Imaging Genomics in the Age of Bioinformatics

Modern medicine has traveled far enough along the road of personalized medicine that we have reached a crossroads between genomics and imaging. Rather than taking a turn at this important intersection, experts are calling on both disciplines to create a map of the vast heterogeneity of each individual tumor using a fleet of imaging techniques, including CT, MRI and PET.

The term radiogenomics is popularly used to describe the growing field of study at this cross-section of imaging-based molecular phenotyping and genetic assays derived from biopsy. Poring over the human genome and epigenetic switches has taken researchers into a new frontier of disease diagnosis, risk stratification, therapy assessment and prognostic medicine. However, the truth is that single biopsy assays are not ideal representations of a tumor. A bit of tissue from that tumor may not even come close to mirroring the full range of genetic expression therein. This is where radiogenomics races in to capture a spectrum of scan traits that tip off specific genes. This, one day, could tell clinicians what they are working with and how best to approach treatment for each patient, or more likely inform new drug models. Further down the road, it is imaging, perhaps, that may one day overtake biopsy and invasive genetic testing.

“Personalized medicine, a patient-centered approach, is becoming the standard in medical care,” says Hans-Ulrich Kauczor, MD, PhD, a professor of diagnostic and interventional radiology at the University Hospital Heidelberg, Germany, during the press conference of the recent 2014 European Congress of Radiology (ECR) in Vienna, Austria. “Recent advances in omics technology should not overshadow the importance of medical imaging and its increasing contribution to [personalized medicine].”

Two California-based experts, Olivier Gevaert, PhD, assistant professor of medicine and biomedical informatics research at Stanford University, and Michael Kuo, MD, from the University of California, Los Angeles, shared a panel addressing radiogenomics at the Radiological Society of North America (RSNA) 2013 annual meeting. Molecular Imaging Insight caught up with both researchers to gain insight into where this discipline is going and what concrete benefits have been seen in the preliminary research of this multidisciplinary field.

Maps of the future

The human genome was first published about 10 years ago. Since that time, the interpretation of genes with molecular phenotyping has translated best to oncology. While completely non-invasive, imaging based sequencing of genes is not yet a reality, researchers can cluster similar genetic profiles together that signal specific tumor characteristics such as texture and morphology. Subclusters within a tumor type can run the whole gamut of cancer heterogeneity. Scientists are able to amass these subclusters into visual association maps that reveal the little universe harbored inside each tumor, with all of its varied functions and over- and under-expression of genes, says Kuo.

Certain characteristics, such as the way the edge of a tumor is shaped, can tell clinicians about potential survival of the patients due to the behavior of the cells that create that particular trait. In fact, up to 85 percent of the genome can be predicted by imaging alone, he adds. This is both qualitative and quantitative imaging as a high art.

Mathematical texture, characteristics of shape and intensity and other semantic features totaling about 180 possible features of disease per patient are drawn into sharp relief with digitization. Informatics databases are being developed to correlate research data with image data from new acquisitions. These features are modeled to produce molecular meta-genes, and predictive maps of meta-gene expression are what suggest prognoses such as poor survival. In fact, four semantic features are associated with poor survival, and one with a happier prognosis. In addition to the edge of a tumor, its sharpness from outside to inside and tumor axis also are considered predictive.

Navigating cancer and more

Bioinformatics pulls together disparate data from numerous modalities, including molecular imaging systems like PET/CT. High expression versus low expression of genes as mapped by advanced imaging could one day be used not just for oncology, but other disorders.

“It’s very scalable,” asserts Kuo. “In theory, it could work for a number of diseases in which imaging is useful, including imaging of dementia, degenerative disorders of the brain and musculoskeletal system and metabolic disorders.”

Many of the radiogenomics studies to date have maximized retrospective data from massive tissue donations from patients who also had imaging data on file. “Essentially these methods are not limited to applications in cancer, but it is very difficult to get access to data sets that have both molecular and imaging data from the same patients. Cancer is one of the areas where you have better access to data,” explains Gevaert.
For now, a number of oncologic applications are leveraging concepts in radiogenomics.

One of these applications is in predicting radiotherapy toxicity as a result of exposure to ionizing radiation. Prior research has suggested that single nucleotide polymorphisms (SNPs) in reactive oxygen species, cytokines and DNA repair genes are related to higher sensitivity and radiation-related inflammation (Clin J Oncol Nurs. 2014 Apr 1). These findings also are reflected in another radiotherapy study (Cancer Discov. 2014 Feb).

As far as mapping tumor heterogeneity, there are some specific kinds of cancer research that have championed radiogenomics, or vice versa. Kuo recently coauthored commentary on clear cell renal cell carcinoma research highlighting particular imaging features that correspond to genetic mutations typical of the disease (Radiology. 2014 Feb). Here researchers recognized eight qualitative and five quantitative image features and five common clear cell renal cell mutations, namely PBRM1, SETD2, BAP1, KDM5C, and VHL in a radiogenomic analysis of 233 patients. Just one of the links found between the genetic information and CT imaging data was that genetic mutations KDM5C and BAP1 correlated with images of renal vein invasion. Interestingly, both of these variants are active in transcriptional control and chromatin remodeling and are considered bedfellows with an aggressive phenotype of tumor (Eur Urol 2013).

When marrying genetic and radiological data, it is essential to not only understand the biology associated with imaging phenotypes, but to know how that process is mimicked through imaging. In another study, the authors provide an example of how something like the presence of internal arteries can alter the epigenetics of genes in hepatocellular carcinoma. In another example, researchers were able to point out an MR imaging phenotype that is intensely linked to active angiogenesis and hypoxia in glioblastoma multiforme. Researchers are now able to pull together DNA copy number variation, messenger RNA expression and MR imaging (Radiology. 2013 Oct 13).

What about PET/CT data? In the case of non-small cell lung cancer (NSCLC), there were a total of 243 statistically significant correlations between imaging NSCLC meta-genes and genetic features. The predictive accuracy of these meta-genes was 59 percent to 83 percent. Of 180 CT image features united with their PET standardized uptake value, 114 were predicted in terms of meta-genes and found to be 65 to 86 percent accurate. As previously suggested, when the predicted image features were mapped to corresponding genetic data sets, tumor size, edge shape, and sharpness ranked highest for prognostic significance (Radiology. 2012 Aug). Other molecular systems could be just as useful for mapping genetic features.

“We are doing CT, MRI and PET already. It’s an approach—in theory this could be used for optical imaging or reporter gene imaging,” explains Kuo. In this way, radiogenomics is inclusive and imaging technology non-specific.

Roadblocks and caution signs

This area of research is still in its infancy and is not yet being used clinically outside of investigational studies. As more data accumulate, the possibility of using these tools in the clinic will mount. But first, those studies need to be standardized for optimal results. The Radiogenomics Consortium has produced an 18-item checklist for reporting radiogenomics studies. STROGAR (strengthening the reporting of genetic association studies in radiogenomics) is the checklist’s official title. The checklist notes that all statistical models must be defined; that specific DNA source and isolation methods in the genotyping strategy be disclosed; that researchers specify how all patient and treatment assessments were made and what impact these had on defined phenotypes and how they arrived at outcomes, whether through clinical evaluation or patient reporting; and that the clinical utility of the study’s radiogenomics be spelled out for that given patient population (Radonc. 2013 Aug).

“A lot of the difficulty is technological,” says Kuo. “It requires a lot of integration. Each [imaging] field is developing independently but linked together, so that is where the challenge lies … being able to bring them into a common language.”

On the plus side, this area of research is built upon a steady foundation of tried and true technology, not investigational technology or requiring new probes or regulatory approval. In a way, the promise of radiogenomics is all there. Current studies are just distilling readily available imaging information, giving it genetic context and making it comprehensive and seamless.

Replacing tissue biopsy?

“At the end of the day, if we can get information without having to stick a needle in a patient, that is the goal. But that is a broader question of imaging technology,” Kuo says. “It’s just one approach to gaining diagnostic information without requiring human tissue.”

Extracting characteristics from molecular perfusion imaging, for instance, and applying computer algorithms to shape that data is where radiogenomics is at, just as it was with structural data from CT. One of the goals of radiogenomics is to look at both molecular and structural imaging. If you overlay the pixel information you can see molecular phenotypes emerging from the static. However, not only is there huge heterogeneity in cancer, but also in imaging and bioinformatics systems. 

“There are so many different scanners, manufacturers, devices, settings and of course these create a bias,” says Gevaert. “You could extract features in one imaging center and have it be different in another center.”

To counter this, researchers use normalization algorithms and correct for patch effects that indicate how a group of patients are imaged in a way that deviates from the norm.

Clinical practice

Is radiogenomics moving into general clinical use? Not so much. Because researchers are using these techniques to look for molecular biomarkers that mimic genetic events, they may end up more applicable to improving drug development rather than being used in the exam room between doctors and patients. Just how close to clinical application we are depends on who you ask.

“I think we are coming close,” says Gevaert. “There are more and more researchers in this area and we are picking up at this point.”

One other challenge that follows along the line of needing a common language between imaging data is the need for more congruity between not only imaging datasets but molecular data and its corresponding image data. Apparently they are not an exact match. This is due to the fact that researchers are often not sure exactly where inside the tumor a tissue sample is coming from. There have been reports of huge molecular heterogeneity in some cases, such as one in which five biopsies from a single tissue sample reflected different molecular profiles. Gevaert says this crisis can be averted with image-guided biopsies to properly characterize the heterogeneity of cancers. This, of course, adds another level of imaging to the mix.

“The idea is that imaging would actually help characterize the heterogeneity,” says Gevaert. Diagnostic imaging data represent the whole tumor. The biopsy represents only a part of the tumor. When there is immense heterogeneity, one or even multiple biopsies may not capture the driving genetic factors, such as EGFR (epidermal growth factor receptor), a well-known target in blood cancer, which is another type of cancer that is ripe for radiogenomics applications. If you have image-guided biopsies, you could capture the heterogeneity by taking samples of each area of the tumor and recording it.

“Due to the focus on different aspects of omics, the importance of the right location and phenotype is currently undervalued,” continues Kauczor. “The combination of image-based phenotyping and genomic analysis via radiogenomics provides a comprehensive approach in [personalized medicine]. In this context, medical imaging also has become essential in guiding tissue sampling for molecular profiling. In the next step, personalized diagnostic imaging and therapeutic approaches will be integrated through theranostics.”

The term theranostics represents an old concept that is only just now making real headway. This refers to molecular compounds that can be used for both diagnostic and therapeutic purposes to achieve increasingly targeted treatments and more precise disease staging and therapy monitoring. All of the omics and now nostics require advanced medical imaging to realize their full potential.