Network-based approach identifies key genes associated with tumor heterogeneity in HPV positive and negative head and neck cancer patients

Abstract

Head and Neck Squamous Cell Carcinoma (HNSCC) is the seventh most prevalent cancer worldwide and is classified as human papillomavirus (HPV) positive or negative. Substantial heterogeneity has been observed in the two groups, posing a significant clinical challenge. In the disease context, global transcriptional changes are likely driven by a few key genes that reflect the disease etiology more accurately compared to differentially expressed genes (DEGs). We implemented our network-based tool PathExt on 501 TCGA-HNSCC samples (64 HPV positive & 437 HPV negative) to characterize central genes in two subtypes, where in subtype 1, HPV-positive samples were considered as cases and negative as controls, and vice versa in subtype 2. We also identified DEGs and performed several analyses on multiple benchmarking datasets to compare the biology of central genes with DEGs. PathExt key genes performed better with respect to DEGs in both subtypes in recapitulating disease etiology. Gene ontology analysis using central genes revealed shared biological processes such as "epithelial cell proliferation" as well as subtype-specific processes (immune- and metabolic-related processes in subtype 1 and peptide-related processes in subtype 2). However, in the case of DEGs, no subtype-specific processes were seen. Additionally, PathExt central genes did better than DEGs on external validation datasets that were specific to HNSCC and included HNSCC-specific cancer driver genes, FDA-approved therapeutic targets, and pan-cancer tumor suppressor genes. Unlike DEGs, central genes exhibit significant expression in various cell types, enrichment for cancer hallmarks, and mutated protein systems. Central gene expression-based machine learning model shows better performance than DEGs in classifying responders/non-responders with 0.74 AUROC. Lastly, the top 10 potential therapeutic targets and drugs were proposed. Overall, we observed PathExt as a complementary approach to DEGs, characterizing common and HNSCC subtype-specific key genes associated with distinct HNSCC molecular subtypes.

Keywords: Cancer hallmarks; Differentially expressed genes; Head and neck cancer; Human papilloma virus; Network-based approach; Tumor heterogeneity.

Fig. 1

Workflow of the current study. (A) The figure represents the overall approach implemented in this study to perform comparative study between PathExt and DEGs based approaches. (B) A flowchart of PathExt pipeline used to characterize key genes from each sample is provided.

Fig. 2

PathExt capture biologically relevant genes. (A) Venn diagram of top100 PathExt central genes and DEGs characterized in subtype 1 and subtype 2 showing common and unique genes; (B) Top20 enriched significant biological processes associated with common 17 PathExt central genes among subtype 1 and 2; (C) Log fold change comparison of gene expression (case compared to control) for the top100 PathExt central genes and DEGs characterized in subtype 1; (D) Log fold change comparison of gene expression (case compared to control) for the top100 PathExt central genes and DEGs characterized in subtype 2.

Fig. 3

PathExt central genes shows distinct processes in HNSCC subtypes. ClusterProfiler 4.0 package was used to obtain significant enriched terms associated with the top100 PathExt key genes and DEGs. Here, we have shown top20 enriched biological processes associated with (A) top100 PathExt central genes in subtype 1; (B) top100 PathExt central genes in subtype 2; (C) top100 upregulated DEGs in subtype 1; (D) top100 upregulated DEGs in subtype 2. Here “Gene Ratio” is the fraction of the input genes in the GO term and ‘Count’ represents the number of key genes common in the gene list of the given term. Only significant processes (FDR ≤0.05) are shown here.

Fig. 4

Comparative performance of PathExt and DEGs on cancer hallmarks and external datasets. (A) Fisher’s odds ratio showing the ability of subtype 1 top100 PathExt central genes and upregulated DEGs in recapitulating HNSCC specific PathExt central genes in multiple external datasets. (B) Fisher’s odds ratio showing the ability of subtype 2 top100 PathExt central genes and upregulated DEGs in recapitulating HNSCC specific PathExt central genes in multiple external datasets. (C) Enriched cancer hallmarks associated with subtype 1 top100 PathExt central genes. (D) Enriched cancer hallmarks associated with subtype 1 top100 upregulated DEGs. (E) Enriched cancer hallmarks associated with subtype 2 top100 PathExt central genes. (F) Enriched cancer hallmarks associated with subtype 2 top100 upregulated DEGs.

Fig. 5

Single cell analysis. PathExt top100 central genes shows higher expression in cell types compared to upregulated top100 DEGs in (A) Subtype 1 and (B) Subtype 2.

Fig. 6

PathExt classifies responder and non-responder with high accuracy. Performance of machine learning models on independent dataset by (A) Subtype 2 PathExt central genes and (B) Subtype 2 DEGs. Performance measure is in the terms of AUROC.