Harvard Medical School (HMS; Boston, MA). In 1987, Cantley broke out his physical chemistry training and dusted off his differential equations to mathematically describe the flux of phosphoinositide (PI) intermediates his lab observed in cell extracts (Chahwala et al., 1987). The calculated rate constants showed that kinases, rather than lipases or phosphatases, were the key enzymes for regulation of PI turnover. The discovery sent Cantley's lab and the field in a new direction.

The impact of that study is still being felt, says Cantley. “It convinced me that modeling was useful and that this kind of approach could be really critical for making decisions on where to intervene with drugs.”

Cantley and his colleagues Tim Mitchison and Marc Kirschner decided to create a department dedicated to integrating the knowledge gleaned from 30 years of reductionist biology into formal models. The culmination of the reductionist approach has resulted in hordes of information available about biological pathways—genomes, transcriptomes, proteomes, interactomes, databases of genetic perturbations, and corresponding phenotypes.

The question now is how to make sense of it all and, more importantly, whether meaningful hypotheses and discoveries will emerge. Those setting their sights on systems biology say formalized mathematical modeling, simulations, and more quantitative measurements of cellular phenomena in real-time will help clarify the complexity of biological systems. Skeptics argue reductionism hasn't lost its power yet. Both factions may be right.

“We realize we can continue to describe individual components of a system,” says Cantley. “But we'll never really understand it until we can mathematically model it.” The faculty forming systems biology groups describe their approach as a way to formalize the increasingly complex “cartoon” models. Some early studies claim surprising answers about how simple systems work—answers not easily gained from a traditional study.

But the jury is definitely still out on an approach that challenges decades of “in the trenches” biology. After years of building up models of particular systems through integrating data molecule by molecule and interaction by interaction, it is hard for many biologists to see how a computational model will add anything new. And some feel that jumping to quantitative modeling is a bit premature given the limitations of current techniques for measuring cellular phenomena quantitatively.

“We can't yet measure simultaneously and quantitatively cellular events in four-dimensional space,” says Ira Mellman, cell biologist at Yale University School of Medicine in New Haven, CT. “If we can do that, then we can have a data set from which we can write models.” Many systems biologists wouldn't argue with that notion, but say model building and technology development can proceed together.

In fact, they acknowledge it will take coordinated efforts of pushing technology forward and training a new generation of scientists to mix disciplines as well as a few breakthrough “success stories” using this approach to get the systems biology movement off the ground. “We're not there yet,” says Stanislav Shvartsman, a chemical engineer at the Lewis Sigler Institute for Integrative Genomics (Princeton University, Princeton, NJ). “But systems biology is waiting to take off and it will happen quickly.”

There is one research domain—the investigation of network organization and behavior—where systems biology is the only logical approach. Investigators in this area hope to discover the design principles used by Mother Nature to engineer a cell, a tissue, or an organism. Perhaps there are motifs or mechanisms that evolution hit on time and again, and perhaps they will use similar principles as those used by electrical or mechanical engineers to solve problems.

So far, a few studies have shown that mechanisms can be discerned from the behavior of some very well-characterized systems. For instance, the bacterial chemotaxis pathway has a dozen or so known components starting with the chemoattractant and ending with a flagellum motor protein that turns clockwise or counterclockwise. The components can be studied individually, but only by studying the entire pathway as a unit could Stanislas Leibler and colleagues (Princeton University and now Rockefeller University, New York, NY) show that the pathway was robust across varying concentrations of the chemoreceptor (Alon et al., 1999). They determined that this network property of robustness was due to the system's ability to adapt precisely back to steady-state.

A second general network property—how much noise there is in a typical network—was studied by Michael Elowitz (Caltech, Pasadena, CA) and colleagues at Rockefeller University. They used cyan fluorescent protein (CFP) and yellow fluorescent protein (YFP) to build strains of bacteria that let them measure intrinsic and extrinsic noise in gene expression systems (Elowitz et al., 2002). The reporter genes were controlled by identical promoters, so in the absence of intrinsic noise, cells express the same amount of each protein and appear yellow, although extrinsic noise causes each cell to be varying intensities of yellow. When intrinsic noise is present, then the amounts of the two proteins inside a cell become unmatched and cells appear either red or green (Fig. 1). “That is a huge question in biology,” says Pam Silver, a member of the new HMS department. “How much noise can a cell tolerate?”

Rather than dealing with whole-network properties, Uri Alon (formerly with Leibler and now at the Weizmann Institute of Science in Rehovot, Israel) is breaking down networks into functional components. He recently showed that, across networks as different as gene transcription units, neuronal processing circuits, and ecological food webs, distinct “network motifs” accounted for a majority of interactions (Milo et al., 2002). These motifs, such as a feed-forward loop, must therefore be considered the basic building blocks of networks.

But Harlan Robins, a mathematician at the newly created Center for Systems Biology at the Institute for Advanced Study, Princeton, NJ, says that the similarities across networks do not mean that mathematicians have it easy. “Biological networks are different than anything we've seen before,” he says. As such, he says, there is a need for new types of mathematics to faithfully mimic their feedback, feed-forward, and other motifs.

Systems biologists are not only interested in general network properties. They say that systems approaches will be so powerful that everyone will be using them routinely within 5–10 years. What approach gives them such confidence? Systems biology has suffered an identity crisis since the term first started making the rounds. Some view it as number crunching of large, system-wide datasets. Others insist it reinvents physiology—integrating molecular and cellular information to learn how tissues, organs, and organisms behave. Finally, some see the extension of systems biology to include synthetic biology, using principles from natural systems to build engineered systems.

The systems scientists interviewed here, however, all share a common vision. The systems they study are no different from the signaling networks, developmental processes, or organelles tackled by their “nonsystems” colleagues. The difference is in the level of inquiry or “comfort zone” as Alan Aderem, director of the Institute for Systems Biology (ISB) in Seattle, WA, puts it. “The physicist enjoys the beauty of reducing a problem to an atomic level, whereas the cell biologist is enthralled by watching the behavior of an entire cell,” he says. “In order for physicists and cell biologists to collaborate in a meaningful way, they must be prepared to leave their comfort zone and enter that of the collaborator.”

Cell biologists will also have to expand their view beyond the terrain of one or two proteins. In many systems experiments, researchers start with as much of the known picture as possible—the components, their interactions, and system perturbation effects—and formulate each piece into mathematical expressions to build a working model. That data-rich, computational model, they say, will lead to new insights and intuition about how systems work. The hypotheses suggested by the model will be tested in the laboratory, and the experimental results will inform the next round of modeling.

Systems biologists emphasize the need for continued wet biology. “No one will be sitting in a room coming up with theoretical ideas and then hoping someone will test them,” says Cantley. “We want people who can do the experiments and go back and change the model.”

Systems people say this type of reiterative modeling will in the long-run be as popular and easy as BLAST searching. Imagine websites where you enter the known components of a system and how you believe they interact, all in a familiar biological language of “represses,” “activates,” or “coregulates.” Then hit the “model” button and out will come an analysis of how the system should behave. The underlying differential calculus (as with the BLAST algorithm) will be neatly tucked away in the program code.

Systems biologists put so much faith in computational modeling because they say it is the only way to efficiently integrate the large datasets coming out of the genomic revolution (not to mention the overwhelming amount of information generated by three decades of reductionist biology).

“If I have five things that interact, I could draw by hand all the possible ways in which they could interact,” says Radhika Nagpal, a computer scientist and fellow at the new HMS department. “But when I have 25 things that interact, it's overwhelming to draw a picture.” She says formal models give scientists a way to verify each other's “stories” or intuition of how a system interacts. And they can be handed off to others for testing and rearranging “in a way that has been very powerful in other areas,” she says.

John Aitchison, a cell biologist at the ISB, uses large datasets in his study of an organelle system. He follows the biosynthesis of peroxisomes in yeast grown on oleic acid. He looks after switching cells from glucose for changes in the transcriptome and compares that to the proteome of mature peroxisomes. He notes that the two datasets don't always match up—some proteins in the proteome are not induced in the expression data.

“Naysayers would say, ‘What are we supposed to do with these lists? There are 300 genes and only 30 are peroxisomal. This is going to lead people in the wrong direction,’” he says. To make sense of the lists, Aitchison says, systems biologists must be as rigorous about gathering multiple lines of evidence as are cell biologists who typically do “five experiments to prove a point.” The integration of those different lines of evidence would of course have to take place on a different scale, and Aitchison says developing the computational tools to do that will be the real challenge.

“It's absolutely critical because we know any point by itself could be misleading—mass gathering of data is error prone,” he says. Integrating the large datasets would strengthen real effects and reveal spurious data.

Aitchison's transcriptome and proteome lists exemplify the second reason why system models will become indispensable tools—the power of faster, better hypothesis generation. When Aitchison realized there were proteins in the peroxisome not induced during biogenesis, he looked to his model and asked what testable hypothesis could explain this subset of proteins. Were there seed or scaffolding proteins already in place? Or were catalytic proteins acting in tiny, undetectable amounts? The wealth of information available from his systems data—on the existence, abundance, and dynamics of multiple proteins—gave Aitchison a better basis for a model.

“Analyzing computationally can help us devise good, reasonable hypotheses, rather than pulling one out of a hat,” he says. “It limits the number of experiments you have to do and that will move the whole field forward more quickly.” Cantley agrees that new intuition sprung from mathematical modeling will replace the current “seat of the pants” hypothesis generation in cell biology.

A new intuition example comes from Timothy Galitski's laboratory down the hall at ISB. Galitski's group studies differentiation of yeast from the familiar budding form to the filamentous, invasive form. His team, which includes software engineers and physicists, integrated data on gene expression changes, filamentation genes known by phenotypic studies, and molecular interactions into a modular abstraction network (Prinz et al., 2004).

Each of the 47 network modules represented an underlying cluster of interactions responsible for a particular cellular function (Fig. 2). The software names module nodes after the gene in the cluster with the most interactions (e.g., the ACT1 module represents the genes involved in reorganizing the actin cytoskeleton during filamentation). The abstracted network visually represents roughly 1,000 “filamentation” genes and their 700 known interactions.

“Instead of looking at long lists or doing a haphazard gene search, this approach brings forward information in the network that is most likely to inspire specific molecular hypotheses,” says Galitski. In fact, one module that emerged was the protein that makes the lid for the cell's trash can, the proteasome. Although the protein only connected to two other proteins in the network, its prominence as a module node meant that it was probably coordinating proteasome function during filamentation. This led the group to hypothesize that cell cycle proteins entering the proteasome for degradation may be regulated by the lid protein rather than the ubiquitination step. Using classic genetics and cell biology experiments, the group showed that was indeed the case.

Galitski says as a cell biologist he's no longer interested in just knowing all the molecular players and how they connect. “I'm interested in having the computer do that to get me quickly and systematically to a level in the data where I can extract insights [about the system].”

Silver says she “grew up in the era of ultra-reductionist biology.” But now, she says, with sequenced genomes and microarray data, it makes sense for people to rethink their approach to biological questions. At the same time, she warns, “no one should think [reductionism] is being substituted by systems biology.” The ideal analysis of a system, Silver, Galitski, and others say, combines a top-down and bottom-up approach. For instance, in Galitski's study, a top-down view gave the 47 network modules important for filamentous growth. A bottom-up study of one module clarified how the lid protein and its cluster partners regulate growth.

Whether top-down or bottom-up, the conceptual design of a systems study may still be reductionist in character. “I think the ‘reductionist/holist’ dichotomy is a red herring,” says Jeremy Gunawardena, a mathematician and director of the Virtual Cell Program in the new HMS department. Systems biology, he says, “is going to be just as reductionist because any experiment has to be an exercise in reductionism—focusing on some features of the system, while carefully controlling all others.”

The ideal analysis approach in systems experiments will involve looking in single cells in real time. “Why would you expect that what goes on in a single cell would be the same in a population of cells?” asks Gunawardena. “They behave quite differently.” In one recent example, Philippe Cluzel (University of Chicago, IL), and colleagues found that the signal transduction process in bacterial chemotaxis shows greater temporal fluctuations if viewed at the level of single cells (Korobkova et al., 2004). Gunawardena says the “single most important hurdle to systems biology” will be developing the experimental techniques that will allow scientists to make real-time, quantitative measurements on single, living cells.

One approach to this problem is fluorescence correlation spectroscopy (FCS). In FCS, researchers measure the fluorescence of labeled proteins as they move into, through, and out of a tiny volume. The amplitude of any fluctuation in the fluorescence of this volume depends (inversely) on the average number of molecules in that volume, and thus the concentration of the molecules. The persistence time of the fluctuations gives a measure of molecular mobility, which is in turn dependent on whether other proteins are interacting with the labeled protein.

Leibler's group has used FCS in further studies of the chemotaxis pathway. They fused an inducible chemotactic signaling protein to green fluorescent protein (GFP) and measured its cellular concentration using FCS (Cluzel et al., 2000). At the same time, they attached beads to the bacterium's flagella and recorded its direction of movement by microscope. The output of individual motors could be measured as a function of the cellular concentration of the signaling protein (Fig. 3). The study tied together the macroscopic output of an entire pathway with the fluctuations of an individual molecule, and revealed that the motor molecules provide the signal amplification seen in chemotaxis.

Although many agree that FCS is a step in the right direction, it will be imperative to move beyond simply tagging two or three proteins in isolated cells. And better technologies will be needed to measure posttranslational modifications quantitatively. John Bergeron, president-elect of the Human Proteome Organization (McGill University, Montreal, Canada), cautions not to rush to the conclusion that we know most or even much of the cell “parts list.” “That's certainly not true for the yeast or human genomes. With all the variants of genes and protein modifications, [each cell] has a personality of its own.”

There are also challenges in measuring cellular responses in three-dimensional tissues. “I'd like to see more attention paid to developing better imaging methods,” says Mellman. “There should be more analogue before we digitize into equations.” He wants researchers to work on measuring multiple parameters simultaneously in live tissues and organs, and thinks those parameters should include macroscopic measures such as cell movement, orientation, and division. But solving the imaging problem won't be the end of it. “As soon as you take one cell and put it with a bunch of neighbors talking to each other, [then you have] simultaneous differential equations whose coefficients we don't have the capacity to measure,” he says. Meanwhile, data variability vexes analysis of everything from array results to imaging.

Aderem, director of the ISB, takes a slightly different view. “We can always say there is something we cannot [yet] measure,” he says. But he is bringing together people from physics, computer science, math, and biology so that they can simultaneously do the systems biology and develop the necessary technology.

He says it's not enough to bring together people from different disciplines under one roof—they have to work on the same projects. And for this reason, he sees wet bench work and computational work as highly complementary. “Somebody in a wet lab would say, ‘I would be able to answer this question, except I cannot get the rate constants with technology available,’” he explains. “Next thing you know, a physicist is developing an adjustment to a microscope that permits those measurements.”

To complete the cycle, that technology would hopefully also enable other biological systems questions to be answered. Shvartsman at the Sigler Institute says not only do you have to bring people together to work on projects, but now the field needs people who are willing to “combine skills in themselves and speak several [scientific] languages.”

And although he admits both the technologies and workforces are not quite ready yet for a systems biology explosion, Shvartsman believes it is only a matter of time before enough “success stories” will win over the skeptics. Mark Ptashne (Sloan-Kettering Institute, New York, NY), molecular biologist and pioneer in gene regulation, says he's not convinced. He says no fundamental discoveries were ever made from systems analyses, but instead usually came from left field.

“We don't know what we don't know,” he says. He explains that work already done in lambda phage, bacteria, and yeast has given us the basic mechanisms and so it is not hard to believe when those mechanisms come out of a systems study. “But without the genetics, I don't know how confident you are of the patterns [emerging].”

The systems scientists are very confident, however, that their combined approach of computational modeling backed up by experimental data will catch on—some even compare it to the molecular biology tools that changed the way biologists asked their questions. Others say it may be an approach only a few people will be comfortable with pursuing and that there's still plenty of work to be done in the reductionist tradition. However, perhaps the view taken by the faculty starting these systems groups is the happiest medium. “It's not that we'll never have to do experiments,” explains Nagpal. “But designing experiments is an act of ingenuity and we have to decide how to set up the best experiments.” Who could argue with that?

Alon, U., M.G. Surette, N. Barkai, and S. Leibler.
. Robustness in bacterial chemotaxis.
Chahwala, S.B., L.F. Flieschman, and L. Cantley.
. Kinetic analysis of guanosine 5′-O-(3-thiotriphosphate) effects on phosphatidylinositol turnover in NRK cell homogenates.
Cluzel, P., M. Surette, and S. Leibler.
. An ultrasensitive bacterial motor revealed by monitoring signaling proteins in single cells.
Elowitz, M.B., A.J. Levine, E.D. Siggia, and P.S. Swain.
. Stochastic gene expression in a single cell.
Korobkova, E., T. Emonet, J.M.G. Vilar, T.S. Shimizu, and P. Cluzel.
. From molecular noise to behavioural variability in a single bacterium.
Milo, R., S. Shen-Orr, S. Itzkovitz, N. Kashtan, D. Chklovskii, and U. Alon.
. Network motifs: simple building blocks of complex networks.
Prinz, S., I. Avila-Campillo, C. Aldridge, A. Srinivasan, K. Dimitrov, A.F. Siegel, and T. Galitski.
. Control of yeast filamentous-form growth by modules in an integrated molecular network.
Genome Res.