Papillary thyroid carcinoma (PTC) represents a heterogeneous disease with diverse clinical outcomes highlighting a need to identify robust biomarkers with clinical relevance. We applied non-negative matrix factorization-based deconvolution to publicly available gene expression profiles of thyroid cancers in the Cancer Genome Atlas (TCGA) consortium. Among three metagene signatures identified, two signatures were enriched in canonical BRAF-like and RAS-like thyroid cancers with up-regulation of genes involved in oxidative phosphorylation and cell adhesions, respectively. The third metagene signature representing up-regulation of immune-related genes further segregated BRAF-like and RAS-like PTCs into their respective subgroups of immunoreactive (IR) and immunodeficient (ID), respectively. BRAF-IR PTCs showed enrichment of tumor infiltrating immune cells, tall cell variant PTC, and shorter recurrence-free survival compared to BRAF-ID PTCs. RAS-IR and RAS-ID PTC subtypes included majority of normal thyroid tissues and follicular variant PTC, respectively. Immunopathological features of PTC subtypes such as immune cell fraction, repertoire of T cell receptors, cytolytic activity, and expression level of immune checkpoints such as and PD-L1 and CTLA-4 were consistently observed in two different cohorts. Taken together, an immune-related metagene signature can classify PTCs into four molecular subtypes, featuring the distinct histologic type, genetic and transcriptional alterations, and potential clinical significance.

Mutation hotspots are either solitary amino acid residues or stretches of amino acids that show elevated mutation frequency in cancer-related genes, but their prevalence and biological relevance are not completely understood. Here, we developed a Smith-Waterman algorithm-based mutation hotspot discovery method, MutClustSW, to identify mutation hotspots of either single or clustered amino acid residues. We identified 181 missense mutation hotspots from COSMIC and TCGA mutation databases. In addition to 77 single amino acid residue hotspots (42.5%) including well-known mutation hotspots such as IDH1 (p.R132) and BRAF (p.V600), we identified 104 mutation hotspots (57.5%) as clusters or stretches of multiple amino acids, and the hotspots on MUC2, EPPK1, KMT2C, and TP53 were larger than 50 amino acids. Twelve of 27 nonsense mutation hotspots (44.4%) were observed in four cancer-related genes, TP53, ARID1A, CDKN2A, and PTEN, suggesting that truncating mutations on some tumor suppressor genes are not randomly distributed as previously assumed. We also show that hotspot mutations have higher mutation allele frequency than non-hotspots, and the hotspot information can be used to prioritize the cancer drivers. Together, the proposed algorithm and the mutation hotspot information can serve as valuable resources in the selection of functional driver mutations and associated genes.