HNSCC Treatment Response Biomarkers

What this research found

Head and neck squamous cell carcinoma (HNSCC) responds unpredictably to treatment, so tumour expression profiles from 270 patients were searched for a signature separating those whose disease progressed from those whose did not. No single gene survived multiple-testing correction — 0 of 31,330 genes reached a false discovery rate below 0.05 — yet a random forest built on the 20 top-ranked genes classified held-out patients with a ROC-AUC of 0.7558. Whatever signal exists here is combinatorial and weak rather than attributable to any individual marker.

• No individual gene distinguished the two outcome groups after correction. Zero of 31,330 genes reached a false discovery rate below 0.05, 563 (1.8%) passed an uncorrected p-value of 0.01, and fold changes among the top 20 spanned only 0.03 to 0.34 on a log2 scale.
• A random forest on those 20 genes reached a ROC-AUC of 0.7558, accuracy of 0.6852, sensitivity of 0.7037 and specificity of 0.6667 on 54 held-out patients, correctly calling 19 of 27 who progressed and 18 of 27 who did not.
• Cross-validated performance was lower than the held-out figure at a ROC-AUC of 0.7019 ± 0.0550, and the gap against a mean training score of 0.9968 indicates the model overfits its training data.
• The three most informative features were LOC645946, an uncharacterised locus, at an importance of 0.078; SCYE1, also known as EMAP-II, a pro-inflammatory cytokine with context-dependent effects on tumour angiogenesis, at 0.071; and C9ORF126, now annotated as NMRK1, an enzyme in NAD+ salvage biosynthesis, at 0.065.
• Outcome labels were derived from progression-free survival rather than radiological assessment: 137 patients with no progression event during follow-up were treated as complete responders and 133 with a progression event as progressive disease, a near-even 50.7% to 49.3% split.

How it was done

Head and neck cancer expression data was retrieved from the Gene Expression Omnibus under accession GSE65858 — 270 tumours across 31,330 genes, already log2-normalised. No India-specific HNSCC dataset carrying treatment response data was found in the repository, so this general cohort was used instead. Patients were labelled by progression-free survival event, Welch's t-test was run across every gene with Benjamini-Hochberg correction, and the 20 genes with the lowest raw p-values were carried forward as model features. A random forest was tuned by grid search over 24 hyperparameter combinations under 5-fold cross-validation, then evaluated on a held-out set of 54 patients, and the highest-ranked features were annotated for biological plausibility. Supporting outputs included a volcano plot, a hierarchically clustered heatmap of the 20 genes across all 270 samples, and a written paper.

Data sources

• Gene Expression Omnibus accession GSE65858 — 270 head and neck cancer tumours profiled across 31,330 genes on the Illumina HumanHT-12 V4.0 platform (GPL10558), with 137 non-progressors and 133 progressors

Figures from this analysis

Outputs produced

• clinical_data.csv

  Clinical metadata for 270 HNSCC samples with CR/PD classification

  0.081 MB

• expression_data.csv

  Normalized gene expression matrix (log2-transformed)

  57 MB

• dataset_summary.txt

  Comprehensive dataset summary and validation

• dea_results.csv

  Complete differential expression analysis results for all genes

  4.1 MB

• top_20_biomarkers.csv

  Top 20 candidate biomarker genes for ML modeling

• best_model.joblib

  Trained Random Forest classifier (optimized via grid search)

• X_train.csv

  Training features (20 biomarker genes)

• y_train.csv

  Training labels (CR=1, PD=0)