Abstract
Traditional analysis of genomic data from bulk sequencing experiments seek to group and compare sample cohorts into biologically meaningful groups. To accomplish this task, large scale databases of patient-derived samples, like that of TCGA, have been established, giving the ability to interrogate multiple data modalities per tumor. We have developed a computational strategy employing multimodal integration paired with spectral clustering and modern dimension reduction techniques such as PHATE to provide a more robust method for cancer sub-type classification. Using this integrated approach, we have examined 514 Head and Neck Squamous Carcinoma (HNSC) tumor samples from TCGA across gene-expression, DNA-methylation, and microbiome data modalities. We show that these approaches, primarily developed for single-cell sequencing can be efficiently applied to bulk tumor sequencing data. Our multimodal analysis captures the dynamic heterogeneity, identifies new and refines subtypes of HNSC, and orders tumor samples along well-defined cellular trajectories. Collectively, these results showcase the inherent molecular complexity of tumors and offer insights into carcinogenesis and importance of targeted therapy. Computational techniques as highlighted in our study provide an organic and powerful approach to identify granular patterns in large and noisy datasets that may otherwise be overlooked.