NEW YORK (GenomeWeb) – A team led by researchers at University College Dublin has developed a model of the c-JUN N-terminal kinase (JNK) protein signaling network that can serve as a biomarker for predicting survival in neuroblastoma patients.
The work, detailed in a paper published this month in Science Signaling, offers a proof of concept for the notion that dynamic models of cell signaling could serve as more effective biomarkers than the sort of discrete molecular markers that have traditionally been the focus of much biomarker research.
"I have used modeling for 30 years to [for instance] make mechanistic models to understand what we know about a system and to understand what we don't know," Boris Kholodenko, a UCD researcher and author on the paper, told GenomeWeb. "But now it appears that these models can actually add to the predictive and prognostic power of biomarkers."
Protein signaling networks are important for the regulation of a variety of cellular processes, and as such are key targets in both biomarker and drug development. However, as these networks are dynamic systems, static measurements of protein markers can fail to fully capture the information contained within them.
For instance, Kholodenko said, "gene expression [measurements] can only give us a snapshot of the system. But the modeling of signaling pathways can predict the behavior of the system."
As a proof of concept, Kholodenko and his colleagues sought to model the JNK pathway, which is involved in mediating cell death. Specifically, they aimed to develop a model of JNK signaling that was able to correlate the JNK response dynamics of neuroblastoma patients to survival data.
As the researchers noted, amplification of the transcription factor MYCN is associated with poor prognosis in neuroblastoma, but a number of patients without this amplification also fare poorly, for reasons that are not well understood.
One potential cause is the failure of these patients to manifest the ultrasensitive JNK activation that leads to apoptosis. Using a variety of techniques including knockdown and protein interaction experiments and phosphorylation assays, the researchers pieced together the components of JNK signaling and their responses to various stimuli. They then used rule-based modeling using ordinary differential equations to incorporate this experimentally derived information into a model of the JNK network.
This model in hand, they tested it in three different cell lines, finding that it could accurately predict the JNK response to various stimuli.
The researchers then applied the model to actual patients, using mRNA expression data to generate personalized models of the JNK response for each patient in order to test whether patients in whom the model predicted an apoptosis-generating ultrasensitive response would have a better prognosis than those predicted to have a low-amplitude response. And, indeed, prediction of a low-amplitude JNK response was linked to poor survival both in patients with and without MYCN amplification.
Having demonstrated the method's potential, Kholodenko said he and his colleagues are now applying it to additional systems, including selection of breast cancer patients for chemotherapy treatment after surgery.
They are also using it to investigate drug response and guide therapy, said Kholodenko's UCD colleague and study co-author Walter Kolch.
"By studying how the network [changes] with the drug [treatment], you can anticipate or rationalize what combinations to use and what the most effective agents will be," he said. He noted that in previous work modeling RAS-BRAF-MEK-ERK signaling, the researchers were able to predict the causes of resistance to BRAF inhibitors in melanoma patients.
"In principle, models are ideally posed to predict what might happen in the system in terms of rewiring or expressing of genes that were never expressed before you give a drug," Kholodenko said. "And this is much less time-consuming than to do, say, cell-based experiments where you have to give a drug to a cell line for six months before you see some resistance."
One obvious challenge to the modeling approach, Kholodenko noted, is gathering the necessary information to piece together the signaling pathways of interest, though, he said, the advent of modern high-throughput genomic and proteomic techniques has made this somewhat easier than it might have been in the past.
Researchers' understandings of their target systems "needs to be fairly detailed for the type of work which we pursued in the paper, because you really need the mechanistic understanding of the pathway to where you can understand the model and relationship between molecules," he said. "Unfortunately, we don't have that depth of understanding for many pathways. So I think it is important to first identify which pathways you need to model in great depth, and then you can really take on those pathways. It's work, but it is doable."
And, Kolch noted, such models are never "complete replicas" of the systems being studied.
"You make assumptions," he said, but added that these can be tested by comparing a model's predictions against experimental observations in cell lines or patient samples. This process of testing a model's assumptions can raise additional questions or illuminate previously unexamined aspects of the signaling networks of interest.
"The model has the power to guide you toward questions to address that would be very difficult to achieve [solely] experimentally," he said.
Kholodenko said the researchers are now interested in exploring where their approach could fit into actual clinical practice.
"What we need to do now is go to clinicians to [learn] what they need, exactly what the clinic demands and what the pressing clinical problems are," he said.
One area where clinical implementation of the method will likely differ from that used in the Science Signaling study is the type of molecular data used. In the study, the researchers were able to use patient mRNA data. However, Kholodenko said, while this data had good concordance with actual protein levels in the case of the JNK pathway, this is not typically the case and so, generally speaking, it would be preferable to measure proteins directly as opposed to using mRNA as a proxy.
One potentially suitable method could be reverse phase protein arrays, which are commonly used in protein signaling studies. However, Kholodenko said, "in the long run, if one wants to bring this to the clinical arena, one needs to use methods that are in the clinical laboratory."
With that in mind, he suggested ELISA and quantitative immunohistochemistry as the likeliest options.
Unlike RPPA or mRNA analyses, these techniques are typically able to measure only a few markers at a time. But, said Kholodenko, once the model has been developed, predictions can be made using a small number of protein measurements.
"Once you have the model you can then determine which proteins are really important in controlling behavior," he said. "And then when you go to patients you do not need to measure everything in the whole model. You only need to measure the control nodes, because these will be the ones that make the difference."
Pathways as biomarkers potentially "have a lot more robustness" than discrete markers, said George Mason University researcher Emanuel Petricoin, whose research also focuses considerably on the use of protein signaling for guiding patient therapy and identifying disease subtypes.
"Given the heterogeneity of genomics and signaling, where one patient may have an aberration in one part of the pathway and another patient may have it in a different part of the same pathway, relying on the [overall] pathway as the marker would ostensibly be more robust," he said, calling the Science Signaling work, which he was not involved in, "a really cool way of doing patient-specific simulations."
He suggested that he and his colleagues might employ similar approaches in their work within the iSPY-2 Breast Cancer Clinical Trial, which is exploring the utility of protein signaling data, along with other types of molecular information, for guiding patient therapy.
To date, Petricoin noted, the researchers have been looking primarily for predictive signatures in pretreatment biopsy samples.
"But there is also a lot of [potential utility] in looking at the dynamic signaling changes that occur in the first month of treatment or so," he said. "Can you predict a patient responding in the first couple weeks after therapy? That dynamic change may be even more informative than just a basal start point."