In the Neugent et al. (2022) study, a total of 50 urine samples were collected from patients divided into two groups: 25 with UTI relapse and 25 in remission. Additionally, 25 urine samples were obtained from healthy individuals as controls. All samples in this study were from female participants. The study participants had no underlying issues affecting their urinary tract, immune system, and did not use indwelling or intermittent catheters. To retrieve the required data, we utilized the prefetch v2.10.7 tool from the National Center for Biotechnology Information (NCBI), enabling us to download the necessary datasets for our analysis. Figure 1A outlined the process, providing a visual representation of the entire workflow involving data collection and subsequent processing steps.

To ensure the sequencing data’s quality, we utilized Trimmomatic v0.39 (Bolger et al., 2014) to remove adapter sequences and low-quality bases. The parameters used were as follows: ILLUMINACLIP:TruSeq3-PE-2.fa:2:30:10:8:true TRAILING:20 MINLEN:60. Following quality control, an additional processing step was employed to eliminate human genomic sequences. For this purpose, we employed bowtie v2.4.4 (Langmead and Salzberg, 2012) and used the T2T-mY-rCRS genome (Nurk et al., 2022), which can be found at https://github.com/marbl/CHM13. This genome includes hard-masked PARs on chrY replaced with “N” and mitochondrion replaced with rCRS. This approach effectively removed any human genomic contamination.

After removing human genomic sequences, we proceeded to align the remaining reads against the virulence factor database (VFDB) (Liu et al., 2022). This database, which can be accessed at http://www.mgc.ac.cn/VFs/main.htm, contains a comprehensive collection of virulence factors. For the alignment process, we utilized bowtie v2.4.4 ¹³ and performed subsequent analysis using samtools v1.13 (Li et al., 2009). By aligning the reads against the VFDB_setB_nt database (available at http://www.mgc.ac.cn/VFs/download.htm), we were able to identify the number of reads corresponding to VFs in each sample. Subsequently, for any sample N, we calculated the abundance as follows (Equations (1) and (2)), (Ma et al., 2020; Chang et al., 2021):

Step 1: Calculation of the copy number of each gene:

Step 2: Calculation of the relative abundance of gene i

a_i: the relative abundance of gene i, b_i: the copy number of gene i from sample N, L_i: the length of gene i, x_i: the number of mapped reads.

This valuable information will be utilized in subsequent analyses to gain a comprehensive understanding of the involvement of oral microbiome-encoded VFGs in rUTI.

To develop a diagnostic model for rUTI, we adopted a two-tiered strategy, utilizing both support vector machine (SVM) (Huang et al., 2018) and random forest (RF) classifiers (Rigatti, 2017). Initially, we trained a suite of standard classifiers: random forest, gradient boosting, SVM, and logistic regression. These classifiers were trained with default parameters to establish a baseline performance. Their accuracy was subsequently assessed on test data, with results collated in a dictionary for straightforward comparison. The SVM classifier demonstrated superior accuracy in differentiating between “No UTI History” and the combined “rUTI Relapse” and “rUTI Remission” categories, guiding our choice of SVM for the binary classification task.

The statistical analyses in this study were conducted using RStudio. We used the vegan package to calculate alpha diversity measures, such as the Shannon and Simpson indices. The Bray distance was directly computed on the VFG profiles. For principal coordinate analysis (PCoA), we utilized the ade4 package in R (Zapala and Schork, 2006). To assess the significance of group differences, we performed Adonis analysis using the vegan package. To test for differential abundances of VFGs, we employed the Kruskal rank-sum test. The p-values were adjusted using the Benjamini–Hochberg (BH) procedure, with a significance threshold set at a p-adjust value of <0.05. We used the ggplot2 package to create boxplots and PCoA plots. The pheatmap package was utilized to construct heatmaps visualizing the patterns of VFG abundances. The ggtern was used for the plotting of ternary diagrams (Hamilton and Ferry, 2018). Furthermore, we utilized the pROC package to generate ROC curves, which were employed to evaluate the performance of diagnostic models.

According to the workflow depicted in Figure 1A, we obtained the relative abundance of VFGs in all samples. Subsequently, we calculated the alpha diversity measures of the three groups, including observed VFGs, Simpson, and Shannon indices based on the abundance of VFGs. Consistent results were observed, indicating that the alpha diversity of VFGs in the urobiome of patients with rUTI was significantly higher compared to those without a history of UTI (Figure 1B, P-values, observed VFGs: 8.5e−6, Simpson: 4.3e−6, Shannon: 9.6e−6). Furthermore, the alpha diversity was significantly higher in patients with recurrent UTI compared to those with a history of UTI but without recurrence (p-values, observed VFGs: 0.00076, Simpson: 3.4e−5, Shannon: 0.00012). There were no notable differences between individuals without a history of UTI and those without recurrence. These results suggest that VFGs may contribute to the recurrence of UTI.

First, we presented the top 20 abundant VFGs (Figure 1C and Supplementary TableS1) in each group, which we considered as core VFGs. We found that VFGs in urine mainly originated from bacterial genera such as Gardnerella, Streptococcus, Corynebacterium, and Staphylococcus. However, patients with rUTI had more virulence factors associated with Escherichia coli O157:H7. Furthermore, we identified unique VFGs in each group, including 552, 671, and 791 unique VFGs in the groups without history of UTI, non-recurrent UTI, and rUTI, respectively (Figure 1D). We presented the top 10 features in rUTI, which were distinctly associated with two typical pathogens, Klebsiella and E. coli O157:H7. This finding once again highlights the presence of unique VFGs in the urine of rUTI patients.

In order to further determine if there are compositional differences in VFGs among the three groups, we conducted PCoA analysis. Based on the Adonis test, we found significant differences in the composition of VFGs from a beta diversity perspective (Figure 2A, p = 0.001, F = 3.2732). To identify these differential VFGs more explicitly, we represented them using ternary plots with adjusted p-values below 0.05 (Figure 2B). We discovered that the majority of differentially abundant VFGs were enriched in women with rUTI, strongly suggesting their potential contribution to UTI recurrence. Additionally, we presented the top 50 VFGs (Figure 2C and Supplementary Table S2), such as VFG043077(gb|WP_000885860). VFG043119(gb|WP_000942326) and VFG043088(gb|WP_000983602). We found that these VFGs are primarily associated with E. coli O157:H7 (Supplementary Table S2). It is evident that individuals with a history of UTI and those prone to recurrence exhibit distinct VFG characteristics.

To further elucidate potential mechanisms, we investigated the associations between VFGs abundance and clinical data, including age, BMI, and urine pH. We identified six significant correlations using Spearman correlation analysis and canonical correspondence analysis (Figures 3A,B and Supplementary Table S3). One correlation was found with urine pH (Figure 3B and Supplementary Table S3, VFG048953(gb|WP_012967596)), one with age (Figure 3B and Supplementary Table S3, VFG035997(gb|WP_000556543)), and the remaining four with BMI (Figure 3B and Supplementary Table S3, VFG034470(gb|WP_001131106), VFG045826(gb|WP_021526504), VFG048844(gb|WP_014838945) and VFG048845(gb|WP_004122486)).

Further investigation is warranted to determine the potential of VFGs profiles in identifying patients with a history of UTI. We developed a machine learning model utilizing SVM to distinguish between females without a UTI history and those with a UTI history. Our findings suggest that the model can achieve an accuracy of 83% (Figure 4A). Moreover, our objective is to create a predictive model that utilizes urobiome information, specifically VFGs, to anticipate UTI recurrence. To accomplish this, we employed a random forest model to evaluate the diagnostic performance of VFGs in detecting rUTI, yielding an accuracy of 75% (Figure 4B). Overall, VFGs at the level of urobiome hold promise as indicators for both UTI history and recurrence. Our combined application of SVM and RF classifiers facilitated the creation of a comprehensive rUTI diagnostic model. The SVM classifier adeptly categorized individuals based on UTI history, while the RF classifier further refined the diagnosis by differentiating relapse and remission states. This dual approach ensures a thorough and precise diagnosis, setting the stage for tailored therapeutic interventions. Future endeavors might incorporate additional features or investigate alternative machine learning algorithms to augment the model’s precision.