This research was approved by the Ethics Committee of Qingdao Eye Hospital (2019-16, 11/Dec/2019) and registered on the Chinese Clinical Trial Registry (ChiCTR2000031875, 14/Apr/2020). All procedures complied with the tenets of the Declaration of Helsinki. Altogether, 69 healthy subjects and 80 patients with different MK were included in this study. Informed consent was obtained from all participants. Detailed information about the subjects is shown in Table S1 .

Bacterial keratitis (BK) was preliminarily considered if patients met at least two of the following criteria: (Bourcier et al., 2003; Ruiz Caro et al., 2017; Chidambaram et al., 2018) ① a history of corneal abrasion during contact lens wearing; ② diabetes mellitus; ③ immunosuppressive treatment; ④ sudden onset of ulceration; and ⑤ an ulcer with a well-defined border on slit-lamp microscopy (except for wreath-like infiltrates in Nocardia keratitis) ( Figure S1A ).

Patients were suspected of having fungal keratitis (FK) if they conformed to two or more of the following conditions: (Bharathi et al., 2003; Jain et al., 2007; Alkatan and Al-Essa, 2019; Ren et al., 2020) ① a history of eye trauma related to plants or soil; ② chronic systemic or local use of antibiotics or corticosteroids; ③ corneal ulceration presenting with a dry appearance; ④ elevated lesions; ⑤ satellite lesions or pseudopods; and ⑥ presence of dense hypopyon or endothelial plaques on slit-lamp microscopy ( Figure S1B ).

Cases with any two or more of the following conditions were regarded as suspected viral keratitis (VK): (Kim et al., 2018; Valerio and Lin, 2019; Erdem et al., 2020), ① repeated recurrences; ② constitutional symptoms of viral infection, such as influenza symptoms and fever; ③linear and branching corneal dendrites with terminal bulbs; ④irregular geographic ulcer or iris atrophy on slit-lamp microscopy ( Figure S1C ); ⑤ feeble response to empiric antibacterial, anti-fungal, and anti-parasitic treatment; ⑥ abundant blisters under the corneal epithelium or dendritic cells on in vivo confocal microscopy (IVCM); and ⑦ characteristics of viral infection in tissue sections of the corneal lesions stained with hematoxylin-eosin.

During the recruiting, based on diagnostic and exclusion criteria, all definitely diagnosed patients with MK and healthy volunteers were included in this study. Subjects with any recent diagnosis and treatment at other eye care institutions, recent usage of any eyedrops (such as artificial tears and low-concentration atropine), any systemic drug therapy, concurrent eye diseases, or any previous oculopathy or ophthalmic/corneal refractive surgery were excluded.

For patients with MK, the conjunctival cotton swab from all diseased eyes were collected and the fellow eyes were untreated. Subjects lay supine in a disinfected operating room (conforming to GB50333-2013-I standard of China), with the facial area except the examined eye covered with a sterile surgical drape after eyelid disinfection with iodophor. A drop of oxybuprocaine hydrochloride eye drops was used for anesthesia before sampling. (Ren et al., 2020) The superior palpebral conjunctiva, superior bulbar conjunctiva, inferior bulbar conjunctiva, inferior fornical conjunctiva, and inferior palpebral conjunctiva of the examined eye were rubbed lightly in sequence using one sterile cotton swab (Katzka et al., 2021) (from the supply department of Qingdao Eye Hospital of Shandong First Medical University according to T/CSBME013-2019 standard of China) soaked with oxybuprocaine hydrochloride eye drops. (Ren et al., 2020) The meibomian orifices, eyelashes, and cornea were avoided.

For control group, a randomly chosen single eye of each healthy subject was treated in the same way. Besides, three sterile swabs were exposed to laminar flow for 3 minutes; after intraday samples were obtained, the remaining oxybuprocaine hydrochloride eye drops were sent for contamination analysis by high-throughput 16S rRNA gene sequencing. If negative results were generated, the collected swabs were used for DNA extraction and sequencing.

The relative content of each bacterial genus and the composition of bacterial gene functions in a sample were used to establish identification models based on the random forest algorithm, which can combine numerous randomized decision trees and aggregate all predictions by averaging. (Biau and Scornet, 2016) The random forest algorithm has excellent performance when variables are much larger than the number of observations. (Biau and Scornet, 2016) Moreover, it is versatile enough to be applied to large-scale problems. (Biau and Scornet, 2016)

Tenfold cross-validation was adopted to reduce the risk of overfitting and bias. (Lin et al., 2020) Briefly, all samples were randomly divided into 10 subsamples with stratified sampling. One single subsample served as the testing data; the remaining 9 subsamples were retained as the training data to construct an identification model. The process was repeated 10 times (folds) and generated an average accuracy to evaluate the final identification model, with each subsamples used exactly once as the testing data.

A confusion matrix, which helps to quickly visualize the proportion of categories that are misclassified into other categories, and a receiver operating characteristic (ROC) curve were used to intuitively demonstrate the performance of the identification models. For a ROC curve, the closer the area under the ROC curve (AUC) is to 1, the better the authenticity of the diagnostic model is.

The precision, recall, and F1-score were employed for statistical analyses of each class. Precision is the ratio of the number of correctly classified samples to the total number of classified samples [True Positive/(True Positive + False Positive)]. Recall is the ratio of the number of correctly identified samples to the number of samples that should have been identified [True Positive/(True Positive + False Negative)]. The F1-score is generally used to measure the accuracy of unbalanced data, and it takes into account both the accuracy and recall of classification models. That is, the F1-score can be regarded as a weighted average of model accuracy and recall rate, with a maximum value of 1 and a minimum value of 0.

Macro-F1, micro-F1, weight-F1, sum recall and sum accuracy were used to statistically describe the performance of the established model. Macro-F1 is the average value of the F1-score in all classes. Micro-F1 is suitable for the condition of multicategory imbalance, in which data are extremely imbalanced and will affect the results. Weighted-F1 is the average of each F1-score multiplied by the proportion of the corresponding class. Sum accuracy indicates the ratio of the number of samples correctly classified by the model to the total number of samples.

According to the results of traditional diagnostic approaches to MK (representative images are presented in Figure S1 ), a total of 149 conjunctival swabs were obtained for further study, including 69 from healthy eyes, 22 from eyes with BK, 35 from eyes with FK, and 23 from eyes with VK. The detailed information for each subject is shown in Table S1 . The total number of OTUs in this study was 28,659, with the number of OTUs in each sample ranging from 232 to 2873. The content and representative sequence of each OTU in each sample are shown in Supplementary Data . The relative bacterial compositions of each sample are presented in Figure 1 .

The Shannon–Wiener index, an estimator of species richness and evenness that delineates within-community characteristics, differed significantly (P=2.29e−10, Kruskal–Wallis) among the four groups ( Figure 2A ). Based on dimensionality reduction, the PCoA plot (Bray Curtis algorithm; P=0.001, R^{2 =} 1, F. Model=8.269, total SumsOfSqs=43.842, and Df=148), an approach to exhibit between-community characteristics, revealed samples from the eyes with BK, FK and VK gathered in a specific region ( Figure 2B ). Figure 2B also illustratively suggested the conjunctival bacterial microbiota in the three types of MK presented minor intergroup differences but major between-group differences.

Furthermore, based on the Kruskal–Wallis algorithm, we used PICRUSt2 with KEGG database results to screen differentially expressed genes in different samples and to generate a heatmap ( Figure S2 ). The weighted nearest sequenced taxon index (NSTI) of each sample is given in Table S3 , with a mean value of 0.13 ± 0.02. Although the predictive accuracy may not be ideal, the heatmap clustering suggests that the three groups present a tendency of gathering related to grouping. As a result, the conjunctival bacterial microbiota may be used to distinguish MKs.

Two machine learning models were established in this study based on the relative composition of conjunctival bacterial microbiota ( Supplementary Data – Model 1 ) and the composition of bacterial gene functions ( Supplementary Data – Model 2 ). The sum accuracy of the former model was 96.25% ( Figure 3 ); that of the latter model was 93.75% ( Figure S3 ). The evaluation indexes of the machine learning model are provided in Table 1 . As these shown, these two models can both effectively distinguish different types of MK.

The NSTI quantifies how closely a genome database represents the community of OTUs. (Douglas et al., 2018) As PICRUSt’s inference accuracy has negative relevance with regard to NSTI values, the weighted NSTI value can be applied to determine the confidence in functional inferences. (Douglas et al., 2018) The accuracy of PICRUSt decreases rapidly with increasing weighted NSTI within an NSTI below 0.2, specifically, R² drops below 0.5 for genomes with weighted NSTIs above 30%. (Louca et al., 2018) According to the estimation from Stilianos Louca et al., (Louca et al., 2018) the accuracy of PICRUSt is approximately 70% when the weighted NSTI is 0.13 ± 0.02. Therefore, the relatively low accuracy of the model based on the composition of bacterial gene functions, to some extent, can be attributed to the substantial additional errors imported during the procedure of PICRUSt.

According to LEfSe ( Figure 4 ), Actinobacteria was the biomarker (LDA score >4) in the control group. Alphaproteobacteria, Sphingobacteriales, Sphingobacteriia, Sphingomonadales, Chitinophagaceae, Sphingomonadaceae, and Sphingomonas were biomarkers in the BK group. Bacteroidales, Bacteroidia, Campylobacterales, Epsilonproteobacteria, Helicobacteraceae, Helicobacter, Lactobacillales, Lachnospiraceae, Lactobacillaceae, Lactobacillus, Ruminococcaceae, Staphylococcaceae, and Staphylococcus were biomarkers in the FK group, and Clostridia, Clostridiales, Bacillaceae, Prevotella, and Clostridiaceae were biomarkers in the VK group. Interestingly, the average relative content of Actinobacteria was reduced by 35.84%, 22.41% and 40.76%, respectively, when the above three types of MK occurred.

Figure 5 reveals a predicted intricate interaction network of the top 30 bacterial genera based on the Spearman correlation coefficient. Lactobacillus, Clostridium, Helicobacter, and Sphingomonas were speculated to have significant relatedness (|SpearmanCoef|> 0.8 and P <0.01) in the interaction network of the conjunctival bacterial microbiome. Lactobacillus exhibited a strong negative correlation with Clostridium, whereas Helicobacter might exert a synergistic effect on Lactobacillus.

The current clinical application of high-throughput 16S rRNA gene sequencing generally aims to verify bacterial infection and to reveal causative bacteria in airtight tissue, such as articular cavity and cardiac valve tissue. (Vondracek et al., 2011; Larsen et al., 2018) However, high-throughput 16S rRNA gene analysis cannot verify bacterial infection in tissues with a microbiota. High-throughput 16S rRNA gene sequencing of conjunctival swabs can achieve a nearly one hundred percent success rate, while the positivity rate of bacterial culture for the ocular surface is only approximately 60% and that of fungal culture ranges from 3% to 65%. (Huang et al., 2016; Wang et al., 2019; Ren et al., 2020) In addition, conjunctival and corneal microbiome have very similar microbial taxonomic profiles with different proportion. (Matysiak et al., 2021) Besides, previous studies have confirmed that different types of pathogens can invoke various ocular pathophysiological responses. (Suzuki et al., 2010; Lobo et al., 2019) Our present work is thus aimed to extend the clinical application of high-throughput 16S rRNA gene sequencing, especially for the tissue diagnosis associated with a microbiota.

In this study, BK, FK, and VK impose diverse selective pressures on the conjunctival bacterial microbiome, and specific bacteria may survive in such a specific circumstance. Besides, the conjunctival bacterial microbiota in three types of MK presented minor intergroup differences and major between-group differences, though many other factors are likely involved, including age, (Zhou et al., 2014; Wen et al., 2017) gender, (Wen et al., 2017) season, (Zhou et al., 2014) and contact lens wear (Shin et al., 2016). That is, the selective pressures of different types of corneal infection on the conjunctival bacterial microbiome appear to be more influential than the aforementioned individual factors, which also suggests microbiome-based machine learning possesses commendable robustness (Lesne, 2008) for differentially diagnosing MK.

A previous study has shown that well-covered human samples (such as gut samples) possess the lowest NSTI (mean NSTI = 0.03 ± 0.02) but that the NSTI of soils is 0.17 ± 0.02 and that of other mammal samples is 0.14 ± 0.06. (Langille et al., 2013) That is, PICRUSt does not offer outstanding performance for conjunctival samples, as it does for gut samples. Undoubtedly, shotgun metagenomic sequencing or metagenomic whole-genome sequencing can theoretically generate a more accurate deduction of bacterial gene functions. Nevertheless, host DNA contamination can overwhelm the low biomass of microbial signals and decrease sensitivity for microbial detection, (Heravi et al., 2020) especially for infected ocular samples, which usually contain abundant host cells. As a result, though the accuracy of the identification models using the bacterial gene functional composition inferred by PICRUSt are conservative, the identification models using the structure of the bacterial microbiota may have a broader practicability in clinical application.

In 2014, Yoshio Nakano et al. (Nakano et al., 2014) reported the first combined clinical application of the microbiota and machine learning, presenting a model to classify oral malodor based on the microbiota in saliva samples. Two years later, Kentaro Iwasawa et al. (Iwasawa et al., 2018) found that dysbiosis of the salivary microbiota can be used to diagnose pediatric-onset primary sclerosing cholangitis. Since 2020, previous studies have combined the microbiota and machine learning for the disease diagnosis, classifying and prognosis forecasting. (Willis et al., 2020) Regardless, the present study adds to our knowledge by first using this approach to distinguish various types of infections in tissues that have a microbiota.

With the development of nanopore sequencing, on-site sequencing has gradually been implemented in clinical applications. For example. MinION is a portable (only weighs 90 g and can plug into any computer with a standard USB 3.0 port), real-time device for DNA and RNA sequencing. In addition, the identification model of machine learning has a very low marginal cost (an increase in the total cost caused by newly generated results). (Ting et al., 2017) Therefore, approaches that combine the microbiota and machine learning are promising with regard to accessibility and affordability.

To better describe and analyze the microorganisms and key genera in an environment, researchers proposed the concept of the “core microbiome”, which means that the existence of microorganisms in this group does not depend on factors such as the environment, lifestyle or physiological differences. (Turnbaugh et al., 2007; Willcox, 2013; Huang et al., 2016) Initially, the conclusion of core microbiome was deduced mainly based on the conjunctival microbiome on healthy eyes. For example, our research team (Huang et al., 2016) identified 10 genera (Corynebacterium, Pseudomonas, Staphylococcus, Acinetobacter, Streptococcus, Millisia, Anaerococcus, Finegoldia, Simonsiella and Veillonella) constituting the conjunctival core microbiome; Dong et al. (Dong et al., 2011) found 12 genera (Pseudomonas, Propionibacterium, Bradyrhizobium, Corynebacterium, Acinetobacter, Brevundimonas, Staphylococci, Aquabacterium, Sphingomonas, Streptococcus, Streptophyta, and Methylobacterium) constituting the conjunctival core microbiome, and Ham et al. (Ham et al., 2018) found 8 (Corynebacterium, Streptophyta sp, Bradyrhizobiaceae sp, Sphingomonas, Ralstonia, Neisseriaceae sp, Acinetobacter, and Pseudomonas). However, with the development of microbiome research, the concept of the core microbiome has gradually evolved from the initial “shared members of microbial communities in different samples” to “members of microbial communities that play a key role in the local ecosystem, host-microbial interactions, microbiome function, and microbiome persistence and stability”. (Chun-Bo et al., 2019) It seems that the conjunctival core microbiome needs to be defined again to determine at least some components that can guarantee ocular surface homeostasis. (Aragona et al., 2021)

Invoking the community theory of agricultural ecosystems, microbial taxa within complex ecological networks are usually categorized by their universal roles based on their level of connectivity with other taxa. (Toju et al., 2018) Highly connected taxa (kinless hubs) within an ecological network are theoretically expected to support higher levels of ecosystem functions than less connected taxa (peripherals). (Shi et al., 2020) Kinless hubs are considered to create ecological niches, i.e., the minimum habitat and resource thresholds necessary for the survival of each microorganism, for other taxa. (Shi et al., 2020) As a result, kinless hubs are of paramount importance to maintain productivity in human-managed agricultural ecosystems such as cropland and orchard. In consideration of the limitation of the concept of “core microbiome”, this study aimed to introduce the concept of “kinless hubs” and “peripherals” into the human microbiome for the first time. Accordingly, we deduced the bacterial interaction network on the conjunctiva as well as significant bacterial genera within the network according to alteration of the conjunctival bacterial microbiome related to different types of infection. Actinobacteria, Lactobacillus, Clostridium, Helicobacter, and Sphingomonas were found to constitute the core interaction network of the human conjunctival bacterial microbiome and can be considered kinless hubs of the conjunctival microbiota.

Actinobacteria, a common bacterial genus, can synthesize numerous secondary metabolites that have antimicrobial, antiviral, antiparasitic, antioxidant, anticancer, and neurological activities. (Valliappan et al., 2014; Dalitz et al., 2017) As shown by LEfSe, compared with healthy subjects, Actinobacteria was significantly reduced in the conjunctival microbiota of MK eyes, and our previous study (Ren et al., 2021) and that of Shivaji et al. (Shivaji et al., 2021) also reported that Actinobacteria is significantly reduced in the conjunctival microbiota of the diseased eyes of BK patients. Hence, Actinobacteria may have a potential value on the conjunctiva, similar to the performance of Corynebacterium mastitidis on the mouse ocular surface, which is a validated ocular probiotic for mouse that can tune ocular immunity and protecting the eye against pathogenic infection. (St Leger et al., 2017)

Furthermore, based on Spearman correlation coefficient analysis, Sphingomonas, as a biomarker for BK, Lactobacillus and Helicobacter, as biomarkers for FK, and Clostridium, as a biomarker for VK, might have significant relationships in the interaction network of the conjunctival bacterial microbiome. According to a previous in vitro study, (Monteiro et al., 2019) Lactobacillus has antimicrobial activity against Clostridium, which is in line with the predicted interaction network in our series. Helicobacter can both cause peptic ulcer disease and alleviate Barrett’s esophagus, and the predicted interaction network showed that it may cooperate with Lactobacillus, which is recognized as a probiotic. (O’Callaghan and O’Toole, 2013; Malnick et al., 2014; Chen et al., 2019) Nonetheless, the network in this study has not been fully verified in vitro and in vivo and further investigation is required. A completely decoded interactional network of the ocular surface microbiome will definitely better facilitate the prevention, diagnosis, and treatment of eye diseases.

There are limitations in this study. First, the sample size was inadequate to establish a versatile identification model for actual clinical application concerning different geographic areas or human races. Second, Acanthamoeba, an infrequent pathogenic wate-rborne parasite of MK that accounts for less than 1%, was not identified. (Zeng et al., 2008; Cohen, 2015; Austin et al., 2017) Third, the degree of acceptance of artificial intelligence algorithms in physicians and patients remains unclear. (Xiang et al., 2020) However, with microbiome data from around the world being deposited in public repositories, an ideal program for clinical use is promising.