Replicator framework reveals how vaginal microbiome defends against pathogens

In a development that is drawing attention across the fields of microbiology, ecology, and women’s health, researchers at the University of Lisbon and the Universitat Politècnica de Catalunya have unveiled a new mathematical model that predicts bacterial vaginosis with an accuracy of 94 percent, matching sophisticated machine-learning systems while doing something those systems cannot: explaining why the prediction works in ecological terms. The study, published in the journal Microbiome, replaces the black-box logic of modern artificial intelligence with a transparent framework rooted in classical population biology, and in doing so offers what may become a new way of thinking about how microbial communities defend the human body against disease.

The work centers on the vaginal microbiome, a delicate ecosystem whose composition has long been known to influence reproductive and urogenital health. A community dominated by Lactobacillus bacteria is generally associated with a healthy state, producing lactic acid that maintains an acidic environment hostile to many pathogens. When that dominance erodes, the community shifts toward a more diverse mixture of anaerobic bacteria, a condition clinically diagnosed as bacterial vaginosis, or BV. BV is the most common vaginal condition in women of reproductive age, and its consequences extend well beyond discomfort: it increases susceptibility to sexually transmitted infections, including HIV, and is linked to yeast infections, urinary tract infections, and adverse pregnancy outcomes.

Clinicians have long relied on the Nugent score, a microscope-based grading system of vaginal smears, to classify samples as healthy, intermediate, or BV-positive. And in recent years, machine-learning classifiers have become remarkably good at predicting those scores from DNA-sequencing data that catalog which bacteria are present. The problem, the researchers argue, is that a prediction is not an explanation. A neural network may flag a community as diseased with high confidence, but it cannot say which ecological forces produced the breakdown. “Machine-learning approaches can accurately predict BV from microbiota profiles, but they provide limited biological insight into the mechanisms linking community structure to clinical outcomes,” the authors note in their paper.

The team, composed of Tomás Ferreira Amaro Freire, Marina Garcia-Romero, and Erida Gjini, set out to close that gap with a mechanistic model built on replicator dynamics, a mathematical framework originally developed in evolutionary game theory to describe how the relative frequencies of competing strategies or species change over time. In the replicator formulation, each microbial taxon grows at a rate determined by its own intrinsic fitness and by its interactions with every other taxon in the community. Species that fare better than the community average increase in frequency; those that fare worse decline. The mathematics captures, in a compact set of equations, the essence of ecological competition, cooperation, and inhibition.

What makes the new study distinctive is the way the team adapted this classical framework to the specific biology of the vagina. Rather than treating all bacterial species as equals in the model, they made a deliberate representational choice: the four vaginally relevant Lactobacillus species were retained as separate entities in the equations, while the remaining taxa were grouped into broader taxonomic units. This design decision reflected the biology. The four Lactobacillus species play distinct protective roles, producing different profiles of lactic acid and other metabolites, and collapsing them into a single category would have erased exactly the information that matters most for clinical outcomes. The remaining organisms, which include the diverse anaerobes associated with BV, were represented at coarser resolution without sacrificing predictive power.

Fitted to microbiota data from a previously published cohort of 394 North American women, the model achieved something that surprised even its developers: it predicted BV status with 94 percent accuracy, matching the performance benchmarks set by machine-learning methods while remaining fully interpretable. Every parameter in the model corresponds to a biological quantity — a growth rate, an interaction strength, an invasion resistance threshold — so researchers can read the equations and understand which organisms are protecting the community and which are undermining it.

Central to the analysis is the concept of invasion resistance, an ecological property describing how well an established community repels a newcomer. In the vaginal context, the “invaders” of interest are pathobionts — organisms that live harmlessly at low abundance but can proliferate into disease when given the opportunity. The replicator framework allowed the team to quantify, for each community composition, how resistant that composition would be to invasion, and to decompose that resistance into contributions from individual taxa and from nonlinear effects arising from pairwise interactions between species. This revealed that community protection is not simply the sum of its parts. Two organisms that individually contribute modestly to stability can, through their interaction, produce effects on invasion resistance that are either synergistic or antagonistic, and these nonlinear terms turned out to be essential for accurate prediction of clinical state.

The framework also clarified the ecological pathways separating healthy, intermediate, and BV-positive states. Healthy communities dominated by protective Lactobacillus species occupy a stable equilibrium from which pathobionts cannot invade; the models show this stability emerging from the strong growth and interaction structure of the lactobacilli themselves. Intermediate communities sit near a tipping region of the dynamics, where modest perturbations can tip the system toward BV. BV-positive states represent a fundamentally reorganized community, one in which the previously dominant protectors have been displaced and the interaction network has been rewired around a different set of dominant organisms. In this sense, bacterial vaginosis is not merely the absence of lactobacilli but an alternative stable configuration of the ecosystem — a perspective with direct implications for treatment, since simply adding lactobacilli back may not suffice if the ecological conditions that sustain them have changed.

The implications reach well beyond the clinic. The authors emphasize that their approach demonstrates the broader potential of replicator-based ecological formalisms for modeling microbiota-invader dynamics across host-associated microbial systems. Any body site — the gut, the skin, the oral cavity — hosts a microbial community whose structure modulates invasion by pathogens, and the same mathematics of fitness, interaction, and invasion resistance applies. Where machine learning will likely continue to dominate pure prediction tasks, mechanistic models of this kind offer the interpretability needed to design rational interventions: probiotic cocktails chosen for their interaction properties, not just their identity, or ecological strategies to tip a vulnerable community back into a protective configuration.

The study also carries a methodological message for the scientific community at a moment when data-driven models are sweeping through biomedicine. The comparison here was not a contest between old and new, but a demonstration that the two approaches can converge on the same answer while asking different questions. The machine-learning benchmark confirmed that the signal exists in the data; the replicator model revealed what the signal means. When a mechanistic model can match the predictive performance of a black-box classifier, the argument for choosing interpretability becomes difficult to ignore.

The research was supported by the Fundação para a Ciência e a Tecnologia in Portugal through the Models4Invasion project and related funding, and the authors have made their data processing and modeling code publicly available through a GitHub repository, alongside supplementary datasets and extended validation on synthetic data. The study used only previously published, de-identified data, and the authors report no competing interests.

For a condition that affects roughly one in three women of reproductive age worldwide at any given time, bacterial vaginosis has long suffered from a gap between diagnostic capability and mechanistic understanding. This new work suggests that the gap can be bridged with tools that have existed in theoretical ecology for decades, provided they are adapted thoughtfully to the biology at hand. The vaginal microbiome, it turns out, obeys rules that equations written in the 1970s can capture — and reading those equations may ultimately matter more than any prediction they produce.