Meeting Report

It's the network

Ask a systems biologist what they do, and you could be in for a long conversation. Speaking at the June 2, 2009, meeting of the Academy's Systems Biology Discussion Group, Andrea Califano of Columbia University admitted as much. "Nobody has an elevator pitch about what systems biology is, and in fact many people disagree on what the term actually means," said Califano.

Cancer is a problem tailor-made for systems biology.

Nonetheless, some common themes permeate the field. Systems biologists generally study molecules, whole cells, tissues, and organisms simultaneously, and they rely heavily on computers to map the complex, layered networks of signals that connect these levels of organization.

At the June meeting, four prominent systems biologists took up a biological problem tailor-made for this research strategy: cancer. As their results showed, the systems approach can reveal both interesting biology and promising new therapeutic targets for one of the world's deadliest diseases.

ARACNe unravels signaling webs

Andrea Califano followed his brief overview of systems biology with a brisk tour of his laboratory's recent work, which entails reverse-engineering cellular signaling pathways. The idea is to build a detailed map of the cell's communication network, revealing new ways to influence complex processes such as tumor development.

One of the lab's major accomplishments was developing the Algorithm for the Reconstruction of Accurate Cellular Networks, or ARACNe. The algorithm takes signal transduction data as an input, and uses information theory to determine how signals travel through the cell. "The idea was that if some information was exchanged between two gene products in the cell, you have to find some kind of a pathway using these two gene products to get there," said Califano.

An ARACNe analysis of a signaling pathway yields the minimal network that can explain all of the observed information flow. That's not necessarily how the real network works, though—the data going into the model could be incomplete, or evolution could have chosen a more convoluted solution. To address that possibility, Califano and his colleagues check ARACNe's predictions with biochemical experiments. Sometimes, the results are surprising.

Computational methods for reverse-engineering are starting to achieve accuracy and sensitivity comparable to the best high-throughput experimental assays.

For example, an analysis of developmental signaling in embryonic stem cells revealed numerous genes that were likely to be regulated by transcription factors and other DNA-binding proteins. Biochemists had already identified some of these regulators, but others were new to science—or perhaps just artifacts of ARACNe. To test some of these previously unknown regulatory interactions, Califano's team used chromatin immunoprecipitation onto microarrays, or ChIP–Chip, a common technique for probing protein–gene interactions. Of the ten most likely interactions ARACNe had highlighted, ChIP–Chip only confirmed one, suggesting that ARACNe was predicting interactions that didn't really occur.

Probing further, though, the researchers performed a detailed biochemical analysis of the ARACNe-predicted interactions, and determined that the algorithm was actually correct. "We think that actually current ChIP–Chip ... analysis methods underestimate the number of real sites where transcription factors bind by about an order of magnitude," said Califano.

The investigators have also developed several other algorithms for refining their signaling maps, and have unleashed these new tools on glioblastoma multiforme, a highly aggressive brain cancer. The algorithms revealed several master regulatory proteins that control more than 75% of the gene expression program driving tumor growth. Cell, mouse, and human data all back up those results, revealing novel targets for a new generation of drugs to treat this disease.

The ups and downs of oncogenes

Galit Lahav of Harvard Medical School is taking a slightly different approach to the systems biology of cancer. Instead of trying to uncover new signaling molecules, she and her colleagues started with one of the best-characterized oncogenes in all of molecular biology: p53.

"It's an ideal system for us to work with," said Lahav, adding that "we want to start by working with a network where a lot is known about it, and there's a lot of information about the different proteins and connections." Despite decades of intense scrutiny, though, p53 still has plenty of mysteries for systems biologists to ponder.

For example, exposing cells to either UV irradiation or γ-irradiation stimulates p53 activity, but the two types of radiation produce subtly different types of p53 responses. UV light causes a slow, steady rise in p53 levels, while γ-rays cause p53 levels to rise and fall in a series of diminishing oscillations. These differences may explain why UV-irradiated cells are more likely to undergo programmed cell death, while γ-irradiated cells are more likely to remain alive, but pause their cell division cycles.

Individual cells within a culture or tissue may have different p53 levels at any given moment, but traditional cell biology experiments will only show the average activity for the whole culture. To get around that, Lahav's team tracked levels of p53 and a negative regulator called Mdm2 in individual cells. In a typical experiment, the researchers digitally analyzed dozens of hours of video microscopy data, producing detailed graphs of each cell's molecular changes.

Fluorescent labeling of p53 and Mdm2 allows tracking of these proteins in individual cells.

The results showed that γ-irradiation actually induces a constant oscillation between p53 and Mdm2, not a diminishing oscillation; as one protein's level rises, the other's falls, and then the cycle reverses itself. Some cells go through one or two oscillations, then shut down p53 and continue their normal life cycle, while other cells undergo multiple p53/Mdm2 oscillations and remain in cell cycle arrest.

Through a combination of computer modeling and laboratory testing, Lahav and her colleagues have now uncovered more components of this molecular oscillator. In their current model, DNA damage stimulates pulses of expression of a kinase called ATM, which triggers the oscillations in p53 and Mdm2, while a negative regulator called Wip1 shuts the system down once the DNA damage is repaired. The model also predicts that another signal, which is still unknown, must provide positive feedback to p53. "We're in the process now of going back to the bench and looking for this positive feedback," said Lahav.

Different similarities

Chris Sander of Memorial Sloan-Kettering Cancer Center approaches cancer research with a technique that dates to the dawn of modern cell biology: perturbing cells with various chemicals, and observing their responses. Unlike traditional cell biologists, though, Sander analyzes data on numerous perturbations and outcomes, builds computational models of the results to predict additional outcomes, and then tests those predictions in the next round of experiments.

With this iterative approach, called combinatorial perturbation analysis, or CoPIA, Sander and his colleagues hope to develop a strategy for truly rational drug design. "If this system is good enough in its predictive power ... you can ask 'what of the imaginable possible set of drugs or existing set of drugs might give you the desired effect downstream?'" said Sander.

At first glance, it's hard to imagine such a general approach working for a disease as diverse as cancer. Besides the tremendous differences between different types of tumors, researchers have discovered that even a single type of tumor can show an astonishing amount of variation from one patient to the next. "There's a huge diversity of molecular alterations in cancer," said Sander. As an example, he pointed to ovarian cancer, where "every chromosome in every ovarian tumor sample has some alteration."

Tumors have diverse gene regulation defects.

That diversity has led many researchers to focus on developing individualized cancer therapies, which would require each patient to get a customized treatment. While such an approach may work, it will be extremely complicated and expensive.

But there may be a better way. Applying CoPIA to glioblastoma multiforme, Sander and his colleagues found specific clusters of signaling proteins that are amplified in many patients' tumors. The regularity of the patterns suggests that the right drug regimen could treat large segments of the patient population, without having to individualize each treatment. "I think the useful step forward ... will be to actually look at what combination of targets might be available slightly more upstream, and ... see whether or not that would give one a systematic way of not treating every individual separately but actually addressing the disease or subtypes of disease."

The meaning of LIF

Arnold Levine of the Simons Center for Systems Biology brought the discussion back to p53, a protein whose signaling he has studied for decades. While his laboratory's original work on p53 used classical cell biology techniques, the team has now turned to systems biology to pursue the next question. "How do you move from understanding these things in cell culture and in animals to humans themselves?" asked Levine.

The human genome project, and the closely related effort to map millions of single nucleotide polymorphisms (SNPs), or single-base differences between individuals, provided part of the answer. In theory, comparing the SNP profiles of people with and without a complex disease should reveal the genes involved, if researchers sample enough people.

In practice, "enough people" for such a whole-genome screen is a number in the thousands, exceeding the budgets and capabilities of most laboratories. Instead, Levine and his colleagues decided to focus on the subset of SNPs that affect p53 and its partners, such as the negative regulator Mdm2 and the downstream growth factor LIF. That allowed them to get statistically meaningful results from far fewer patients.

A single base change in p53 has far-reaching effects.

For p53, one change is particularly significant: the presence of either an arginine or a proline sequence at codon 72. About 80% of healthy Caucasians are homozygous for the arginine residue, while about 98% of healthy Africans are homozygous for the proline.

In Mdm2, a SNP in the enhancer region of the gene's first intron, with either a G or a T nucleotide, also shows racial differences. About 43% of Caucasians are homozygous for the G nucleotide, while Africans are overwhelmingly homozygous for the T. Examining the haplotypes in more detail, Levine and his colleagues determined that the African versions of p53 and Mdm2 are the more ancient ones, suggesting that the Caucasian versions have undergone some sort of positive selection.

Based on the low fertility of the lab's p53-deleted mice, the researchers hypothesized that the positive selection could be related to reproduction. Indeed, analyzing the SNP distribution in patients from a fertility clinic revealed a disproportionate number with proline at p53 codon 72, and a similar skew favoring the G residue in Mdm2's enhancer. A SNP in LIF also appears to be involved in many cases of infertility.

Besides being potentially useful in directing patients to the correct fertility treatment strategy, the finding provides an unusually satisfying validation of both classical cell biology and cutting-edge systems approaches. "It's very rare that you know a pathway ahead of time, you look for SNPs, and you can start to get information about what they really do, and then start to put that together," said Levine.

Open Questions

Can algorithms that predict single-cell signaling patterns be scaled up to predict signaling across whole tissues and organs?

What proportion of a computer-predicted signaling network has to be checked in the lab to validate it?

What signaling molecule provides positive feedback in the p53/Mdm2 oscillation cycle?

Is oscillation a general phenomenon in cell signaling molecules?

Will combinatorial algorithms point to clinically useful combinations of chemotherapies?

What genetic polymorphisms are most important for driving cancer progression?