Journal of Vascular and Interventional Radiology
Volume 18, Issue 7 , Pages 821-830, July 2007

Radiogenomic Analysis to Identify Imaging Phenotypes Associated with Drug Response Gene Expression Programs in Hepatocellular Carcinoma

  • Michael D. Kuo, MD

      Affiliations

    • Department of Radiology, University of California-San Diego Medical Center, San Diego, CA 92103
    • Corresponding Author InformationAddress correspondence to M.D.K.
    • ,
  • Jeremy Gollub, PhD

      Affiliations

    • Department of Biochemistry, Stanford University, Stanford, Calif
    • ,
  • Claude B. Sirlin, MD

      Affiliations

    • Department of Radiology, University of California-San Diego Medical Center, San Diego, CA 92103
    • ,
  • Clara Ooi, MD

      Affiliations

    • Department of Radiology, University of Hong Kong, Shatin, Hong Kong
    • ,
  • Xin Chen, PhD

      Affiliations

    • Department of Pharmacy Biopharmaceutical Sciences, University of California-San Francisco, San Francisco, Calif.

Received 6 January 2006; received in revised form 23 April 2007; accepted 23 April 2007.

Article Outline

Purpose

To determine whether conventional contrast-enhanced computed tomography (CT) could be used to identify imaging phenotypes associated with a doxorubicin drug response gene expression program in hepatocellular carcinoma (HCC) by using an integrated imaging-genomic approach.

Materials and Methods

Thirty HCCs were analyzed and scored individually across six predefined imaging phenotypes. Unsupervised and supervised bioinformatics analyses were performed to correlate the imaging scores with the corresponding tumor microarray data (each microarray contained gene expression measurements across ∼18,000 genes) to identify relationships between the imaging traits and underlying tumor gene expression. Enrichment for a predefined doxorubicin-response gene expression program was then performed against the imaging phenotype–associated genes and enrichment determined.

Results

An imaging phenotype related to tumor margins on arterial phase images demonstrated significant correlation with the doxorubicin-response transcriptional program (P < .05, q < 0.1). It was also significantly associated with HCC venous invasion and tumor stage (P < .05, q < 0.1). Tumors with higher tumor margin scores were more strongly associated with the doxorubicin resistance transcriptional program and had a greater prevalence of venous invasion and worse stage. Tumors with lower tumor margin scores, however, demonstrated a converse relationship.

Conclusions

It is possible to identify HCC imaging phenotypes at CT that correlate with a doxorubicin drug response gene expression program. Given the role of doxorubicin in regional therapies for HCC management, it is possible that such an approach could be used to guide HCC therapy on a tumor-by-tumor basis on the basis of underlying tumor gene expression patterns.

Abbreviations: cDNA, complementary DNA, FDR, false discovery rate, HCC, hepatocellular carcinoma, SAM, significance analysis of microarrays

 

MICROARRAY analysis of gene expression is a powerful tool that enables one to survey, in parallel, the expression of thousands of genes at once (1, 2, 3, 4). As a result of the ability to identify differential changes in the expression level of many genes simultaneously, thematic expression patterns can emerge that are indicative of underlying biologic processes and can provide insights into the transcriptional state of a cell. Because cancer is fundamentally a disease of genetic instability, functional genomics approaches have naturally lent themselves to the study of cancer, where they have been used to delineate genetic programs and molecular markers associated with tumor biology and patient prognosis for a large variety of human cancers on a tumor-by-tumor basis (5, 6, 7, 8, 9). Researchers have also begun to use this technology to identify gene expression programs associated with therapeutic outcome (10, 11, 12).

Hepatocellular carcinoma (HCC) is the fifth most common cause of cancer deaths (13). It is a molecularly heterogeneous disease with an unpredictable natural history and treatment response profile. Although local-regional therapies such as transarterial chemoembolization have demonstrated substantial survival benefits in patients with unresectable disease, treatment response is not uniform, with some patients clearly responding better than others to the same therapy (14, 15). It is likely that much of this differential tumor-to-tumor drug response variability can be accounted for by underlying intrinsic molecular diversity within the tumor (11, 16, 17). Thus, identification of genetic programs that may help predict tumor response as afforded by genomic approaches could be of clinical benefit to patients with this disease and would facilitate the transition to personalized medicine.

Although HCC gene expression profiling can reveal its cause and prognosis, such analyses are ultimately dependent on fresh tissue specimens and specialized equipment, thus limiting its widespread clinical adoption (18, 19). Diagnostic imaging, conversely, is routine in clinical practice but perceived to lack molecular detail. Accordingly, noninvasive means that could extract elements of the large-scale gene expression information of a tumor could be of potential clinical benefit, particularly if these molecular surrogates contained information that could have an effect on disease management (eg, prediction of tumor drug response). We have previously demonstrated that noninvasive imaging phenotypes reflect underlying genomics and can serve as molecular surrogates of gene expression programs (20, 21). Given that gene expression programs can serve as a common language for cellular states and that clinical imaging can provide noninvasive portraits of tumor physiology and structure, we sought to identify whether a drug response gene expression program could be associated with a noninvasive imaging phenotype obtained with a standard clinical imaging modality.

In this study, we proposed that specific imaging traits of HCC obtained with conventional contrast medium–enhanced computed tomography (CT) could be associated with drug response gene expression programs defined by gene expression profiling. Because doxorubicin is a routine component in the chemotherapeutic regimen of transarterial chemoembolization for the treatment of HCC, we focused on identifying imaging phenotypes that could potentially be correlated with a previously defined doxorubicin gene expression program (22). Accordingly, we examined the genomic profiles of 30 HCCs against their matched biphasic CT scans to identify relationships between the tumor imaging appearance and the underlying tumor genomics related to this previously identified doxorubicin response gene expression program.

Back to Article Outline

Material and Methods 

This study was performed in accordance with institutional review board guidelines and approval.

CT Protocol 

All CT examinations consisted of conventional contrast-enhanced CT and were performed with a multidetector helical CT scanner (GE Medical Systems, Milwaukee, Wis) that generates four sections per gantry rotation. Low-molecular-weight nonionic iodinated contrast medium (120 mL of Isovue; Bracco Diagnostics, Princeton, NJ) was administered in a standard fashion by means of intravenous injection at a rate of 3–4 mL/sec through an antecubital vein. Initial unenhanced images were obtained, followed by contrast-enhanced images acquired at both arterial (2.5-mm-thick sections) and portal venous (5-mm-thick sections) phases with either timed runs or bolus tracking software. All images were retrospectively obtained and acquired within 1 month before surgical resection. None of the tumors included in this study received chemotherapy before resection.

Image Trait Selection and Evaluation 

Six predefined imaging traits were evaluated for their association with the doxorubicin response gene expression program. Imaging traits were selected a priori to reflect and incorporate qualitative aspects of tumor vascularity, texture, and margins as assessed with CT and are described in Table 1. Evaluated images consisted of matched contrast-enhanced CT scans of the arrayed HCC tumors obtained before resection and according to the CT protocol described earlier. Images of the 30 HCCs were evaluated and scored across these imaging traits by two board-certified radiologists in consensus (M.D.K. and C.B.S.).

Table 1. Summary of the Six Evaluated Imaging Traits, Trait Definitions, and Scoring System
Imaging TraitDefinitionScoring System
Internal arteriesThe presence or absence of discrete enhancing tubular structures within the tumor0 = absent, 1 = present
Texture heterogeneity, arterial phaseA measure of the internal structure and textural heterogeneity of the tumor evaluated at the arterial phase of imaging. For example, a score of 0 indicates that the internal structure and texture of the tumor was perfectly homogenous, and a score of 4 indicates that the tumor had a highly heterogeneous internal structure and texture.0–4
Wash-in, maximumThe maximum wash-in (relative increase in attenuation of the tumor from unenhanced to arterial phase imaging) of any portion of the tumor0–4
Washout, maximumThe maximum washout (relative decrease in attenuation of the tumor from arterial phase to portal venous phase imaging) of any portion of the tumor0–4
NecrosisThe presence of necrosis within the tumor, where necrosis is defined as an area with near total absence of enhancement on any contrast-enhanced phase of imaging0 = absent, 1 = present
Tumor margin score, arterial phaseA qualitative assessment of the transition zone of the tumor to liver evaluated at the arterial phase of imaging and scored from 0 to 4, where 0 is indicative of a perfectly demarcated tumor with a sharply defined transition between tumor and liver and 4 is indicative of an infiltrating morphology with a broad ill-defined transition along the entire periphery of the tumor.0–4

Clinical Outcomes 

Tumor TMN stage (n = 29) and the presence or absence of venous invasion (n = 30) by histopathology were evaluated in standard fashion as previously described (23).

Microarray Gene Expression Data 

HCC gene expression profiling data, extracted from two-color complementary DNA (cDNA) microarrays containing ∼23,000 clones and representing ∼18,000 genes, were extracted from the published dataset as previously described by Chen et al (23) for 77 HCC samples, including the 30 imaged HCCs. Data were retrieved from the Stanford Microarray Database (24) and filtered for spot quality as follows: within-spot, pixel-to-pixel correlation between channels larger than 0.4, no contamination, polymerase chain reaction (PCR) failure or software flag, and foreground signal at least 1.5 times background in each channel. Log ratios from probes (cDNA clones) representing the same UniGene cluster (UniGene build 184) were then averaged together. The resulting vectors for each gene were centered by mean and filtered out if fewer than 70% of values were present across either the entire set of 77 HCC samples (53 values required) or across the 30 imaged samples (21 values required). Missing values were replaced by using K nearest neighbors imputation before any further analysis was performed.

Gene Expression Programs 

Four predefined gene expression programs were used for radiogenomic analysis of HCC imaging and global gene expression profiling. The doxorubicin response gene expression program was curated from Gyorffy et al (22) as previously described. In addition, predefined multidrug resistance, liver-specific, and HCC venous invasion gene expression programs were similarly obtained (23, 25). All clones or probes from the four gene expression programs were mapped to genes by using UniGene build 184.

Data Analysis 

An overview of our radiogenomic integrative approach is outlined in Figure 1.

  • View full-size image.
  • Figure 1. 

    Overview of radiogenomic analysis used to identify imaging phenotypes associated with drug response and HCC-associated gene expression programs and pathologic outcomes.

Data analysis was performed by using R version 2.2.0 (26). Significance tests of gene association to imaging traits and clinical endpoints were performed by using the Significance Analysis of Microarrays (SAM) package for R, version 1.22 (27). Missing data were imputed by using the Impute package for R, version 1.0-4 (28). Agglomerative hierarchical clustering was performed with the Cluster 3.0 software by using uncentered Pearson correlation and centroid linkage (29).

Local background-corrected, log-ratio data for UniGene clusters and unmapped clones were tested for significance of association to each of six image traits and two clinical endpoints by using SAM analysis with 500 permutations. Linear least squares regression was used against actual values for continuously scored traits, and a t statistic was used for binary scores. TMN stage was treated as binary measures by classifying it as low (stages I and II) or high (stages III and IV). SAM produces false discovery rates (FDRs) for each gene; for the purpose of the gene expression program–level analysis described below, P values for each gene were determined by comparing the SAM score for the gene to the null distribution of SAM scores generated for all genes in the course of the permutation process. For consistency with the FDR analysis, genes with positive and negative scores were segregated for the purpose of computing P values. Thirty imaged HCC samples were included in tests of the image traits, whereas 77 HCC samples were included in tests of the clinical outcomes. Resulting per-gene values are included in supplemental data.

The significance of the relationship between each of the four gene expression programs and the image traits and outcomes was assessed by performing a hypergeometric test of the number of genes in each program (list) with a P value of less than .01 versus all genes in the data with a P value of less than .01. FDRs at the level of the gene expression program versus trait or outcome were determined by means of permutation. Because the liver-specific and venous invasion gene expression programs were determined from significance tests on a superset of the same data used here, the genes have a much higher level of nonindependence in these data than do those in the doxorubicin and multidrug resistance programs. Accordingly, different methods were used to determine the FDR. For the doxorubicin and multidrug resistance programs, we permuted scores 100 times, performing SAM analysis as above for each trait or outcome and assessing the significance as above of 400 randomly selected lists of genes—100 each of four sizes matching the actual gene expression programs—for each permutation of scores. Expected FDR values for 32 hypotheses (four programs × eight traits and/or outcomes) were calculated from these permutations for each of the P values corresponding to the resistance programs. A similar approach was taken for the liver-specific and venous invasion programs, but program membership was not permuted to preserve the dependency structure of the genes, producing a more conservative estimate of FDR. The relationship between image trait scores and clinical outcomes was determined by using SAM as above.

Doxorubicin resistance scores for the 30 imaged tumors were determined from measured gene expression log ratios and data presented by Gyorffy et al (22). Nine genes present in our data were either included in the predictive model for doxorubicin resistance or sufficiently well characterized for us to assign a weight of ±1, where a positive score indicates that the gene is upregulated in resistance samples and vice versa. For each of our samples, the resistance score was the sum of the log ratios for these nine genes weighted by these positive or negative weights. A positive score was indicative of a greater predicted resistance, whereas a negative score was indicative of a greater predicted sensitivity.

Supporting data with SAM analysis results and gene expression program membership are available in supplemental data.

Back to Article Outline

Results 

In total, 30 imaged and pathologically confirmed HCC in 30 patients were evaluated across each of the six evaluated imaging traits, and 8,364 measured genes passed our quality filters from DNA microarray analysis.

Doxorubicin Gene Expression Program 

The doxorubicin gene expression program consists of 61 genes (79 individual clones corresponding to 61 UniGene clusters) identified from gene expression profiling of a number of different doxorubicin resistance and sensitive human cell lines, as previously described (22). On the basis of our data filters and when accounting for intermicroarray platform differences, 22 of the 61 genes were present in our dataset and thus available for radiogenomic analysis.

To determine whether any of the six preselected imaging phenotypes could be associated with the doxorubicin gene expression program, we set out to analyze the imaging traits against the global gene expression profiles of each imaged tumor. The six imaging traits were evaluated for an association to individual gene expression patterns, followed by analysis for enrichment of significance in the doxorubicin resistance gene expression program as described earlier in Materials and Methods. Of the six imaging traits examined in this study, only one, “tumor margins score, arterial phase” (tumor margins score), was found to show significant correlation with the doxorubicin gene expression program (P = .0056, FDR = 0.076, Table 2). This imaging trait is a qualitative assessment of the interface between the tumor and adjacent parenchyma. Thus, an HCC imaging phenotype identified with conventional biphasic CT appeared to correlate with a predetermined drug response gene expression program. Furthermore, there appeared to be relative specificity in this imaging trait–gene expression program association because only one imaging trait showed significant correlation with the gene expression program.

Table 2. FDR Values for Each Imaging Trait and Pathologic Outcome for the Four Evaluated Gene Expression Programs
ParameterDoxorubicin ProgramMultidrug Resistance ProgramLiver-Specific ProgramVenous Invasion Program
Imaging trait
Tumor margin score, arterial phase0.076⁎0.4010.003⁎0.036⁎
Wash-in, maximum1.0001.0000.3771.000
Washout, maximum1.0001.0000.2501.000
Texture heterogeneity, arterial phase0.7201.0000.0470.361
Necrosis1.0001.0000.0471.000
Internal arteries0.4011.0000.003⁎0.079
Pathologic outcome
Venous invasion0.5661.0000.003⁎<0.003⁎
TMN stage0.3180.7380.005⁎0.003⁎

⁎Statistically significant (q < 0.1).

We next sought to evaluate the structure of this relationship. To visualize the organization of the doxorubicin gene expression program in relationship to the tumor margin score imaging trait, we first applied agglomerative hierarchical clustering to the 22 doxorubicin gene expression program genes and 30 tumor samples (Fig 2). Clustering enabled us to separate the tumors into two major classes on the basis of their gene expression by cutting the cluster dendrogram at the first bifurcation. We assessed the apparent doxorubicin resistance of each tumor by using a simplified classifier based on the Prediction Analysis of Microarrays (PAM) analysis performed by Gyorffy et al (22), as described in Materials and Methods. We then evaluated each tumor’s tumor margin scores and predicted doxorubicin resistance against their class assignment as defined by clustering. For each tumor’s tumor margin score and predicted doxorubicin resistance score, we performed a one-sided Wilcoxon rank-sum test, finding significant segregation of low and high scores in the two classes of tumors defined by the cluster dendrogram (P = .001 for predicted doxorubicin resistance, P = .002 for tumor margin score). It is interesting that tumors associated with relatively greater doxorubicin resistance on the basis of gene expression tended to have higher tumor margin scores, whereas tumors associated with relatively greater doxorubicin sensitivity tended to demonstrate lower tumor margin scores. Thus, a relationship between a specific imaging trait of the tumor captured at contrast-enhanced CT and expression level of genes implicated in doxorubicin response began to emerge from this analysis.

  • View full-size image.
  • Figure 2. 

    Hierarchical clustering of genes and tumor samples from the HCC doxorubicin gene expression program. Terminal branches of the dendrogram are colored according to the two major subgroups, as labeled. Tumor margin scores and doxorubicin gene expression resistance scores for each HCC are indicated below, with color intensity representing score magnitude. Bright red is indicative of a higher score and bright green is indicative of a lower score. Gene expression of the doxorubicin gene expression program is represented in table format Rows represent genes, and columns represent samples. Magnitude of log ratio gene expression is represented by color intensity, as indicated in the key, relative to the mean for each gene. A subset of named genes in the doxorubicin gene expression program is labeled. Representative CT scans of HCCs with high and low tumor margin scores are depicted to the left.

Multidrug Resistance Program 

Tumors have the ability to acquire resistance to drugs that share different chemical structures and mechanisms of action, more broadly known as multidrug resistance, through the activation or repression of certain genes or pathways (30). To determine the specificity encoded in our newly identified drug response gene expression program–imaging trait association, we next tried to determine whether the tumor margin score was specific for the doxorubicin gene expression program or if it also encapsulated a broader multidrug resistance gene expression program. Accordingly, we analyzed the association between the tumor margin score imaging trait and a predefined multidrug resistance gene expression program. This program consisted of 52 genes that were associated with multidrug resistance, as previously described (25). Thirty-three of 52 genes in the multidrug resistance program were present in our dataset.

We analyzed gene expression associated with the tumor margin score imaging trait against the multidrug gene expression program to determine if there was a correlation. On the basis of this analysis, no significant association was identified between tumor margin score and expression of the multidrug resistance gene expression program (Table 2). Because these two gene expression programs are relatively distinct in terms of their gene composition, with no overlapping genes present (with UniGene build 184), it is possible that other imaging traits may be distinctly associated with the multidrug resistance program and not the doxorubicin program. Therefore, we evaluated whether any of the other five imaging traits correlated with the multidrug resistance program. None of the other five imaging traits correlated with expression of the multidrug resistance gene expression program. Thus, on the basis of our analysis, the tumor margin score appears to be relatively specific for the doxorubicin gene expression program and is not associated with the multidrug resistance gene expression program, although it is important to note that the multidrug resistance gene expression program was derived from experiments on a very different microarray platform (Affymetrix GeneChips vs two-color spotted cDNA microarrays for our data and the doxorubicin program).

Liver-Specific Gene Expression Program 

Poorly differentiated HCC tumors tend to be more aggressive and invasive than well differentiated HCC, which in general is associated with worse outcome (31). Similarly, tumors that respond to treatment are usually correlated with improved patient outcome compared with nonresponders. Given these relationships, we next sought to characterize the relationship between our newly uncovered tumor margin score–doxorubicin gene expression program association and liver differentiation as gauged by a liver specific gene expression program highly correlated with hepatocyte differentiation (23). This liver-specific program consists of 333 genes (313 present in our dataset) that are enriched for liver-specific metabolic enzymes, clotting factors, complement components, and apolipoproteins. In general, this program has been observed to be upregulated in nontumor liver and downregulated in HCC consistent with classical observations of liver differentiation. Furthermore, downregulation of the liver-specific program in HCC is broadly related to upregulation of proliferation-associated genes; thus, poorly differentiated tumors tend to exhibit a more aggressive phenotype based on gene expression.

To examine this association, we first analyzed the relationship between the tumor margin score imaging trait and liver differentiation at the level of expression of the genes in the liver-specific program. Our analysis revealed that the tumor margin score was highly correlated with the liver-specific gene expression program, as shown in Table 2 (P < .001, FDR < 0.01). It is interesting that tumors with high tumor margin scores demonstrated poor differentiation at the gene expression level, with coordinate downregulation of the liver-specific program, whereas tumors with low tumor margin scores demonstrated better liver differentiation at the gene expression level, with coordinate upregulation of the liver-specific program (Fig 3). Analysis of the other five imaging traits against the liver-specific program revealed that the imaging trait “internal arteries” also showed significant correlation with the liver-specific gene expression program (P < .001, FDR < 0.01); however, the significance of its association was less than that with the tumor margin score.

  • View full-size image.
  • Figure 3. 

    Hierarchical clustering of genes and tumor samples from the HCC liver-specific gene expression program. Terminal branches of the dendrogram are colored according to the two major subgroups and labeled on the basis of the overall expression level of the liver-specific gene expression program genes. Good liver-specific function is indicated by black and poor liver-specific function is indicated by blue. Tumor margin scores are indicated below, with color intensity representing score magnitude as for Figure 2; log ratio gene expression values are shown below that.

Next, to determine the relative relationship between the tumor margin score, liver-specific program, and doxorubicin program, we analyzed the intersection of their respective genes (Fig 4). Although four genes in our dataset were in both the liver-specific and doxorubicin programs, only two—CYP27A1 and CYP4V2—were found to be significantly associated to the tumor margin score (FDR < 0.1). Both genes, however, are members of the cytochrome p450 superfamily, which encodes for drug metabolizing and detoxifying enzymes. Thus, the relative contribution of the genes associated with the tumor margin score to the doxorubicin program in terms of liver differentiation appears relatively small (two of 22 genes shared among all three groups), particularly when compared against the background of the association of the tumor margin score with the liver-specific program (94 of 313 genes in the liver-specific program significantly associated with tumor margin score [FDR < 0.1]). It is unclear, however, what the relative contribution of these two particular genes is to the doxorubicin program and, in particular, what their relative weight is to this program’s phenotypic manifestation in tumors.

Pathologic Outcomes Analysis and Venous Invasion Gene Expression Program 

Having identified a relationship between the tumor margin score imaging phenotype and two separate gene expression programs in HCC, we next sought to evaluate whether this imaging trait was also correlated with and reflected in more conventional disease markers as gauged by means of histopathologic examination. Accordingly, we analyzed our imaging traits against pathologic outcomes in our dataset. Microscopic venous invasion status was known for all 30 tumors, and TMN stage was known for all but one tumor. Venous invasion was graded for the presence or absence of histologic evidence of vascular invasion, whereas TMN stage was analyzed by using a two-class system as low (stages I and II) or high (stages III and IV).

Analysis demonstrated a significant association between the tumor margin score and venous invasion, as shown in Table 3 (P = .0047). Not surprisingly, the internal arteries imaging trait, which also showed significant correlation with the liver-specific program, was also significantly correlated with venous invasion (P = .0054), reinforcing, at a gene expression level, the classic observation of a relationship between liver differentiation and microscopic venous invasion. However, the tumor margin score was also correlated with TMN stage (P = .0019) in our analysis, whereas the internal arteries phenotype was not. Therefore, these results suggest that the tumor margin score captured additional information beyond that indicated by the presence of internal arteries.

Table 3. Relationship between Imaging Traits and Pathologic Outcome
Pathologic Outcome
Imaging TraitVenous InvasionTMN Stage
Tumor margin score, arterial phase.0047⁎.0019⁎
Texture heterogeneity, arterial phase.063.5027
Internal arteries.0054⁎.1408
Necrosis.448.8304
Wash-in, maximum.9915.0889
Washout, maximum.921.135

Note.—Data are given as P values.

⁎Statistically significant.

To define whether these associations seen at histopathologic examination were also reflected at the gene expression level, we first analyzed the relationship between the tumor margin score and a venous invasion gene expression program. Briefly, this program, defined by Chen et al (23), consists of 82 genes highly associated with microscopically determined venous invasion and is highly enriched in genes associated with both cellular proliferation and the liver-specific gene expression program. Our analysis demonstrated a statistically significant correlation between the tumor margin score and the venous invasion gene expression program (P < .001, FDR = 0.04) (Fig 5, Table 2) consistent with findings at histopathologic examination. The internal arteries image trait was not, surprisingly, also associated with the venous invasion gene expression program (P = .001, FDR= 0.08); however, none of the other image traits were significantly associated to it. Furthermore, TMN stage was also highly correlated with both the venous invasion and liver-specific gene expression programs (P < .001, FDR < 0.01 for both programs). Similarly, the presence of microscopic venous invasion in our dataset was also strongly correlated with the liver-specific program and, not surprisingly, the venous invasion program (P < .001, FDR < 0.01 for both programs). Thus, the results of these analyses confirmed a number of relationships between several HCC gene expression programs and phenotypic manifestations at the level of imaging and histopathology.

  • View full-size image.
  • Figure 5. 

    Hierarchical clustering of genes and tumor samples from the HCC venous invasion gene expression program. Histopathologic evidence of venous invasion is shown in yellow (presence of invasion) or blue (absence of invasion). Tumor margin scores and log ratio gene expression values are indicated below as in Figure 2.

Integrative Analysis of Tumor Margin Score Imaging Trait 

To explore the structure and organization of the associations between the different HCC phenotypes (tumor margin score and pathologic outcomes) and HCC gene expression programs, we integrated the previous analyses by applying two-way clustering to the outcomes and overall gene expression for each analyzed program. As can be seen, a general relationship emerges between the tumor margin score, pathologic outcomes, HCC-specific gene expression programs (liver-specific and venous invasion), and the doxorubicin gene expression program (Fig 6). In general, tumors with high tumor margin scores tended to group with gene expression patterns associated with doxorubicin resistance, poor liver-specific gene expression, and venous invasion and demonstrate pathologic outcomes associated with high TMN stage and histologic venous invasion, whereas the converse was true for tumors with low tumor margin scores. Thus, tumor margin score, in this analysis, can be broadly associated with a number of measures at both the gene expression program and pathology levels that are associated with a more aggressive HCC phenotype.

  • View full-size image.
  • Figure 6. 

    Hierarchical clustering of scores and summary expression values. Indicators of poor prognosis (venous invasion, advanced TMN stage, predicted doxorubicin resistance, high tumor margin score, and associated gene expression) largely occur together. Tumor margin score and predicted doxorubicin resistance are shown on a continuous scale as in other figures, with bright green indicating the lowest value and bright red the highest. Histopathologic evidence of venous invasion is indicated by green (negative) or red (positive). TMN stage is indicated by bright green (stage I), dim green (stage II), dim red (stage III), or bright red (stage IV). Liver-specific program gene expression is summarized as a simple average of the log ratios shown in Figure 3; green indicates less liver differentiation and red indicates more liver differentiation. Venous invasion program gene expression is shown as a weighted average, where log ratios were weighted by +1 or −1 depending on their direction of regulation as in Chen et al (23); green indicates downregulation of the venous invasion program (associated with the absence of invasion) and red indicates upregulation (associated with the presence of invasion).

Back to Article Outline

Discussion 

Functional genomic approaches can be used to identify gene expression programs associated with tumor treatment response (11, 22, 25, 32, 33). Herein, we demonstrated, by applying a radiogenomics approach, proof-of-concept that imaging phenotypes captured by means of conventional noninvasive imaging methods can be associated with certain drug response gene expression programs. By examining an HCC dataset with CT and gene expression profiling data, we identified a single imaging trait—the tumor margin score—that correlated with a previously defined doxorubicin response gene expression program. Moreover, the tumor margin score correlated with a simplified doxorubicin resistance classifier, demonstrating a general relationship between tumor margin score severity and predicted tumor doxorubicin resistance. Further analysis revealed a significant association between the tumor margin score and both liver differentiation and venous invasion at the gene expression level as well as tumor TMN stage. Results of histopathologic examination helped confirm the association with venous invasion. Thus, the imaging-genomic associations uncovered herein do not appear to be random and, in fact, support and lend further basis to classical observations of HCC behavior.

The results of this study reinforce the concept that there is information embedded in the images produced by medical imaging that can be associated with diverse genetic programs. Because cancer is a genetically heterogeneous disease, it is possible that different genetic programs associated with treatment response can be uncovered for other drugs or malignancies and associated with particular imaging phenotypes. Although it is unlikely that all imaging traits have coherent underlying biologic meaning or that imaging traits can be associated with every gene expression program, these results suggest that further studies investigating these associations and their extent are warranted.

Drug resistance is complex and diverse in nature. A number of genes have been implicated in drug resistance (30). ABCB1, which encodes for PGP, a member of the adenosine triphosphate–dependent membrane transport proteins, is one such gene that has received much attention in multidrug resistance. It has been associated with drug resistance in a number of malignancies, including HCC (34). It is interesting that ABCB1, although present in our dataset, was not characteristically overexpressed in the doxorubicin-resistant group as would be predicted. However, several enzymes involved in drug metabolism and detoxification—notably CYP27A1, a member of the cytochrome p450 family—were characteristically downregulated in the doxorubicin-resistant group. Furthermore, TIMP2, which is involved in extracellular matrix metabolism and cell migration, was also characteristically upregulated in the doxorubicin-resistant group in our analysis. Given that drug resistance can occur with a number of different mechanisms, it is possible that, in this gene expression model of drug resistance, the drug efflux pathway is not a significant contributor to HCC drug resistance but that other mechanisms may play a more substantial role.

Our results overall suggest a general association between severity of the tumor margin score imaging trait and an aggressive HCC phenotype that is, at least in part, manifested through broad activation or repression of a number of fundamental HCC gene expression programs (liver-specific and venous invasion). Genes found in these programs include a disproportionate number of liver differentiation, extracellular matrix metabolism, and cellular proliferation genes. The genes involved in the doxorubicin gene expression program exhibit a relatively diverse compositional nature with membership to a number of different functional classes such as cellular migration, transport, signaling, and detoxification, with little overlap with the HCC gene expression programs investigated in our analysis (22). It is therefore difficult to assess from this analysis the mechanistic associations, if any, that exist between the tumor margin score and its association with the doxorubicin-resistance, liver-specific, and venous invasion gene expression programs. Undoubtedly, additional studies will be necessary to more clearly define the molecular basis for the association between the doxorubicin-resistance gene expression program and the tumor margin score imaging trait.

It is important to stress that our investigations were exploratory in nature. The CT scans were not prospectively acquired, and the imaging traits examined were ultimately qualitative in nature. Additional independent studies with larger numbers will be necessary to confirm our findings. Nonetheless, it is promising that, by using imaging modalities currently used in the assessment and planning of HCC treatment, we were able to integrate large-scale genomic information in a coherent manner with imaging findings to address a focused question of potential treatment response at the gene expression level. These preliminary results offer promise that it may be possible in the near future to detect intertumoral differences in treatment response to certain drugs at the imaging level and, thus, help guide personalized treatment.

Back to Article Outline

References 

  1. Schena M, Shalon D, Davis RW, Brown PO. Quantitative monitoring of gene expression patterns with a complementary DNA microarray. Science. 1995;270:467–470
  2. DeRisi JL, Iyer VR, Brown PO. Exploring the metabolic and genetic control of gene expression on a genomic scale. Science. 1997;278:680–686
  3. Scheel JR, Kuo MD. Exploring the human genome in cancer using genomic approaches. J Vasc Interv Radiol. 2006;17:1225–1233
  4. Hsiao A, Kuo MD. High throughput biology in the post-genomic era. J Vasc Interv Radiol. 2006;17:1077–1085
  5. Alizadeh AA, Eisen MB, Davis RE, et al. Distinct types of diffuse large B-cell lymphoma identified by gene expression profiling. Nature. 2000;403:503–511
  6. Golub TR, Slonim DK, Tamayo P, et al. Molecular classification of cancer: class discovery and class prediction by gene expression monitoring. Science. 1999;286:531–537
  7. Lapointe J, Li C, Higgins JP, et al. Gene expression profiling identifies clinically relevant subtypes of prostate cancer. Proc Natl Acad Sci U S A. 2004;101:811–816
  8. Perou CM, Sorlie T, Eisen MB, et al. Molecular portraits of human breast tumours. Nature. 2000;406:747–752
  9. Pomeroy SL, Tamayo P, Gaasenbeek M, et al. Prediction of central nervous system embryonal tumour outcome based on gene expression. Nature. 2002;415:436–442
  10. Staunton JE, Slonim DK, Coller HA, et al. Chemosensitivity prediction by transcriptional profiling. Proc Natl Acad Sci U S A. 2001;98:10787–10792
  11. Holleman A, Cheok MH, den Boer ML, et al. Gene-expression patterns in drug-resistant acute lymphoblastic leukemia cells and response to treatment. N Engl J Med. 2004;351:533–542
  12. Troester MA, Hoadley KA, Sorlie T, et al. Cell-type-specific responses to chemotherapeutics in breast cancer. Cancer Res. 2004;64:4218–4226
  13. Edwards BK, Brown ML, Wingo PA, et al. Annual report to the nation on the status of cancer, 1975-2002, featuring population-based trends in cancer treatment. J Natl Cancer Inst. 2005;97:1407–1427
  14. Llovet JM, Real MI, Montana X, et al. Arterial embolisation or chemoembolisation versus symptomatic treatment in patients with unresectable hepatocellular carcinoma: a randomised controlled trial. Lancet. 2002;359:1734–1739
  15. Lo CM, Ngan H, Tso WK, et al. Randomized controlled trial of transarterial lipiodol chemoembolization for unresectable hepatocellular carcinoma. Hepatology. 2002;35:1164–1171
  16. Paez JG, Janne PA, Lee JC, et al. EGFR mutations in lung cancer: correlation with clinical response to gefitinib therapy. Science. 2004;304:1497–1500
  17. Mellinghoff IK, Wang MY, Vivanco I, et al. Molecular determinants of the response of glioblastomas to EGFR kinase inhibitors. N Engl J Med. 2005;353:2012–2024
  18. Lee JS, Chu IS, Heo J, et al. Classification and prediction of survival in hepatocellular carcinoma by gene expression profiling. Hepatology. 2004;40:667–676
  19. Iizuka N, Oka M, Yamada-Okabe H, et al. Differential gene expression in distinct virologic types of hepatocellular carcinoma: association with liver cirrhosis. Oncogene. 2003;22:3007–3014
  20. Kuo MD, Sirlin CB, Chan BK, Matcuk G, Ooi C, Chen X. Genome wide imaging characterization of hepatocellular carcinoma. 2004;Presented at the 90th Scientific Assembly and Annual Meeting of the Radiological Society of North America, Chicago, Ill, December 2
  21. Kuo MD, Wang DS, Nardini C, Diehn M, Cha S, Chan BK. Identification of alterations in global gene expression in glioblastoma multiforme with magnetic resonance imaging. 2004;Presented at the 90th Scientific Assembly and Annual Meeting of the Radiological Society of North America, Chicago, Ill, December 1
  22. Gyorffy B, Serra V, Jurchott K, et al. Prediction of doxorubicin sensitivity in breast tumors based on gene expression profiles of drug-resistant cell lines correlates with patient survival. Oncogene. 2005;24:7542–7551
  23. Chen X, Cheung ST, So S, et al. Gene expression patterns in human liver cancers. Mol Biol Cell. 2002;13:1929–1939
  24. Stanford micoarray database. Stanford University web site. http://smd.stanford.edu. Accessed January 2006.
  25. Gyorffy B, Surowiak P, Kiesslich O, et al. Gene expression profiling of 30 cancer cell lines predicts resistance towards 11 anticancer drugs at clinically achieved concentrations. Int J Cancer. 2006;118:1699–1712
  26. R Development Core Team (2005). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing, ISBN 3-900051-07-0, URL http://www.R-project.org. Accessed January 2006.
  27. Tusher VG, Tibshirani R, Chu G. Significance analysis of microarrays applied to the ionizing radiation response. Proc Natl Acad Sci U S A. 2001;98:5116–5121
  28. Hastie T, Tibshirani R, Narasimhan B, Chu G. Impute: imputation for microarray data. R package version 1.0-4. Accessed January 2006.
  29. Eisen MB, Spellman PT, Brown PO, Botstein D. Cluster analysis and display of genome-wide expression patterns. Proc Natl Acad Sci U S A. 1998;95:14863–14868
  30. Gottesman MM, Fojo T, Bates SE. Multidrug resistance in cancer: role of ATP-dependent transporters. Nat Rev Cancer. 2002;2:48–58
  31. Ng IO, Lai EC, Chan AS, So MK. Overexpression of p53 in hepatocellular carcinomas: a clinicopathological and prognostic correlation. J Gastroenterol Hepatol. 1995;10:250–255
  32. Chang JC, Wooten EC, Tsimelzon A, et al. Gene expression profiling for the prediction of therapeutic response to docetaxel in patients with breast cancer. Lancet. 2003;362:362–369
  33. Kang HC, Kim IJ, Park JH, et al. Identification of genes with differential expression in acquired drug-resistant gastric cancer cells using high-density oligonucleotide microarrays. Clin Cancer Res. 2004;10:272–284
  34. Goldstein LJ, Galski H, Fojo A, et al. Expression of a multidrug resistance gene in human cancers. J Natl Cancer Inst. 1989;81:116–124

 From the 2006 SIR Annual Meeting.None of the authors has identified a conflict of interest.

PII: S1051-0443(07)00776-2

doi:10.1016/j.jvir.2007.04.031

Journal of Vascular and Interventional Radiology
Volume 18, Issue 7 , Pages 821-830, July 2007

Article Tools

* Image Usage & Resolution