Integrative transcriptomic and single-cell analyses identify Candida-responsive MMP10, CA9, and STC2 candidates in oral squamous cell carcinoma

Background Oral squamous cell carcinoma (OSCC) arises at a microbially exposed mucosa, but the relationship between tumor transcriptional programs and host responses to Candida albicans is uncertain. We sought direction-consistent genes shared with short-term C. albicans stimulation and assessed their diagnostic, cellular, pathway, and tissue contexts. Methods We directionally integrated four paired C. albicans -stimulated/control oral-epithelial experiments (GSE56091) with 40 paired OSCC/adjacent tissues (GSE37991). Four machine-learning methods prioritized candidates, and independent bulk validation used 167 OSCCs and 45 controls (GSE30784). An exploratory ridge score was fixed in the discovery cohort and applied to the external cohort. Cellular localization and state were examined in paired single-cell profiles from seven patients (GSE198315) and an external oral-cavity subset of 10 carcinomas and three normal-mucosa samples (GSE181919). Qualitative PAS and IHC assessment used archival oral tissues from three groups, with four cases per group. Results Candida stimulation yielded 508 exploratory differentially expressed genes, and OSCC yielded 1,850; 26 of 35 overlapping genes changed concordantly. The four-method intersection comprised MMP10 , CA9 , and STC2 ; external single-gene AUCs were 0.968, 0.714, and 0.907, respectively. The three-gene score achieved an external AUC of 0.959 (95% CI 0.935-0.983), but at the discovery-derived cutoff sensitivity was 0.605 and specificity was 1.000; calibration was poor, and discrimination did not exceed MMP10 alone (DeLong P = 0.365). In GSE198315, paired epithelial pseudobulk supported MMP10 and CA9 under prespecified targeted three-gene correction (FDR = 0.0455 for both), whereas STC2 was directional but nonsignificant (FDR = 0.0665). Genome-wide FDR was approximately 0.892 for each gene, so the single-cell findings support localization and directional consistency rather than definitive genome-wide differential expression. The q95 calls showed modest overlap with CopyKAT and inferCNV (Jaccard=0.141 and 0.254, respectively). MMP10 and CA9 showed stronger pooled descriptive localization than STC2 ; CA9 remained hypoxia-associated. Fungal-element-compatible PAS staining provided morphological context but could neither identify species nor quantify burden. Conclusions MMP10 and CA9 had the strongest multi-level support as direction-consistent Candida -responsive OSCC candidates, while STC2 remains exploratory. The three-gene combination is not yet a calibrated clinical model, and the 26-gene overlap defines a shared host-response pattern rather than evidence that C. albicans causes the OSCC transcriptional state.