Introduction

Evolutionary theory and systems biology have both transformed the understanding of biology, the first over many decades and the second more recently. Still, although systems biologists routinely learn about genetic networks by comparing them between related species, relatively little is known about how evolution shapes those networks. At the same time, although theoretical studies have explained many features of evolution, they often assume a simple relationship between genes and reproductive success, neglecting the rich biological networks that act during development to transform genetic potential into phenotypic reality. A July 1, 2008, symposium sponsored by the Systems Biology Discussion Group illuminated what these fields have to offer each other by highlighting research in evolutionary systems biology.

Evolving gene regulatory networks

Aviv Bergman recently founded the Department of Systems and Computational Biology at the Albert Einstein College of Medicine. He has explored the interactions between systems and evolutionary biology in contexts ranging from abstract mathematical systems to cohorts of human centenarians.

Bergman contrasted the role of variability in evolutionary and systems biology. Molecular and developmental biologists, he said, often take advantage of the fact that phenotypes of natural populations are surprisingly robust, in spite of significant genetic variation. In contrast, Bergman said, evolutionary biologists "inhale variation. We cannot survive without variation."

"There is a deep connection between developmental genetic complexity and insensitivity to mutation."

Bergman and his collaborators studied a mathematical model to explore the interaction between networks and evolution, which act on very different time scales. The researchers represented expression levels of various genes as a vector whose steady-state values serve as the phenotype on which selection acts. The underlying network was represented as a matrix whose elements capture the effect of each gene product on the expression of each other gene. "I can create practically any transcription regulatory architecture with this type of interaction," Bergman stated. Evolution occurs as new generations are created by swapping elements in the current generation of matrices and through mutations. Bergman noted that the difference between sexual and asexual reproduction "does not really matter" for this process.

The researchers explored the robustness of the resulting systems by examining changes in response to mutations. "If the system is robust, then a mutation will not cause any perturbation to the outcome," Bergman commented. This comparison is meaningful even in cases where there is no selective pressure, he stressed. "We can assess [robustness] even when there is no optimum to check against."

Bergman and his coworkers found that robustness increased spontaneously over time, even in the absence of selection. In work with Mark Siegal, he found that more complex systems—those with a higher fraction of nonzero matrix elements—evolved to be more robust than less complex systems. "There is a deep connection between developmental genetic complexity and insensitivity to mutation," Bergman observed.

Clues to longevity

Technical advances have made human biological data cheaply available, Bergman said. "This economy makes it possible for humans to become a model organism," he claimed, one whose phenotypic variations are much better understood than those of traditional model species. "The availability of data will close the gap between the phenotypic and the genotypic knowledge."

Possibilities for changes in the frequency of a functional genotype with age.

Einstein College of Medicine researchers have been tracking a large cohort of very-long-lived Ashkenazi Jews, using their not-as-old offspring as controls. "The goal of this project is to identify the genetic and epigenetic components contributing to increased lifespan and involved in age-related disease," Bergman said.

Both aging and evolution winnow deleterious genes over time, Bergman observed, although aging acts on the current population while evolution modifies subsequent generations. "When a population is aging, the killing genes, those that kill you, are going to be depleted," he said, "while the longevity genes might increase in the population."

In the study, however, some gene variants that contribute to age-related diseases declined at first but then increased in abundance in the very old population. "The hypothesis that we are going to work with is that centenarians are endowed with robustness, or a buffering mechanism, which suppresses the deleterious effect of age-related diseases and allows the accumulation of deleterious allelic variation." Although this mechanism makes it harder to assign genes as beneficial or deleterious, it clearly demonstrates that robustness is an important side effect of evolution.

Bergman also explored how a biological network's ability to harbor genetic variation might affect cancer. Functionally compromising any gene that is a member of a complex regulatory network can phenotypically reveal this hidden genotypic variation. Bergman and colleagues hypothesized that cancers, specifically of the head and neck, result from the breakage of robustness, and devised mechanisms to identify the biological elements responsible for the disease.

When robustness breaks

Several years ago Mark Siegal of New York University collaborated with Bergman to study how robustness breaks down, reiterating Waddington's half-century-old observation that wild-type organisms are less variable than mutants. This robustness of phenotype persists in spite of external and internal variations, which Siegal divided into four types. Organisms may have genetic differences, for example, or their overall environment may differ. Even if their overall conditions are identical, the "micro-environment" of each individual could vary. Finally, many biological processes, such as the partitioning of a small number of biomolecules among two daughter cells, have intrinsic stochasticity to which the cells must respond. This raises an important question for evolutionary biologists, Siegal said: "How do new traits emerge if organisms are so robust to begin with?"

"How do new traits emerge if organisms are so robust to begin with?"

Specific genes may help to buffer the effect of variation. For example, impairment of the heat-shock protein Hsp90 is well known to increase phenotypic variation, leading Rutherford and Lindquist in 1998 to designate it a "phenotypic capacitor." "The inference is that Hsp90 is normally buffering whatever genetic differences exist between these strains and making sure that development happens reliably in these strains despite these genetic differences," Siegal commented. He and Bergman found that many other genes have the same effect when they are knocked out.

Siegal and his postdoc Sasha Levy decided to systematically survey phenotypic capacitors in a model organism. "We wanted to screen every gene in an organism to see whether knocking out that gene increases phenotypic variation," as it does for Hsp90, he said.

The researchers chose to study the yeast Saccharomyces cerevisiae, for which a comprehensive library of single-gene knockouts is available and readily manipulated. Indeed, the Ohya and Morishita labs at the University of Tokyo had already published a survey of 4718 knockout strains, measuring 220 quantitative morphological traits through automated analysis of microscope images.

What makes a capacitor?

Levy and Siegal re-analyzed the Japanese data to quantify, for each knockout, the increase in phenotypic variance in response to micro-environmental and stochastic variations. To do this, they had to develop a way to distinguish intrinsic variance changes from those driven by changes in mean values and to narrow the original 220 measures of phenotype to 70 reasonably independent measures. They then compared the variability changes to those produced by randomly swapping parameters, quantifying their significance using a metric that they call the "phenotypic potential."

"By using these permutations," Siegal said, "we can estimate that there are hundreds of genes that have a higher phenotypic potential than you would expect to see by chance." Levy and Siegal also performed their own microscopy experiments to validate the analysis. With this many examples in hand, Siegal said, "we can start asking questions about what makes them phenotypic capacitors."

Phenotypic capacitor (PC) duplicates are protein–protein interaction hubs (but PC singletons are not).

The observations rule out some leading candidate attributes, Siegal observed. For example, "you get a different constellation of variation" for each knockout, "so it's not like there's one high-level process called robustness that there are different ways of knocking out." In addition, he said, "we see a diverse set of functional processes represented," although some processes are more abundant in the data set.

Because phenotypic capacitors are important but not essential, the researchers explored whether they were partially redundant, for example by having a paralog elsewhere in the genome resulting from a past gene duplication. However, most of the capacitors were singletons, having no paralog. Based on known interactions, these singleton phenotypic capacitors tend to lie within modules of genetic and protein–protein interaction networks, as indicated by a high clustering coefficient.

In contrast, phenotypic capacitors that had a paralog in the genome tend to be very highly connected, linking modules with distinct but related functions. The duplication was generally evolutionarily old, and the expression of the capacitor and its paralog diverge from one another even more than their age would predict.

"We think that these two classes of genes, the singleton and the duplicate capacitors," Siegal said, "are revealing the architecture of robustness at two different levels of biological organization."

A relatively simple cancer

The dynamics of cancer growth has many ingredients of traditional evolution, said Franziska Michor of the Memorial Sloan Kettering Cancer Center. Natural selection, she said, requires "a population that's undergoing replication. While undergoing replication it can accumulate mutations, so there is variation in the population, and on this variation, then, selection can work." All of these conditions occur during cancer development.

Michor focused on chronic myeloid (or myelogenous) leukemia, which she described as an especially simple type of cancer to study because a single genetic change causes its initial stages. "This is probably not true for any other cancer," she noted. The so-called Philadelphia chromosome results when normal chromosomes 9 and 22 exchange segments, resulting in a "fusion oncogene" denoted BCR/ABL. The resulting fusion oncoprotein, is a constitutively active tyrosine kinase that promotes uncontrolled cell proliferation.

"The cancer retains the differentiation hierarchy we see in the normal blood system."

Because this protein is unique to cancer cells, the drug imatinib (called Gleevec in the U.S., Novartis) that targets it "will have almost no effect in any normal cell," Michor said. "It's the first cancer chemotherapeutic we have that causes virtually no side effects and reduces the number of cancer cells by a large extent."

As part of a large, phase 3 trial, Michor and her colleagues monitored the declining number of cancer cells in patients receiving Gleevec treatment. "In every single patient the leukemic cell burden is depleted exponentially at two different slopes," she said, with decay rates averaging 5% and 0.8% per day. "The existence of different slopes points to the existence of different subpopulations of cancer cells."

In the normal blood system, immortal stem cells give rise to progenitor cells, which give rise to the precursors of the myeloid and lymphoid lineages. Finally, these precursors produce the terminally differentiated cells, such as red blood cells, lymphocytes, and platelets, that are found in the peripheral blood. Michor modeled the response to Gleevec treatment by assuming that this hierarchy was replicated among the cancer cells that bear the Philadelphia chromosome. "This is new," she commented, in that the cancer cells are not a homogenous population. "The cancer retains the differentiation hierarchy we see in the normal blood system."

Evolutionary dynamics of cancer

Evaluating treatment responses requires understanding which populations correspond to the observed exponential decay times, and in particular, whether cancer stem cells are killed along with their progeny. Unfortunately, in three patients who stopped the treatment, the levels of cancer cells rapidly rebounded to the original levels or even higher. "The cancer stem cell is not depleted by the Gleevec therapy," Michor concluded. Possible explanations for this insensitivity include the multidrug resistance efflux pumps found in normal stem cells and the relatively low rate of cell division of these cells. Alternatively, the stem cells may not express the fusion oncoprotein.

Michor extended the mathematical model to analyze the development of resistance to Gleevec, which results from point mutations in the kinase domain of the fusion oncoprotein. "A single base change is sufficient to render the most advanced chemotherapeutic we have completely useless," she observed.

The stochastic model includes a third hierarchy of resistant cells, but also must go beyond a deterministic dynamics, Michor noted. "If we want to predict probabilities of resistance we have to go into stochastic modeling."

"The cancer stem cell is not depleted by Gleevec therapy."

The model provides a tool for predicting the chances that resistant cells are already present, Michor said. "The probability of resistance increases during disease progression," Michor said. "If this risk is large, then maybe we should go immediately to bone-marrow transplant, rather than trying to treat [with Gleevec], because the longer you wait with bone-marrow transplant, the lower the rate of successful treatment becomes."

Michor is currently using a similar analysis to determine the level of the hierarchy where the original mutation leading to chronic myeloid leukemia occurs. The convergence of evolutionary theory with systems biology, like its well known convergence with developmental biology (EvoDevo), should enrich both fields and provide new tools for understanding how complex networks affect, and are affected by, evolution.

Open Questions

How do biological networks differ when they result from sexual reproduction instead of asexual?

Does selective pressure change the robustness of evolving networks?

What are the best ways to identify buffering genes and their effects in human populations?

How do new traits emerge in spite of the robustness of phenotype?

Do the mechanisms that buffer micro-environmental and stochastic variation also buffer macro-environmental or genetic variation?

Can genes that suppress phenotypic variation be systematically found in multi-cell organisms?

What is the dynamics of cancer stem cells during therapy?

In which level of the stem-progenitor-differentiated cell hierarchy do the mutations leading to cancer occur?