Cancer’s biological behavior was exceedingly complex, driven by the nonlinear interplay of thousands of genes. When observations were restricted, such as in treatment-resistant HPV+ head and neck squamous cell cancer, this intricacy had become a key barrier to deriving mechanistic findings. Deep learning approaches like generative adversarial neural networks (GANs) have shown promise as reliable solutions for nonimage data augmentation. For a study, researchers sought to determine if using different GANs to highlight gene expression variations between treatment susceptible (TS) and treatment-resistant (TR) HPV+ head and neck squamous cell carcinoma was useful. On 17 patients with HPV-positive oropharyngeal squamous cell carcinoma (OPSCC), investigators performed RNAseq. TR patients were characterized as those who had locally progressed OPSCC at the time of presentation and recurred quickly after receiving standard of care therapy (n=11).
In most cases, patients did not react to salvage therapy and died quickly. On the other hand, TS patients were classified as those who presented with locally advanced OPSCC, had the standard of care therapy, and were followed for five years without recurrence (n=6). These two cohorts were utilized for training two different GANs, producing 500 strong synthetic samples of each phenotype for downstream analysis. Pathway enrichment analysis (PAE) was hampered because only 100 differentially expressed (DE) genes passed FDR cutoffs in 17 datasets. Between synthetic samples, DE testing generated 1,196 DE genes. PAE indicates these genes correspond to coherent pathways, confirming the deep learning technique. The VEGFA-VEGFR2 pathway, in particular, was significantly enriched in the TR group (P=1.5×10-4). The TS population was also enriched for the NKT cell signature (P=3.0×10-62) and the gene library of T-cell receptor stimulation of apoptosis (P=5.4×10-9). Finally, when comparing the full cohort of renal papillary n=534 vs. clear renal cell n=295 samples, DE testing on GAN augmented cohorts of renal papillary vs. renal clear cell samples from The Cancer Genome Atlas, trained on 10 samples per condition, reliably predicted DE genes with an accuracy of 93% and F1 score of 88%. To the best of the study groups’ knowledge, this was the first study to use independent GANs to boost structural gene expression alterations in unusual clinical circumstances like treatment-resistant HPV+ OPSCC. This approach revealed differences in angiogenesis and T-cell activity between HPV+ OPSCC that were therapy responsive and resistant.