Whole-transcriptome sequencing and machine learning detect molecular signatures of endometrial cancer in non-invasive vaginal swabs

OBJECTIVE: Current tissue-based methods for ruling out endometrial cancer in symptomatic women are highly invasive. We explored whether non-invasive vaginal samples could be used to detect endometrial cancer. METHODS: Women undergoing hysterectomy were enrolled in an exploratory study, PNK001, wherein vaginal swabs, ectocervical swabs, endocervical cytobrushes, and endometrial tissue were obtained. Additionally, we secured RNA sequencing data from The Cancer Genome Atlas and endometrial tissues from the Cooperative Human Tissue Network. We generated sequencing data from swab, cytobrush, and tissue samples representing 27 PNK001 participants and 46 Cooperative Human Tissue Network samples. We analyzed differential expression, cell type signatures, and expressed somatic variants. We performed machine learning analyses on tissue samples and data from 20 PNK001 participants using expressed features and surgical pathology as reference labels. RESULTS: Machine learning classifiers trained on and applied to tissue samples achieved receiver operating characteristic area under the curve values of 0.97 and 0.98 on an independent test set. Significant differential gene expression and elevated somatic variant counts were present in endocervical cytobrush samples, ectocervical swabs, and vaginal swabs from women with endometrial cancer. Classifiers trained using expressed genes and variant counts distinguished 5 benign from 15 malignant cases in cytobrush, ectocervical, and vaginal swab samples, with average receiver operating characteristic area under the curve values of 0.6 to 0.96 in cross-validation. CONCLUSIONS: Vaginal swabs provided sufficient signal to detect endometrial cancer using RNA-based machine learning. We therefore selected the vaginal swab as the primary sample type for a second study (PNK002) to develop and validate a test for endometrial cancer. A non-invasive vaginal swab test with high sensitivity and negative predictive value could potentially rule out cancer in symptomatic women instead of invasive workup.