Mutation Profiles of Urothelial Cancer: Will Genomics Change Radiology or Vice Versa?

See also the article by Alessandrino and Williams et al in this issue.

Dr Choyke is chief of the Molecular Imaging Program at NCI in Bethesda, Md, and professor of radiology at the Uniformed Services University of the Health Sciences, also in Bethesda. His research focuses on the use of molecular imaging of cancer for diagnosis and image-guided therapy. A fellow of ACR and the Society of Abdominal Radiology, he is the principal investigator of clinical studies involving MRI and molecular imaging at NCI.

In this issue of Radiology, Alessandrino et al (1) performed targeted genomic analysis of bladder cancers in 103 patients who were also staged with cross-sectional imaging. Unlike most studies that attempt to link imaging features to genomic features, commonly referred to as radiomics or radiogenomics, Alessandrino et al linked genomic features to imaging features in the hope of improving the interpretation of the latter by providing enhanced a priori knowledge of the pattern of tumor spread associated with a specific genomic profile (2,3). Unlike radiogenomics, which has been many years in development and has an uncertain future, this reverse approach, which might be termed genomic radiology, has immediate and useful implications for oncologic imaging. The results of the genomic analysis of bladder cancers allowed the cancers to be characterized as high or low risk based on mutational profiles. The high-risk patient group was further defined as having mutations in TP53 and/or RB1 and/or KDM6A, while the low-risk group was defined as harboring ARID1A, FGFR3, PIK3CA, STAG2, or TSC1 mutations without any high-risk mutations.

Among the high-risk mutations, TP53 mutation was associated with a higher frequency of nodal (relative risk = 1.7) and osseous (relative risk = 1.9) metastases, while RB1 mutations were associated with peritoneal metastases (relative risk = 5.9). High-risk mutational burden of a tumor was independently associated with shorter metastasis-free survival (hazard ratio = 3.5) and overall survival (hazard ratio = 3.1; P = .02) compared with low-risk mutational burden. Meanwhile, presence of a low-risk mutation, such as ARID1A mutation, was associated with prolonged survival (hazard ratio = 3.1). Clearly, such information could be helpful to radiologists interpreting the imaging findings of such patients.

Increasingly, genomic profiles of cancers are recorded in routine histories of patients and are thus available to radiologists through the electronic medical record embedded in picture archiving and communication systems. As these kinds of data become more widely available, it is important that radiologists understand the potential value of the data in helping them interpret studies more accurately. However, this is no easy task given the diversity of cancers and genomic profiles. Therefore, the potential improvements suggested by Alessandrino et al will require substantial investment in matched genomic and radiologic data sets for a wide spectrum of cancers.

Sequencing of the cancer genome is becoming less expensive, recently dipping below $1000 for whole-genome sequencing, and is now routine in many cancer centers (4). However, much of this information is not currently used, least of all by radiologists. Such genetic data can dictate optimal therapies in some patients, commonly referred to as precision medicine. Also, not all mutations are of equal importance. Some mutations are considered passengers, that is, they have no direct impact on selective growth of the tumor, while others are considered drivers, that is, they have profound impact on selective growth (5). Combinations of driver mutations can lead to aggressive growth as whole molecular pathways are hijacked to accelerate tumor growth.

In the past decade, interest has increased in the use of cancer genomics as cost and availability have become less of a barrier. Early examples of reliance on genomics for drug selection are identification of amplification of HER2, indicating the use of trastuzumab in breast cancer, and mutation of EGFR in lung cancer, indicating the use of erlotinib and related tyrosine kinase inhibitors. More genetic mutations are being reported, and improved understanding of their impact on patient care is needed.

As radiologists, we can no longer ignore this tectonic shift in medical practice. If we learn to use genetic information in our practices, the value of our interpretations will be increased. In bladder cancer, Alessandrino et al showed that specific mutations can point to specific “hotspots” on the CT scan that deserve additional attention. Unlike generic histories typically provided on imaging request forms (eg, bladder cancer), genomics offers a more nuanced and specific description of disease. We have tended to ignore it up until now, but Alessandrino et al point out how such information could be of practical use.

This revolution in mutational analysis is happening across all cancers. However, some recurring themes exist. Several “bad actor” genetic mutations tend to show up recurrently, such as PTEN, TP53, RB1, and BRCA1 and BRCA2. To understand the full impact of these mutations, we will need studies with large numbers of patients (>10 000) using whole-genome sequencing (approximately 25 000 genes). While this is a big ask, there is ample precedent for it. The Cancer Genome Atlas (https://www.genome.gov/Funded-Programs-Projects/Cancer-Genome-Atlas) was a project begun in 2006 to explore the entire spectrum of genomic changes involved in human cancer. It generated more than 2.5 PB of data, which are now publicly available, and has spawned numerous bioinformatic studies. Most cases included imaging data, which were stored in a companion database known as The Cancer Imaging Archive (https://www.cancerimagingarchive.net/). However, the imaging data were not systematic, so gaps now exist in the current database. By building on these archives, it should be possible to provide genome-based predictions of radiologic findings for the broad spectrum of human cancers.

The approach of Alessandrino et al has some limitations. In their study, a select number of mutations were tested with Oncopanel, an institutional gene testing panel of 237 genes used at the Dana-Farber Cancer Institute (Boston, Mass). Other such panels sample only a select number of well-known genes. Many such panels exist, and we have not yet reached consensus on what constitutes an adequate screen of the genome. Thus, standardized reporting of these mutations is lacking. To make matters worse, because not all mutations have equal impact on biologic processes, subtyping of genetic abnormalities will become increasingly complex. As complete whole-genome sequencing becomes economically more feasible, it should be possible to standardize genomic reports to include all known genes presented in a standardized manner. This should allow natural language processing of the electronic medical record to extract relevant mutations and to provide radiologists with rapid and reliable look-up tables of the radiologic implications of these mutational profiles. However, tumors are inherently heterogeneous, and most decision making will be based on biopsy material, which is susceptible to sampling error. Thus, even in this new world of genomic characterization, our knowledge about biologic tumor characteristics will be incomplete. Additionally, although we are all excited about the role genomics will play in precision medicine due to the seductive logic that understanding brings solutions, it should be noted that its current impact on patient outcomes has been modest at best, and it remains to be proven that our intensive commitment to genomics will provide us with better outcomes.

This concept of using medical history, both personal and genomic, is not new in radiology. It is standard for residents to learn of hereditary cancer syndromes, such as von Hippel–Lindau disease and Lynch syndrome, that carry well-characterized germline mutations. These mutations result in syndromes that direct our attention to specific organs and educate our eye to seek specific abnormalities. In the new era, we will find many more somatic genetic defects resulting in a plethora of phenotypes. This will quickly surpass the ability of training programs to educate residents or busy practitioners to recall information from memory. Fortunately, another major disruptive force, artificial intelligence, is also on the horizon. Artificial intelligence excels at taking complex sets of data and reducing them to important and actionable consequences. Together with artificial intelligence, the arrival of genomic radiology will further increase the value of oncologic imaging.