Machine learning developed an intratumor heterogeneity-driven macrophage prognostic signature in hepatocellular carcinoma via integrated bulk and single-cell transcriptomics
Year: 2026
Type: article
Abstract: Introduction Intratumor heterogeneity (ITH) contributes to treatment resistance and poor outcomes in hepatocellular carcinoma (HCC), but its cellular basis and prognostic relevance remain incompletely defined. This study aimed to develop an ITH-centered prognostic framework by integrating bulk and single-cell transcriptomics. Methods We integrated bulk transcriptomes from 363 TCGA-LIHC tumors with single-cell RNA sequencing data from GSE149614. DEPTH2 was used to quantify ITH, and Scissor analysis mapped the high-ITH phenotype onto single-cell populations. Candidate genes were identified by intersecting ITH-associated, disease-associated, and key cell population-associated differentially expressed genes, followed by univariate Cox regression and a 101-combination machine-learning pipeline. A five-gene random survival forest (RSF) signature was constructed and evaluated in the independent GSE14520 cohort. Results High ITH was associated with shorter overall survival, and Scissor localized the high-ITH signal predominantly to macrophages. The final five-gene RSF signature (FABP5, PKM, HPX, TAT, and HRG) achieved an apparent C-index of 0.771 (95% CI: 0.733–0.810) in the training cohort and was validated in the independent GSE14520 cohort (C-index = 0.636, 95% CI: 0.574–0.698), with 1-, 3-, and 5-year AUCs of 0.670, 0.671, and 0.681 in external validation. The risk score remained independently associated with survival after clinical adjustment, and a risk score-stage nomogram achieved a C-index of 0.778 (95% CI: 0.737–0.818). High-risk tumors showed greater copy-number burden, distinct pathway activity, increased immune checkpoint expression, stronger predicted immune evasion, and differential predicted drug sensitivity. Human Protein Atlas data supported the expression and prognostic patterns of the five genes, and qRT-PCR in five paired clinical specimens confirmed their expected tumor expression changes. Molecular docking nominated tranilast as a candidate PKM-binding compound, although a 100-ns molecular dynamics simulation indicated an unstable binding pose. Discussion These findings support an ITH-associated, macrophage-linked five-gene framework for HCC risk stratification and provide cautious, testable hypotheses for biomarker-guided therapeutic investigation. Larger multicenter prospective cohorts and mechanistic experiments are needed.
Authors: Xinhao Wang(Investigation, Methodology, Writing – original draft, Writing – review & editing), Chengjian He(Formal analysis, Writing – original draft), Yang Liu(Investigation, Methodology, Writing – original draft), Naijian Ge(Conceptualization, Funding acquisition, Supervision, Writing – review & editing)
Language: English
Subfield: Hepatology
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $3,801
Award: G4471957152
OpenAlex ID: W7233808639
