Dear Editor,
We read with great interest the letter “Critical flaws in the molecular classification of hepatocellular carcinoma (HCC) based on metabolic zonation” submitted by Xie et al. [
1] We appreciate their thoughtful comments and agree with several points raised, but would also like to clarify and supplement some important aspects that may have been overlooked.
First, regarding the clustering methodology: Our classification was not predetermined based on biological characteristics [
2]. Rather, we used principal component analysis (PCA) and silhouette coefficients to systematically and objectively identify combinations of gene sets that yielded the highest degree of separation in our dataset. In other words, the combination of metabolic functions and signaling pathways presented in our study was selected mechanistically according to the silhouette coefficient, as the most statistically robust approach for our analysis. However, we recognize that dividing the glycolysis subclass into phosphoinositide 3-kinase (PI3K)/mammalian target of rapamycin (mTOR) High and NOTCH/transforming growth factor (TGF)-β High does not always accurately reflect previously reported molecular subtypes (such as Hoshida S1/S2) [
3,
4] or the tumor immune microenvironment (TIME) and prognosis. Thus, we acknowledge the need to further explore improved classification methods that do not rely solely on the silhouette coefficient.
To further address concerns about the methodological validity of our clustering approach, we performed a meta-analysis of silhouette coefficients across all six HCC cohorts, consisting of one training and five validation cohorts (
Fig. 1). The training cohort (n=136) was optimized to maximize the silhouette coefficient, and all validation cohorts also demonstrated consistently high silhouette coefficients (range 0.46–0.56). The pooled mean silhouette coefficient was 0.53, indicating stable cluster separation across diverse datasets. Importantly, the difference between the training and validation cohorts was modest, suggesting minimal overfitting and supporting the reproducibility and robustness of our clustering approach in external datasets. Nevertheless, we recognize that statistical reproducibility does not necessarily address the biological complexities and limitations highlighted by Xie et al. [
1] We agree that further efforts are needed to refine subclassification methods by integrating both biological insight and computational rigor.
Next, regarding the biological overlap and spectrum: As discussed by Halpern et al. [
5], the concept of zonation itself represents an inherently continuous spectrum. In recent years, it has also been reported that primary liver cancers can display intermediate features between HCC and intrahepatic cholangiocarcinoma (iCCA), and thus these entities are now increasingly considered to exist along a biological continuum [
6,
7]. Just as complete separation between zone 1 and zone 2, or between iCCA and HCC, is inherently unachievable due to their biological continuum, it is also unavoidable that some degree of overlap will remain between groups or individual cases, no matter how distinct the clustering appears. This reflects the fundamental limitation of attempting to categorize inherently continuous biological spectra into discrete groups.
We are well aware that there is biological crosstalk and overlap among pathways. In fact, the possibility of overlap between the interleukin (IL)6–JAK–signal transducer and activator of transcription (STAT)3 and Kirsten rat sarcoma viral oncogene homolog (K-RAS) groups was also described in our manuscript [
2], and these groups were not treated as completely independent, but rather as part of a potential continuous spectrum.
Third, regarding the Wnt/β-catenin subtype and prognosis: The paper by Montironi et al. [
8] is cited as evidence that Wnt/β-catenin-mutated HCCs show early recurrence and thus have a poor prognosis. However, it is important to clarify that Montironi et al. did not directly report prognosis or survival outcomes for these immune phenotypes. Instead, their study demonstrated that while approximately 70–85% of Wnt/β-catenin-activated HCCs exhibit an immune-excluded phenotype with low immune cell infiltration, about 15–30% display inflamed features, such as increased CD8+ T cell infiltration and upregulation of Chemokine (C-C motif) ligand (CCL)5 [
8]. Thus, their findings highlight the dual nature of Wnt/β-catenin-activated HCC, but do not provide data on differences in prognosis or survival between these subgroups.
It has long been recognized that Wnt/β-catenin-mutated HCCs exhibit a dual nature [
9]. In our previous studies as well, we emphasized this duality and warned against the fixed notion that all Wnt/β-catenin-activated HCCs are immune-desert [
10,
11]. Nevertheless, we are concerned that, in recent years, the interpretation of these findings has often been overstated in the literature, leading to the misconception that Wnt/β-catenin-activated HCCs are uniformly responsive to immune checkpoint inhibitors (ICIs).
Nevertheless, we fundamentally agree with the issues raised by the authors. These are indeed limitations of analyses using bulk ribonucleic acid (RNA) sequencing data, and we concur that systems biology approaches, including the integration of spatial transcriptomics and proteomics, will be essential in future studies. We are currently pursuing research in these directions ourselves. Once again, we thank the authors for their constructive comments and hope that this exchange will contribute to further progress in the field.
FOOTNOTES
-
Authors’ contribution
Conceptualization: T.A. and N.N.; Software: T.A.; Writing – original draft preparation: T.A.; Writing – review & editing: N.N. and M.K.; Supervision: N.N. and M.K.
-
Acknowledgements
This work was supported in part by a Grant-in-Aid for Scientific Research from the Japan Society for the Promotion of Science (25K11194, T. Aoki) and a grant from SRF (T. Aoki).
-
Conflicts of Interest
T.A.: No relevant conflicts of interest to disclose. N.N.: No relevant conflicts of interest to disclose. N.N. is an Editorial Board member of CMH. M.K.: M.K. has received grants from Taiho Pharmaceuticals, Chugai Pharmaceuticals, Otsuka, Takeda, Sumitomo Dainippon-Sumitomo, Daiichi Sankyo, AbbVie, Astellas Pharma, and Bristol-Myers Squibb. He has also received grants and personal lecture fees from Merck Sharpe and Dohme (MSD), Eisai, and Bayer, and is an adviser for MSD, Eisai, Bayer, Bristol-Myers Squibb, Eli Lilly, Chugai, AstraZeneca and ONO Pharmaceuticals.
Figure 1.Meta-analysis of silhouette coefficients for clustering reproducibility across six HCC cohorts. Forest plot showing the mean silhouette coefficient and 95% confidence interval for each cohort. The training cohort (n=136) was optimized to maximize the silhouette coefficient and showed the highest separation (mean=0.61). The remaining five cohorts served as independent validation cohorts, all demonstrating reproducibly high silhouette coefficients (range 0.46–0.56). The overall pooled mean silhouette coefficient was 0.53 (95% CI 0.52–0.55), supporting the robustness and reproducibility of the clustering approach across diverse datasets. CI, confidence interval; HCC, hepatocellular carcinoma.
Abbreviations
Intrahepatic cholangiocarcinoma
REFERENCES
- 1. Xie Y, Zhu X, Liang Q. Critical flaws in the molecular classification of HCC based on metabolic zonation: Letter to the editor on “Molecular classification of hepatocellular carcinoma based on zoned metabolic feature and oncogenic signaling pathway”. Clin Mol Hepatol 2026;32:e144-e145.
- 2. Aoki T, Nishida N, Kurebayashi Y, Sakai K, Fujiwara N, Tsurusaki M, et al. Molecular classification of hepatocellular carcinoma based on zoned metabolic feature and oncogenic signaling pathway. Clin Mol Hepatol 2025;31:981-1002.
- 3. Hoshida Y, Nijman SM, Kobayashi M, Chan JA, Brunet JP, Chiang DY, et al. Integrative transcriptome analysis reveals common molecular subclasses of human hepatocellular carcinoma. Cancer Res 2009;69:7385-7392.
- 4. Hoshida Y. Nearest template prediction: a single-sample-based flexible class prediction with confidence assessment. PLoS One 2010;5:e15543.
- 5. Halpern KB, Shenhav R, Matcovitch-Natan O, Toth B, Lemze D, Golan M, et al. Single-cell spatial reconstruction reveals global division of labour in the mammalian liver. Nature 2017;542:352-356.
- 6. Jeon Y, Kwon SM, Rhee H, Yoo JE, Chung T, Woo HG, et al. Molecular and radiopathologic spectrum between HCC and intrahepatic cholangiocarcinoma. Hepatology 2023;77:92-108.
- 7. Komuta M, Ueno A, Sakamoto M. The spectrum of primary liver cancers: heterogeneity and continuity. A foundation for diagnosis and treatment of cancer. Hepatology 2023;77:10-12.
- 8. Montironi C, Castet F, Haber PK, Pinyol R, Torres-Martin M, Torrens L, et al. Inflamed and non-inflamed classes of HCC: a revised immunogenomic classification. Gut 2023;72:129-140.
- 9. Inagawa S, Itabashi M, Adachi S, Kawamoto T, Hori M, Shimazaki J, et al. Expression and prognostic roles of beta-catenin in hepatocellular carcinoma: correlation with tumor progression and postoperative survival. Clin Cancer Res 2002;8:450-456.
- 10. Aoki T, Nishida N, Kurebayashi Y, Sakai K, Morita M, Chishina H, et al. Two distinct characteristics of immune microenvironment in human hepatocellular carcinoma with Wnt/β-catenin mutations. Liver Cancer 2023;13:285-305.
- 11. Aoki T, Nishida N, Kudo M. Clinical significance of the duality of Wnt/β-catenin signaling in human hepatocellular carcinoma. Cancers (Basel) 2022;14:444.
Citations
Citations to this article as recorded by
