[¹⁸F]FDG PET/CT multiomics identifies Hedgehog-driven HPV-negative head and neck squamous cell carcinoma

Abstract

Head and neck squamous cell carcinoma (HNSCC) is a highly heterogeneous malignancy with limited predictive markers to guide personalized treatment, particularly in human papillomavirus (HPV)-negative cases, which exhibit poor outcomes. Identifying reliable biomarkers for prognosis and therapeutic response remains a critical challenge. In a retrospective cohort of 51 patients with primary HPV-negative HNSCC, we investigated the prognostic significance of the Hedgehog (HH) signaling pathway and its association with imaging biomarkers. Genomic and transcriptomic analysis revealed that HH pathway activation correlated with distinct [¹⁸F]FDG PET/CT radiomic features, notably the PET-derived “histogram:ih.max”, a surrogate for peak [¹⁸F]FDG uptake that was associated with inferior survival outcomes. Functionally, pharmacologic inhibition of HH signaling demonstrated anticancer efficacy across multiple models, including HNSCC cell lines, patient-derived tumoroids, and in vivo xenograft models. Importantly, HH pathway inhibition induced reproducible changes in imaging characteristics in xenografts, including a measurable reduction in [¹⁸F]FDG uptake, closely mirroring patterns observed in patient tumors. Together these findings demonstrate that integration of multi-level molecular profiling with functional imaging captures HH-driven tumor biology in HPV-negative HNSCC. Our study underscores the value of [¹⁸F]FDG PET/CT multiomics in linking tumor biology with imaging features, providing a framework for biologically-informed patient stratification and hypothesis-driven evaluation of treatment response. These results support further translational validation of HH pathway inhibition in HPV-negative HNSCC within appropriately designed preclinical and clinical studies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12943-026-02607-8.

Background

Head and neck squamous cell carcinoma (HNSCC) represents a heterogeneous group of malignancies originating from the mucosal linings of the oral cavity, pharynx, larynx, and nasal cavities, contributing significantly to morbidity and mortality worldwide [1]. While treatment de-escalation strategies have been explored for HPV-positive HNSCC, HPV-negative HNSCC remains a clinically aggressive disease with a five-year overall survival rate below 50%, limited therapeutic options, and high rates of recurrence [2].

While genomic profiling, including panel sequencing and whole-exome sequencing (WES), has expanded our biological understanding of HNSCC, its clinical translation into personalized treatment decisions remains challenging. In particular, genomics-guided recommendations by molecular tumor boards (MTBs) are still largely infrequent in this entity [3, 4]. This limited clinical actionability may reflect, in part, the substantial genomic and transcriptomic heterogeneity that characterizes HPV-negative HNSCC. Nevertheless, WES remains a powerful research tool for elucidating recurrent somatic mutations and pathway-level alterations that provide insights into tumor behavior and can inform future therapeutic strategies [5, 6].

To bridge the gap between genomic data and clinical interpretation, several computational frameworks have been developed to assess the pathogenicity of individual mutations. These include sorting intolerant from tolerant (SIFT), polymorphism phenotyping v2 (PolyPhen-2), evolutionary model of variant effect (EVE), and combined annotation-dependent depletion (CADD) scores—providing a robust basis for assessing the deleterious effects of mutations [7–10]. The comprehensive nature of the EVE and CADD model has led to their increased prominence and widespread adoption in genomic research, facilitating a deeper understanding of the molecular underpinnings of cancer. While these tools help predict mutational impact, transcriptomic profiling is increasingly used to evaluate the downstream biological consequences of such mutations, including pathway activation states and gene expression programs that are not discernible from DNA-level data alone. This complementary role has positioned RNA-seq as an essential component of contemporary cancer genomics.

Alongside molecular profiling, radiomics has emerged as a non-invasive approach to characterize tumor biology using routinely acquired medical imaging modalities, such as positron emission tomography (PET) and computed tomography (CT). By leveraging high-dimensional feature extraction, radiomics facilitates the discovery of quantitative imaging biomarkers that reflect the tumor’s underlying biological processes. Radiomic features—encompassing geometric shape descriptors, first-order intensity statistics, and advanced texture-based metrics—offer a detailed, reproducible characterization of intratumoral heterogeneity and spatial complexity [11–13]. For instance, increased fluorodeoxyglucose (FDG) uptake in an [¹⁸F]FDG PET image is indicative of heightened glycolytic activity, a hallmark frequently linked to aggressive tumor phenotypes [14, 15]. Integrating radiomic features with genomic and transcriptomic profiles allows for the development of radiogenomic models that map imaging traits to specific molecular alterations. This multi-modal approach holds promise for non-invasive tumor characterization, patient stratification, and therapy response monitoring, thereby supporting further investigation of functional imaging as a component of integrative oncology research.

Among the signaling pathways implicated in HNSCC, the Hedgehog (HH) pathway plays a pivotal role in regulating cell proliferation, tissue regeneration, and stemness. Aberrant activation of this pathway has been associated with cancer progression, immune evasion, and therapy resistance [15–17]. At the core of this pathway is GLI1, a zinc finger transcription factor that governs the expression of oncogenic programs and supports the maintenance of cancer stem cells (CSCs)—a subpopulation of therapy-resistant, tumor-initiating cells that contribute to recurrence [18–23]. The presence of CSCs in HPV-negative HNSCC has been linked to aggressive disease biology, positioning the HH-GLI1 axis as a biologically-relevant pathway with therapeutic relevance.

In this study, we demonstrate that specific radiomic features extracted from [¹⁸F]FDG PET/CT imaging quantitatively reflect the transcriptomic and proteomic activation state of the HH pathway in HPV-negative HNSCC. By integrating WES-based molecular profiling with transcriptomic signatures and functional imaging data, we identify a radiogenomic signature capable of stratifying tumors based on pathway activity, biological aggressiveness and clinical outcome. In a preclinical setting we show that GLI1 inhibition—either pharmacologically or through genetic knockdown—not only attenuates tumor proliferation and stem-like properties but also induces measurable changes in imaging-derived features, thus offering a non-invasive readout for treatment response. Collectively, these findings establish an imaging-guided, biology-informed framework that supports future investigation of precision oncology approaches for patients with HPV-negative HNSCC.

Materials and methods

Patient characteristics and human samples

Initially, 127 patients diagnosed with HNSCC between the 8th of June 2006 and the 31 st of July 2015, with whole-body [¹⁸F]FDG PET/CT scans acquired at the General Hospital Vienna, were retrospectively enrolled in this study (institutional review board approval with ethics ID 1649/2016). All patients gave written consent, and experiments were conducted following the Declaration of Helsinki. In total, 73 patients were excluded from the study due to several reasons: lesion sizes below 64 voxels (N = 2), the presence of a second primary tumor (N = 4), a positive HPV status (N = 10), insufficient DNA amount for WES analysis (N = 57), and rare localization (Sinonasal squamous cell carcinoma, N = 1). This resulted in a final cohort size of 51 HPV-negative HNSCC patients for further analysis. Further clinical and imaging-related details are provided in the Supplementary Methods.

Genomic data acquisition and processing

DNA isolation from formalin-fixed paraffin-embedded material

DNA was extracted from HNSCC formalin-fixed paraffin-embedded (FFPE) samples. Tumor areas were annotated by a board-certified pathologist (L.K.), and from these annotated regions, DNA was isolated. WES was performed on a NovaSeq 6000 system (Illumina). Additional methodological details regarding the DNA isolation procedure and normalization approach are provided in the Supplementary Methods.

Whole exome sequencing analysis

Variant calling and filtering

The variant calling and filtering were conducted as Spielvogel et al.[24] described. In brief, raw reads were aligned to the GRCh38 genome using Burrows—Wheeler Alignment (BWA) [25]. Small variants were detected with Strelka2 [26] and VarDict [27], then merged and annotated with the Variant Effect Predictor (VEP) [28], including CADD scores [8]. Germline variants were filtered based on Sukhai et al. [29], retaining only somatic variants present in less than 10% of samples, with specific read thresholds. Filtering utilized population variant databases, keeping variants with a minor allele frequency below 1% in non-Finnish Europeans and excluding those classified as "benign" or "likely benign" in ClinVar database [30].

Pathway-level alteration scores and pathway selection

To derive higher-level genomic features, alteration-level CADD scores (RRID:SCR_018393) were aggregated to compute gene-level alteration scores by summing CADD scores across all variants per gene. Genes were then assigned to pathways using the Kyoto Encyclopedia of Genes and Genomes (KEGG) database (RRID:SCR_012773, [31]). For downstream analysis, only pathways associated with the KEGG network term “Pathways in Cancer (hsa05200)” were selected, yielding 21 cancer-related KEGG pathways. Patients were stratified into “low” or “high” pathway alteration groups based on the median alteration score for each of the 21 pathways. Log-rank survival analyses were performed in an exploratory, hypothesis-generating framework to rank pathways according to their prognostic relevance. The pathway yielding the strongest separation between the groups, defined by the smallest nominal p-value, was selected for downstream analysis and in vitro/in vivo validation. No formal multiple-testing correction was applied, as the analysis was intended for pathway prioritization rather than confirmatory inference, with independent molecular and functional validation serving as orthogonal support. Further technical details and the full list of included pathways are provided in the Supplementary Methods and Supplementary Table 2.

Copy number alteration detection from the whole exome sequencing data

Two independent tools, CNVkit (RRID:SCR_021917, [32]) and GATK-CNV [33], were used to identify copy number variants (CNVs) from the WES reads, and the results were merged using a custom integration approach. CNVs were annotated using gene models from the KEGG pathway database to focus on cancer-relevant regions. Additional methodological details, including filtering criteria, annotation sources, and CNV post-processing steps, are provided in the Supplementary Methods.

Transcriptome data acquisition and processing

RNA isolation from formalin-fixed paraffin-embedded material

First, tumor areas on FFPE samples were annotated by a board-certified pathologist (L.K.), and from these annotated regions, RNA was isolated. RNA sequencing was performed on a AVITI System (Element Biosciences). Additional methodological details regarding the RNA isolation procedure and normalization approach are provided in the Supplementary Methods.

RNA sequencing data analysis

The differential gene expression analysis was calculated based on the gene counts produced using featureCounts from Subread package v2.0 (RRID:SCR_009803, [34]) and further analyzed by Bioconductor package DESeq2 v1.34.0 (RRID:SCR_015687, [35]). Data generated by DESeq2 with independent filtering were selected for the differential gene expression analysis due to its conservative features and to avoid potential false positive results. Genes were considered as differentially expressed based on a cut-off of adjusted p-value < 0.05 and log2(fold-change) ≥ 1 or ≤ −1.

Gene Set Enrichment Analysis (GSEA) was performed using the R bioconductor package clusterProfiler v4.0.0 (RRID:SCR_016884, [36]). The source of annotation was the KEGG database (RRID:SCR_012773) and either all pathways (n = 344) or only pathways in the KEGG network term “Pathways in Cancer (hsa05200)” (n = 21) were used.

Image acquisition and radiomic feature extraction

Image acquisition and radiomic feature extraction and prioritization were performed as previously described by Spielvogel et al. [24]. Briefly, tumors were delineated by two board-certified nuclear medicine specialists, and consensus segmentations were used to define volumes of interest (VOIs). Radiomic features (n = 112) were extracted from normalized standard uptake value (SUV) maps using an International Biomarker Standardization Initiative (IBSI)-compliant in-house framework [37]. The analyses were conducted in an exploratory, hypothesis-generating manner to assess associations between imaging-derived features and HH pathway disruption, to identify potential and biologically plausible links between genomic alterations and imaging phenotypes. Accordingly, radiomic features were evaluated for consistency with molecular and functional findings in an exploratory framework, without formal multiple-testing correction. Additional technical details regarding the delineation procedure, software tools, and feature extraction workflow, as well as a full list of considered features, are available in the Supplementary Material.

In silico mining of external cohorts

Bulk RNA sequencing data

RNA-seq data for GLI1 and various CSC markers (e.g., SOX2, ALDH1A1) were retrieved from the TCGA-HNSC dataset (RRID:SCR_003193) and analyzed to investigate differential expression across subgroups defined by HPV status. Subgroup information was obtained via the TCGAbiolinks R package (RRID:SCR_017683).

To complement the TCGA data, gene expression information for GLI1 was also extracted from the Cromer Head-Neck dataset via the Oncomine™ Research Premium database. Further gene expression data on GLI1 and key CSC markers in primary HNSCC cell lines were obtained from the DepMap database (release 23Q3). Additional processing details, and plotting methods are provided in the Supplementary Methods.

Single-cell RNA sequencing data

Publicly available single-cell RNA sequencing data (dataset GSE181919 [38]) was retrieved from the GEO database and analyzed using the Seurat package (RRID:SCR_016341, [39]). The analysis focused on cells annotated as Primary Cancer (CA), and dimensionality reduction was performed using UMAP based on a predefined list of tumor-initiating cell (TIC)-associated genes. Gene expression of HH pathway members was visualized using UMAP and compared across cellular subpopulations. Additional processing parameters, filtering steps, and tools are detailed in the Supplementary Methods.

Staining of human samples

All patient samples, where enough residual tumor tissue was left after (i) the DNA extraction/sequencing and (ii) the RNA isolation/sequencing were used for staining of protein targets. For each staining a 3 µm thick tissue slice was used.

Immunohistochemistry

Immunohistochemical staining for protein expression was performed as previously described [40]. Staining for Ki67 and SOX2 was conducted using the following antibodies (Ki67: Cell Signaling #12,202, 1:300 dilution, RRID:AB_2620142; SOX2: Cell Signaling #14,962, 1:300 dilution, RRID:AB_2798664), and slides were scanned with the Olympus Slideview VS200 scanner (RRID:SCR_024783).

Immunofluorescence

Immunofluorescence (IF) staining was performed to detect nuclear GLI1 protein expression. Tissue slides were incubated with an anti-GLI1 antibody (NBP1-78,259, 1:200 dilution, RRID:AB_11030198), and nuclei were counterstained with DAPI (Merck, #508,741, 1:50,000 dilution). Stained slides were scanned using the Olympus Slideview VS200 scanner.

The analyses of marker expression on the stained slides was performed in QuPath (Version 0.4.3, RRID:SCR_018257). Detailed analyses parameters and thresholds are provided in Supplementary Table 3.

In vitro methodologies

Detailed cell numbers, seeding densities, treatment concentrations, and incubation durations are provided in the Supplementary Methods.

Cultivation of human HNSCC cell lines and keratinocyte cells

Human HNSCC cell lines (FaDu, CAL27, SCC25, SCC154) and keratinocyte lines (HaCaT, primary oral keratinocytes) were cultured under standard conditions in DMEM or DermaLife K Medium supplemented with 10% FBS and 1% Penicillin/Streptomycin. Cells were maintained at 37 °C and 5% CO₂, regularly tested for mycoplasma contamination, and counted using a CASY cell counter.

Viability assay

To assess cell viability upon JK184 treatment or GLI1 knockdown, resazurin-based viability assays were performed on CAL27, SCC25, FaDu, and SCC154 cell lines. Measurements were taken after 24 or 72 h (h) of treatment using a TECAN Synergy H1 plate reader.

Migration assay

A wound healing-based migration assay was conducted using IBIDI inserts in CAL27 cells. JK184 was applied after insert removal, and migration was visualized and quantified over time using a TECAN Spark reader and ImageJ analysis.

Colony formation assay

Colony formation capacity was tested after JK184 treatment or GLI1 knockdown. CAL27 and SCC25 cells were seeded at low densities, and colonies were stained with crystal violet after 14–21 days and manually counted.

Immunoblotting

Whole protein lysate preparation and immunoblot analyses were performed as described previously [40]. A short version of the procedure and a list of all primary antibodies is available in Supplementary Materials and Supplementary Table 4, respectively.

Fluorescence-Activated Cell Sorting (FACS)

Annexin V staining—detection of apoptosis

To assess cell apoptosis after treatment with JK184 or GLI1 knockdown induction, a FACS Annexin V/7-AAD assay was performed as previously described [40].

DAPI cell cycle staining—detection of cell cycle state

To determine the cell cycle phase distribution, JK184-treated HNSCC cells were processed as in the Annexin V assay. Briefly, 5 × 10⁵ cells were fixed dropwise in 70% ethanol (4.5 mL added to 0.5 mL cell suspension while vortexing) and incubated overnight at 4 °C. After washing with PBS, cells were stained with 1 µg/mL DAPI (Thermo Fisher, #62,248) in 200 µL MACS buffer for 30 min (min) at room temperature (RT) and analyzed via flow cytometry.

Assessment of marker expression

To analyze CSC-related marker expression, cells were harvested 96 h post-seeding, with doxycycline (Dox, 100 ng/mL) added at seeding and again at 48 h of incubation. Cells were incubated for 20 min at 4 °C in 25 µL MACS buffer containing the following antibodies (1:50 dilution): APC-CD24 (BioLegend, #311,118, RRID:AB_2072735), FITC-CD44 (BioLegend, #338,804, RRID:AB_1501197) and PE-c-MET (R&D Systems, FAB3582P, RRID:AB_1026293). After staining, cells were washed twice in MACS buffer and resuspended in 100 µL MACS containing either 5 µL 7-AAD (Abcam, ab228563) or 1 µg/mL DAPI for viability gating.

Flow cytometry acquisition and analysis

All samples were acquired on a BD FACS Canto II (Becton Dickinson) and analyzed using FACSDiva software (version 8.0.1). Standard gating strategies were applied to exclude debris and dead cells, and to quantify marker-positive populations.

Genetic engineering of HNSCC cell lines

HNSCC cells (3–4 × 10⁵ per well in 6-well plates) were transduced with 30 µL of concentrated viral supernatant in the presence of 10 µg/mL polybrene (Merck, #TR-1003-G), diluted in DMEM + +. Plates were centrifuged at 1,000 × g for 1 h at RT. Three days post-transduction, successfully transduced cells were selected using 2 µg/mL puromycin (for Tet-pLKO-puro and mCherry constructs), 300 µg/mL hygromycin (for pMSCV constructs) for 48–72 h, or FACS sorted using a FACSMelody™ Cell Sorter (Becton Dickinson).

The virus production and cloning strategy for generating GLI1-targeting constructs, overexpression plasmids, and mCherry controls are detailed in the Supplementary Methods.

RNA sequencing and GSEA of HNSCC cell lines

RNA isolation, sequencing, and data processing were performed as described [41]. The GSEA was performed in R using the MSigDB terms related to “stem cell” and “progenitor” in the C2:CGP network and the HH pathway-associated terms in the CP:REACTOME network (packages: fgsea [1.20.0, RRID:SCR_020938] and msigdbr [7.5.1, RRID:SCR_022870]).

Organoid culture and screening

Cultivation of human HNSCC organoids and viability assessment

Patient-derived HNSCC organoids were generated from tumor biopsies collected at Antwerp University Hospital, with ethics approval and informed consent (UZA Ethical Committee, ref. 20/09/090). Organoid lines (HNSCC_001, HNSCC_006, HNSCC_007, HNSCC_010) were cultured following the protocol of Driehuis et al. [42]. Drug screening with JK184 (1 nM-10 µM) was performed using an automated high-throughput pipeline at DrugVision.AI (University of Antwerp) over five to seven days. Organoid viability was quantified using the OrbitsOncology label-free image analysis platform, and normalized drug response (NDR) was calculated.

The full screening and viability quantification protocol, including NDR formula and classification criteria, is available in the Supplementary Methods.

In vivo methodologies

All animal procedures were approved by the Medical University of Vienna and the Austrian Ministry of Education and Science (ethical protocol: 2022–0.626.080). Tumor volumes never exceeded 20 mm in any dimension.

Murine xenograft model

Both male and female NSG (NOD.Cg-Prkdc^scid Il2rg^tm1Wjl/SzJ, RRID:IMSR_JAX:005557) mice were subcutaneously injected in the hind flanks with 4 × 10⁶ CAL27, SCC25, or SCC25 shGLI1_2 cells. When tumors reached ~ 50–150 mm³, mice were randomized to receive either (a) vehicle or JK184 (5 or 10 mg/kg daily, i.p., for 16–17 days), or (b) vehicle or Dox (500 mg/kg in drinking water for 22 days). Tumor volumes were measured every 3–4 days using calipers and calculated as: volume (mm³) = (length × width²)/2. Mice were euthanized at endpoint; tumors and organs were collected. Experimental details and randomization strategy are outlined in the Supplementary Methods.

Preclinical PET/CT imaging

Three randomly selected mice per group underwent [¹⁸F]FDG-µPET/CT imaging. 100 µL [¹⁸F]FDG was injected via the tail vein, followed by static µPET and µCT scans 60 min post-injection (Siemens Inveon). Imaging protocol mirrored human clinical HNSCC PET/CT standards. Radiomic features were extracted in compliance with the IBSI from the manually segmented tumor regions of interest (ROIs) using 3D Slicer (RRID:SCR_005619). Radiomic analysis in this setting was performed in an exploratory, hypothesis-driven manner to assess whether GLI1 knockdown-induced biological perturbations are reflected in imaging-derived features and whether these changes recapitulate patterns observed in patient tumors. Given the high dimensionality and intercorrelation of radiomic features, no formal multiple-testing correction was applied, as the objective was feature prioritization and biological signal detection rather than confirmatory biomarker testing. To limit overinterpretation related to multiplicity, emphasis was placed on effect size, directionality, and consistency across independent data (genomic, transcriptomic, and imaging), as well as orthogonal functional validation in genetic and pharmacological GLI1 perturbation models. Concordance between patient and preclinical radiomic patters was therefore interpreted as supportive biological evidence, and findings from radiomic analyses are framed as exploratory and hypothesis-generating, with formal validation requiring future prospectively designed studies. Further experimental details are outlined in the Supplementary Methods.

In vivo fluorescence imaging

Fluorescence imaging was conducted using an IVIS® Lumina™ S5 (Perkin Elmer) with a mCherry filter. Mice were imaged on day 0 and day 14 using standardized acquisition settings. A native NSG mouse served as a background control. Signals were normalized and quantified with Living Image® software (v4.8.0).

Histology and immunohistochemistry

Tumors and organs were fixed in 4% phosphate-buffered formaldehyde (Roti® Histofix, Carl Roth), paraffin-embedded, and sectioned at 2 µm. H&E staining was performed on all tissues. Immunohistochemistry for Ki67 (Cell Signaling, #12,202, 1:400 dilution, RRID:AB_2620142), cleaved caspase-3 (CC3, Cell Signaling, #9664, 1:150 dilution, RRID:AB_2070042) and SOX2 (Cell Signaling, #14,962, 1:100 dilution, RRID: AB_2798664) was conducted on tumor sections following the same protocols described in [40]. Imaging and quantification procedures followed those described in the Supplementary Material and Supplementary Table 3.

Blood parameter assessment

Blood was collected via cardiac puncture into Mini-Collect K3EDTA tubes (Greiner Bio-One). Plasma was isolated by centrifugation at 5000 × g for 20 min. Serum levels of aspartate aminotransferase (AST), alanine aminotransferase (ALT), and blood urea nitrogen (BUN) were quantified using a VetTest 8008 chemistry analyzer (IDEXX Laboratories).

Statistical methods and software

Statistical analyses were performed using R (v4.1.3, R Foundation for Statistical Computing, Vienna, Austria) and GraphPad Prism (v9.0, GraphPad Software, RRID:SCR_002798) with a general α of 0.05. All p-values in the context of the association of CADD values with prognosis and radiomic feature association are to be interpreted exploratorily.

The survival analyses was conducted using the survminer package (v0.4.9, RRID:SCR_021094) with the log-rank test. In vitro and in vivo data are presented as mean ± standard deviation (SD) or standard error of the mean (SEM), respectively. Differences in gene/protein expression between two or more groups were assessed using a one-sided t-test or two-way ANOVA.

For radiomic analyses, Mann–Whitney U tests and Pearson correlations were applied to account for non-normal distributions and ensure feature comparability. GSEA of RNA-seq data was performed using the weighted Kolmogorov–Smirnov test. Gene expression correlations were analyzed using Spearman correlation.

Tumor growth in xenograft experiments was assessed via two-way ANOVA or mixed-effects models (REML) with stacked matching. Tumor weight and marker levels were analyzed using one-way ANOVA (CAL27) or one-sided t-tests (SCC25 WT vs. shGLI1_2). Venn diagrams were generated using InteractiVenn [43], and figure composition was done in Affinity Publisher (Serif, https://affinity.serif.com/de/).

Results

Genomic, transcriptomic and in vivo imaging characterization of HNSCC tumors reveals Hedgehog signaling as a key prognostic indicator

To identify genomic traits with prognostic relevance, WES on 51 HPV-negative HNSCC tumors was performed. It revealed recurring mutations in genes, such as MUC4 (63%), TTN (35%), and TP53 (31%) (Fig. 1A), consistent with prior studies [5]. We utilized the CADD model to elucidate the functional implications of these mutations. However, no prognostic stratification was evident at the gene level relevant to at least 15% of the cohort (Fig. 1A, Supplementary Table 6). To capture broader biological relevance, we aggregated CADD scores across 21 KEGG cancer-related pathways, following Spielvogel et al. [24]. The HH signaling pathway showed the strongest association with disease-free survival (p = 0.009, Fig. 1B-C) and also remained the strongest independent predictor in multivariate Cox regression (Supplementary Fig. S1A). This prognostic value was confirmed in the TCGA-HNSCC cohort (Supplementary Fig. S1B). Further analysis of patient samples revealed significantly elevated transcriptional activity of the HH pathway in the “HH high” group (Fig. 1D-E, Supplementary Fig. S1C). This finding was corroborated by an enrichment of somatic mutations and structural variants associated with HH activation (Supplementary Table 7–8). Furthermore, IF staining demonstrated significantly elevated nuclear expression of GLI1 (p = 0.0002, Fig. 1F and Supplementary Fig. S1D), the key activator of the HH pathway, in HH-high tumors. This increase in nuclear GLI1 was accompanied by a significantly higher fraction of Ki67-positive cells, indicating enhanced proliferative activity in this group (p = 0.036, Supplementary Fig. S1E). Next, we explored potential associations between the alteration state of the HH pathway and PET/CT radiomic features. The alteration state of the HH pathway correlated significantly with specific imaging features, notably the first order CT features “stat.P10” and “stat.mean”, and many PET texture features, including the first order feature (“ih.uniformity”) (Fig. 1G and Supplementary Fig. S1F). In addition, a surrogate marker for peak FDG uptake also showed a slight positive correlation with the alteration state of the HH pathway (Supplementary Fig. S1H). Importantly, several radiomic features showed independent prognostic value (Supplementary Fig. S1I and Supplementary Table 9), underscoring the complementary prognostic value of imaging and genomic data. These findings support previous evidence that integrating radiomic and genomic features enhances patient stratification beyond either modality alone [24].

Fig. 1.

Multiomics profiling of 51 HPV-negative HNSCC tumors identifies Hedgehog signaling as a prognostic marker. A Heatmap showing genes (rows) mutated in ≥ 15% of tumors (≥ 8 patients), clustered hierarchically by mutation patterns. Columns represent patient samples, also hierarchically clustered. Clinical annotations (e.g., sex, treatment status) are shown above. Right panel: dot plot of log-rank p-values assessing prognostic impact of each gene. B Binary heatmap of CADD-derived pathway disruption scores (median as cutoff) across 21 cancer-related KEGG pathways (columns) per patient (rows), with prognostic significance (log-rank p-values), which was used as ranking, indicated above. C Kaplan–Meier curve showing reduced survival in patients with high HH pathway alteration. D Gene set enrichment analysis of six “high” and six “low” patient samples across 21 cancer-related KEGG pathways. E Enrichment plot for “Hedgehog signaling” of “high” vs “low” patients. F Nuclear GLI1 protein expression (red) in HNSCC patients of the in-house cohort, grouped by HH alteration status: “low” (N = 9) and “high" (N = 9). Cell nuclei stained with DAPI (cyan), stromal regions outlined in white. Scale bar: 50 µm. G Chord diagrams showing correlations (R values, color-coded) between the top genomic feature and CT- or PET-derived radiomic features. Patients grouped by HH pathway status as in (C)

GLI1 is overexpressed in HPV-negative tumors

To investigate the oncogenic potential of HH signaling in HPV-negative HNSCC, we assessed the expression of GLI1, the pathway’s key transcriptional activator. GLI1 was found to be significantly upregulated in HPV-negative tumor versus normal tissue and HPV-negative vs HPV-positive tumor in two independent datasets (Crome Head & Neck and TCGA-HNSC, p-value < 0.001, Supplementary Fig. S1J-K), supporting its potential relevance in malignant transformation. Furthermore, GLI1 expression alone exhibited a trend toward an association with poorer patient survival in the TCGA-HNSC cohort (Supplementary Fig. S1L).

Hedgehog pathway blockade inhibits proliferation, induces apoptosis and cell cycle arrest in HNSCC cell lines and patient-derived tumoroids

In light of the observed consistent upregulation of GLI1 in HPV-negative HNSCC, we evaluated the therapeutic potential of its inhibition using the small-molecular inhibitor JK184 [44]. Treatment of four HNSCC cell lines with JK184 significantly reduced proliferation, induced G₁ cell cycle arrest (Fig. 2A, Supplementary Fig. S2A), and decreased GLI1 protein levels (Fig. 2B). Apoptosis was also induced across all cell lines (Fig. 2C, Supplementary Fig. S2B), with sensitivity correlating to baseline GLI1 expression (Supplementary Fig. S2C). JK184 had minimal effects on non-malignant epithelial cells (HaCaT and primary oral keratinocytes), suggesting a therapeutic window (Fig. 2D, Supplementary Fig. S2D-E). Efficacy of GLI1 blockade was further confirmed in four patient-derived tumoroids (Fig. 2E-F), illustrating efficacy in complex patient-derived 3D models of HNSCC. Noteworthy is that the observed effects were more substantial in HPV-negative lines (Fig. 2A-C, Fig. 2E, and Supplementary Fig. S2A-B).

Fig. 2.

JK184 selectively inhibits growth and induces apoptosis in HPV-negative HNSCC cells. A Dose–response curves showing JK184 sensitivity in three HPV-negative (FaDu, CAL27, SCC25) and one HPV-positive (SCC154) HNSCC cell line after 72 h. IC₅₀ values are indicated. B Immunoblots and quantification of GLI1 protein expression in HNSCC cells treated with JK184 for 24 h, normalized to loading controls (β-TUBULIN or GAPDH) and untreated cells. C Apoptosis analysis (Annexin V/7-AAD) showing proportions of viable, early apoptotic, late apoptotic, and dead cells after 24 h of JK184 treatment. Data are presented as mean ± SD. D JK184 dose–response curves and IC₅₀ values in CAL27, HaCaT (immortalized keratinocytes), and OralKera (primary gingival keratinocytes). E Normalized drug response (NDR) curves for four patient-derived HNSCC tumoroid lines treated with JK184 for 168 h. F Representative images of untreated and JK184-treated [22 nM] HNSCC organoids after 120 h of treatment. Magenta: Orbits® label-free tumoroid segmentation; Green: Cytotox green fluorescence signal (cell death)

GLI1 knockdown reduces proliferation, induces apoptosis, and modulates the cancer stem cell phenotype in HPV-negative HNSCC cell lines

To further dissect GLI1’s role in HPV-negative HNSCC, we used an Dox-inducible short hairpin (shRNA)-mediated knockdown approach to precisely study the role of GLI1 in CAL27 and SCC25. Dox treatment reduced GLI1 protein levels, inhibited proliferation, and induced apoptosis (Fig. 3A-C). Given the established HH pathway’s role in modulating the plasticity and stemness of cancer cells across a variety of cancers [17, 18, 21, 45], we assessed the impact of GLI1 knockdown on key CSC markers. Consistent with GLI1's role as stemness oncogene, the expression of ALDH1A1, OCT4, SOX2, CD44, and c-MET decreased, while CD24—a marker linked to reduced inflammation and favorable prognosis [46]—increased (Fig. 3D-E). This was accompanied by a reduction in colony-forming ability (Fig. 3F).

Fig. 3.

GLI1 knockdown reduces proliferation, induces apoptosis, and alters cancer stem cell markers in vitro. A Immunoblotting results of GLI1 and SUFU in CAL27 and SCC25 cells post 96 h of cultivation with or without Dox [100 ng/mL]. B Growth curves of two inducible GLI1 knockdown clones (GLI1_1 and GLI1_2) in CAL27 (left) and SCC25 (right). Arrows indicate time points of Dox addition. C Annexin V/7-AAD assay showing proportions of viable, early apoptotic, late apoptotic, and dead cells in WT and GLI1 knockdown CAL27 and SCC25 cells after 96 h ± Dox. D-E Expression of cancer stem cell markers in SCC25 WT and shGLI1_2 cells after 96 h ± Dox: (D) CD24, CD44, c-Met; (E) ALDH1A1, OCT4, SOX2. (F) Colony-formation assay post priming in DMEM + + with or without Dox for 96 h: Formed colonies were visualized on day 21 with crystal violet

Similar effects were observed in CAL27 cells treated with JK184, as confirmed by GSEA, and reduced migration, colony formation, and CSC marker expression (Supplementary Fig. S3A-E). Conversely, GLI1 overexpression induced opposing phenotypes (Supplementary Fig. S3F). Correlation analysis in DepMap and TCGA datasets further linked GLI1 expression to CSC markers like CD44, POU5F1, and SOX2 (Supplementary SFig. 3G-H)). Additionally, single-cell RNA-seq data revealed that tumor-initiating cell clusters (TICs) exhibited strong HH pathway activity and high CSC marker expression (Supplementary Fig. S3I).

Together, these data highlight a critical link between HH activation and a malignant CSC phenotype in HNSCC, underscoring the therapeutic potential of targeting this pathway in managing aggressive subtypes of this tumor.

Genetic GLI1 inhibition attenuates tumor growth and mimics in vivo imaging characteristics of patients with aggressive HNSCC

To validate the in vitro efficacy of GLI1 inhibition in vivo, we subcutaneously injected inducible SCC25 shGLI1_2 or SCC25 WT control cells into NSG mice. Upon tumor formation (~ 50 mm³), GLI1 knockdown was induced via Dox (Fig. 4A, Supplementary Fig. S4A). Over three weeks, GLI1 suppression significantly reduced tumor growth and weight (Fig. 4B-4E, Supplementary Fig. S4B-4E), without any signs of systemic toxicity (Supplementary Fig. S4J-K). Significantly, Dox treatment drastically reduced GLI1 levels in the SCC25 shGLI1_2 tumors (Fig. 4F). At the same time, GLI1 expression was largely unaffected in the parental SCC25 tumors (Supplementary Fig. S4F). Notably, GLI1 knockdown also reduced nuclear SOX2 expression (Fig. 4G), an effect not observed in WT tumors (Supplementary Fig. S4G).

Fig. 4.

GLI1 knockdown suppresses tumor growth and mirrors in vivo imaging features observed in HPV-negative HNSCC patients. A Experimental schematic showing interventions including treatment and tumor measurement. B-C Tumor growth curves and representative images (individual tumor photographs were taken independently and subsequently put together in a collage at an appropriate scale) of subcutaneous SCC25 shGLI1_2 xenografts in vehicle-treated (“No Dox”) vs. Dox-treated NSG mice (Day 22). Scale bar: 1 cm. D Tumor weight at day 22: vehicle (n = 5) vs. Dox [500 mg/kg] (n = 11). E In vivo mCherry fluorescence on days 0 and 14. F Immunoblot of GLI1 and β-ACTIN from tumors (n = 3/group) with quantification normalized to vehicle. G SOX2 expression in tumor Sects. (1 area shown per group; quantified from 5 regions per tumor, 3 tumors per group). Scale bar: 50 µm. H Overview of the in vivo imaging and analysis pipeline. I-J Heatmaps of radiomic features from [¹⁸F]FDG µPET (I) and CT (J) scans (n = 3/group) with selected group-specific features. K Overlap of radiomic features linked to HH status in patients and features discriminative in the preclinical model. “Not applicable” denotes unmatched features between datasets. Tabular summary of the overlapping radiomic feature

To evaluate treatment-induced radiomic changes, [¹⁸F]FDG µPET/CT scans were performed on three randomly selected mice per group(Fig. 4H). Hierarchical clustering of averaged features revealed distinct PET (Fig. 4I) and CT (Fig. 4J) profiles between groups. GLI1 knockdown reduced “PET: Maximum” and increased “PET: Joint energy”, indicating lower [¹⁸F]FDG uptake and greater signal homogeneity, respectively (Fig. 4I). CT-derived first order features, such as “CT: Skewness”, and texture features, such as “CT: Correlation”, were also modulated, illustrating the broader impact of GLI1 knockdown on tumor radiomics (Fig. 4J).

Importantly, 4 of 6 (66.7%) PET-derived and 6 of 11 (54.5%) CT-derived features that correlated with HH signaling in patients also aligned with those seen in GLI1-suppressed tumors in our preclinical models (Fig. 4K). In contrast, Dox treatment of SCC25 WT tumors did not result in any strong group-specific clustering of the radiomics features (Supplementary Fig. S4H-I).

Pharmacological GLI1 inhibition impairs in vivo growth of HPV-negative HNSCC tumors

To assess the in vivo efficacy of JK184, CAL27 cells were injected into NSG mice, followed by daily treatment with JK184 (5 or 10 mg/kg) or vehicle for 16 days once tumors became palpable (Fig. 5A). JK184 significantly inhibited tumor growth from day 3 onward (Fig. 5B-D) without systemic toxicity (Supplementary Fig. S5F-H). Moreover, JK184-treated tumors showed reduced GLI1 protein levels (Fig. 5E), indicating a target-specific effect of the inhibitor. JK184 also decreased Ki67 and increased CC3 expression, indicating reduced proliferation and enhanced apoptosis (Fig. 5F, Supplementary Fig. S5A). Similar results were observed in SCC25 xenografts (Supplementary Fig. S5B-E).

Fig. 5.

Pharmacological Hedgehog inhibition reduces tumor growth and affects clinical, pathological, and imaging features of HNSCC. A Experimental schematic showing treatment and tumor monitoring. B-C Tumor growth curves and representative images of CAL27 xenografts in vehicle- vs. JK184-treated NSG mice at day 16. Scale bar: 1 cm. D Tumor weights on day 16: vehicle (n = 8), JK184 5 mg/kg (n = 6), JK184 10 mg/kg (n = 8). E Immunoblots of GLI1 and GAPDH from tumors (n = 3/group), with GLI1 quantified relative to GAPDH and normalized to vehicle. F Ki67 staining in representative tumor regions (1 shown per group; quantification from 5 areas per tumor, 3 tumors/group). Scale bar: 50 µm. G Schematic model summarizing HH-driven HNSCC growth and the effects of genetic or pharmacological blockade of GLI1-dependent HH signaling on HNSCC tumor growth. The figure was created in BioRender.com

Overall, our findings demonstrate that active HH signaling correlates with poor prognosis and elevated [¹⁸F]FDG and more heterogeneous imaging patterns in HPV-negative HNSCC tumors. By integrating radiomic profiling with genomic and transcriptomic analyses, we identified a consistent molecular and imaging signature of HH pathway activation. Through pharmacologic and genetic perturbation, we further show that HH inhibition reduces tumor growth and stemness, increases apoptosis, lowers [¹⁸F]FDG uptake, and results in more homogeneous imaging features. Notably, GLI1 knockdown produced radiomic patterns resembling those of “HH alteration low” tumors in our in-house cohort (Fig. 5G), highlighting concordance between molecular perturbation and imaging-derived features. Together, these results establish a preclinical framework that directly links imaging, multi-omics molecular profiling, and tumor biology, providing a basis for biologically-informed tumor stratification and future translational evaluation of imaging biomarkers in HNSCC and potentially other cancers.

Discussion

This study demonstrates that integrating molecular profiling with functional imaging offers a precision oncology framework for HPV-negative HNSCC, linking oncogenic HH pathway activity to distinct, quantifiable imaging phenotypes and potential therapeutic vulnerabilities. By complementing genomic profiling with transcriptomic analyses, we captured the functional downstream consequences of HH activation, providing orthogonal validation of pathway engagement beyond mutational status alone. Together, these multi-omics datasets define a biologically coherent HH-active tumor subset, characterized by distinct survival outcomes, concordant genomic, transcriptomic, and imaging signatures. Through genomic analyses, we identified recurrent mutations in negative regulators of the HH pathway, associated with increased GLI1 expression and nuclear localization—a molecular signature of active pathway signaling. Transcriptomic profiling confirmed these findings by demonstrating significant enrichment of the HH pathway and a broad upregulation of HH target genes, aligning with the observed protein-level activation of GLI1. Notably, the “basal cell carcinoma” expression signature—which is known to be driven by aberrant HH signaling—was also significantly enriched in HH-high tumors, providing orthogonal transcriptomic validation for pathway activation in this subgroup. High GLI1 levels, in turn, correlated with enhanced proliferation, stem-like phenotypes and worse patient survival across two independent cohorts of HPV-negative HNSCC, consistent with previous reports linking GLI1 to adverse outcomes across multiple malignancies [22, 47–50].

Importantly, these molecular alterations manifested in distinct imaging phenotypes: tumors with high HH pathway activity demonstrated increased [¹⁸F]FDG tracer uptake and pronounced intratumoral heterogeneity, alongside CT texture abnormalities such as altered “stat.P10” values. These associations extended to PET-derived texture features like “ngt.contrast” and “cm.diff.avg,” and first order metrics such as “ih.max”, which were particularly associated with HH-driven tumor biology. These imaging features reflect aggressive tumor biology and increased metabolic turnover, aligning with established PET/CT correlates of proliferation and cellular disorganization [51–53]. Consistent with this interpretation, RNA-seq analysis revealed that HH-active tumors display significant upregulation of cell cycle and proliferation-associated transcripts, providing a molecular explanation for the observed imaging signatures. Moreover, Ki67 staining confirmed the hyperproliferative phenotype of HH-high tumors, providing additional biological validation for these radiomic signatures.

To validate the biological relevance of these imaging-derived features, we employed both pharmacological inhibition (JK184) and doxycycline-inducible GLI1 knockdown systems in vitro and in vivo. Both strategies significantly reduced proliferation, induced apoptosis, and suppressed stemness markers (CD44, ALDH1A1, SOX2, among others), consistent with prior studies on the role of HH signaling in maintaining a stem-like, therapy-resistant phenotype[21, 54–57]. Notably, GLI1 inhibition was more effective in HPV-negative HNSCC models with higher baseline GLI1 levels, supporting the potential of HH pathway activation as a candidate biomarker for therapeutic response in this subset of HNSCC.

We next asked whether these molecular perturbations translated to observable imaging changes. In GLI1 knockdown xenografts, [¹⁸F]FDG PET/CT scanning revealed significantly reduced FDG uptake (“PET:Maximum”) and increased homogeneity (“PET:Joint energy”)—imaging traits previously linked to favorable biology and improved outcomes in clinical studies [14, 15, 58, 59]. JK184-treated tumors also displayed decreased tumor volumes and reduced GLI1 and Ki67 expression, validating these findings across independent therapeutic modalities and providing preclinical evidence that functional imaging fingerprints can non-invasively reflect treatment response.

Collectively, these findings outline an integrative, biology-driven precision oncology framework in HPV-negative HNSCC, composed of three integrated components: 1) Molecular profiling identifies tumors with active HH signaling. 2) Functional imaging captures this biologic signal through quantifiable radiomic features. 3) Pathway inhibition induces measurable imaging and molecular responses. Through the integration of genomic, transcriptomic, and imaging data, this framework connects genomic alterations with their downstream transcriptional and phenotypic effects, providing a multidimensional view of tumor biology that may inform future minimally-invasive, imaging-guided translational studies.

To facilitate clinical translation, several next steps are critical. First, the reproducibility and robustness of HH-associated radiomic features should be validated in prospective clinical trials with harmonized PET/CT protocols. Second, integrating radiomic features with complementary liquid biopsy improve specificity and enhance patient stratification even further. Third, although WES was successfully performed on 51 FFPE tumor samples, the inherently variable quality of and tissue quantity on archival FFPE material limited the number of samples suitable for RNA-seq, and only 12 samples met stringent quality criteria with comparable read counts. While this selective inclusion ensured analytical robustness, it represents a practical limitation that future studies using prospectively collected, high-quality tissue can address. Fourth, regulatory qualification will require dedicated biomarker-based therapeutic trials that selectively enroll patients with evidence of HH pathway activation, thereby optimizing the likelihood of therapeutic benefit. In this context, we chose GLI1 immunofluorescence staining as a direct protein-level readout of HH pathway activation, offering immediate spatial and subcellular resolution of its activity. This approach provides experimentally grounded functional validation, thereby circumventing the need for computational inference of protein activity via tools such as VIPER [60], which rely on high-quality transcriptomic networks that are difficult to derive from small and FFPE-constrained RNA-seq datasets. Finally, machine learning-based radiogenomic models, such as that developed by Spielvogel et al. [24] could be leveraged to automate prediction of pathway activity from clinical imaging, enabling future investigations of adaptive treatment strategies and early response assessment.

Despite inherent limitations related to cohort size, sample source and anatomical heterogeneity, the strength and reproducibility of our findings are supported by validation in independent patient cohorts and functional in vitro and in vivo modeling. We also note that sample size limitations affected the in vivo experiments, and that larger cohorts would further increase statistical power and reinforce the robustness of the observed effects.

In combination our results; however, highlight the potential of multiomics-integrated [¹⁸F]FDG PET/CT imaging to capture tumor-intrinsic biological programs, offering a path toward more biologically-informed and personalized therapeutic investigations in HPV-negative HNSCC and beyond.

Conclusion

This study demonstrates that distinct radiomic features derived from [¹⁸F]FDG PET/CT imaging can identify a subset of HPV-negative HNSCC patients with particularly aggressive and proliferative tumors driven by HH pathway activation. Integrated genomic and transcriptomic analyses confirmed that these imaging-defined phenotypes correspond to transcriptionally active HH signaling programs. We further demonstrate that GLI1 activation produces a characteristic imaging fingerprint, and that both pharmacologic or genetic inhibition of HH signaling reduces proliferation, induces apoptosis, and reshapes in vivo imaging features in ways that mirror human tumor phenotypes. These findings lay the groundwork for minimally-invasive, imaging-based approaches to identify actionable biological vulnerabilities and assess treatment response in preclinical models, supporting further translational validation of HH-associated radiomic signatures within precision oncology.

Abbreviations

AJCC

American Joint Committee on Cancer

ATCC

American Type Culture Collection

BWA

Burrows-Wheeler Alignment

CA

Primary Cancer

CADD

Combined annotation-dependent depletion (score)

CNVs

Copy number variants

CSC(s)

Cancer stem cell(s)

CT

Computed tomography

DFS

Disease-free survival

DMEM

Dulbecco’s modified eagle’s medium

Dox

Doxycycline

EGA

European Genome-phenome Archive

FACS

Fluorescence-activated cell sorting

FDG

Fluorodeoxyglucose

FFPE

Formalin-fixed paraffin-embedded

GEO

Gene Expression Omnibus

GLI1

Glioma-associated oncogene homolog 1

GSEA

Gene set enrichment analysis

H

Hour(s)

HH

Hedgehog

HNSCC

Head and neck squamous cell carcinoma

HPSCC

Hypopharyngeal squamous cell carcinoma

HPV

Human papillomavirus

i.p.

Intraperitoneal

IBSI

Imaging Biomarker Standardization Initiative

IF

Immunofluorescence

KEGG

Kyoto Encyclopedia of Genes and Genomes

LSCC

Laryngeal squamous cell carcinoma

MTB

Molecular tumor board

MUV

Medical

NDR

Normalized drug response

NSG

NOD scid gamma

OPSCC

Oropharyngeal squamous cell carcinoma

OralKera

Oral keratinocytes

OSCC

Oral squamous cell carcinoma

P/S

Penicillin/Streptomycin

PET

Positron emission tomography

PIL

Preclinical imaging lab

PolyPhen-2

Polymorphism phenotyping v2

RNA-seq

RNA sequencing

ROIs

Regions of interest

RT

Room temperature

SD

Standard deviation

SEM

Standard error of the mean

shRNA

Short hairpin RNA

SIFT

Sorting intolerant from tolerant

TICs

Tumor-initiating cells

TMA

Tissue microarray

VOIs

Volumes of interest

WES

Whole exome sequencing

Authors’ contributions

Conceptualization: S.S. and L.K.; Methodology: S.S., D.P., C.S.P., E.G., M.S., D.H., C.P., D.P.E., V.B., K.T., H.V., H.Z., A.L., C.D., G.H., M.D., F.A., D.S.; Software: S.S., C.S.P., D.H., V.B., K.T.; Validation: S.S., C.S.P., D.H., V.B., K.T., L.K.; Formal analysis: S.S., D.P., C.S.P., D.H., V.B., K.T., H.V., H.Z., A.L., C.D.; Investigation: S.S., D.P., C.S.P., D.H., V.B., K.T., H.V., H.Z., A.L., CD; Resources: R.M., H.A.N., G.H., A.R.H., L.K.; Data curation: S.S., G.H., A.R.H; Writing—original draft: S.S.; Writing—review & editing: S.S., D.P., C.S.P., E.G., M.S., D.H., C.P., D.P.E., V.B., K.T., H.V., H.Z., A.L., C.D., G.H., M.D., F.A., D.S., A.R.H., L.K.; Visualization: S.S.; Supervision: L.K.; Project administration: S.S., L.K., A.R.H.; Funding acquisition: S.S., A.R.H., L.K..

Funding

The financial support by the Austrian Federal Ministry for Digital and Economic Affairs and the National Foundation for Research, Technology and Development is acknowledged. Moreover, the joint funding and scientific support of the Christian Doppler Research Association and Siemens Healthineers, making the Christian Doppler Laboratory for Applied Metabolomics (CDL-AM) possible, is gratefully acknowledged. L.K. acknowledges support from the Christian Doppler Research Association (CDL-AM, CD10277102) and the Austrian Science Fund (FWF grant: P29251 and P34781). In addition, S.S. gratefully acknowledges the financial support through the Alexander Karl Prize of the Stiftung Tumorforschung Kopf-Hals (STKH). For the purpose of open access, the author has applied a CC BY public copyright license to any accepted manuscript version arising from this submission.

Data availability

RNA sequencing and WES data from patient samples generated in this study are available through controlled access at the European Genome-phenome Archive (EGA) under accession numbers EGAD50000002225 (RNA-seq) and EGAD50000002226 (WES). Access is granted to qualified researchers upon approval by the relevant Data Access Committee, in accordance with patient consent and ethical regulations. Processed RNA-seq data have additionally been deposited in the Gene Expression Omnibus (GEO) under accession number GSE316548. Radiomic feature correlation scores with HH pathway disruption and pathway-level alteration scores generated in this study are included in the Supplementary Material - Part 2 or are available from the corresponding author upon reasonable request. All analysis scripts and pipelines used in this study are publicly available at [https://github.com/BioIT-CEITEC] (https://github.com/BioIT-CEITEC). The repository contains curated workflows for demultiplexing, quality control, alignment, quantification, differential expression analysis, pathway enrichment, and DNA sequence alignment. Input files required to reproduce the analyses are provided as Supplementary Data where possible. Additional experimental source data, including Western blot quantifications and flow cytometry data, are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

All patients gave written consent, and experiments were conducted following the declaration of Helsinki and under the institutional review board approval with ethics ID 1649/2016.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.