Metabolic phenotyping combined with transcriptomics metadata fortifies the diagnosis of early-stage Hepatocellular carcinoma

Abstract

Introduction: The low sensitivity of alpha-fetoprotein (AFP) renders it unsuitable as a stand-alone marker for early hepatocellular carcinoma (eHCC) surveillance. Therefore, additional blood-based biomarkers with enhanced sensitivities are required.

Objectives: In light of the metabolic changes that are distinctive to eHCC development, the current study presents a panel of serum metabolites that may serve as noninvasive diagnostic indicators for patients with eHCC.

Methods: Serum samples obtained from normal control (NC), cirrhosis, and eHCC patients were analyzed by four different metabolomic platforms. A meta-analysis of very early-stage HCC transcriptomic datasets retrieved from public sources supports the integrated interpretation with metabolic changes.

Results: A total of 94 metabolites were significantly correlated with a progressive disease status. Integrated analysis of the significant metabolites and differentially expressed genes from meta-analysis emphasized metabolic pathways including bile acid biosynthesis, phenylalanine and tyrosine metabolism, and butanoate metabolism. The 11 metabolites associated with these pathways were compiled into a metabolite panel for use as diagnostic signatures. With an accuracy of 81.8%, compared with 45.4% for a model trained solely on AFP, the model enhanced its ability to differentiate between the three groups by incorporating a metabolite panel and AFP. Upon examining the trained models using receiver operating characteristic curves, the AFP and metabolite panel combined model exhibited greater area under the curve values in comparisons between NC and eHCC (1.000 versus 0.810) and cirrhosis and eHCC (0.926 versus 0.556). The result was consistent in an independent validation cohort.

Conclusion: This study emphasizes the role of circulating metabolite markers in the diagnosis of eHCC.

Keywords: Carcinoma, Hepatocellular; Early diagnosis; Gene expression profiling; Machine learning; Metabolomics.

Graphical abstract

Fig. 1

Meta-analyses of in silico datasets identified transcriptional differences between very early-stage HCC and controls. A. Volcano plot of meta-analyzed data of seven Affymetrix microarray GEO datasets. B. Volcano plot of meta-analyzed data of three Illumina microarray GEO datasets. C. The number of differentially expressed genes from the meta-analyzed data. Red indicates increased expression in the cancer group, whereas blue indicates decreased expression. D. Network analysis by differentially expressed genes from the meta-analyzed data of seven Affymetrix microarray GEO datasets. E. Network analysis by differentially expressed genes from the meta-analyzed data of three Illumina microarray GEO datasets. For D and E, pink-colored circles indicate pathways related to the metabolism of endogenous metabolites.

Fig. 2

The application of metabolomics facilitated the identification of several metabolic pathways that may have been altered, as evidenced by the supportive findings obtained from a meta-analysis. A. Detection information of the metabolites identified by multiple platforms. C18 and aHILIC stand for C18 RPLC and amide HILIC, respectively. Positive and negative ionization electric fields are expressed as ‘pos’ and ‘neg’, respectively. B-D depicts grouped metabolites with significant increasing or decreasing trends along with the liver status. *, **, and *** indicate p < 0.05, p < 0.01, and p < 0.001, respectively. Quartiles of the data are represented by box plots, and the whiskers indicate the maximum and minimum values. B. Significant metabolites in bile acid biosynthesis. C. Significant metabolites in phenylalanine and tyrosine metabolism. D. Significant metabolites in butanoate metabolism. E–G represent pathways involving multiple metabolites with increasing or decreasing trends. Regarding the two horizontally attached squares underneath the gene symbols, the left square indicates the log2FC of the meta-analyzed data from the Affymetrix microarray, whereas the right square indicates the log2FC of the meta-analyzed data from the Illumina microarray. The symbol † denotes the consistency of the HCC proteomics dataset obtained from the Proteomics Data Commons. E. Bile acid biosynthesis. F. Phenylalanine and tyrosine metabolism. G. Butanoate metabolism.

Fig. 3

A metabolite panel significantly improved the discriminatory capabilities of machine learning algorithms. A. Confusion matrix of Random Forest-based discrimination using AFP. B. Confusion matrix of Random Forest-based discrimination using a combination of AFP and the metabolite panel. C. Variable importance scores of the model constructed by AFP and the metabolite panel. D. Confusion matrix of Random Forest-based discrimination using a combination of AFP and the metabolite panel on AFP-negative eHCC (<20 ng/mL). E. The accuracy of each model from A, B, and D. F. Receiver operating characteristic curve discriminating eHCC versus NC. G. Receiver operating characteristic curve discriminating eHCC versus cirrhosis.

Fig. 4

The trained model validated its performance on an independent cohort. A. Receiver operating characteristic curve discriminating eHCC versus cirrhosis in validation set. B. Receiver operating characteristic curve discriminating AFP-negative eHCC versus cirrhosis in validation set.

Supplementary Fig. 1

Supplementary Fig. 2

Supplementary Fig. 3

Supplementary Fig. 4

Supplementary Fig. 5

Supplementary Fig. 6

Supplementary Fig. 7

Supplementary Fig. 8

Supplementary Fig. 9

Supplementary Fig. 10