We read with great interest the recently published study proposing an immune-related radiotranscriptomic signature (IRS) to predict prognosis and response to immunochemotherapy in intrahepatic cholangiocarcinoma (ICC). The integration of bulk, single-cell, and spatial transcriptomics with radiomic features represents a commendable step toward resolving the spatial complexity of the tumor immune microenvironment. Of particular note is the identification of a 3-gene IRS—including PLAUR, CD40LG and FGFR4—that stratified patients by survival and predicted benefit from anti-PD-1/PD-L1-based therapies [1]. However, certain findings in the results section merit deeper investigation to further solidify the clinical implications of the proposed model.

First, the mutual exclusivity observed between PLAUR and FGFR4 expression in malignant cells is intriguing but not fully understood. The authors suggest a functional divergence, with PLAUR associated with macrophage recruitment and immunosuppression, whereas FGFR4 is enriched in epithelial tumor regions and correlates with anti-tumor immunity. However, this observation raises key mechanistic questions: are these mutually exclusive subpopulations genetically distinct ICC clones, or do they reflect dynamic phenotypic plasticity under immune selection pressure? Recent work in hepatobiliary tumors has shown that metabolic heterogeneity can drive mutually exclusive immune profiles and therapeutic responses [2]. To clarify this, lineage tracing or trajectory analysis using single-cell data (e.g., RNA velocity or pseudotime ordering) could reveal potential differentiation hierarchies or transitions between FGFR4high and PLAURhigh tumor cell states.

Second, the study highlights that the IRS score correlates positively with M0 macrophages and neutrophils and negatively with CD8⁺ T cells and naive B cells, suggesting a shift toward an immunosuppressive tumor immune microenvironment. While consistent with a role for PLAUR in myeloid recruitment, the authors do not analyze how this cellular composition evolves across spatial compartments. Interestingly, digital spatial profiling (DSP) analysis revealed that IRS scores were relatively stable across the tumor core, intermediate zone, and invasive margin - but differed significantly between epithelial and stromal compartments. This suggests that immune suppression may be primarily epithelial-centric, raising the possibility that tumor intrinsic factors-rather than stromal elements-dominate immune exclusion. Given that gradients of immune cell infiltration (particularly CD8⁺ T cells) along invasive margins have been predictive of ICI efficacy in other tumors, further quantification of spatial immune gradients within the DSP data may enhance the biological resolution of the IRS.

Third, the authors report that the 3-feature radiomic signature correlated most strongly with PLAUR expression and macrophage infiltration in the tumor epithelium. These features were derived entirely from arterial phase CT using wavelet transforms, suggesting that tumor perfusion heterogeneity may encode immune context. While plausible, it is important to consider that PLAUR expression is not only immune-related, but also associated with extracellular matrix remodeling and protease activity, which may in turn influence local vascular permeability and enhancement. Thus, the strong correlation between arterial phase texture and PLAUR may reflect a mixture of immune and stromal remodeling signals. To disentangle this, voxel-level imageomics correlation with immunohistochemical markers of vasculature, stroma, and myeloid cell density would provide further interpretability.

Fourth, in the immunochemotherapy cohort, the radiotranscriptomic signature outperformed PD-L1 expression in predicting objective response and disease control. The reported AUCs (0.84 and 0.81) are very encouraging. However, this result must be interpreted with caution due to the small sample size (n=36) and the lack of prospective biomarker-stratified treatment data. More importantly, the study did not report how IRS-high and IRS-low patients differed in terms of conventional histological or molecular characteristics, such as tumor grade, lymphovascular invasion, or IDH1/FGFR2 mutation status. Previous studies have suggested that immunotherapy response in ICC may cluster around distinct molecular subtypes, and it remains uncertain whether radiotranscriptomic risk groups are independent of these known classifiers [3]. Multivariable analysis adjusting for these clinicopathologic and genomic features may strengthen claims of additive predictive value.

Finally, the preclinical validation of uPAR as a therapeutic target—especially its combinatorial effect with anti-PD-1 therapy—is a notable strength. The authors show that uPAR blockade reduces tumor proliferation and macrophage infiltration and increases CD8⁺ T cell recruitment in PDX models. However, whether these effects are dependent on specific macrophage phenotypes (e.g., M2 polarization) remains unverified. Given the known role of uPAR in monocyte migration via CCL2/CCR2 and its interaction with integrins, future studies could investigate whether targeting uPAR reprograms macrophages towards an M1-like, immunostimulatory phenotype. In addition, uPAR expression in non-tumor tissues may raise safety concerns. Spatial proteomic profiling in treatment-naïve versus treated tumors could elucidate potential off-target immune or stromal activation, which is essential for clinical translation.

In summary, while the study presents a promising IRS model with encouraging preclinical and translational findings, several mechanistic and methodological aspects warrant further investigation before widespread clinical implementation. In particular, this work exemplifies the growing potential of radiogenomics to non-invasively capture the spatial and molecular heterogeneity of tumors, bridging the gap between imaging phenotypes and immunobiology. As spatial transcriptomics and AI-driven image analysis mature, radiotranscriptomic models such as the IRS may serve not only as predictive biomarkers, but also as dynamic tools for longitudinal immune monitoring and treatment response assessment. Future efforts to integrate multi-omic data across large, prospectively annotated cohorts will be critical to establish the clinical utility and generalizability of such approaches in immuno-oncology.

Citations to this article as recorded by