AI and the Vaginal Microbiome: A New Frontier in Cervical Cancer Prediction

Cervical cancer screening today relies almost entirely on two data points: HPV genotype and cytology. A new perspective review, published in Frontiers in Network Physiology and co-authored by Daye's Michelle Gomes alongside researchers from UCL, Queen Mary University of London, and LSHTM, argues that a third signal — the vaginal microbiome — could meaningfully sharpen who actually needs a colposcopy referral and who doesn't.

Why HPV-positive isn't the whole story

Most HPV infections clear on their own within one to two years. The problem is that current screening pathways can't reliably tell, among HPV-positive women, who will clear the infection and who will progress toward high-grade lesions and cancer. In England alone, the NHS Cervical Screening Programme issues roughly five million invitations a year, generating large volumes of follow-up tests and colposcopy referrals for a comparatively small number of high-grade cases.

What the microbiome adds

The review synthesises a growing body of evidence that the composition of the vaginal microbiome is itself a risk signal. Communities dominated by Lactobacillus crispatus are consistently protective, supporting an acidic pH and a healthy mucosal barrier. Communities lacking Lactobacillus dominance, or dominated by the more unstable Lactobacillus iners, are linked to higher rates of HPV persistence and progression to cervical intraepithelial neoplasia (CIN2/3). Despite this evidence, no validated cervical cancer risk model currently incorporates microbiome data.

“This perspective review examines artificial intelligence approaches for cervical cancer prediction and evaluates the emerging role of the vaginal microbiome as a complementary biomarker within these interconnected physiological networks.”

Gomes et al., Frontiers in Network Physiology (2026)

Where AI fits in

Existing AI models for cervical cancer fall into three buckets: screening and triage, diagnostic classification, and prognosis. Tabular models — logistic regression, random forests, gradient boosting — dominate when the inputs are demographic and clinical variables, while imaging-based models rely on convolutional neural networks to read Pap smears and colposcopy images. The authors' review of the published literature found that virtually all of these models rely on demographic, clinical, or imaging inputs alone, with no validated cervical cancer risk calculator yet built on microbiome data.

Their proposed fix isn't a new AI paradigm, but an added feature set: community state type, relative abundance of key taxa, and simple ecological markers like vaginal pH and diversity, layered onto the clinical and HPV-genotype variables that existing models already use. For tabular models this means new inputs to the same logistic regression or gradient-boosting pipeline; for imaging models, it means extending the network with additional inputs alongside image data.

A pathway that starts with self-sampling

The translational case is where the paper connects most directly to Fern's own thesis. The authors outline a workflow in which self-collected samples — the same category of sample already used for at-home HPV testing — are processed for both HPV genotype and microbiome profile in the same lab step, with results returned on the same timeline as current HPV tests. No new sample collection is required, only an added analytical layer on the specimen already being collected.

That detail matters because it's the same infrastructure Fern-backed Daye has already built and scaled: a clinically validated, at-home diagnostic that removes the friction of an in-clinic visit from cervical screening. A microbiome-informed AI layer wouldn't require reinventing sample collection — it would ride on top of the access model that already exists.

What's still missing

The authors are candid about the gaps: most existing microbiome studies are small, cross-sectional, and drawn from limited, non-diverse cohorts, particularly within the UK. Closing that gap will require large, prospective, multi-ethnic cohorts that track microbiome and HPV status over multiple screening rounds, alongside rigorous bias auditing before any tool reaches NHS screening pathways.

If that evidence base builds out as proposed, the prize is a meaningfully more precise triage system — one that sends the right women to colposcopy sooner, and spares the rest an unnecessary referral. It's exactly the kind of clinically grounded, evidence-first innovation Fern looks for: proof that in women's health, the biggest gains often come from measuring something that was there all along, but never built into the model.