Can AI identify promising vaginal probiotic strains?

When considering a vaginal probiotic, we may read a species name off the label and assume it carries a defined benefit. A new study ¹ challenges that assumption directly: does the functional potential of Lactobacillus crispatus belong to the species as a whole, or to individual strains we've been lumping together far too generously?

One species, forty-seven versions of it

The team collected 1,126 bacterial isolates from vaginal swabs collected from 36 healthy women aged 18 to 30. Of these, 639 turned out to be Lactobacillus crispatus. Phylogenomic analysis grouped these 639 isolates into 47 distinct clades, revealing substantial diversity within the species. Across the 639 isolates, the pangenome contained 4,482 gene families, of which 1,579 belonged to the core genome. This highlights meaningful genomic diversity within L. crispatus, despite an overall conserved genomic structure.

This diversity isn't just academic: strains differed in the presence and copy number of genes encoding pullulanase (PulA), an enzyme involved in glycogen degradation that may help L. crispatus exploit this carbon source in the vaginal environment. Existing general probiotic screening tools may struggle to capture these strain-level differences. When the researchers ran the strains through an existing tool called iProbiotics, nearly every one scored well, averaging 97.58%, which illustrates the challenge of distinguishing strain-level functional potential using existing general probiotic prediction approaches.

The phenotypes that distinguish higher-performing strains

The researchers then tested 67 representative strains on five in vitro traits that may matter clinically and considered relevant to vaginal probiotic potential: how well they grow, how much they acidify their environment, how much lactic acid and hydrogen peroxide they produce, and how well they block vaginal pathogens. The results varied widely. Lactic acid output ranged from 2.02 to 10.29 g/L, and hydrogen peroxide levels from 9.39 to 18.3 µmol/L, even though all the strains belonged to the same species. Final pH ranged from 3.8 to 4.45, so every strain made its environment more acidic, just not to the same degree.

Strains also differed in how well they blocked Gardnerella vaginalis, a bacterium linked to bacterial vaginosis : the inhibition zone averaged 15.35 mm but varied by several millimeters from one strain to the next. The researchers combined these five measurements into a single score, giving the heaviest weight, 30%, to how well a strain fought off BV-related pathogens. They then split the strains at the median into a higher functional-potential group, labeled Lcris-SFS, and a lower functional-potential group.. The three top-performing strains are now being tested in animals and in early clinical studies.

A model that reads function from the genome

To avoid months of lab culture work, the researchers built a computer model called VLCPredictor. It reads short DNA snippets from each strain's genome, narrows 349,184 of these snippets down to the 534 most informative ones, and uses that pattern to predict a strain’s functional potential, helping prioritize candidates for further experimental testing. Eight prediction methods were tested, and the best performing model, a random forest, achieved a mean AUC of 0.742 for distinguishing strains with higher versus lower functional potential.. Applied to 103 published genomes, the model flagged about 76% of strains from healthy women as having higher predicted functional potential, compared with only 31% of strains from women with bacterial vaginosis. The model isn't perfect: it was trained on genomes from a single region and ethnic group, and only 16 genomes came from women with bacterial vaginosis, so it still needs more validation. The importance of strain-level selection is also illustrated by previous clinical studies cited by the authors: one L. crispatus-based probiotic reduced bacterial vaginosis recurrence following antibiotic treatment, whereas another Lactobacillus-based probiotic combination did not improve cure rates in a separate trial. 

The clinical message is not that genomic prediction can already select a probiotic for an individual patient. Rather, this study reinforces the importance of strain-level characterization: belonging to a health-associated species such as L. crispatus does not guarantee identical functional properties. VLCPredictor could help researchers prioritize promising candidates, but in vivo and clinical validation remain essential before translating these predictions into practice.