The in vitro growth kinetics of Lactobacillus strains under anaerobic conditions at pH 6.0 were evaluated, a characteristic that is important for commercial scale-up. The primary vaginal Lactobacillus isolates grew variably, between and within species (Fig 1). L. crispatus and L. jensenii strains grew the best, as quantified by calculation of the area under the curve (AUC), followed by L. vaginalis, L. mucosae and lastly L. gasseri strains.

The in vitro ability of isolates to grow at lower pHs was considered as important, since these conditions are optimal for the maintance of a healthy vagina [37,38]. All L. crispatus strains showed reduced growth at pHs 4.5, 4.0, and 3.5 compared to pH 6.0, although they appeared to tolerate low pHs better than other Lactobacillus spp. (Fig 1). In contrast, L. jensenii, L. vaginalis, L. gasseri, and L. mucosae strains appeared to be less acid tolerant, with the number of strains able to grow decreasing with lower pHs. Notably, there were no differences in growth at any pH between vaginal Lactobacillus strains isolated from women without BV or STIs versus those originating from women with BV and/or STIs (Fig 1). The majority of vaginal L. crispatus, L. vaginalis and L. jensenii strains grew better than their respective ATCC strain at all pHs tested. In comparison, the ten probiotic strains (including L. reuteri, L. rhamonosus and L. acidophilus) showed less growth than L. crispatus strains at pH 6.0, and the majority of probiotic strains showed some tolerance to pHs 3.5–4.5 (Fig 1).

All vaginal Lactobacillus strains tended to have similar kinetics in their ability to lower pH under anaerobic conditions, with the lowest pH reached being pH 3.7, while none of the probiotic isolates lowered the pH <4.2 (S1 Fig). Overall, L. crispatus lowered the pH significantly more than L. gasseri, L. mucosae, L. vaginalis and any of the probiotic strains at 48 hours, after adjusting for multiple comparisons (Fig 2). Growth of individual isolates at pH 6.0 did not significantly predict their ability to lower culture pH at 48 hours (Spearman rho = -0.17, p = 0.165), indicating that differences in growth between strains did not account for differences in their ability to lower pH.

L- and D-lactate are considered important metabolic by-products of Lactobacillus strain metabolism that have anti-viral and anti-bacterial activity, respectively [30–32]. Therefore, the ability of vaginal Lactobacillus strains (46/57, due to assay limitations) to produce L- and D-lactate were evaluated under anaerobic conditions (Fig 3A and 3B). Vaginal Lactobacillus spp. tended to produce lower amounts of L- than D-lactate (median 10.6 ng/μL [IQR 5.6–19.4] vs. 35.8 ng/μL [30.4–41.9], p<0.0001), although all appeared to produce both isomers. Some vaginal strains produced high levels of L- but almost no D-lactate (such as L. crispatus 95.34PA), while others produced large amounts of D- but little L-lactate (such as L. jensenii 96.45PA). Again, STI and BV status of the donor did not influence L- and D-lactate levels measured (Fig 3A and 3B). Further, the majority of vaginal Lactobacillus isolates produced more of both isomers than their respective ATCC reference strains. With the exception of two probiotic L. acidophilus isolates (which produced similar levels of both isomers), all probiotic strains produced higher levels of L- than D-lactate, although concentrations of both were significantly lower than those of the vaginal strains (p<0.0001).

The ability of strains to produce L- and D-lactate did not correlate (Spearman rho = 0.00, p = 0.9592), neither did production of D-lactate (Spearman rho = 0.05, p = 0.7250) or L-lactate (Spearman rho = -0.23, p = 0.0838) with growth. There was also no correlation between ability of isolates to lower culture pH and their ability to produce L- or D-lactate (Spearman rho = 0.15, p = 0.2604 for L-lactate; rho = -0.02, p = 0.8960 for D-lactate), indicating that lactate was not the only cause of lowering culture pH.

Earlier studies suggested that H₂O₂ was the most influential Lactobacillus spp.-produced metabolite to inhibit pathogens in the lower FGT, although these studies were conducted aerobically [39,40]. To mimic in vivo conditions in the lower FGT, all experiments in this study were performed under anaerobic conditions. All of the vaginal L. gasseri and L. vaginalis strains produced small quantities of H₂O₂ (ranging from 0.5–15 μM), while some of the L. jensenii, L. crispatus, and L. mucosae strains did not produce any measurable H₂O₂ at all (Fig 3C). Of all previously evaluated in vitro characteristics, H₂O₂ only correlated with L-lactate levels positively (Spearman rho = 0.61, p<0.0001), but not with D-lactate (Spearman rho = -0.20, p = 0.1504).

The ability of abiotic culture supernatants from a subset of vaginal Lactobacillus strains (25/57; selected based on performance in previous assays) to inhibit the growth of a panel of P. bivia and G. vaginalis strains was determined under anaerobic conditions (Fig 4). Most of the vaginal Lactobacillus isolates (besides L. gasseri 94.98PB, L. mucosae 86.30PA and 87.21PA) showed some degree of inhibition towards all clinical P. bivia strains, although the extent of inhibition differed by Lactobacillus and P. bivia strain (Fig 4). L. crispatus tended to inhibit P. bivia growth to a greater extent than other species, inhibiting all P. bivia strains > 90%. Vaginal Lactobacillus strains inhibited P. bivia similarly to their respective ATCC and the probiotic Lactobacillus strains (Fig 4).

In contrast, the inhibition of G. vaginalis strains was generally more variable and significantly lower (Fig 4). Only three vaginal Lactobacillus (L. crispatus 100.16A, L. jensenii 88.33PA and 92.27PA) and the L. crispatus ATCC strain inhibited the growth of the full panel of G. vaginalis strains. Inhibition appeared to be highly G. vaginalis strain-specific, as almost all Lactobacillus strains inhibited the vaginal G. vaginalis isolate 3H1 (Fig 4; bottom row), while only 9/25 inhibited the G. vaginalis ATCC strain (Fig 4; top row). Overall, L. crispatus strains tended to exhibit the strongest inhibitory activity against G. vaginalis, although this was not statistically significant. The ATCC reference Lactobacillus strains generally showed broader and better inhibitory activity against G. vaginalis than vaginal Lactobacillus isolates (Fig 4). Notably, the abilities of vaginal Lactobacillus strains to inhibit the pathogens tested were highly correlated (G. vaginalis and P. bivia: Spearman rho = 0.814, p<0.0001), indicating that individual Lactobacillus strains have a similar capacity to inhibit various BV-associated bacteria. Similarly to the previously assessed characteristics, BV/STI status of the donor did not influence inhibitory abilities.

Metronidazole and clindamycin are SOC for BV treatment and penicillin and amoxicillin are commonly used antibiotics. All vaginal Lactobacillus isolates (n = 57) were resistant to metronidazole, while the susceptibility to clindamycin varied within and between Lactobacillus spp. (Fig 5A). L. vaginalis and L. jensenii isolates were the most susceptible to clindamycin, while L. gasseri strains were the least. The inhibition zones of the probiotic strains were smaller than those of most clinical strains (with the exception of L. gasseri), indicating that these might be less susceptible to clindamycin than the majority of vaginal Lactobacillus strains.

All vaginal Lactobacillus isolates showed similar susceptibility to penicillin, with the exception of L. gasseri isolates, which tended to be less susceptible (Fig 5A). Probiotic L. acidophilus strains were more susceptible to penicillin than most vaginal strains, while L. rhamnosus isolates tended to be similarly susceptible, and L. reuteri less. The inhibition zones for amoxicillin were highly comparable to those measured for penicillin, as was the intra-species variability (Fig 5A). L. crispatus isolates were significantly more susceptible to amoxicillin than L. rhamnosus (adj. p = 0.0497), and L. mucosae than L. reuteri (adj. p = 0.0128) and L. rhamnosus (adj. p = 0.0033), indicating that L. reuteri and L. rhamnosus but not L. acidophilus strains may be less susceptible to amoxicillin than the vaginal isolates.

The rifamycins rifampicin and rifabutin are commonly used to treat tuberculosis (TB) and form part of the six-month antibiotic regimen administered to ~500,000 people in South Africa who develop TB annually [41], and have previously been explored as a treatment option for BV [42,43]. All Lactobacillus isolates were highly susceptible to rifampicin (<12.5 μg/ml) and rifabutin (<10.0 μg/mL) and susceptibilities of the probiotic isolates were similar to the vaginal strains (Fig 5A).

Broader antibiotic resistance profiling to inhibitors of protein, cell wall/membrane and DNA synthesis was conducted on a subset of 14 vaginal and six probiotic Lactobacillus isolates (Fig 5B). Vaginal L. gasseri strains appeared to be susceptible to a wider range of antibiotics than other Lactobacillus spp. L. crispatus strains showed exactly the same antibiotic susceptibility, while the intra-species variability was higher for other vaginal Lactobacillus spp. There were no clear differences between vaginal and probiotic Lactobacillus spp. in antibiotic susceptibility profiles. Based on the clinical MIC breakpoints defined by the EUCAST (2019), all vaginal isolates would be considered sensitive to ampicillin (MIC ≤4 μg/mL), and the majority to chloramphenicol (MIC ≤8μg/mL). For penicillin, (sensitivity MIC≤0.25 μg/mL, resistance MIC >0.5μg/mL), all clinical L. gasseri, L. mucosae and L. vaginalis strains were sensitive, all L. crispatus strains and L. jensenii 88.33PA strains had an intermediate sensitivity, and L. jensenii 92.27PA was resistant. L. crispatus 100.16A, L. gasseri 117.73PA and 100.46PA, L. jensenii 92.17PA and 95.31PA, L. vaginalis 79.24PA and L. mucosae 102.33PA were susceptible to clindamycin (MIC ≤1 μg/mL), while all other isolates had MICs >1μg/mL. Since the breakpoint for clindamycin has been defined as >4μg/mL, accoding to the EUCAST 2019 guidelines, it was not possible to conclude with the current data whether these isolates are resistant.

The selected Lactobacillus strains should strongly adhere to host cells [21,22] but not influence cell viability or cause inflammatory changes [44,45]. All vaginal L. crispatus (10/10), and the majority of L. vaginalis (7/8), L. jensenii (16/17) and L. mucosae (10/12) and two-thirds of the L. gasseri (6/9) strains adhered to Ca Ski cells to varying extents, independently of the BV/STI status of the donor (Fig 6). Overall, vaginal L. crispatus strains tended to adhere the strongest followed by L. mucosae, L. jensenii, L. vaginalis, and L. gasseri, and several vaginal isolates (2/10 L. crispatus, 2/9 L. gasseri, 6/8 L. vaginalis, and 11/18 L. jensenii) adhered better than their respective ATCC reference strain (Fig 6).

For a subset of vaginal Lactobacillus isolates that included three strains per vaginal Lactobacillus spp. and six probiotic strains, the effect on cervical epithelial Ca Ski cell viability and cytokine expression was determined (Fig 7). None of the vaginal strains decreased the viability of Ca Ski cells below 80%, as measured by trypan blue straining. After grouping the cytokines into their biological categories, all bacterial isolates up-regulated regulatory cytokines (including IL-1RA and IL-10) 1.5-fold (L. crispatus 100.16 a) to 10-fold (L. reuteri Z1006) compared to the control. In contrast, only five Lactobacillus strains (L. gasseri 100.46 PA, L. mucosae 80.23 a, L. vaginalis 79.24 PA, L. rhamnsus P0154 and C21134) induced adaptive cytokines (including IL-2/4/5/13/15/17 and IFN- γ) ~ 2-fold more than the control, while the remaining 16 strains downregulated adaptive cytokines up to 1.6-fold. Only for L. gasseri 100.46 PA slightly higher levels of growth factors and haematopoietic cytokines (including IL-7/9, FGF-basic, G-CSF, GM-CSF, PDGF-BB, and VEGF) were measured compared to the control, and the remaining strains downregulated this group of cytokines up to 4-fold. For about half of the strains (10/21) up to 4.2-fold higher levels of chemokines (including IL-8, Eotaxin, IP-10, MCP-1, MIP-1α/β and RANTES) were measured than for the control, while for the other half up to 24-fold lower levels were found. Four strains (L. gasseri 100.46 PA, L. mucosae 80.23 a, L. vaginalis 79.24 PA, L. rhamnosus C21134) induced higher inflammatory responses (1.02-, 2.44-, 5.84-, 5.88-fold, respectively, including IL-1β/6/12(p70) and TNF-α) than the control, while the remaining 17 Lactobacillus strains downregulated inflammatory responses up to 47-fold. In summary, the majority of bacterial strains induced cytokine responses comparable to the control and did not induce inflammatory responses. However, for L. gasseri 100.46 PA, L. mucosae 80.23 and L. vaginalis 79.24 PA higher levels of cytokines were measured in all cytokine categories, indicating that the inflammatory potential of these Lactobacillus strains requires further investigation.

To identify the most promising vaginal Lactobacillus strains for the development of a probiotic for vaginal health, a weighted scoring system was devised based on PPP characteristics, including: their ability to grow well in vitro, lower pH, adhere to cervical epithelial cells, produce L- and D-lactate and H₂O₂, and inhibit growth of BV-associated species (G. vaginalis and P. bivia). Cytokine and antibiotic susceptibility profiles were not considered, as data were available only for a subset of vaginal strains. For this ranking, a total of 36/57 Lactobacillus strains (for which complete characteristic evaluations were available) were included. To allow objective comparisons, the performance of vaginal isolates was compared with their respective ATCC reference and the probiotic isolates. Because some of the characteristics included in the ranking correlated, a weighted scoring system was developed to account for co-linearity between characteristics (Fig 8A). Including all characteristics in the scoring matrix (model A; Fig 8B), L. mucosae 90.13PA, L. crispatus 70.6PA and 100.16A, L. vaginalis 80.23B, and L. jensenii 88.33PA and 92.27PA were the six highest scoring Lactobacillus strains, with similar points as the ATCC L. crispatus and ATCC L. gasseri strains, while all commercially available probiotics ranked lower. Further step-wise sensitivity analyses were performed to test the robustness of the selection. Model B: Because the concentrations of H₂O₂ produced by Lactobacillus were all below the level found to be microbicidal [30], a ranking model was considered in which H₂O₂ was excluded, which did not change the ranking of the top six strains (Fig 8C). Model C: In addition to excluding H₂O₂ production, the inhibitory effect on G. vaginalis and P. bivia were grouped to avoid skewing towards strains that performed generally well in the inhibition assays (Fig 8C), which resulted in two L. gasseri strains being included in the top-10 strains.

The sensitivity analysis suggested that the scoring system was quite robust and showed that isolates from both STI/BV-negative and women with STIs or BV were among the top-performing strains, indicating that even Lactobacillus strains that originate from women with BV or STIs can have promising probiotic characteristics. Although all commercially available probiotic strains were included in the scoring system, none scored in the top ten in any of the models considered.

Prior to selecting vaginal Lactobacillus strains for the development of a probiotic, we selected well-performing strains from a range of Lactobacillus spp. for whole genome sequencing and genome analysis to assess the presence of antimicrobial resistance determinants, mobile elements and prophages, including L. mucosae 90.13PA (ranked consistently in the top 2, from BV/STI-positive women), L. crispatus 70.6PA (highest ranking L. cripsatus strain from BV/STI-negative women) and L. crispatus 73.55a (highest ranking L. cripsatus strain from BV-positive women), L. gasseri 94.98PB (highest ranking L. gasseri strain, from BV/STI-negative women), and L. jensenii 95.1PA (from BV/STI-negative women).

The whole genomes of five selected strains were sequenced to a median depth of 117X (range 102-142X). Full assembly statistics are shown in S1 Table. Screening for the presence of well characterised antimicrobial resistance determinants revealed no matches to the Resfinder, CARD and AMRFinder databases for any of the assemblies. However, analysis of the RAST annotated assemblies revealed that all five strains harboured open reading frames (ORFs) with homology to the ldb0660 and copY-like genes in Lactobacillus delbrueckii subsp. bulgaricus that respectively encode a copper-translocating P-type ATPase and a negative transcriptional regulator, both of which play a role in copper homeostasis. In addition, three of the five strains (L. crispatus 70.6 PA, L. crispatus 73.55 a and L. mucosae 90.13 PA) harboured genes encoding putative TetM-type ribosomal protection proteins. Two strains (L. gasseri 94.98 PB and L. jensenii 95.1 PA) possessed potential beta-lactamase class A proteins, while L. crispatus 73.55 a harboured an ORF encoding a putative streptothricin acetyltransferase. The latter was flanked by genes encoding predicted mobile element-associated transposases. Finally, two classes of potential multidrug transporters were identified. All strains except L. gasseri 94.98 PB harboured genes encoding putative multidrug and toxin extrusion (MATE) family efflux pumps with homology to the MepA protein in Staphylococcus aureus and a further three strains (L. crispatus 70.6 PA, L. crispatus 73.55 a and L. jensenii 95.1 PA) harboured genes encoding putative major facilitator superfamily (MFS) class efflux pumps with homology to the Lde efflux pump in Listeria monocytogenes.

Putative intact prophages were present in the genomes of two of the five strains. L. crispatus 70.6 PA harboured a single predicted 43.9Kbp prophage, while L. mucosae 90.13 PA harboured three predicted prophages of 16.9Kbp, 37.5Kbp and 40.1Kbp, respectively. Complete CRISPR-Cas loci (i.e. one or more signature cas genes located alongside a CRISPR array) were identified in the genomes of three of the strains. L. crispatus 73.55 a harboured a type I-E locus with spacers showing matches to a lactococcal phage Tuc2009. L. mucosae 90.13 PA harboured two distinct loci encoding a type I-C and type III-A system, respectively. Spacers present in the type I-C locus showed homology to phage-like regions on the L. mucosae LM1 plasmid and the Streptococcus thermophilus phage Sfi19, while those in the type III-A locus showed homology to several plasmids (L. mucosae LM1, C. botulinum pCBH, and an unnamed plasmid in Lactobacillus fermentum DR9) as well as the Streptococcus phage Javan623. L. jensenii 95.1 PA harboured a single type II-A locus with spacers showing homology to several phages (an enterococcal phage, EFRM31 and the Lactobacillus phages Lf1 and Lj928).

Vaginal Lactobacillus isolates were derived from women (16–22 year old) enrolled in the Women’s Initiative in Sexual Health (WISH) study [68], approved by the Human Research Ethics Committees of the University of Cape Town (UCT HREC #267/2013). Women who were ≥18 years old provided written informed consent and those <18 years provided written assent, and consent was obtained from their parent(s) or legal guardian(s). Eligibility criteria included being HIV negative, generally healthy, not pregnant or menstruating at the time of sampling. Participants were tested for BV (Nugent scoring) and STIs (including C. trachomatis, N. gonorrhoeae, T. vaginalis, Mycoplasma genitalium, HSV-2, Hemophilus ducreyi, Treponema pallidum and Lymphogranuloma venerum). The vaginal microbiota was characterised by 16s rRNA gene sequencing [36]. Cervicovaginal secretions were collected using menstrual cups, diluted 1:4 in PBS and stored with glycerol (20%v/v) at -80°C for isolation of Lactobacillus strains.

Commercially available probiotics included Reuterina Femme (oral capsules, Ascendis Pharma, South Africa), Provacare Probiotic Vaginal Care (vaginal capsules, Provacare, Canada), Vagiforte Plus (oral capsules and vaginal tablet, Bioflora CC, South Africa), Vagiforte Plus Combo Pack (oral capsules and vaginal spray, Bioflora CC, South Africa), Gynophilus (vaginal pessary, biose, France) and Muvagyn (vaginal pessary, Hälsa Pharma GmbH, Germany).

Cervicovaginal secretions or commercial probiotics dissolved in De-Man-Rogosa-Sharpe (MRS) were streaked onto MRS agar plates and incubated anaerobically (37°C, 48 hours). Single colonies with distinct morphologies were picked and re-streaked until homogenous. Pure colonies were inoculated into MRS broth, grown anaerobically (37°C, 48 hours), and stocks stored with glycerol (20% v/v) at –80°C. Matrix Assisted Laser Desorption/Ionization time-of-flight (MALDI-TOF, MALDI Biotyper, Bruker Daltonik, USA) was used to identify Lactobacillus spp. To determine bacterial concentrations at a standardized optical density (OD), overnight cultures were adjusted to OD_600nm 0.1±0.01, serial dilutions plated onto MRS agar plates that were incubated anaerobically (37°C, 48 hours), colonies were counted, and CFU/mL calculated (S2 Fig).

To confirm MALDI-TOF species identification, single colonies were incubated (37°C for 20 minutes, followed by 90°C for 15 minutes) in Tris-EDTA buffer (pH 7.4) with 20ng/mL proteinase K (BioLabs, USA). PCR reactions contained 8% (v/v) template, 0.5uM of F27 and R5 primers [69], 0.2mM PCR nucleotide mix (Promega, USA), 1x GoTaq Flexi Buffer (Promega, USA), 1.25 units GoTaq G2 Flexi DNA polymerase (Promega, USA), 2.5mM MgCl2 (Promega, USA) and nuclease-free H₂O. PCR conditions were 96°C for 5 minutes; 30 cycles at 94°C for 30 seconds, 55°C for 30 seconds and 72°C for 1.5 minutes, and a final elongation at 72°C for 7 minutes. Products were Sanger sequenced at Macrogen Europe. 16S rRNA sequences were aligned to National Centre for Biotechnology Information (NCBI) database sequences with the Basic Local Alignment Search Tool (BLAST) algorithm, selecting the highest percentage identity match over 99% as the species identity.

Cultures standardized to OD_600nm 0.1±0.01 in MRS broth adjusted to pH 3.5, 4.0, 4.5 or 6.0 using HCl were incubated in triplicate under anaerobic conditions in 96-well plates at 37°C for 48 hours. OD_600nm was measured at 0, 3, 6, 12, 18, 24, 36 and 48 hours and area under the curve (AUC) was calculated for each isolate. Experiments were repeated three times.

Cultures adjusted to OD_600nm 0.1±0.01 in MRS broth were grown anaerobically at 37°C, and culture supernatant pH measured at 3, 6, 12, 24 and 48 hours using a calibrated pH meter. Experiments were performed in duplicate.

Cultures adjusted to OD_600nm 0.1±0.01 were incubated anaerobically for 24 hours at 37°C. Supernatants were filtered (0.2 μM) and concentrations of D- and L-lactate measured using colorimetric assays (Sigma-Aldrich, USA). H₂O₂ concentrations were measured using a fluorometric kit (Sigma-Aldrich, USA), according to the manufacturer’s instructions. Experiments were performed in duplicate.

G. vaginalis (ATCC 14018 and five vaginal isolates) and P. bivia (ATCC 29303 and five vaginal isolates) were grown in brain heart infusion (BHI) supplemented with 5% horse blood, standardised to OD_600nm 0.4±0.01, and 100 μL was added in triplicate to 96-well plates. Abiotic Lactobacillus-conditioned MRS (100 μL; same as used for lactate measurement) was added and incubated anaerobically at 37°C for 48 hours. OD_600nm was measured at 0, 3, 6, 12, 18, 24, 36 and 48 hours. As control, growth of G. vaginalis and P. bivia strains in the presence of plain MRS, at the same ratio as abiotic culture supernatants, was measured. AUC for each growth curve was calculated and percentage of growth inhibition in the presence of Lactobacillus-conditioned MRS compared to the control calculated. Experiments were repeated three times.

Ca Ski cells (ATCC CRL-1550) were grown to 80% confluence in 10% fetal calf serum (FCS) Dulbecco’s Modified Eagle Medium (DMEM) in 24-well cell culture plates (37°C, 5%CO₂). Lactobacillus isolates (4.2x10⁶ CFUs) were added in triplicate and co-cultured for 3 hours at 37°C (5%CO₂), after which unbound bacteria were gently removed by washing the cells three times with PBS. Cells and bound bacteria were detached using 0.1% Triton X-100, serially diluted, plated onto MRS agar plates, and incubated anaerobically at 37°C for 48 hours. The percentage of adhered bacteria was calculated as the number of bound cells compared to total number of bacterial cells initially added. Experiments were repeated three times.

Ca Ski cells were grown in 10% FCS DMEM to 80% confluence at 37°C in 5% CO₂. Lactobacillus strains (4.2x10⁶ CFUs) were added in duplicate and incubated for 24 hours (37°C, 5% CO₂). Ca Ski cells incubated without Lactobacillus strains were included as control. Supernatants were harvested and filtered (0.22 μm, Corning Costar Spin-X, Sigma-Aldrich, USA) to exclude bacterial cells. A Bio-Plex Pro Human Cytokine 27-Plex Luminex kit (Lot 64064139, Bio-Rad, USA) was used to measure concentrations of cytokines, chemokines and growth factors in the culture supernatant, according to the manufacturer’s instructions. Data were acquired using a Bio-Plex Suspension Array Reader (Bio-Rad Laboratories Inc, USA), and a 5PL regression line was used to determine concentrations from standard curves. All values below the detection limit were recorded as half of the lowest measured concentration for each cytokine. Viability of Ca Ski cells after co-culture was evaluated using trypan blue staining (0.4%, Sigma-Aldrich, USA).

Double-layer disc diffusion was used to determine the susceptibility to metronidazole (5μg), clindamycin (2μg), penicillin (2μg) and amoxicillin (10μg; Davies Diagnostics, South Africa), as described previously [58]. These experiments were performed in duplicate. For rifampicin and rifabutin (Sigma-Aldrich, USA), minimal inhibitory concentrations (MICs) were determined using two-fold serial dilutions according to European Committee on Antimicrobial Susceptibility Testing (EUCAST) 2019 guidelines, with concentrations ranging from 5–0.00488μg/mL for rifabutin and 25–0.024μg/mL for rifampicin. MICs below the lowest or above the highest concentration tested were assigned a MIC half of the lowest or twice the highest concentration tested, respectively. Experiments were performed in duplicate. For a subset of Lactobacillus strains (n = 20), broader antibiotic susceptibility profiles were determined using Sensititre GPALL1F plates (including ampicillin, cefoxitin, chloramphenicol, ciprofloxacin, clindamycin, daptomycin, erythromycin, gentamicin, levofloxacin, linezolid, moxifloxacin, nitrofurantoin, oxacillin, penicillin, quinupristin/dalfopristin, rifampin, streptomycin, tetracycline, tigecycline, trimethoprim/sulamethoxazole and vancomycin; Thermo Fisher Scientific Inc., USA), according to the manufacturer’s instructions.

Genomic DNA was extracted from Lactobacillus isolates using the Quick-DNA Miniprep Kit (Zymo Research, USA) following the manufacturer’s instructions after a pre-lysis step which consisted of resuspension of the cell pellet in 200 μl PrimeStore Molecular Transport Medium (Longhorn Vaccines-Diagnostics, USA) and storage at -80°C for 24 hours. Libraries for sequencing were prepared using the Nextera DNA Flex kit (Illumina, USA), according to the manufacturer’s instructions, and sequenced using the Illumina MiSeq Reagent Micro Kit v2 on an Illumina MiSeq instrument in a paired-end, dual indexed 2 x 151 cycle sequencing run. Draft de novo assemblies were prepared using a customised assembly pipeline. Briefly, raw reads were trimmed using Trimmomatic v0.3.9 [70] using a sliding window approach (Phred quality cut-off of 15 averaged across 4 bases), a leading/trailing quality cut-off of 3 and the removal of reads below 30bp. Trimmed reads were assembled using SPAdes v3.13.0 [71] using the ‘—careful’ flag and omitting the initial read error correction step. To improve the draft assemblies, gap closure and contig extension were performed using GMcloser v1.6.2 [72]. The quality of the final assemblies for each strain was assessed using Quast v5.0.2 [73]. Rapid Annotation of microbial genomes using Subsystems Technology (RAST) was used to annotate the de novo assemblies [74].

Annotated draft assemblies were screened for the presence of putative antimicrobial resistance determinants using Resfinder v3.2. [75], CARD [76] and AMRFinder [77]. Putative prophages were identified using the online PHASTER server [78] [https://phaster.ca/]. CRISPR-Cas systems were identified using the CRISPRCasFinder online tool [79] [https://crisprcas.i2bc.paris-saclay.fr/] and spacer targets were identified using CRISPRTarget [80].

GraphPad Prism6 (GraphPad Software, USA), STATA version 11.0 (StataCorp, USA) and RStudio were used to generate graphs and analyse data. The Mann-Whitney U test was used for non-parametric variables, and Spearman Rank test was used for non-parametric correlations. The Kruskal-Wallis one-way analysis of variance was used to compare variables with three or more groups. A false-discovery rate step-down procedure [81] was used to adjust p-values for multiple comparisons, and adjusted p<0.05 were considered significant.

To rank Lactobacillus strains according to PPP criteria, a scoring system was developed. For each characteristic, Lactobacillus isolates were ranked into quartiles relative to the whole group: Scored “0” for characteristics <25^th percentile (relative to all isolates); “1” for characteristics between 25–50^th percentile; “2” for characteristics between 50–75^th percentile, and “3” for characteristics between >75^th percentile. Because some characteristics may correlate (e.g. ability to lower pH and H₂O₂ production), all scores were weighted to account for co-linearity: Weight = 1-(sum of column Spearman Rho÷sum of column score).