Single-cell transcriptomic analysis was performed using OSCC scRNA-seq data from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/) under accession numbers GSE173468 (26), GSE188737 (27) and GSE234933 (28). To define a strict OSCC cohort, sample selection was limited to primary tumors localized in the oral cavity, specifically the tongue, buccal mucosa, floor of mouth, hard palate, alveolar ridge and lip, as verified by clinical metadata. Specimens from extraoral head and neck sites (such as the larynx or pharynx) were discarded.

Data processing was performed using the Seurat R package (v5.3.0; Posit Software, PBC). Seurat object was constructed using the ‘CreateSeuratObject’ function. The ‘PercentageFeatureSet’ function was employed to calculate the mitochondrial gene percentage per cell. Quality control thresholds were applied to retain cells expressing between 300 and 6,000 detected genes while exhibiting a mitochondrial gene content of <15%. Genes detected in <3 cells were excluded from subsequent analyses. Batch effects were corrected using the Harmony algorithm. Successful integration was confirmed by Uniform Manifold Approximation and Projection (UMAP) visualization, in which cells were clustered by biological cell type rather than by the dataset of origin, and canonical marker expression patterns for major cell lineages were preserved. The top 3,000 highly variable genes were identified using the ‘FindVariableFeatures’ function. Principal component analysis was executed with the ‘RunPCA’ function. Cell clustering analysis was performed on the initial top 20 principal components using the ‘FindClusters’ function, followed by dimensionality reduction and visualization via UMAP implemented in the ‘RunUMAP’ function. Marker genes defining cell clusters were identified using the ‘FindAllMarkers’ function. Cell clusters were annotated based on the expression profiles of canonical marker genes.

High-dimensional weighted gene co-expression network analysis (hdWGCNA) was performed on scRNA-seq data for epithelial cells using the ‘hdWGCNA’ package. Genes expressed in ≥5% of total epithelial cells were selected for metacell matrix construction via the ‘MetacellsByGroups’ function. The optimal soft-thresholding power was determined using the ‘TestSoftPowers’ function. Following soft threshold selection, an unsigned topological overlap matrix network was constructed via the ‘ConstructNetwork’ function. Module eigengenes were extracted using the ‘ModuleEigengenes’ function. Spearman's correlation coefficients were calculated to assess the associations between module eigengenes and hypoxia/glycolysis scores, with statistical significance defined as a false discovery rate-corrected P<0.05.

Bulk RNA-seq data were derived from The Cancer Genome Atlas (TCGA) Head and Neck Squamous Cell Carcinoma dataset (n=522; http://tcga-data.nci.nih.gov/tcga), in which data from OSCC samples, including tongue, buccal mucosa and lip, were extracted and assembled into the TCGA-OSCC cohort. A total of 50 cancer-related hallmark scores were quantified using single-sample gene set enrichment analysis (ssGSEA) with curated gene sets from the Molecular Signatures Database [accession number: H (hallmark gene sets); https://www.gsea-msigdb.org/gsea/msigdb/index.jsp]. Hierarchical clustering analysis was performed based on the hallmark ssGSEA scores using the gene set variation analysis algorithm. Kaplan-Meier recurrence-free survival (RFS) curves were generated using the ‘survminer’ R package. Differential gene expression analysis between OSCC tissues and normal control tissues was conducted using the ‘limma’ R package. Spearman's correlation coefficients were computed between gene expression levels using R (v4.3.1).

The TCGA-OSCC cohort was used as the training cohort, while the GSE65858 (HNSCC samples; n=270) (29) dataset served as the validation cohort; only samples containing complete survival information were included in either cohort. Hypoxia- and glycolysis-associated genes identified by hdWGCNA were used as input features within the Mime machine learning framework (25). Prior to model construction, the expression profiles of the input genes were independently extracted and subjected to Z-score standardization in both training and validation cohorts. This framework integrates 10 distinct algorithms: Random survival forests, elastic network (Enet), stepwise Cox (StepCox), CoxBoost, partial least squares regression for Cox, supervised principal components, generalized boosted regression models, survival support vector machine, Ridge and least absolute shrinkage and selection operator.

Briefly, the ‘ML.Dev.Prog.Sig’ function was employed to establish combined prognostic models based on the hypoxia- and glycolysis-associated genes using the training cohort (TCGA-OSCC cohort), and the resulting models were validated using the independent validation cohort (GSE65858). The performance of the combined models was evaluated based on the concordance index (C-index) within the validation cohort, where a higher C-index signified superior predictive accuracy and reliability. Patients were stratified into high- and low-risk groups based on the median risk score calculated using the combined model with the highest validation C-index. Kaplan-Meier curves depicting RFS were generated to analyze survival outcomes. Core feature genes were identified from the candidate genes using the ‘ML.Corefeature.Prog.Screen’ function. Cox proportional hazards regression was used to calculate hazard ratios and 95% confidence intervals for the core feature genes.

H3K27ac ChIP-seq data of OSCC cells (HSC3 and SAS) were obtained from the GSE205455 dataset (30). Reads were aligned to the hg19 reference genome using Bowtie2 (v2.4.1; Ben Langmead, Johns Hopkins University), followed by sorting and filtering using the Picard tool to remove unmapped, multimapped and duplicate reads. H3K27ac peaks were identified using the ‘findPeaks’ function in HOMER (v5.1; http://homer.ucsd.edu/homer/) with -style histone parameter, where adjacent peaks within 12.5 kb were merged. SE and TE classification was performed via the SE tool in the HOMER algorithm, defining points with a tangent slope >1 as SEs, and those with a slope ≤1 as TEs. Visualization of the H3K27ac enrichment profiles was conducted using the Integrative Genomics Viewer (https://igv.org).

The Toolkit for Cistrome Data Browser (http://dbtoolkit.cistrome.org/) was used to identify the putative transcription factors for DYNC1H1.

Human OSCC cells (HSC3 and SAS) were obtained from the Japanese Collection of Research Bioresources Cell Bank. Cells were cultured in Dulbecco's Modified Eagle Medium (DMEM; Gibco; Thermo Fisher Scientific, Inc.) supplemented with 10% fetal bovine serum (Gibco; Thermo Fisher Scientific, Inc.) and 1% penicillin-streptomycin (100 U/ml penicillin; 100 µg/ml streptomycin) at 37°C. Under normoxic conditions, cells were maintained in 21% O₂, 5% CO₂ and 74% N₂. Under hypoxic conditions, cells were cultured in 1% O₂, 5% CO₂ and 94% N₂ for 24 h.

Lentiviral constructs mediating the KD of DYNC1H1 (KD-DYNC1H1), OE of DYNC1H1 (OE-DYNC1H1), KD of FOSL2 (KD-FOSL2), OE of FOSL2 (OE-FOSL2) and corresponding negative control vectors [KD-control (ctrl) and OE-ctrl] were purchased from Shanghai GenePharma Co., Ltd. For OE, the full-length coding sequences of DYNC1H1 and FOSL2 were cloned into the pLV-CMV-MCS-PGK-Puro lentiviral vector backbone (Shanghai GenePharma Co., Ltd). The corresponding OE control (OE-ctrl) was the pLV-CMV-MCS-PGK-Puro backbone without insert. For KD, short hairpin RNA (shRNA) sequences targeting DYNC1H1 or FOSL2 were inserted into the pLV-U6-shRNA-PGK-Puro lentiviral vector (Shanghai GenePharma Co., Ltd). The KD control (KD-ctrl) was the pLV-U6-shRNA-PGK-Puro backbone inserted with nontargeting shRNAs.

Recombinant lentiviruses were produced using a 3rd generation system. Briefly, 293T cells (Cell Bank of the Chinese Academy of Sciences) were co-transfected with 10 µg of transfer plasmids (pLV-CMV-MCS-PGK-Puro for OE or pLV-U6-shRNA-PGK-Puro for KD; Shanghai GenePharma), 7.5 µg psPAX2 and 2.5 µg pMD2.G using Lipofectamine 3000 (Invitrogen; Thermo Fisher Scientific, Inc.) according to the manufacturer's protocol. Transfection was performed at 37°C for 8 h, after which the medium was replaced. Lentiviral particles were harvested from the culture supernatant at 48 h post-transfection, filtered through 0.45-µm membranes, and concentrated by ultracentrifugation.

HSC3 and SAS cells were transfected with lentiviral particles at a multiplicity of infection of 10 for 24 h. Stable polyclonal populations were selected using 2 µg/ml puromycin for 7 days and subsequently maintained in DMEM containing 1 µg/ml puromycin. All downstream experiments were performed at least 48 h after the selection process was completed. The efficiency of target gene KD or OE was evaluated using quantitative PCR (qPCR) and western blot analyses. The shRNA sequences were as follows: KD-DYNC1H1, sense, 5′-GCAGAUAAACCCGUGUCUU-3′, and antisense, 5′-AAGACACGGGUUUAUCUGC-3′; KD-FOSL2, sense: 5′-GGAUUAUCCCGGGAACUUU-3′, and antisense, 5′-AAAGUUCCCGGGAUAAUCC-3′; KD-ctrl, sense, 5′-UUCUCCGAACGUGUCACGU-3′ and antisense, 5′-ACGUGACACGUUCGGAGAA-3′.

Total RNA from HSC3 and SAS cells was isolated using the TRIzol^® reagent (Invitrogen; Thermo Fisher Scientific, Inc.), followed by cDNA synthesis using the PrimeScript RT Reagent Kit (Takara Bio, Inc.) according to the manufacturer's protocol. qPCR amplification was performed on a 7,500 Real-Time PCR System (Applied Biosystems; Thermo Fisher Scientific, Inc.) using the SYBR Green qPCR Master Mix Kit (Thermo Fisher Scientific, Inc.) with the following cycling conditions: Initial denaturation at 95°C for 30 sec; 40 cycles of 95°C for 5 sec and 60°C for 34 sec, and 95°C for 15 sec. Relative gene expression was normalized to GAPDH and calculated via the 2^−ΔΔCq method (31). The primer sequences used for qPCR were as follows: DYNC1H1, forward, 5′-TTGGGCACTAGGAAATTGATGC-3′, and reverse, 5′-GCAGGGTTGATACGCCACA-3′; mediator of RNA polymerase II transcription subunit 1 (MED1), forward, 5′-GAGGGCATCAACATTTGGTCA-3′, and reverse, 5′-AGATGAGAGCCCAGTCCATTC-3′; RNA polymerase II subunit A (POLR2A), forward, 5′-GGGTGGCATCAAATACCCAGA-3′, and reverse, 5′-AGACACAGCGCAAAACTTTCA-3′; SMARCC2, forward, 5′-ACTGCCGATCAAATGTTTCCT-3′, and reverse, 5′-ACAGGCAATTATTCTGCACCAAG-3′; FOSL2, forward, 5′-CAGAAATTCCGGGTAGATATGCC-3′, and reverse, 5′-GGTATGGGTTGGACATGGAGG-3′; RNA polymerase II subunit B (POLR2B), forward, 5′-AAAGGCTTGGTTAGACAACAGC-3′, and reverse, 5′-ATCGTGGCGGTTCTTCAACTT-3′; 5′-3′ exoribonuclease 2 (XRN2), forward, 5′-CACACATGAACCGAACTTTACCA-3′, and reverse, 5′-GCACAAGGAAGACTATCGGCAA-3′; mRNA-decapping enzyme 1A (DCP1A), forward, 5′-TCTGGACACAAGCATCTGACG-3′, and reverse, 5′-GGGTGGTGATTTCAGGCTGG-3′; ASXL1, forward, 5′-CGCGCCTGGTATTAGAAAACT-3′, and reverse, 5′-GCATCCTTCTTGAGCGTGAAAAG-3′; RING finger protein 2 (RNF2), forward, 5′-CAAACGGAACTCAACCATTAAGC-3′, and reverse, 5′-CCACTTCTAAGGGCTGTGATG-3′; peroxisome proliferator-activated receptor δ (PPARD), forward, 5′-AGGGCTGACTGCAAACGA-3′, and reverse, 5′-CTGCCACAATGTCTCGATGTC-3′; RBPJ, forward, 5′-CGGCCTCCACCTAAACGAC-3′, and reverse, 5′-TCCATCCACTGCCCATAAGAT-3′; Myc-associated factor X (MAX), forward, 5′-CAATCTGCGGCTGACAAACG-3′, and reverse, 5′-GCACTTGACCTCGCCTTCT-3′; GAPDH, forward, 5′-GGAGCGAGATCCCTCCAAAAT-3′, and reverse, 5′-GGCTGTTGTCATACTTCTCATGG-3′.

HSC3 and SAS cells were lysed in RIPA buffer (MilliporeSigma). Protein concentrations were quantified using the BCA Protein Assay Kit (Beyotime Biotechnology). Lysates were denatured at 95°C for 10 min. Denatured samples (40 µg per lane) were resolved using 10% sodium dodecyl sulfate-polyacrylamide gel electrophoresis and electrotransferred onto polyvinylidene fluoride membranes (MilliporeSigma). Membranes were blocked with 5% skim milk for 2 h at room temperature, followed by overnight incubation at 4°C with primary antibodies. After washing with TBST (0.1% Tween-20), the membranes were incubated with horseradish peroxidase-conjugated goat anti-rabbit secondary antibody (1:5,000; cat. no. ab205718; Abcam) for 1 h at room temperature. Protein bands were visualized using an enhanced chemiluminescence reagent (SuperSignal West Pico PLUS Chemiluminescent Substrate; Thermo Fisher Scientific, Inc.). The following primary antibodies were used: Anti-DYNC1H1 (1:2,000; cat. no. ab245554; Abcam), anti-FOSL2 (1:2,000; cat. no. 19967; Cell Signaling Technology, Inc.), anti-ASXL1 (1:2,000; cat. no. ab50817; Abcam) and anti-GAPDH (1:2,000; cat. no. ab8245; Abcam).

HSC3 and SAS cells were crosslinked with 1% formaldehyde for 15 min at room temperature and quenched with 125 mM glycine solution. After washing with phosphate-buffered saline, cells were lysed in lysis buffer and sonicated on ice using a Q700 sonicator (Qsonica LLC) for 15 cycles (15 sec on, 45 sec off). The sonicated lysates were immunoprecipitated with anti-H3K27ac (cat. no. ab4729; Abcam) and anti-FOSL2 (cat. no. 19967, Cell Signaling Technology, Inc.) overnight at 4°C. Antibody-chromatin complexes were captured by incubation with Protein A/G magnetic beads (Thermo Fisher Scientific, Inc) at 4°C for 2 h, followed by a series of salt washes (including one wash with low-salt buffer, one wash with high-salt buffer, one wash with LiCl buffer and two final washes with TE buffer). Immunoprecipitated chromatin was eluted in elution buffer at 65°C for 30 min, treated with RNase A and proteinase K, and purified by phenol-chloroform-isoamyl alcohol extraction. Purified DNA was analyzed using qPCR as aforementioned. The following primer sequences were used (based on human genome assembly hg38): TE (chr14: 101948350-101948574), forward, 5′-GTGTGACCCGTCTGGTGTAG-3′, and reverse, 5′-GTCCACGTCATTGGGAGAGG-3′; SE1 (chr14: 101964657-101964898), forward, 5′-TCTCATCGCTCCTGGAAGGT-3′, and reverse, 5′-AAGGAACTTGCGCATCTGCT-3′; SE2 (chr14: 101972663-101972979), forward, 5′-GGACTGCGCAATTTCTGTGT-3′, and reverse, 5′-AGCCACACACGATTACAACCT-3′; SE3 (chr14: 101979574-101979881), forward, 5′-GAGGGTAACGCTAGTGAGCC-3′, and reverse, 5′-TGGCTTTTCTCCACGCTCAT-3′.

Cell proliferation was assessed using the CCK-8 (Beijing Solarbio Science & Technology Co., Ltd). Briefly, HSC3 and SAS cells were seeded into 96-well plates at a density of 5×10³ cells/well. At 0, 1, 2 and 3 days post-seeding, 10 µl of CCK-8 solution was added to each well and incubated for 2 h at 37°C. Absorbance was measured at 450 nm using a microplate reader (Thermo Fisher Scientific, Inc.).

Transwell inserts with an 8 µm pore size (Corning, Inc.) were pre-coated with Matrigel (BD Biosciences) and incubated at 37°C for 1 h to allow polymerization. HSC3 and SAS cells were resuspended in serum-free DMEM (Gibco; Thermo Fisher Scientific, Inc.) and seeded in the upper chambers at 3×10⁵ cells/well in 200 µl. The lower chambers contained 600 µl of DMEM supplemented with 10% fetal bovine serum (Gibco; Thermo Fisher Scientific, Inc.). Following 12 h incubation at 37°C, non-invaded cells were removed from the upper membrane surface. Invading cells on the lower membrane surface were fixed with 4% paraformaldehyde for 20 min and stained with 0.1% crystal violet for 30 min at room temperature. The stained cells were quantified by counting five randomly selected fields per membrane under an optical microscope (Olympus Corporation) at ×200 magnification.

HSC3 and SAS cells were seeded into 6-well plates at 3×10⁶ cells/well and cultured for 48 h at 37°C. Glucose uptake was quantified using a Glucose Uptake Assay Kit (cat. no. ab136955; Abcam). Lactate production was measured using an L-Lactate Assay Kit (cat. no. ab65331; Abcam).

Quantitative data are presented as mean ± standard deviation from at least three biologically independent experiments. Statistical analyses were performed using R (version 4.3.1; R Foundation for Statistical Computing) and GraphPad Prism (version 10.4.0; Dotmatics). Comparisons between two groups were assessed using unpaired Student's t-tests. Comparisons among ≥3 groups were conducted using one-way analysis of variance followed by Tukey's post hoc test. P<0.05 was considered to indicate a statistically significant difference.

To delineate the relationships among cancer-related hallmarks, a hierarchical clustering dendrogram of 50 cancer-related hallmark gene sets was generated based on the ssGSEA score matrix. The ‘hypoxia’ and ‘glycolysis’ hallmark gene sets clustered closely in the dendrogram while maintaining greater distance from other hallmark gene sets such as ‘fatty acid metabolism’, ‘oxidative phosphorylation’ and ‘bile acid metabolism’ (Fig. 1A). Patients were stratified into high- and low-score groups based on the optimal cutoff values derived from hypoxia or glycolysis scores. Kaplan-Meier analysis revealed that patients with higher hypoxia scores had significantly worse RFS (Fig. 1B). Although glycolysis scores showed no significant association with RFS in OSCC, a trend toward worse prognosis was observed in patients with high scores (Fig. 1C). Collectively, these results indicate that hypoxia is a risk factor for worse RFS in OSCC.

Unsupervised clustering was performed on single-cell transcriptomic data from 32 OSCC tissues by integrating the GSE173468, GSE188737 and GSE234933 datasets (Fig. 2A). A total of nine distinct cell clusters were identified based on the expression of canonical cell type markers (Fig. 2B). Cell clusters were identified using the following marker genes: KRT6A⁺ for epithelial cells, CD3D⁺ for T/natural killer cells, COL1A1⁺RGS5⁻ for fibroblasts, RGS5⁺ for pericytes, LYZ⁺ for macrophages, CD79A⁺ for B/plasma cells, IRF4⁺ for dendritic cells, VWF⁺ for endothelial cells, and TPSAB1⁺ for mast cells (Fig. 2C and D). Given that OSCC originates from the oral epithelium (32), subsequent analyses on epithelial cells were focused on.

To identify the genes associated with hypoxia and glycolysis, hdWGCNA was performed on single-cell transcriptomic data from epithelial cells. A co-expression network was constructed using a soft threshold power of 12 (Fig. S1). A total of 18 non-gray modules were identified (Fig. 3A). Subsequently, the correlation between module eigengenes and hypoxia and glycolysis scores was assessed. Modules M11 and M13 exhibited significant associations with both the hypoxia and glycolysis scores, with correlation coefficients higher than those of other modules (Fig. 3B). Consequently, M11 and M13 modules were selected as the hypoxia- and glycolysis-associated modules. The 104 genes comprising these two modules were defined as the hypoxia- and glycolysis-associated genes (Table SI).

To identify the critical hypoxia- and glycolysis-associated genes, the Mime framework was employed to construct prognostic models based on the 104 genes. The TCGA-OSCC cohort served as the training cohort, and the GSE65858 dataset was used as the validation cohort. A total of 117 combined models were evaluated (Fig. S2). These combined models were ranked according to the C-index in the validation cohort, and the top 20 models are shown in Fig. 3C. The ‘StepCox(backward) + Enet (α=0.8)’ combined model demonstrated superior performance, achieving the highest C-index in the validation cohort (Fig. 3C). Risk scores were then calculated based on this optimal combined model. In both the training and validation cohorts, patients in the high-risk group exhibited significantly worse RFS (Fig. 3D). Subsequently, core signatures were selected using multiple algorithms to identify four core genes: HSPG2, IGSF3, DUS1L and DYNC1H1 (Fig. 3E and F). To clarify their individual prognostic roles, a Cox regression analysis was performed. The results indicated that HSPG2 and DYNC1H1 were risk factors, whereas IGSF3 and DUS1L were associated with favorable outcomes (Table SII).

Collectively, HSPG2, IGSF3, DUS1L and DYNC1H1 were identified as the critical hypoxia- and glycolysis-associated genes.

To further elucidate the mechanisms regulating the expression of critical hypoxia- and glycolysis-associated genes, the H3K27ac signals at the HSPG2, IGSF3, DUS1L and DYNC1H1 loci in HSC3 and SAS cells were analyzed using the GSE205455 dataset. TEs were identified at the HSPG2, IGSF3 and DUS1L loci in both HSC3 and SAS cells (Fig. 4A). The DYNC1H1 locus contained a TE and an SE in both cell lines (Fig. 4A). Thus, DYNC1H1 was identified as a putative SE-regulated gene, based on the presence of a broad H3K27ac domain.

Subsequently, the H3K27ac peaks were profiled flanking the DYNC1H1 locus in HSC3 and SAS cells using the GSE205455 dataset to annotate the TE and SE regions (Fig. 4B). DYNC1H1 expression was significantly elevated in OSCC tissues compared with that in normal controls (Fig. 4C). To further demonstrate that DYNC1H1 is an SE-regulated gene responsive to hypoxia, H3K27ac enrichment in the TE and SE regions of DYNC1H1 in OSCC cells under normoxic and hypoxic conditions was measured. Under normoxic conditions, significant H3K27ac enrichment was observed in the TE and SE regions (SE1-3; Fig. 4D). Compared with normoxia, hypoxia significantly increased H3K27ac enrichment in these enhancer regions (Fig. 4D). Furthermore, hypoxia upregulated DYNC1H1 mRNA and protein levels in HSC3 and SAS cells compared with hypoxic cells (Fig. 4E and F).

To assess the functional role of DYNC1H1 in the hypoxic response of OSCC cells, DYNC1H1 was knocked down in HSC3 and SAS cells (Fig. 4G and H; Fig. S3A and B). CCK-8 assays revealed that hypoxia significantly promoted the proliferation of HSC3 and SAS cells, whereas this effect was counteracted by DYNC1H1 KD, with reversal rates of 129% in HSC3 cells and 125% in SAS cells (Fig. 4I). Transwell assays revealed that hypoxia markedly enhanced the invasive capacity of OSCC cells (Fig. 4J). By contrast, hypoxia failed to promote cell invasion upon DYNC1H1 KD (Fig. 4J). The rescue rate of DYNC1H1 KD in hypoxia-induced invasion was 99% in both HSC3 and SAS cells (Fig. 4J).

Glycolysis is a critical metabolic pathway in tumor cells under hypoxia, characterized by increased glucose uptake and lactate production (10,13). In the present study, glucose uptake and lactate generation under normoxic and hypoxic conditions was measured. Compared with normoxic conditions, hypoxia significantly increased glucose uptake, an effect that was partially reversed by DYNC1H1 KD, with reversal rates of 93% in HSC3 cells and 91% in SAS cells (Fig. 4K). Consistently, hypoxia significantly upregulated l-lactate production, which was partially counteracted by DYNC1H1 KD, with reversal efficiencies of 73% in HSC3 cells and 76% in SAS cells (Fig. 4L).

Collectively, DYNC1H1 was identified as a hypoxia-responsive SE-regulated gene, and its KD attenuated hypoxia-induced oncogenic phenotypes in OSCC.

To identify hypoxia-responsive transcription factors that potentially regulate DYNC1H1, potential transcription factors binding to the enhancer regions of DYNC1H1 were predicted, with the top 20 shown in Fig. 5A. The correlation between these transcription factors and DYNC1H1 expression using bulk RNA-seq data from the TCGA-OSCC dataset was then analyzed. Since FAIRE was not included in the TCGA-OSCC dataset, the remaining 19 transcription factors were analyzed. Among these, 12 transcription factors (MED1, POLR2A, SMARCC2, FOSL2, POLR2B, XRN2, DCP1A, ASXL1, RNF2, PPARD, RBPJ and MAX) were significantly positively correlated with DYNC1H1 expression (Fig. 5B). Thus, these 12 transcription factors were prioritized for further validation based on the strength of their correlation with DYNC1H1 expression.

The mRNA levels of 12 transcription factors in HSC3 and SAS cells were assessed under normoxic and hypoxic conditions. The transcription of MED1, POLR2A, SMARCC2, POLR2B, DCP1A, RBPJ and MAX remained hypoxia-insensitive in both cell lines (Fig. 5C). XRN2 and RNF2 transcription was significantly elevated in HSC3 cells under hypoxia but exhibited no significant changes in SAS cells (Fig. 5C). Hypoxia significantly downregulated PPARD transcription in SAS cells but not in HSC3 cells (Fig. 5C). Hypoxia significantly upregulated FOSL2 and ASXL1 transcription in both cell lines (Fig. 5C). Furthermore, hypoxia increased the FOSL2 protein level in both HSC3 and SAS cells (Fig. 5D). Hypoxia elevated the ASXL1 protein level in SAS cells but downregulated it in HSC3 cells (Fig. 5D).

To further validate FOSL2-driven transcriptional activation of DYNC1H1, ChIP-qPCR was performed to assess FOSL2 occupancy in the TE and SE regions of DYNC1H1 in OSCC cells under normoxic and hypoxic conditions. Compared with normoxia, hypoxia significantly increased FOSL2 enrichment in the TE and SE regions of DYNC1H1 in both HSC3 and SAS cells (Fig. 5E). Subsequently, FOSL2 was overexpressed in HSC3 and SAS cells (Fig. 5F and G; Fig. S3G and H). FOSL2 OE upregulated DYNC1H1 mRNA and protein levels under normoxic conditions (Fig. 5H and I).

Collectively, FOSL2 was characterized as a hypoxia-responsive transcription factor that activates DYNC1H1 transcription by binding to its TE and SE regions.

To investigate the functional role of the FOSL2/DYNC1H1 axis under hypoxic conditions, FOSL2 KD and DYNC1H1 OE in HSC3 and SAS cells was performed (Fig. S3C-F). Furthermore, it was confirmed that under normoxic conditions, FOSL2 was effectively downregulated in the FOSL2 KD cells, while DYNC1H1 expression was upregulated in the DYNC1H1 OE group compared with their respective controls (Fig. 6A-D). CCK-8 assay demonstrated that the KD of FOSL2 completely abolished hypoxia-induced proliferation, with an elimination efficiency of 128% in both cell lines (Fig. 6E). The anti-proliferative effect of FOSL2 KD was partially rescued by DYNC1H1 OE under hypoxia, with rescue rates of 75% in HSC3 cells and 77% in SAS cells (Fig. 6E). Transwell assays revealed that the KD of FOSL2 completely suppressed hypoxia-induced invasion in both cell lines, with 100% inhibition in both cell lines (Fig. 6F). DYNC1H1 OE under hypoxia led to a near-complete rescue of this inhibitory effect, with rescue rates of 98% in HSC3 cells and 104% in SAS cells (Fig. 6F).

The glucose uptake and lactate production in HSC3 and SAS cells was further quantified. FOSL2 KD nearly completely attenuated hypoxia-stimulated glucose uptake, with inhibition rates of 99% in HSC3 cells and 95% in SAS cells (Fig. 6G). DYNC1H1 OE almost completely rescued the inhibitory effect of FOSL2 KD on glucose uptake under hypoxic conditions, with a rescue rate of 96% in both cell lines (Fig. 6G). Similarly, FOSL2 KD partially reduced hypoxia-stimulated lactate generation, with inhibition rates of 88% in HSC3 cells and 84% in SAS cells (Fig. 6H). Under hypoxic conditions, the reduction in lactate generation caused by FOSL2 KD was partially reversed by DYNC1H1 OE, with reversal rates of 60% in HSC3 cells and 64% in SAS cells (Fig. 6H).

Collectively, these results indicated that the FOSL2/DYNC1H1 axis drove hypoxia-induced proliferation, invasion, glucose uptake and lactate production in OSCC cells.