Genomic Characterization of Cervical Tumors Driven by Rare High-Risk HPV Subtypes Utilizing a Deep Learning Framework

Abstract

As many as 30-50% of patients with locally advanced cervical cancer (LACC) experience recurrence after chemoradiation therapy (CRT). More sophisticated prognostic schemes and targeted therapeutic options are urgently needed. Human papillomavirus (HPV) is responsible for the majority of invasive cervical carcinomas, most commonly HPV 16 and 18 subtypes. In large part due to their rarity, little is known about the biologic behavior of LACC driven by non-16/18 HPV genotypes. We hypothesized that rare HPV subtypes drive distinct gene expression patterns. To aid in the common dilemma of limited sample procurement, we developed a tool which leverages population data and generative adversarial neural network (GAN) modelling to exploit patterns of gene expression of small phenotypic groups for downstream analysis. Here, we employed a deep learning approach to identify gene expression patterns unique to tumors caused by rare HPV subtypes. Whole transcriptome analysis of tumor biopsies from an institutional LACC tumor bank were employed including 130 HPV16/18 positive, and at least 3 each of rare HPV subtypes 31, 33, 39, 45, 52, 58, and 59. A Wasserstein GAN model (WGAN) was trained for each rare HPV subtype, and used to generate augmented rare HPV population data for downstream differential expression testing between rare HPV subtypes and HPV16/18. K-means clustering analysis was performed using R stats package. Differential gene expression testing between rare HPV subtype populations was performed using the EdgeR R-package. Pathway Enrichment analysis was conducted using EnrichR software. By minimizing intergroup clustering distance, 3 clusters of HPV driven disease were demonstrated. The first cluster demonstrated biology similar to HPV16 and HPV18. Namely, pathway enrichment analysis performed on lists of differentially expressed genes between GAN augmented rare HPV subtypes (HPV45, HPV52, and HPV58) and classical HPV subtypes 16 and 18, yielded no significant statistical differences. The second cluster demonstrated enriched gene expression in proinflammatory IL-1 and IL-6 family cytokines and focal adhesion signaling (HPV31, HPV33, HPV39) (IL-1 regulation of extracellular matrix, p = 5.0e-6; Oncostatin M signaling, p = 2.0e-7; Integrin Signaling, 2.5e-7). The last cluster demonstrated an increased Th2/myeloid signature (HPV59) (T helper cell surface molecules 1.2e-7, myeloid cell signature 8.5e-9). Utilizing a deep learning architecture to model gene expression of rare HPV subtypes we demonstrate 3 groups of HPV driven cervical cancer biology. These groups include a classical subtype, a high IL-1/IL-6 high subtype, and a myeloid associated subtype. Knowledge of such tumor heterogeneity can aid in the future development of targeted treatment plans and patient prognosis.