An RNA-based signature enables high specificity detection of circulating tumor cells in hepatocellular carcinoma

Abstract

Circulating tumor cells (CTCs) are shed into the bloodstream by invasive cancers, but the difficulty inherent in identifying these rare cells by microscopy has precluded their routine use in monitoring or screening for cancer. We recently described a high-throughput microfluidic CTC-iChip, which efficiently depletes hematopoietic cells from blood specimens and enriches for CTCs with well-preserved RNA. Application of RNA-based digital PCR to detect CTC-derived signatures may thus enable highly accurate tissue lineage-based cancer detection in blood specimens. As proof of principle, we examined hepatocellular carcinoma (HCC), a cancer that is derived from liver cells bearing a unique gene expression profile. After identifying a digital signature of 10 liver-specific transcripts, we used a cross-validated logistic regression model to identify the presence of HCC-derived CTCs in nine of 16 (56%) untreated patients with HCC versus one of 31 (3%) patients with nonmalignant liver disease at risk for developing HCC (P < 0.0001). Positive CTC scores declined in treated patients: Nine of 32 (28%) patients receiving therapy and only one of 15 (7%) patients who had undergone curative-intent ablation, surgery, or liver transplantation were positive. RNA-based digital CTC scoring was not correlated with the standard HCC serum protein marker alpha fetoprotein (P = 0.57). Modeling the sequential use of these two orthogonal markers for liver cancer screening in patients with high-risk cirrhosis generates positive and negative predictive values of 80% and 86%, respectively. Thus, digital RNA quantitation constitutes a sensitive and specific CTC readout, enabling high-throughput clinical applications, such as noninvasive screening for HCC in populations where viral hepatitis and cirrhosis are prevalent.

Keywords: blood biopsy; circulating tumor cells; early cancer detection; hepatocellular carcinoma; predictive modeling.

Fig. 1.

dPCR quantitation of HCC cells after microfluidic enrichment from blood. (A) Schematic representation of the integrated platform for digital RNA-PCR scoring of CTCs. Hematopoietic components are depleted from whole blood through CTC-iChip processing as previously described (5). RNA from the CTC-enriched product is subjected to WTA, encapsulation of cDNA molecules within lipid droplets, and PCR amplification for transcripts of interest. (B) Heat maps derived from microarray (Left) and RNA-sequencing (RNAseq) (Right) datasets comparing expression of 10 liver-specific transcripts in HCC versus other tissues. The microarray dataset compares fetal liver, adult liver, and 10 cases of HCC (JZR samples) versus normal tissues (15 samples shown of 79 tissues tested) and blood components (15, 16). RNAseq compares 10 cases of HCC (17), with WBCs collected from eight independent healthy donor (HD) blood samples processed through the CTC-iChip. (C) dPCR quantitation of ALB transcripts from micromanipulated HepG2 spiked into whole blood and processed through the CTC-iChip. Each data point represents one-sixth of the CTC-iChip product. (D) Pie charts representing the distribution of transcripts for each of the 10 selected liver-specific genes, following dPCR analysis of 1 ng of HepG2-cell RNA. Samples were nonamplified or subjected to WTA (three independent reactions). (E) Total number of transcripts of interest after spiking increasing numbers of HepG2 cells into blood (n = 3), CTC-iChip processing, and dPCR. (F) Pie charts depicting the relative fraction of droplets for each of the 10 target transcripts after spiking increasing numbers of HepG2 cells into blood and CTC-iChip processing, as noted in E (n = 3).

Fig. S1.

WBC gene expression. Relative expression (qRT-PCR) of candidate liver-specific signature genes, amplified from 5 ng of cDNA from healthy blood donor WBCs (buffy coat), normalized to GAPDH (n = 3). The gene targets APOC1, HP, HPR, and SERPINA1 were eliminated from the liver signature due to their high relative expression.

Fig. S2.

WTA characterization. (A) Total number of droplets derived from 1 ng of HepG2 cell RNA for all 10 liver-specific genes. Three independent WTA reactions are compared with a nonamplified cDNA control. (B) Consistent amplification ratio for each liver-specific gene following WTA (three independent reactions), relative to the nonamplified cDNA control.

Fig. S3.

Buffy coat dPCR. Failure of ALB transcript droplet PCR amplification from buffy coat (WBCs and nonenriched CTCs) RNA extracted from blood specimens of nine HCC patients and three healthy donor (HD) controls. The ALB transcript is appropriately amplified from HepG2 RNA. Matched GAPDH transcript quantification is shown as a control for relative RNA content.

Fig. 2.

CTC score from patients with HCC compared with at-risk patients. (A) Heat maps depicting relative signal intensities for each of the 10 liver-specific transcripts across different patient cohorts. Primary droplet numbers are log-10–transformed and scaled to the highest value for each transcript. (Upper) Healthy donors (blood bank, n = 26) and high-risk patients with CLD under active clinical surveillance for HCC (n = 31). Etiologies of CLD include HBV infection (n = 16), hepatitis C virus (HCV) infection (n = 6), alcohol (EtOH) (n = 6), or other causes (n = 3). (Middle) Patients with HCC, classified as untreated (newly diagnosed, n = 16) or receiving ongoing treatment (currently undergoing various therapies, n = 32). Patients are grouped according to Barcelona Clinic Liver Cancer (BCLC) criteria from early clinical stages (0 and A) to advanced clinical stages (B–D). Patients who have completed treatment and have NED are shown (n = 15). Four of these cases represent repeated analysis of patients initially tested before or during treatment (HCC-030_2, HCC-058_2, HCC-060_2, and HCC-064_2). (Lower) Patients with cancers other than HCC (n = 43): intrahepatic cholangiocarcinoma (ICC); pancreatic ductal adenocarcinoma (PDAC); breast, lung, and prostate cancers; melanoma; and cancers of nonhepatic origin metastatic to the liver (MET). Clinical data are listed in Table S6. (B) Box plots representing the integrated CTC score for the patient cohorts above. **P < 0.01, ***P < 0.0001 (χ², degrees of freedom = 5). (C) Receiver operator characteristic (ROC) curves for untreated HCC both without (Left) and with (Right) LOOCV. AUC, area under the curve; FPR, false-positive rate; TPR, true-positive rate.

Fig. S4.

ROC curves for individual genes. ROC curves were derived for each transcript within the liver-specific signature, using univariate logistic regression and all first-draw active HCC patient samples. AUC, area under the curve; FPR, false-positive rate; TPR, true-positive rate.

Fig. S5.

Multigene model parameters and modeling equations. Coefficients and model statistics for are shown for the logistic regression model (table) using all first-draw active HCC patient samples. The formulae used for calculation of the PPVs and NPVs are shown below. Akaike Inf. Crit., Akaike information criteria.

Fig. S6.

CTC score clinical correlates. (A) Nonsignificant correlation between CTC score (all patients with HCC) and the etiology of underlying liver cirrhosis [alcohol-induced (EtOH), HBV infection, hepatitis C virus (HCV) infection and EtOH, and nonalcoholic steatohepatitis (NASH)]. All patients with HCV-induced HCC in our cohort also had significant alcohol exposure. (B) Significant correlation between CTC score and clinical stage (Barcelona criteria: early stage 0 and A versus advanced stage B–D). (C) Trend approaching significance between CTC score and imaging-based (macroscopic) evidence of vascular invasion.

Fig. S7.

ROC curves, multigene model parameters, and model performance for HCC versus other malignancies. (A) Individual gene ROC curves with AUC and P values displayed. (B) Coefficients and model statistics for the logistic regression model using all first-draw active HCC patient samples. (C) Non–cross-validated and LOOCV logistic regression ROC curves for untreated HCC patient draws. (D) Comparison between scores of HCC patients and those patients with other cancers. P = 0.013, Mann–Whitney U test. (E) Transcript count variations across two blood draws on patients HCC.041 and HCC.075 in the absence of clinical intervention.

Fig. 3.

Longitudinal monitoring of patients treated for HCC. (A) Serial blood measurements performed at 1-wk intervals in two patients (HCC-041 and HCC-075) in the absence of therapeutic intervention. Concurrent CTC score (red) and serum AFP (black) measurements are shown. (B) Longitudinal monitoring of two patients (HCC-058 and HCC-060), before (Pre) and after (Post) resection of localized HCC. HCC-060 had NED 1 mo after resection but then developed a recurrence of HCC (Rec). (C) Serial monitoring of a patient (HCC-042) initially treated with the immune checkpoint inhibitor nivolumab (Nivo), followed by radioembolization (Embo) of the residual mass. The tumor mass and postembolization changes are shown by computed tomography scan. Concurrent CTC score (red) and serum AFP (black) measurements are shown.

Fig. 4.

Modeling early detection of HCC using CTC score and AFP measurements. (A) Absence of correlation between CTC score and serum AFP levels in all patients with HCC with concomitant measurements. (B) Proportion of 15 newly diagnosed, untreated patients with HCC identified by CTC score alone, AFP (>100 ng/mL) alone, or both CTC score and AFP. No AFP measurement was available for one patient with untreated HCC. (C) Bar graphs representing all newly diagnosed HCC patients, showing those patients identified by serum AFP (>100 ng/mL), CTC score, or the combination of the two tests (Either). Six of these newly diagnosed patients met the Milan criteria for localized disease amenable to curative liver transplantation [Milan (+)]. Two of six Milan(+) patients were identified by CTC score, but none of five had an AFP level >100 ng/mL [one Milan (+) patient did not have AFP measurement]. (D) PPV and NPV calculations for CTC score alone, AFP (>20 ng/mL) alone, or both, as a function of HCC prevalence. The CTC score model assumes 56% test sensitivity and 95% specificity, as observed in untreated patients with HCC; for AFP (>20 ng/mL), the 53% test sensitivity and 87% specificity are established from a population-based study (19).