The closed eye harbours a unique microbiome in dry eye disease

Cohort

Based on the power and pairwise sample-size estimator for permutational multivariate analysis of variance (PERMANOVA) application Micropower¹⁴, a low abundance 16S rRNA dataset similar to previous ocular microbiome studies¹² would require a minimal sample size of 30 to generate a discriminatory power of 0.8 with a significance level of 0.05. Therefore, we aimed for at least 35 subjects per study arm. Table 1 shows the demographics and clinical characteristics of the enrolled participants. Based on a post-enrolment exam and two clinical surveys, after informed consent, participants were classified as dry eye or controls (Supplemental S1). The resulting 36 control and 36 dry eye subjects were then stratified by clinical severity of eye disease (Fig. 1 and Supplemental Table 1). The study arms were randomized to daily saline eye wash upon awakening or non-intervention. A baseline closed eye tear sample was collected at randomization and the final sample after one month, yielding 144 samples. We did not observe significant differences between the stratifications normal versus mild and moderate versus severe with respect to microbial ecology (Fig. 2), however clear separation was present between the mild and moderate samples, indicating this more granular stratification introduced a false dichotomy in the data. Therefore, we combined subjects stratified as none or mild into one cohort, now termed normal, and we combined subjects stratified as moderate or severe into another cohort, now termed dry eye (Fig. 3). We used this simplified stratification system for all subsequent analyses (Methods).

Sequencing data

To investigate the bacterial microbiome of the closed dry eye, we used metabarcoding of the 16S rRNA gene. Sequence reads were generated using the Illumina MiSeq system (57,022 ± 47,011 counts/sample, Methods). We assigned taxonomy using the Greengenes database¹⁵, producing 1,593 data points/sample. After excluding any reads aligned to chloroplasts or cyanobacteria, we examined the bacterial community composition using the remaining 1,195 data points/sample that were collectively assigned to 185 genera¹⁶. All tear samples had the standard > 1,000 aligned reads/sample, so we retained all samples for further analysis. We used a log₂-transformation of cumulative sum scaling¹⁷ (log₂ CSS) to normalize our dataset, consistently producing 8,700–10,000 reads per sample (Supplemental Fig. S2).

The closed eye microbiome is distinct in dry eye disease

In the taxonomic analysis of the closed eye microbiome, the microbiomes of individuals with chronic dry eye disease and those without form distinct communities. Dry eye microbial communities are more diverse as quantified by richness (ANOVA, f = 4.8, P = 7.5 × 10⁻⁵), evenness (ANOVA, f = 13, P = 2.1 × 10⁻¹²) and Shannon diversity (ANOVA, f = 12, P = 5.9 × 10⁻¹², Fig. 2a). As demonstrated by principal coordinate analysis (PCoA) of Bray–Curtis dissimilarity, these communities form unique clusters as quantified by PERMANOVA (R² = 0.21, P = 0.00033, Fig. 2b). Similarly, multivariate redundancy analysis (RDA, variance = 95.25, f = 3.71, P = 0.001) and canonical correspondence analysis (CCA, chi² = 0.16 f = 3.64, P = 0.001) showed these communities are distinct, and this factor accounts for the majority of variance in the data. As quantified by two-way ANOVA adjusted for false discovery rate (FDR), univariate analysis of the relative abundance of bacterial genera showed individuals with dry eye disease have differences in the relative abundance of 113 genera, with the most significant differences in OPB56, Methylobacteriaceae, Bacteroidetes, Pseudomonas, and Meiothermus (all with FDR-adjusted P < 2 × 10⁻²², Supplemental Table 2). Similarly, mixed effects regression detected the most significant differences in the abundance of OPB56, Bacteroidetes, Pseudomonas, Meiothermus, and Methylobacteriaceae (all with FDR-adjusted P < 3.1 × 10⁻²⁰, Supplemental Table 3) among 49 genera with significant differences in relative abundance. Figure 2c shows relative abundance at the order level. To further investigate the importance of particular bacterial operational taxonomic units (OTUs), we used Spearman network analysis (Fig. 2d).

The closed dry eye microbiome remains distinct despite daily eye wash

Daily eye wash on awakening with sterile saline is a proposed therapy for dry eye disease¹⁸. We tested if the microbial community composition could be normalized by the prescription of daily saline eye wash. However, the microbial communities of dry (Supplemental Fig. S3) and normal eyes (Supplemental Fig. S4) were relatively unaffected by daily eye wash. The diversity of subjects with dry eye remained higher than that of normal individuals (richness, t-test, P = 7.5 × 10⁻⁵, evenness, t-test, P = 2.8 × 10⁻¹⁴ and Shannon diversity, t-test, P = 6.1 × 10⁻¹⁴, Fig. 3a). Multivariate analysis of these communities also remained distinct (PCoA, PERMANOVA, R² = 0.15, P = 0.00033, RDA, variance = 64.75, f = 5.73, P = 0.001 and CCA, chi² = 0.11, f = 5.62, P = 0.001, Fig. 3b). Univariate analysis detected the most significant differences in the abundance of MLE112, Lactobacillaceae, Streptococcus, Sphingobium, Caldicoprobacter and Anaerococcus (ANOVA, P < 0.01, Fig. 3c). Core microbiome analysis highlighted key differences in the distribution of unique genera, and network analysis demonstrated the importance of key taxa (Fig. 3d,e).

The closed eye microbiome in dry eye disease is distinct at baseline

To gain further insight into the microbial community composition of the dry eye, we narrowed our focus to examine the closed eye microbial composition at the time of randomization. As indicated by our more general analysis, both diversity (richness, t-test, P = 0.016, evenness, t-test, P = 5.8 × 10⁻⁶ and Shannon diversity, t-test, P = 8.1 × 10⁻⁸, Fig. 4a) and multivariate clustering of these communities are distinct at baseline (PCoA, PERMANOVA, R² = 1.34 P = 0.00033, RDA, variance = 53.95, f = 7.12, P = 0.001 and CCA, chi² = 0.09, f = 7.12, P = 0.001, Fig. 4b). ANOVA identified differential abundant genera (Fig. 4c). Univariate analysis using negative binomial regression identified Methylobacteriaceae, Pseudomonas, Bradyrhizobium and Allobaculum as the five most differently abundant genera (all with FDR-adjusted P < 1.1 × 10⁻¹¹, Supplemental Fig. S5a and Supplemental Table 4). At the phylum level, Firmicutes and Bacteroidetes remained unchanged, while Verrucomicrobia (FDR P = 4.6 × 10⁻¹¹) and Proteobacteria (FDR P = 6.0 × 10⁻¹⁰) showed the most significant differences. Core microbiome analysis identified 45 genera unique to dry eye with only 14 genera unique to the normal eye (Fig. 4d). Discriminant analysis of principal components at the order level revealed the abundances of orders OPB56 and Rhizobiales are important to discriminate the normal eye while the orders Halanaerobiales, Erysipelotrichales and Anaeroplasmatales are important to discriminate dry eye (Fig. 4e). Network analysis provided further evidence of these distinct communities (Fig. 4f).

Machine learning accurately classifies dry eye samples at baseline

To more precisely identify unique features with the potential to function as biomarkers for patients with dry eye, we used linear discriminant analysis of effect size. We identified 5 genera that reliably identified individuals with dry eye (Pseudomonas, Methylobacteriaceae, Helicobacter, Acetobacter and Stenotrophomonas) with 3 genera that reliably identified normal patients (Leuconostocaceae, Streptococcus and Calothrix, Supplemental Fig. S5b). To further support the uniqueness of microbial communities in dry eye, we built a support vector machine using leave-one-out cross-validation that could identify samples with 94% accuracy. Similarly, when we developed a random forest classifier, variable importance analysis identified the prevalence of the genera Methylobacterium, Megasphaera, Parabacteroides, S247, Bifidobacterium, Streptococcus, Desulfovibrio, Acetobacter, Dialister and Bacillus as particularly useful to identify samples from individuals with dry eye (Importance > 20, Supplemental Fig. S5c).

The closed eye microbiome in dry eye disease diverges after one month

We examined the stability of the dry eye microbiome by focusing on samples collected one month later in the same individuals. The dry eye microbiome remained distinct from the normal microbiome and remained consistent with the community composition noted at baseline (Supplemental Figs. S2, S3). Only slight divergence was noted over the course of a month (Fig. 5). The diversity of dry eye remained higher than that of the normal eye (richness, t-test, P = 5.2 × 10⁻⁵, evenness, t-test, P = 1.9 × 10⁻¹¹ and Shannon diversity, t-test, P = 8.6 × 10⁻¹¹, Fig. 5a). The distinct clustering of these communities on multivariate analysis also persisted (PCoA of Bray–Curtis dissimilarity, PERMANOVA, R² = 1.52, P = 0.00033, RDA, variance = 66.53, f = 8.52, P = 0.001 and CCA, chi² = 0.10, f = 8.14, P = 0.001, Fig. 5b). ANOVA identified differentially abundant genera (Fig. 5c). Similarly, negative binomial regression noted the greatest differences in the abundance of the genera Methylobacteriaceae, Pseudomonas, Azospirillum, Bradyrhizobium and Coriobacteriaceae (all with FDR-adjusted P < 7.0 × 10⁻¹⁸, Supplemental Fig. S6a, Supplemental Table 4). However, core microbiome analysis detected 76 unique genera in dry eye, 69 shared genera and 24 genera unique to the normal eye (Fig. 5d). Similar to baseline, discriminant analysis of principal components at the order level identified OPB58 and Bacteroidetes as important discriminators of the normal eye. In individuals with dry eye, additional orders enabled further discrimination, with Flavobacteriales, Alteromonadales and Actinomycetes, Anaeroplasmatales and Desulfuromonadales functioning as the primary discriminators instead of Halanaeobiales and Erysipelotrichales (Fig. 5d).

Machine learning more accurately identifies dry eye after a month

We used linear discriminant analysis of effect size to identify bacterial genera with the potential function as biomarkers and noted 10 genera unique to individuals with dry eye and 9 unique to individuals with a normal eye, supporting the limited divergence of these communities (Supplemental Fig. S6b). We also developed a support vector machine with leave-one-out cross-validation that identified dry eye samples with 97% accuracy. Variable importance analysis of a random forest classifier noted the prevalence of Methylobacterium, Megasphaera, Parabacteroides, S247, Bifidobacterium, Streptococcus, Desulfovibrio and Acetobacter as important identifiers (Importance > 20, Supplemental Fig. S6c).

The closed eye microbiome at baseline is unaltered by the cellular concentration of the tear fraction

To ensure our microbial ecological analysis was unaltered by the cellular content of the tear fraction, we compared the microbial composition of each cellular fraction collected at baseline. Three samples were excluded from this analysis due to incomplete data (n = 141). Community composition was similar between both high and low cellular fractions (Fig. 6).