In exploratory study PNK001 of women undergoing hysterectomy, RNA-based machine learning found significant differential gene expression and higher somatic variant counts in vaginal swabs (and other lower-genital-tract samples) from endometrial cancer cases, suggesting a non-invasive path to rule out malignancy. Classifiers using expressed genes and variant counts distinguished 5 benign from 15 malignant cases across cytobrush/ectocervical/vaginal samples with cross-validated ROC AUCs ranging 0.6–0.96.

Why It Matters To Your Practice

• Today’s workup to exclude endometrial cancer in symptomatic patients often escalates to invasive sampling; a vaginal swab approach could shift the first-line pathway toward non-invasive triage.

• If validated, a high-sensitivity, high–negative predictive value swab test could reduce unnecessary procedures, speed reassurance, and focus resources on higher-risk patients.

Clinical Implications

• Signal was detectable not only in endometrial tissue but also in endocervical cytobrush, ectocervical, and vaginal swab RNA—supporting the feasibility of “downstream” sampling for a uterine malignancy.

• Performance varied by sample type and model: tissue-only classifiers reached ROC AUC 0.97–0.98 on an independent test set, while lower-genital-tract sample classifiers ranged from 0.6–0.96 in cross-validation—highlighting where prospective validation must focus.

Insights

• The approach combines transcriptomics (differential expression) with expressed somatic variant counts—an example of multimodal features that AI can integrate better than traditional rules-based screening.

• Study design linked model labels to surgical pathology, aligning ML development with the clinical reference standard clinicians already use for decision-making.

The Bottom Line

• This early evidence suggests vaginal swabs can carry enough RNA signal for AI-enabled detection of endometrial cancer; the investigators selected vaginal swabs for the follow-on validation study (PNK002).