Transcriptomic Profiling of Oral Squamous Cell Carcinoma

A reproducible reanalysis of paired tumor and normal RNA-seq from patients with oral squamous cell carcinoma, checked against an independent cohort and explorable in an interactive Streamlit app.

Exploratory research — no clinical claims

Exploratory associations from public data. These are not validated biomarkers, and the project makes no diagnostic or clinical claims.

Up / down in tumor (discovery): 363 / 976
Replicated in an independent cohort: 706 of 1,339
Matched tumor–normal pairs: 3 + 15

Volcano plot of paired differential expression in GSE20116: 976 downregulated and 363 upregulated gene representatives in tumor versus normal tissue, with PTHLH, LAMC2, COL4A6 and MMP11 labelled among upregulated genes and TMPRSS11B, PTGFR, PYGM and CRNN among downregulated genes. — FIG. 02.1 — Paired differential expression · GSE20116 · Figure from the project repository

01Problem

Oral squamous cell carcinoma accounts for most oral cancers, and dental teams are often the first to see a suspicious lesion. This project asks what differs at the level of gene expression between OSCC and matched normal oral tissue from the same patients — and whether those differences hold up in a separate cohort.

02Question

Which gene representatives and pathways differ between OSCC and matched normal oral tissue?

03Data

Discovery: GEO GSE20116 — six RNA-seq samples from three matched patients, using the raw count columns from the original publication's supplementary table (Tuch et al., 2010). Validation: GEO GSE184616 — deposited raw gene counts for 15 matched tumor / adjacent-normal pairs from a separate, HPV-negative cohort.

04Methods

Checked sample identities and patient pairing against GEO SOFT files, with checksum-locked downloads and a provenance manifest.
Selected one gene representative per symbol, independently of fold changes, and filtered low expression (10,541 representatives retained).
Paired PyDESeq2 negative-binomial model (~ patient_id + condition) with median-of-ratios normalization, convergence auditing and Benjamini–Hochberg correction.
Hallmark gene-set enrichment: preranked GSEA with GSEApy, plus over-representation analysis against the tested background.
Independent validation: discovery candidates were frozen before fitting the same paired model to GSE184616 (15 pairs, 14 residual degrees of freedom).
Publication figures, an executable notebook, automated tests, and a Streamlit / Plotly research explorer.

05Tools

Python
PyDESeq2
GSEApy
pandas
Streamlit
Plotly
pytest

06Visualizations

Volcano plot of tumor versus normal differential expression in GSE20116 with up- and downregulated genes highlighted. — FIG. 02.1.1Figure from the project repository ·Volcano plot of paired PyDESeq2 Wald tests, tumor versus normal (BH-adjusted p < 0.05, |log2 fold change| > 1).

Paired expression plots showing normal-to-tumor changes for each patient across four candidate genes. — FIG. 02.1.2Figure from the project repository ·Individual patient trajectories for selected candidates; points are log-normalized expression, with fold changes and q-values from the paired model.

Bar chart of the lowest-FDR Hallmark gene sets by normalized enrichment score, including E2F Targets and Myogenesis. — FIG. 02.1.3Figure from the project repository ·Top Hallmark GSEA terms; positive NES leans toward tumor. Enrichment does not establish pathway activation or causality.

Clustered heatmap of the top differentially expressed genes across tumor and normal samples, annotated by condition and patient. — FIG. 02.1.4Figure from the project repository ·Top 30 genes by adjusted p-value (row z-scores, clustered). This view is descriptive, not independent evidence of separation.

07Findings

Discovery (q < 0.05, |log2 fold change| > 1): 363 upregulated and 976 downregulated gene representatives out of 10,541; 35 features were excluded after an optimizer convergence failure.
Leading upregulated representatives include PTHLH, LAMC2 and COL4A6; leading downregulated include TMPRSS11B, PTGFR and PYGM.
Hallmark GSEA placed E2F Targets toward tumor (NES 2.61) and Myogenesis toward normal tissue (NES −2.57).
Independent cohort (GSE184616): 1,187 of 1,339 frozen candidates were testable; 964 had the same direction (81.2%) and 706 replicated at genome-wide q < 0.05.

08Limitations

Three discovery patients leave only two residual degrees of freedom, so dispersion estimates are uncertain.
Historical SOLiD / hg18 processing and older gene symbols limit coverage; discovery and validation use different quantification units.
Tumor heterogeneity, tissue composition and field effects in normal margins can drive expression differences.
This replicates differential-expression associations — it is not a diagnostic classifier, prognostic model, causal mechanism or validated biomarker.
No qPCR or wet-lab confirmation was performed.