A discovery-stage computational pathology framework has linked a proliferation-dominant malignant-cell transcriptional program with histopathological morphology in hepatocellular carcinoma, supporting internal prognostic stratification within The Cancer Genome Atlas data. Published in the Journal of Translational Medicine, the study addresses biological heterogeneity in hepatocellular carcinoma, where relationships between malignant transcriptional states and histopathology remain poorly understood. Researchers combined single-cell profiling, spatial transcriptomics, bulk transcriptomics, genome-wide association study summary statistics, and whole-slide images to connect microscopic tissue features with cellular gene expression.
To evaluate cellular populations, the investigators analyzed 146,515 cells following quality control. Within this dataset, researchers evaluated 20,153 hepatocyte-lineage cells. The team performed clustering based on expression-deviation scores derived from inferCNV, utilizing copy-number variation patterns to identify abnormal populations. This computational screening identified 9,050 malignant candidate cells, of which 9,026 cells entered non-negative matrix factorization analysis.
Using non-negative matrix factorization, the authors grouped tumor-level expression programs into five malignant-cell metaprograms. Metaprogram 3 emerged as a proliferation-dominant program enriched for cell-cycle and mitotic genes. Molecular comparison revealed that metaprogram 3 exhibited a 68 percent exact top-fifty gene overlap with the union of Hallmark G2M checkpoint and E2F target gene signatures. Across TCGA-LIHC, ICGC-LIRI-JP, GSE14520, and GSE116174 cohorts, metaprogram 3 correlated strongly with these signatures, displaying Spearman correlation coefficients between 0.908 and 0.963.
The prognostic association of metaprogram 3 was attenuated after adjusting for proliferation-related signatures. To connect these transcriptional patterns with tissue morphology, the researchers developed an image-based metric designated the MP3-guided pathology model score. This framework extracted cellular measurements with CellProfiler and deep-learning features using a ResNet-50 network applied to whole-slide images from the TCGA-LIHC cohort.
Evaluation within a fixed patient-level internal held-out test cohort from TCGA-LIHC assessed the prognostic utility of the pathology model score. The model achieved a concordance index of 0.706 in the training cohort of 228 patients and 0.801 in the internal held-out test cohort of 99 patients. In test samples, area under the receiver operating characteristic curve values reached 0.895 at one year, 0.813 at three years, and 0.845 at five years.
Analysis of model calibration showed consistent population-level alignment alongside individual-level variation. At the cohort level, mean predicted three-year survival risk was 0.388 compared to a Kaplan–Meier-observed risk of 0.393. However, a calibration slope of 1.961, with a 95 percent confidence interval spanning 1.330 to 2.591, demonstrated individual-level miscalibration. In a separate pooled complete-case multivariable analysis of 303 patients, the score remained associated with survival after adjustment for available clinical variables.
The authors conducted multiple exploratory internal analyses, including calibration assessments, limited clinical baseline comparisons, decision-curve analyses, all-feature Ridge-Cox modeling, and permutation testing. Immunotherapy response patterns were computationally assessed using the TIDE and SubMap algorithms. Candidate therapeutic compounds were computationally prioritized across four pharmacogenomic collections: the Connectivity Map, the Cancer Therapeutics Response Portal, PRISM, and the Genomics of Drug Sensitivity in Cancer database.
Despite these computational findings, the authors stressed that this discovery-stage framework has unestablished generalizability and clinical applicability. While the system demonstrated internal prognostic stratification within TCGA-LIHC, the findings cannot yet be applied to clinical practice. The investigators concluded that independent multicenter external validation and prospective clinical evaluations are required before any clinical translation can occur.
Subject of Research: Pathology-omics framework linking single-cell transcriptional programs with tumor histology in hepatocellular carcinoma
Article Title: A discovery-stage NMF-based pathology–omics framework linking single-cell transcriptional programs with histopathological imaging in hepatocellular carcinoma
Article References: Liu, L., Sun, X., Huang, X., Xu, S., & Guo, Y. (2026). A discovery-stage NMF-based pathology–omics framework linking single-cell transcriptional programs with histopathological imaging in hepatocellular carcinoma. Journal of Translational Medicine. https://doi.org/10.1186/s12967-026-09239-y
Image Credits: AI Generated
DOI: 10.1186/s12967-026-09239-y
Keywords: hepatocellular carcinoma, single-cell RNA sequencing, spatial transcriptomics, non-negative matrix factorization, metaprogram 3, TCGA-LIHC, CellProfiler, ResNet-50, inferCNV, copy-number variation, Hallmark G2M checkpoint, E2F targets, computational pathology, prognostic stratification, whole-slide images, TIDE, SubMap
News Source: Nathaniel Bowman. (November 16, 2026). Discovery Framework Links Malignant Liver Cancer Programs With Tumor Tissue Morphology. Scienmag.



