Association of metabolic and genetic heterogeneity in head and neck squamous cell carcinoma with prognostic implications: integration of FDG PET and genomic analysis

Cancer is a heterogenous disease at genetic, epigenetic, and phenotypic levels [1]. Cancer progression is driven by a genetic process of clonal evolution, which eventually causes tumor genetic heterogeneity, a tumor with multiple subsets of subclonal mutations [1]. Acquired tumor genetic heterogeneity is caused by the selective pressures during the evolution process and affected by tumor vasculature and immune system in the microenvironment [2]. Furthermore, genetic heterogeneity eventually drives the phenotypic heterogeneity of tumor by interacting environmental factors [3]. Heterogeneous subsets of tumor have different molecular targets, which may result in different levels of resistance to the cancer treatment [4]. Accordingly, tumor heterogeneity is associated with the progression and eventual clinical outcomes of cancer patients [5]. Thus, evaluation of tumor heterogeneity is crucial for selecting anticancer strategies and predicting clinical outcomes [6]. Advances in next-generation sequencing (NGS) have allowed for extensive understanding of tumor genetic heterogeneity and provided useful features to evaluate of tumor heterogeneity [7, 8]. The mutant allele tumor heterogeneity (MATH), a genetic heterogeneity feature, is easily calculated as a percentage of mutant allele frequencies among tumor-specific mutated loci. MATH has been known to have a prognostic value in HNSC and colon cancer [9, 10].

Phenotypical heterogeneity can be noninvasively studied using various imaging techniques including computed tomography (CT), magnetic resonance imaging (MRI), and ¹⁸F-fluorodeoxyglucose positron emission tomography (FDG PET) [11]. FDG PET is a compelling image modality to evaluate metabolic heterogeneity of tumors, a phenotypic tumor heterogeneity [12]. Recently, heterogeneity parameters obtained using FDG PET have been extensively evaluated and reported to have diagnostic and prognostic values in multiple types of malignancies including HNSC, non-small cell lung cancer, and pancreatic cancer [12‐16]. Although the metabolic features evaluated by FDG PET are closely associated with biological factors in the tumor microenvironment [12, 17, 18], it is still unknown whether metabolic heterogeneity is associated with genetic heterogeneity [14].

Herein, we investigated if metabolic heterogeneity based on FDG PET was associated with genetic heterogeneity represented by MATH. Furthermore, we explored the prognostic value of both metabolic and genetic heterogeneity features in predicting the outcomes of patients with HNSC.

We found that the tumor metabolic features estimated by FDG PET showed a mild but statistically significant level of association with tumor genetic heterogeneity. Specifically, tumor metabolic-volumetric and metabolic heterogeneity features of FDG PET were associated with MATH in a mild degree. This finding supports the notion that quantifiable FDG uptake features reflect the tumor heterogeneity at the genomic level in HNSC [10]. Also, there was additive prognostic value when the FDG PET and genetic heterogeneity features were combined. Additionally, both genetic heterogeneity feature (MATH) and glycolysis feature (GlycoS) were independently predictive of OS even after adjusting for clinicopathologic features.

Recently, tumor imaging phenotypes were found to be related to gene expression profiles in HNSC [10, 17, 29‐32]. Specifically, SUV and heterogeneity features estimated by FDG PET were related to 1177 differentially expressed genes in normal and tumor tissues [10]. Previous studies have shown a link between FDG uptake and several specific genes which modulate glucose metabolism [33‐35]. Also, phenotypic whole tumor-level heterogeneity can be noninvasively recognized by various imaging techniques including computed tomography (CT), magnetic resonance imaging (MRI), and FDG PET [11]. However, it has been unclear whether the genetic heterogeneity assessed by a small sample of tumor tissue can reflect the whole tumor-level phenotypic heterogeneity or not [6]. Also, there has been no study to evaluate the association of tumor heterogeneity measured by FDG PET and genomic analysis in patients with HNSC. As cancer cells are evolved in a heterogeneous spatiotemporal environment based on genetic heterogeneity, we hypothesized that genetic heterogeneity might be associated with whole tumor level heterogeneity measured by FDG PET. In this study, by utilizing the database of TCGA and TCIA, we were able to find that there is an association between whole tumor level heterogeneity based on FDG PET and genetic heterogeneity in HNSC. Although the association was statistically significant, the level of association was weak with correlation coefficients of 0.4~0.5. This weak level of association was not a surprise because the methods to measure the tumor heterogeneity were totally different between FDG PET heterogeneity parameters and MATH. MATH was obtained from genetic sequencing data of a small portion of tumor, while FDG PET heterogeneity parameters were calculated from an imaging data reflecting metabolic status of a whole tumor area. It is noteworthy that there was a mild degree of association between MATH and FDG PET heterogeneity parameters, even with this striking difference of the methods to measure the tumor heterogeneity.

MATH is a genetic heterogeneity measure, which can be easily quantified as a percentage of mutant allele frequency among tumor-specific mutated loci. Also, the prognostic value of MATH has been validated in HNSC and colon cancer [9, 36, 37]. In the patients with HNSC, high MATH score was associated with increased mortality [36, 37]. Also, MATH was associated with the risk of metastases in patients with colon cancer [9]. However, the ability of MATH to represent tumor heterogeneity has not been tested by other modalities. We have demonstrated that MATH was highly associated with the representative heterogeneity features from FDG PET (entropy, COV). This result is in line with a recent study by Moon et al. They showed that Shannon’s heterogeneity index was associated with entropy in patients with small cell lung cancer [38]. Furthermore, we found that MATH was predictive of OS in patients with HNSC even after adjusting clinicopathologic features.

Recent meta-analyses showed that various FDG PET features including SUVmax, MTV, and TLG were prognostic factors in multiple types of malignancies [39‐42]. Also, heterogeneity features of FDG PET have shown to be associated with treatment response and clinical outcome in multiple types of malignancies [13, 15, 43‐45]. Among the heterogeneity features, entropy and COV have been widely accepted and proven to be useful for predicting treatment response and clinical outcomes [13, 15, 44]. For example, entropy was predictive of OS in pancreatic cancer [15], and the changes in entropy were independently associated with treatment response in erlotinib-treated non-small cell lung cancer [13]. Also, COV was superior to conventional parameters in predicting therapy response and disease progression in rectal cancer [44]. We found that entropy and COV were strongly associated with MATH, a genetic heterogeneity feature, which re-enforced the genetic background of the features and thus increased possibility of clinical utilization of the features.

MTV and TLG are radiomic features that represent metabolic-volumetric tumor burden. MTV is a measurement of tumor volume with increased glucose metabolism, while TLG is the product of MTV and the mean SUV of the volume. MTV and TLG are considered to be better prognostic factors than simple metabolic feature such as SUVmax [41, 46, 47]. In this study, we also found that MTV and TLG are significantly associated with genetic heterogeneity. As the tumor spatially grows, the larger volume of tumor likely to be more heterogenous reflecting genetic heterogeneity by cancer evolution. Multiple studies have shown that MTV and heterogeneity features of FDG PET such as COV and texture features are associated [48‐50]. Therefore, tumor metabolic-volumetric features are likely to be an indicator of tumor genetic heterogeneity due to cancer evolution. Also, the association of MTV and TLG with genetic heterogeneity may further explain the robustness of the features in predicting the clinical outcomes.

Glycolysis is a crucial pathway regulating oncogenes, tumor suppressor genes, and glycolytic enzymes as well as accelerating cell proliferation in cellular metabolism [51]. Factors of metabolic glycolysis are associated with poor prognosis and tumor resistance to therapy in HNSC [52]. Also, glycolysis gene expression correlates FDG uptake features [30, 33‐35]. However, the previous studies only explored the relationship of representative genes such as glucose transporter (GLUT) or hexokinase (HK). On the other hand, we utilized a novel glycolysis signature, GlycoS, which was derived from multiple glycolysis-associated genes defined by Reactome [19, 53]. We found that metabolic-volumetric features (MTV, TLG) were significantly associated with GlycoS. Unexpectedly, SUVmax and SUVpeak were not associated with GlycoS. One potential explanation is that many glycolysis-associated genes may not influence the intensity of FDG uptake, because the FDG uptake kinetics is primarily determined by glucose transportation by GLUT and phosphorylation by HK [54]. Even though, a higher number of patients may prove the associations between SUVmax and GlycoS, since there was a trend of positive correlation (P = 0.114). Also, we found that GlycoS is predictive of OS in patients with HNSC. Furthermore, GlycoS was predictive of OS even after adjusting MATH and clinicopathologic features. This implies that GlycoS has an additive prognostic value over MATH.

There are several limitations to this study. First, a limited number of samples were available in public archives. Also, we found only a mild degree of association between the genetic and FDG PET heterogeneity, and there were large number of scatters outside of the standard deviation (Fig. 2). Further studies will yield clearer results if analyzed using a larger number of expanded data. Second, FDG PET data from TCIA were applied to different technologies, reconstruction, and attenuation correction methods. So each image is difficult to compare to each other, and even SUVmax values vary, which may affect clinical decision [55]. To solve this problem, we did voxel interpolation to make all images have uniform voxel sizes. Also, we used entropy and COV as a tumor heterogeneity texture features because these are the most reliable upon reconstruction method. Third, although metabolic features and genomic signatures obtained in this study were candidates for future biomarkers, these are not validated precisely. Although we used representative genomic and metabolic features which have clinical implications with prognosis, there might be better features than these eight features. In a future study, more features could be considered for understanding cause and effect through systemic tumor biology. Nonetheless, our results show a correlation between genetic heterogeneity features and metabolic heterogeneity features and prognostic value about each feature.

Tumor genetic heterogeneity was associated in a mild degree with metabolic heterogeneity measured by FDG PET in patients with HNSC. Genetic and metabolic heterogeneity features were predictive of OS, and there was additive prognostic value when the FDG PET and genetic heterogeneity features were combined. Moreover, genetic heterogeneity feature (MATH) and glycolysis feature (GlycoS) were independent predictors of OS.