From association to attribution: translational challenges in carcinogen-informed radiogenomics for pancreatic cancer

Dear Editor, The study by Xie et al proposes an integrative radiogenomic framework linking NNK-associated transcriptional programs, single-cell tumor ecology, and CT-derived radiomic signatures to enable preoperative risk stratification in pancreatic ductal adenocarcinoma[1]. This approach aligns with current efforts to reposition radiomics from purely prognostic modeling toward biologically interpretable surrogates of tumor behavior and environmental influence. As such frameworks increasingly aim to support mechanistic inference and clinical translation, several conceptual and methodological considerations merit further discussion to consolidate specificity, deployability, and biological grounding. This study complies with the TITAN 2025 framework for transparent and ethical reporting of AI-assisted research[2]. A central assumption is that an ITGA3-centered program represents an NNK-specific carcinogenic imprint. However, the available evidence is also consistent with a broader invasive, desmoplastic PDAC phenotype. ITGA3 resides within a tightly coupled EMT and ECM–integrin network that becomes progressively enriched as tumor purity declines and stromal remodeling intensifies, raising the possibility that the attributed carcinogen signal partially reflects fibroinflammatory architecture rather than exposure specificity. In the absence of explicit specificity testing, attribution remains indeterminate, as similar transcriptional configurations may emerge from convergent upstream pressures, including hypoxia, mechanical stress, or broader clinical context, all of which converge on matrix engagement and mesenchymal programs. This distinction is nontrivial, because the translational promise of carcinogen-informed radiogenomics depends on isolating exposure-linked biology rather than recapitulating aggressive phenotype[3]. A falsification-oriented analysis conditioning on invasion and stromal burden would clarify whether the NNK-associated program provides exposure-specific information beyond an invasion-matched baseline. The NIR score is positioned as an actionable CT-based tool, yet its translational trajectory is shaped by failure modes not captured by internal reproducibility metrics. ICC-based robustness reflects stability within a fixed acquisition and annotation ecosystem, whereas real-world deployment is dominated by cross-domain variation arising from heterogeneous scanners, reconstruction pipelines, and contrast timing, each capable of systematically reshaping texture distributions. External validation alone may be insufficient when validation cohorts remain partially aligned with the training domain, since apparent transportability can reflect shared technical conventions rather than biological generalization. Segmentation introduces a second, clinically consequential bottleneck in PDAC, where indistinct tumor margins, peritumoral inflammation, and vascular adjacency generate boundary uncertainty that propagates nonlinearly into radiomic features[4]. Operationalizing deployability through protocol-stratified stress testing and explicit acquisition standards would better align the model with multicenter implementation and regulatory review. Interpretation of the NIR score further relies on inferred immune suppression and stromal engagement, yet PDAC represents a context in which such inferences are particularly vulnerable to structural bias. Extensive desmoplasia and low tumor purity can couple transcript abundance to tissue composition, such that apparent immune depletion may reflect stromal expansion rather than biological exclusion, and macrophage enrichment may arise from shifts in compartment proportion. Without explicit conditioning on tumor purity and stromal fraction, mechanistic claims risk conflating tumor ecology with admixture effects. A parallel limitation concerns the radiomic signal itself: texture features may be mathematically stable while remaining biologically underdetermined if their histopathological substrates are not specified[5]. Given that fibrosis, necrosis, edema, and microvascular heterogeneity can produce overlapping imaging heterogeneity in PDAC, pathology constitutes a critical missing intermediate layer. Linking purity-adjusted immune modeling with quantitative pathology in prespecified subsets would enable falsifiable anchoring of radiomic features to underlying tissue architecture. Taken together, these considerations do not detract from the conceptual ambition of the proposed framework but clarify the requirements for carcinogen-informed radiogenomics to become clinically decisive. For surgical oncology, progress will depend less on additional association than on demonstrated specificity, transportability, and biological anchoring under real-world variation. Methodologically, falsification-oriented controls, domain-shift stress testing, and pathology-linked validation can strengthen causal plausibility while limiting overinterpretation. Translationally, clearly defined acquisition and segmentation standards are prerequisites for multicenter uptake and regulatory evaluation. Ultimately, the clinical value of NIR will be realized only if it informs treatment selection, such as neoadjuvant strategies, perioperative risk mitigation, or trial enrichment, thereby transforming preoperative risk scores into actionable decisions rather than prognostic descriptors.