A new way to look at liver cancer


  • Joseph Locker

    Corresponding author
    1. Department of Pathology, The Marion Bessin Liver Center, Bronx, NY
    2. Albert Einstein Cancer Center, Albert Einstein College of Medicine, Bronx, NY
    • Department of Pathology, Albert Einstein College of Medicine, 1300 Morris Park Ave., Bronx, NY 10461
    Search for more papers by this author
    • fax: (718) 430-3483.

  • See Article on Page 667

Hepatocellular carcinoma (HCC) exemplifies a fundamental limitation of cancer pathology. Under the microscope, no two cancers look exactly the same, even though they may be equivalent by established criteria (Fig. 1). Conversely, two tumors in the same diagnostic category may disseminate at drastically different rates or show completely different responses to therapy. This dilemma is particularly frustrating to the pathologist, because retrospective analysis usually shows some tumor-specific feature that looks like it must have been important. Indeed, the concept of tumor grading is based on finding the particular features that make a difference, but grading usually sorts out the extremes—most tumors tend to rank in the middle of a grading system.

Figure 1.

Two kinds of hepatocellular carcinoma (HCC). Two moderately differentiated HCCs show clear differences from each other. The tumor on the right has larger cells and nuclei and forms larger acini. Such features vary from one tumor to another but do not predict clinical behavior. However, one tumor has a poor-prognosis, high-proliferation, Cluster A gene expression profile, while the other has a good-prognosis, high-apoptosis, Cluster B profile. There are two illustrated morphological features that correspond to the profiles. The left tumor has a mitotic cell (Mi), while the right tumor has an apoptotic cell (Ap).


HCC, hepatocellular carcinoma; AFP, α-fetoprotein.

Recent scientific and technical advances have provided new ways to look at HCC. The Genome Project and high-throughput cloning efforts have defined almost all human genes, which now number about 30,000. It appears that only a few thousand more genes are hiding in the Genome Project data, so this effort is nearly complete. High-throughput preparation of DNA clones or synthesis of oligonucleotides, combined with microprinting technology, has produced high-density microarrays in which thousands of tiny DNA spots are printed on a single glass slide. The slide is then hybridized with fluorescently labeled RNA extracted from a tumor, and the hybridization intensity, a direct measure of gene expression, is quantified by confocal laser scanning. Within a few years, microarrays have progressed from rather crudely assembled collections of a few thousand poorly identified genes to highly organized sets that can screen expression of the vast majority of genes in a single analysis. Microarray studies of tumors have produced detailed pictures of gene expression, information that is not apparent in conventional histopathology. An array produces a very large data set for each tumor, and the data are useful even when the relationships among the tumor-expressed genes have not been worked out. The pattern of gene expression is an empirical but unique picture of a tumor. This utility is fortunate, because the data sets are so large and encompass so many newly described genes that it is difficult (not impossible, but a lot of work) to find relationships or discriminate which expression changes are important.

In this issue of HEPATOLOGY, Lee et al.1 have combined two substantial resources to change the diagnostic paradigm. They analyzed specimens from a large, 90-patient cohort who received partial hepatectomy for HCC and then had long-term follow-up. The specimens were studied with an advanced 21,000-gene oligonucleotide microarray. This effort produced an extremely large set of significant data that is a valuable resource for many kinds of studies. The analysis in the paper stresses only two aspects, but these are important ones: HCC classification and prognosis.

The first conclusion, perhaps not very surprising, is that HCCs have a pattern of gene expression that is a lot like normal liver. Other cancers show a similar relationship to their tissue of origin.2 Only 4,200 genes showed as much as a 2-fold difference in HCC compared to normal liver expression in at least 10% of the tumors, and although the authors did not elaborate this point, about 3-fold more genes usually show consistent expression between tumor and liver. The second conclusion, however, is unexpected: HCC are surprisingly homogeneous and fall into only two categories, “Cluster A” and “Cluster B.” This fact was determined by unsupervised hierarchical clustering analysis—i.e., the tumors were grouped only by the relationships of their abstract gene expression patterns to one another, not by the function of the genes that contributed to the patterns. Function is an important issue that was considered later in the study.

The homogeneity of HCC distinctly contrasts with breast cancer, for example, which shows striking diversity from one tumor to another.3 Such heterogeneity is expected because tumors arise and progress through a highly varied sequence of genetic alterations. Nevertheless, the observation of homogeneity in HCC appears to be robust. A closer inspection of the data (Fig. 1) shows that the clusters have subgroups, but with only limited variation. Moreover, the HCCs were associated with three main etiologies—hepatitis B virus, hepatitis C virus, and nonviral causes—and these etiologies were represented equally in the two clusters. HCC of other etiologies (alcohol, hemochromatosis) also fit into the same two clusters. Finally, in case the homogeneity reflected the limited variety of their own patient cohort, Lee et al. showed similar clustering of expression in published data that represented a much greater variety of specimens.

Their significantly different survival outcomes also confirm that the two clusters are biologically distinct. This result is particularly striking because the clustering analysis was unsupervised. Most other studies have supervised clustering with survival data to find genes associated with survival.4–9 In contrast, faster progression and decreased survival are an intrinsic property of Cluster A tumors compared to those in Cluster B.

The two Clusters cannot be distinguished by histopathological grade or etiology, so how do they differ? Lee et al. identified 406 genes for which a change in gene expression, either positive or negative, predicted survival. This time, they used supervised analysis, although the vast majority of these “survival” genes had already been identified by unsupervised analysis. The functions of the survival genes suggest why tumors in the two clusters behave differently. Forty-five percent of the up-regulated Cluster A survival genes were associated with cell proliferation. Many other Cluster A up-regulations were antiapoptotic. Another up-regulated gene was very important. This gene encodes hypoxia induced factor 1,10 a transcription factor that stimulates adaptive responses to hypoxia, and production of angiogenic factors. Thus, the Cluster A tumor cell is more malignant because it divides more frequently. It is also more resistant to apoptosis, better adapted to hypoxia, and a better stimulator of angiogenesis. Might Cluster A represent a later stage of progression than cluster B? This possibility seems to be ruled out by the retrospective analysis of data from two published studies, since the genes that predicted intrahepatic metastasis9 or early recurrence6 distributed into both Clusters A and B.

The survival genes do not include the usual markers of hepatocytic differentiation—e.g., α-fetoprotein (AFP), albumin, or α1-antitrypsin. This absence of these genes from the survival list is unexpected, because it is generally presumed that the highest-grade tumors will be the most “dedifferentiated” and have the least resemblance to a normal liver cell. If this presumption is correct, then markers of hepatocytic differentiation ought to be survival genes. The absence of AFP from the list of survival genes is particularly surprising, because AFP is a known prognostic marker.11 In this case the answer is complex:Both Cluster A and Cluster B include AFP-positive and -negative tumors. In Cluster B, AFP expression had a strong negative correlation with survival, but in Cluster A, AFP expression had a positive correlation. Thus, the two effects cancelled each other out in the broader analysis. Perhaps, however, the AFP-negative tumors of Cluster A might be the dedifferentiated high-grade tumors that would be expected to do badly.

The array data predicted that proliferation would be higher in Cluster A and apoptosis higher in Cluster B tumors. Lee et al. confirmed these predictions by analyzing tissue sections. They used Ki-67, a marker of cell proliferation, and also directly counted apoptotic cells (see Supplementary Fig. 5 to Lee et al. at http://interscience.wiley.com/jpages/0270-9139/suppmat/index.html). Significantly, the correlation of counts and expression clusters confirmed a study by Ito et al.,12 who reported that proliferation and apoptosis rates were strong and opposite predictors of survival. The relationships might seem obvious: Tumors that grow faster or have less cell death should be more aggressive. In several cancers, however, increased cell proliferation and apoptotic cell death are linked together and a high apoptotic rate sometimes indicates a more aggressive cancer.13

Technology is rapidly evolving, and it is already possible to include expression profiling in the patient work-up. Such testing is not indicated by Lee et al.,1 because their data suggest that the pathologist can provide nearly the same prognostic information by counting Ki67-positive and apoptotic cells. Nevertheless, prognostic information has relatively little impact on treatment, since both Cluster A and Cluster B patients benefited from their resections. The use of advanced testing to determine therapy is a radically different situation, and a recent study of breast cancer provides a model for this approach.4 Gene expression profiles were used to predict which patients might benefit from chemotherapy. However, diagnosis and treatment have not reached the point where the same approach would be useful for HCC.

Even if gene expression profiling is not required for management of HCC patients, the research is very promising. Lee et al.1 focused on unsupervised clustering and prognosis, but there are numerous other important questions that could be approached with further analysis of their comprehensive data sets. Other recent studies have characterized gene expression profiles associated with progression from precancerous lesions,14 recurrence,6 vascular invasion,15 intrahepatic and extrahepatic metastasis,9, 16, 17 hepatitis B or C etiology,5, 7, 18–22 p53 mutation,8 and sensitivity to selected chemotherapeutic agents.23, 24 Though many of these studies should be elaborated with more specimens and comprehensive microarrays, such research will provide a detailed mechanistic model of HCC progression. Most important, expression profiling will reveal specific targets for rational therapy.