PET-Based Radiogenomics Supports mTOR Pathway Targeting for Hepatocellular Carcinoma

2[18F]fluoro-2-deoxy-d-glucose PET and CT (FDG-PET/CT) are diagnostic tools for imaging selective uptake of the radiolabeled glucose analog, FDG, by cancer cells, which are characterized by enhanced glucose metabolism (1). In general, FDG-PET/CT plays a practical role in detecting metastatic foci, rather than identifying intrahepatic tumors, in the staging of patients with hepatocellular carcinoma (HCC; ref. 2–4). Increased PET uptake is likely to be most useful as an oncologic marker for selecting resection or transplant candidates and predicting posttreatment outcomes (5–8).

Radiogenomics has been mainly used to create imaging biomarkers mirroring the genomolecular makeup of malignant disease without needing tissue sampling (9–11). FDG-PET/CT imaging combined with anatomic and functional properties could be an excellent tool for radiogenomic approaches in oncology care (12). Despite the widespread clinical utility of FDG-PET imaging in multiple cancer types, the underlying cellular or molecular bases have only been investigated in a few types such as breast, colon, and lung cancers (13–15). The relevant issues still remain unclear in HCC. Fortunately, the intensity of FDG-PET uptake in HCC lesions varies markedly across tumors and patients (7, 16–18), and this could facilitate radiogenomic studies of the disease.

To address this matter, we examined the prognostic significance of preoperative FDG-PET uptake by the target lesion in patients undergoing surgical resection for HCC, and analyzed the molecular and genotypic correlates of high intensity FDG-PET uptake. The ultimate goal of this study was to identify therapeutic targets based on stratifying patients with HCC in terms of radiogenomic surrogates. To this end, we designed in vitro and in vivo experiment to characterize FDG-avid HCC cell lines.

The current study was approved by the Institutional Review Board (IRB) of Asan Medical Center (No. 2019–1321), and the study was conducted in accordance with the Declaration of Helsinki. Specimens provided by the Bio-resource center were obtained from patients who gave written informed consent. A flow diagram of the overall study design is displayed in Fig. 1.

We first screened a preconstructed panel of 206 patients who received curative hepatectomy for HCC between 2005 and 2011 at Asan Medical Center and for whom there were both whole-exome sequencing and RNA sequencing (RNA-seq) data on resected tumor samples (19–21). We finally curated 117 individuals in whom FDG-PET/CT scans obtained within 4 weeks before surgery were available, and used them as a discovery set (Supplementary Table S1). No patients received prior local or systemic anti-HCC treatment. Most were infected with hepatitis B virus (HBV; 77.8%) and had early-stage disease [i.e., Barcelona Clinic Liver Cancer (BCLC)-0/A; 95.7%]. To confirm the robustness and generality of the results from this set, we gathered a validation set of 81 patients with HCC of different composition who underwent primary hepatic resection between 2009 and 2013, and for whom we had surgical specimens archived in our own Bio-resource center (http://brc.amc.seoul.kr), in addition to preoperative FDG PET/CT scans. The archive is an IRB-approved human biorepository in Asan Medical Center. Over half of the validation cohort had non–early stages of HCC (i.e., BCLC-B/C; 55.6%), and 39.5% of the tumors were of non-HBV origin (Supplementary Table S2). RNA-seq data were newly generated for each surgical sample in this set. See Supplementary Data for details.

To examine the external consistency and retest reliability of the prognostic power of the molecular signatures relevant to FDG-avid tumors, we used the Cancer Genome Atlas Liver Hepatocellular Carcinoma (TCGA-LIHC) and Liver Cancer-RIKEN, Japan (LIRI-JP) datasets comprising 262 and 191 samples, respectively, with pathologically confirmed HCC and full outcome data (22, 23). The two public-access gene expression datasets were downloaded from the cBioPortal (RRID:SCR_014555) and International Cancer Genome Consortium data portal (https://dcc.icgc.org), respectively. In terms of outcome data, both overall survival (OS) and recurrence-free survival (RFS) were available in the TCGA-LIHC database, but only OS was available for the LIRI-JP samples. Details of the demographic and clinical characteristics of each sample contained in the two datasets are provided in Supplementary Tables S3 and S4.

For details of the PET/CT imaging protocols, see Supplementary Data. All PET/CT images were analyzed using a dedicated workstation (Syngo.via VB20A, Siemens, Erlangen, Germany) by two board-certified nuclear medicine physicians working in consensus (J.S. Kim and M. Oh). For quantitative analysis, maximum standardized uptake values (SUVmax) were measured by drawing a volume-of-interest (VOI) for the tumor in the liver (24). In cases of multiple lesions, the SUVmax that showed the highest FDG uptake was recorded. Tumor to non-tumor SUV ratios (TNR) were calculated as the ratios between the SUVmax of tumors and the mean SUV of liver backgrounds determined by averaging the SUVs of three 1-cm spherical lesion-free VOIs– two in the right lobe and one in the left lobe (25).

On the basis of visual analysis of the largest tumor, tumors were classified as FDG-avid or FDG–non-avid depending on whether FDG-PET uptake was higher than background hepatic activity or similar to or lower than background hepatic activity (26). Representative FDG-PET/CT images of each group are presented in Fig. 2A.

DNA- or RNA-seq data were obtained from the largest tumor in subjects in both the discovery and validation cohorts with ≥2 intrahepatic lesions. To identify functional sets of genes associated with FDG uptake, single-sample scores for each gene set were calculated from the transcriptomic data by the single-sample Gene Set Enrichment Analysis (GSEA) method implemented in Genepattern (27, 28). Differentially expressed genes in the FDG-avid group and their protein–protein interaction networks were identified and deduced using the STRING (RRID:SCR_005223) database (https://string-db.org; ref. 29).

We developed an “FDG-avid signature”, which consisted of the top hundred genes significantly upregulated in the FDG-avid group of the discovery cohort in order of P value (Supplementary Table S5), and tested its prognostic value in the external public datasets. See Supplementary Data for the further transcriptomic analyses.

Five well-characterized human HCC cell lines (Huh7, PLC/PRF/5, Hep3B, HepG2, and SNU182) were obtained from the Korean Cell Line Bank, Seoul, Republic of Korea. Sorafenib-resistant Huh7 cells (Huh7-SR) were generated by exposure of Huh7 cells to stepwise increments of sorafenib concentration (Supplementary Figs. S1 and S2). The cell lines were authenticated upon receipt by short tandem repeat (STR) profiling, using an AmpFlSTR Identifier Kit (Applied Biosystems, Foster City, California, Catalog No. 4322288), and based on available STR profiles. Before use, all cells were screened for the presence of Mycoplasma by PCR using a Mycoplasma detection kit (Myco-ReadTM, BioMAX). The reported experiments were performed less than 2 months after thawing early passage cells. Details of in vitro studies, including 18F-FDG uptake testing, cell viability assays, and Western blot analysis are given in the Supplementary Data. PLC/PRF/5 cell line-based xenograft NRG mouse models were used for animal studies. Experimental designs and procedures, including animal FDG-PET imaging and IHC staining, are presented in the Supplementary Data. Omipalisib (GSK2126458; Sellckchem) was used to block the mTOR pathway in both in vitro and in vivo settings. All animal study protocols were approved by the Institutional Animal Care and Use Committee of Asan Institute for Life Sciences (No. 2019–02–341).

For statistical comparisons, the χ² test, Fisher exact test, Student t test, or Mann–Whitney U test was performed when appropriate. RFS was calculated from the date of tumor resection until detection of the first HCC recurrence, death, or the last follow-up. Survival curves were analyzed by the Kaplan–Meier method and compared using the log-rank test. Independent factors for RFS were evaluated by multivariate Cox regression analysis. Significance was defined as two-tailed P < 0.05. All statistical analyses were performed with R Version 3.4.1.

Normalized gene expression data for the discovery and validation cohorts have been deposited in GEO (https://www.ncbi.nlm.nih.gov/geo; accession nos. GSE124751 and GSE164121, respectively).

Of the 117 patients in the discovery cohort, 57 (48.7%) had FDG-avid tumors. The FDG-avid group was associated with worse tumor biology in terms of poorly differentiated tumors, microvascular invasion, and serum α-fetoprotein (AFP) elevation (Supplementary Table S1); similar trends were also observed in the validation set of 81 patients (Supplementary Table S2). RFS time was shorter in the FDG-avid group, with 5-year RFS rates of 48% in the avid group and 68% in the non-avid group, and this difference remained significant after adjusting for all clinical covariates including traditional predictors such as microvascular invasion and serum AFP level [adjusted hazard ratio (HR), 1.69; 95% confidence interval (CI), 1.01–2.83; P = 0.045; Fig. 2B; Supplementary Table S6). The difference was even clearer in the validation set of more advanced tumors [adjusted HR, 2.45; 95% CI, 1.33–4.51; P = 0.004 for the FDG-avid group (60.5%; n = 49) vs. the FDG–non-avid group (39.5%; n = 32); Fig. 2B; Supplementary Table S6].

In terms of the predefined molecular classifications of HCC (30–34), Hoshida S1 (43.9%), Boyault G1-G3 (56.1%), Chiang's proliferation (57.9%), and Lee's (73.7%), and Roessler's A (64.9%) classes, along with proliferation properties, were the most prevalent subclasses in the FDG-avid group of the discovery set (Fig. 2C). In contrast, the FDG–non-avid group members were mainly Hoshida S3 (58.3%) and Boyault G4 (45.0%) tumors, which are known to be well differentiated. In addition, the FDG–non-avid tumors were significantly enriched in the Lee's (70.0%) and Roessler's B (65.0%) clusters associated with favorable survival (32, 34). Gene sets representing poor prognostic signatures for HCC (33–36) were also significantly associated with intense FDG uptake (Fig. 2D).

To see whether HCC-driven FDG uptake reflected the extent of glucose metabolism, we measured the expression of glycolysis-related genes and examined their relationships with SUVs on PET in the discovery samples (Supplementary Fig. S3). The expression of the glycolytic genes, SLC2A1 and HK2, encoding glucose transporter 1 (GLUT1) and hexokinase 2 (37), respectively, was found to be positively correlated with TNR expression (r = 0.50, P < 0.001; and r = 0.32, P < 0.001, respectively). In contrast, an inverse relationship with TNR was observed for the G6PC gene, which is involved in gluconeogenesis, the opposite process to glycolysis (37). As anticipated, expression of both glycolytic genes was significantly higher in the FDG-avid group (Ps < 0.01). These observations demonstrate that the FDG-avid HCCs are metabolically active cells with increased uptake and trapping of FDG, possibly due to upregulation of GLUT1 and hexokinase 2 and downregulation of glucose 6-phosphatase (37).

Single-sample GSEA in the discovery set based on Hallmark databases revealed that FDG-avid tumors were significantly enriched for gene sets related to glycolysis and its regulatory partners in the cell cycle (Fig. 3A). Strikingly, both the MTORC1 and PI3K_AKT_MTOR pathways and their core components such as TSC2, MTOR, RHEB, and EIF-4EB1 were strongly upregulated in PET-positive cases, suggesting that activation of mTOR signaling could be a key element of the glycolytic phenotype in tumor cells (Fig. 3A and B). In addition to the individual involvement in mTOR and glycolysis-related genotypes detected by GSEA, high biological connectivity of key components in the two genetic pathways was also observed by visualizing the protein-networks of differentially expressed genes in FDG-avid HCCs, as expressed by a high average local clustering coefficient of 0.83 (Fig. 3C). We next examined somatic mutations and DNA copy-number alterations, focusing on a list of 17 genes in the canonical PI3K/AKT/MTOR pathway. A number of genes in the pathway were found significantly altered in the FDG-avid group, including PIK3CA, PTEN, TSC2, RICTOR, and MTOR (36.8% vs. 18.3%, P = 0.042; Fig. 3D). On the other hand, genes associated with hepatic physiology including bile acid and fatty acid metabolism, adipogenesis, and coagulation were downregulated in samples with high FDG uptake. Other gene sets including TGFβ signaling- and immune response–related pathways were not significantly associated with FDG-PET avidity (Supplementary Fig. S4). These molecular findings were confirmed in the validation set of more advanced-stage tumors, which might become potential major candidates for mTOR pathway inhibitors (Fig. 3E).

We divided the levels of expression of the FDG-avid signature based on the single-sample GSEA enrichment scores: high for the top 25%, mid for the 25%–75% group, and low for the bottom 25% group. We next examined the prognostic relevance of this grouping in the two public HCC sets, of multiethnic (TCGA-LIHC; n = 262) and Japanese origin (LIRI-JP, n = 191), respectively, as shown in Fig. 4. Tumors with high FDG-avid signatures were significantly associated with the poorest survival outcomes, followed by the mid FDG-avid signature samples in both cohorts. These correlations point to a genetic contribution to the FDG-PET uptake that reflects the level of glucose utilization in cancer cells.

Based on the radiogenomic data identified in clinical settings, we next investigated the changes in metabolic and proliferative behavior upon inhibition of the mTOR pathway in in vitro models of HCC. We first screened six representative HCC cell lines, and, among these, we selected for further experiments PLC/PRF/5, Huh7, and Huh-SR cells, which displayed increased glucose uptake (Fig. 5A). To evaluate the potential of mTOR inhibition for treating FDG-avid HCC, we tested the effect of omipalisib, a second-generation dual PI3K/mTOR inhibitor, in phase I clinical trials against solid tumors (38), in the three selected cell lines, that showed the strongest anticancer efficacy against those cell lines after screening a library of 17 anticancer drugs targeting the PI3K–mTOR pathway (Supplementary Fig. S5). The IC₅₀ values for omipalisib ranged from 0.07 to 1.3 μmol/L in both the parental and sorafenib-resistant cell lines, while those for sorafenib, lenvatinib, and regorafenib, which are first-line or rescue drugs for sorafenib failure currently used to treat HCC in clinical practice (6), were mostly >4.0 μmol/L (Fig. 5B). Interestingly, omipalisib was much more effective in Huh7-SR cells than regorafenib, a standard rescue drug for sorafenib (39). Treatment with increasing concentrations of omipalisib for 72 hours reduced the viability of FDG-avid PLC/PRF/5 and Huh7 to a greater extent than that of non-avid cells such as Hep3B and HepG2 (Supplementary Fig. S6). Glucose uptake was already reduced in all three cell lines shortly after 24 hours of exposure to 0.5 μmol/L omipalisib, and this was accompanied by decreased expression of major modulators of glycolysis such as GLUT1 and hexokinase 2, and these findings were restricted to the FDG-avid cell lines (Fig. 5C and D; Supplementary Figs. S7 and S8). However, a significant drop in cell viability in all these cell lines was evident from 48 hours of treatment (Fig. 5E). Western blotting also showed that omipalisib resulted in decreased phosphorylation of mTOR and its downstream targets in the early phase (24 hours) of treatment (Fig. 5F).

This antitumor effect of mTOR inhibition via suppression of glucose metabolism was next examined in vivo using tumor xenografts derived from PLC/PRF/5 cells (Fig. 6A). Two weeks after injecting the cells, when tumor volume was around 150 mm³, we performed baseline FDG-PET imaging and then randomized the mice to receive either 1.5 mg/kg omipalisib (n = 7) or control vehicle (n = 7). After 3 days, FDG uptake began to decrease dramatically in the omipalisib-treated animals (Fig. 6B and C), and this was eventually followed by a decline in tumor volume (Fig. 6D and E), and remained low over the entire 3-week duration of treatment. IHC staining indicated that omipalisib exerted its antitumor effect by targeting mTORC1-dependent phosphorylation of 4E-BP1 and S6 ribosomal protein rather than phosphorylation of Akt, a sensitive substrate of mTORC2 (Fig. 6F; Supplementary Fig. S9). Collectively, these data suggest that the FDG-PET avidity per se of HCC tumors could dictate their susceptibility to the mTOR inhibitor.

Since the glycolytic switch of cancer cells is controlled by nonmetabolic oncogenic regulators (40), FDG uptake in PET scans could serve as an oncologic surrogate bridging the phenomics and genomics of malignant diseases (1, 41). Our nuclear imaging-molecular correlative study found that aside from cell-cycle checkpoints reciprocally interconnected with metabolic pathways, the signaling network of mTOR cascades was robustly activated in FDG-avid HCCs—which have a poor prognosis—and this did not appear to depend on tumor stage. The same findings were noted for gene mutations in the PI3K/AKT/mTOR circuitry, which are known to be more prevalent in Asians with HCC (42, 43). In addition, the prognostic correlations of a nucleogenomic signature related to FDG-positive lesions in various racial and regional cohorts lend support to the idea that the molecular components involved play key roles in the aberrant glucose metabolism in HCC; these findings also avoid a limitation of the current study, namely that most of the HCC examined were due to HBV (Supplementary Tables S3 and S4; refs. 22, 23). Subsequent investigation showed that the reduction in FDG uptake shortly after inhibiting mTOR was followed by a significant decrease of cell viability and tumor regression specifically in the FDG-avid cell lines, which in turn may point to mTOR-dependent metabolic reprogramming of the HCC.

As a radiogenomic tool, PET has the advantage of allowing assessment of functional and physiologic activity with greater accuracy than anatomical imaging modalities such as CT or magnetic resonance imaging (41). 18F-FDG, an 18F-labeled glucose analogue, is the tracer of choice for detecting the increased glucose turnover in cancerous tissues (27). Features of tumor images extracted from FDG-PET/CT have been directly associated with type-specific genetic alterations in cancers, including epithelial–mesenchymal transition-related genes in lung cancer, MYC target genes in breast cancer, and KRAS/BRAF genes in colon cancer (14, 15, 44). The current study revealed the central metabolic effects of mTOR underpinning the intense FDG uptake in HCC tumors. The same correlation between (phosphorylated) mTOR expression and positive FDG-PET scans has been reported in IHC studies of non–small cell lung cancer and mesothelioma neoplasia (45, 46). The mTOR signaling pathway is actively involved in major aspects of HCC behavior including cell proliferation and spreading, which are energetically fueled by metabolic reprogramming (42). Such aggressive effects of deregulated mTOR linked to FDG uptake could be at least partly responsible for the early tumor recurrence in patients with PET-positive HCC.

Unfortunately, a representative randomized clinical trial of an mTOR inhibitor (i.e., everolimus) failed to demonstrate a survival advantage in the second-line setting of advanced HCC (47). In addition, a recent pan-cancer basket trial including 3 liver cancer cases did not demonstrate a predictive role for response to everolimus of genetic alterations in TSC1, TSC2, or MTOR related to mTOR activation, which were detected in only 6 samples out of our 57 FDG-avid series (48). However, an exploratory study of breast cancer samples from the TAMRAD trial indicated that activation of mTORC1 and its later effectors could be a potential predictive biomarker of everolimus efficacy (49). Indeed, the recent positive results from the REACH-2 trial of ramucirumab after sorafenib in a selected HCC population with serum AFP ≥400 ng/mL have encouraged biomarker-driven approaches to clinical trials in hepatic oncology (50). Based on our translational results, nuclear imaging criteria-based umbrella designs evaluating next-generation mTOR inhibitors (e.g., omipalisib) in patients chosen on the basis of FDG-PET features of HCC would be worth further consideration: these noninvasive protocols have the additional advantage that HCC is usually diagnosed from imaging studies alone without tissue sampling (6). On the other hand, PET imaging has emerged as a useful prognostic tool for evaluating HCC before liver transplantation (LT) due to its ability to predict aggressive biological behavior, as observed in our surgically resected cases (7, 8, 17). An updated meta-analysis demonstrated that mTOR inhibitor–based regimens improved post-LT outcomes, mainly in terms of HCC recurrence, compared with standard immunosuppression protocols, but the difference was only significant during the first few years after LT (51). Most recently, an exploratory analysis of the SiLVER-trial showed that the survival advantage of sirolimus treatment after LT for HCC was most pronounced in patients with active tumors producing AFP, a tumor group which is quite closely related to our FDG-avid series (52). Given the present radiogenomic findings, incorporating mTOR inhibitors into selected recipients with high FDG-avid signatures as measured by our method, or with increased FDG uptake in actual PET scans, is more likely to highlight or boost the anti-HCC effect of immunosuppressants.

In conclusion, the present combined imaging-genetic data indicate that aberrant signaling by the mTOR cascade is the key driving force in activating glucose metabolism in HCC as manifested by FDG uptake in PET images. Use of this nucleogenomic biomarker is likely to facilitate PET-based target curation in clinical trials, as well as precision pharmaceutical care employing tailored medicines.

No disclosures were reported.

J. An: Conceptualization, data curation, software, formal analysis, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. M. Oh: Resources, data curation, software, formal analysis, supervision, visualization, methodology, writing–original draft. S.-Y. Kim: Software, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. Y.-J. Oh: Resources, data curation, software, formal analysis, investigation, visualization, methodology, writing–review and editing. B. Oh: Data curation, validation, investigation, visualization, methodology, writing–review and editing. J.-H. Oh: Validation, visualization, methodology, writing–review and editing. W. Kim: Validation, investigation, visualization, writing–review and editing. J.H. Jung: Data curation, formal analysis, investigation. H.I. Kim: Data curation, supervision, validation, writing–review and editing. J.-S. Kim: Data curation, validation, investigation, methodology, writing–review and editing. C.O. Sung: Conceptualization, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. J.H. Shim: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing.

This study was supported by grants from the Basic Science Research Program through the National Research Foundation of Korea funded by the Ministry of Science, (ICT grant no. NRF-2017R1E1A1A01074298, to J.H. Shim), and the research fund of Hanyang University (grant no. HY-201900000002619, to J. An).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.