Current uses of artificial intelligence in the analysis of biofluid markers involved in corneal and ocular surface diseases: a systematic review

Abstract

Corneal and ocular surface diseases (OSDs) carry significant psychosocial and economic burden worldwide. We set out to review the literature on the application of artificial intelligence (AI) and bioinformatics for analysis of biofluid biomarkers in corneal and OSDs and evaluate their utility in clinical decision making. MEDLINE, EMBASE, Cochrane and Web of Science were systematically queried for articles using AI or bioinformatics methodology in corneal and OSDs and examining biofluids from inception to August 2021. In total, 10,264 articles were screened, and 23 articles consisting of 1058 individuals were included. Using various AI/bioinformatics tools, changes in certain tear film cytokines that are proinflammatory such as increased expression of apolipoprotein, haptoglobin, annexin 1, S100A8, S100A9, Glutathione S-transferase, and decreased expression of supportive tear film components such as lipocalin-1, prolactin inducible protein, lysozyme C, lactotransferrin, cystatin S, and mammaglobin-b, proline rich protein, were found to be correlated with pathogenesis and/or treatment outcomes of dry eye, keratoconus, meibomian gland dysfunction, and Sjögren’s. Overall, most AI/bioinformatics tools were used to classify biofluids into diseases subgroups, distinguish between OSD, identify risk factors, or make predictions about treatment response, and/or prognosis. To conclude, AI models such as artificial neural networks, hierarchical clustering, random forest, etc., in conjunction with proteomic or metabolomic profiling using bioinformatics tools such as Gene Ontology or Kyoto Encylopedia of Genes and Genomes pathway analysis, were found to inform biomarker discovery, distinguish between OSDs, help define subgroups with OSDs and make predictions about treatment response in a clinical setting.

Introduction

Ocular surface diseases (OSD) are conditions affecting corneal and conjunctival structures, tear film characteristics and production, and adnexal gland functions [1, 2]. OSDs are not only associated with significant psychological burden and poor self-perceived health status[3], but also poses a significant economic burden to the individual and society, such as decreased work productivity, absenteeism, and costs of physician visits, ocular lubricants, punctual plugs, and more, reported in the United States [4], Canada [5], and China [6].

Tear fluid homeostasis is central to providing lubrication and nutrients to the ocular surface and is composed of various enzymes, growth hormones, lipids, salts, neuropeptides, mucins [7], which are produced by lacrimal glands, meibomian glands, conjunctival goblet cells, corneal epithelial cells, and vascular sources [8]. The complex protein and metabolite tear film content facilitates a dynamic, wide ranging, individually tailored response to infection and other abnormalities affecting the ocular surface. Tears, which can easily and non-invasively be collected in clinic, have been used to discover biomarkers for determining disease aetiology and risk factor, conversion, severity, or prognosis, and treatment strategy and outcomes, using proteomics [7–9] and metabolomics [10, 11]. Several studies have made use of differences in tear proteomes of various OSDs and corneal diseases such as aqueous deficient dry eye and Meibomian gland dysfunction (MGD) [12], or keratoconus, pterygium, graft-versus-host-disease, and controls [7], to identify differentially expressed proteins and evaluate them as potential biomarkers for diagnosis and treatment.

Aqueous humour obtained in surgery (e.g. keratoplasty, phakic intraocular lens implantation) has been found to correlate with disease progression in keratoconus [13]. For example, abnormal expression of proteome measured via liquid chromatography with tandem mass spectrometry (LC-MS/MS) and analysed by hierarchical clustering, principal component analysis, functional interaction sub-networks, and Gene ontology (GO) analysis were implicated in corneal proteolysis, regulation of hypoxia, of fibrinolysis, response to calcium ions, platelet activation, etc. (e.g., haemoglobin subunit beta, haptoglobin, Ig kappa chain V-I region EU) [13].

Generally, artificial intelligence (AI) refers to the capability of computing systems for pattern recognition, and for reproducing human cognitive characteristics (e.g., generalize, and learn from experience) in large datasets [14]. Machine learning (ML), a type of AI, can be used to extract generalized principles from data to make predictions or classifications by applying algorithms and mathematical modelling based on explicit rules and instructions about the data [15].

With its impressive power to identify patterns, classify, cluster, or make predictions from large datasets, AI is well suited to analyse the massive data output produced by continually advancing novel analytical technologies such as proteomics and metabolomics. In proteomics, the data regarding proteins expression in ocular fluids is analysed using AI and compared against databases containing large amounts of labelled protein sequence information [16]. The end result is a proteomic signature or profile of the fluid, which can not only elucidate molecular mechanisms of ocular diseases, but also be used to diagnose disease or monitor the outcome of therapeutics [17–19]. Metabolomics involves the large-scale study of endogenous and exogenous metabolites in various tissues as to provide an assessment of the metabolic phenotype of a certain state of disease, and in combination with AI/bioinformatics can be used to obtain putative metabolic pathways and biomarkers associated with disease mechanisms and treatment strategies at the level of the individual patient [10, 11, 20]. The combination of these methods has driven major advancements in precision medicine by allowing examination of individual variability in disease prognosis and individualized treatment strategies [18, 21]. As such, a growing number of ophthalmology studies have adopted these methods to analyse biofluids as biomarkers.

Exploration of biofluids using AI and bioinformatics may offer insight into pathophysiology, prognosis, and fuel the discovery of new therapies of OSDs and corneal diseases. Therefore, the current study aims to systematically review the literature describing application of AI and bioinformatics-based analyses using biofluids as biomarkers in OSDs and corneal diseases. The methodology and findings of eligible studies are summarized and appraised with a focus on assessing the potential of clinical implementation of these approaches.

Methods

Study design and registration

The findings from this systematic review are reported in accordance with the Preferred Reporting Items for a systematic Review and Meta-analysis (PRISMA) guidelines [22]. Study protocol details were prospectively registered on PROSPERO (reg. CRD42020196749). The current systematic review is focused on OSDs and corneal diseases and is a part of a series of systematic reviews on analysis of biofluids using AI for various specific eye conditions in ophthalmology, which are reported elsewhere.

Search strategy

Systematic searches of the literature were conducted in five databases including Embase, MEDLINE, Cochrane Central Register of Controlled Trials, Cochrane Database of Systematic Reviews, and Web of Science from the time of database inception through August 11, 2020, with an update of the search strategy performed on August 1, 2021. A comprehensive set of search terms capturing three categories including ophthalmology, AI/bioinformatics, and proteomics/metabolomics/lipidomics terms were used to construct the search strategy (Appendix A). The search was not restricted by language or study design. Hand-searching of the reference list of included studies was also performed in order to identify relevant articles.

Inclusion and exclusion criteria

Studies were included if they referred to intra-ocular or ocular surface conditions using biofluid marker samples to make AI/bioinformatic-based predictions about disease aetiology or risk factors, treatment outcomes or treatment strategies, and diseases conversion or progression. Samples of biofluids from vitreous, aqueous, or tear fluid, as well as plasma or ophthalmic biopsies were deemed eligible. Studies were excluded if they referred exclusively to paediatric eye diseases, non-human subjects, or included only post-mortem biofluid samples. Additionally, we excluded cross-sectional studies that only used the simplest form of AI (simple regression analysis). Abstracts, reviews, systematic reviews, meta-analyses, single case reports, editorials (without adequate study details and data presentation), and any type of non-peer reviewed article were considered ineligible. Lastly, the subset of studies that met the inclusion criteria, and referred to OSDs or corneal diseases were selected for the current review.

Study selection

The titles and abstracts, and then the full texts were independently screened by two review authors (DRP, AP) for relevant articles. Title and abstract screening included any literature that focused on any OSD and biofluid sampling. At this point, articles were included even if it was not clear if an AI analysis was performed. During full-text screening, any articles that did not meet all our specified inclusion criteria were excluded. If a consensus for conflicts could not be reached between the two reviewers, a third reviewer (SK or TF) resolved the conflict.

Data collection and risk of bias assessment

Data extraction of included studies was undertaken by one reviewer (DRP) using a standardized data abstraction form. To ensure accuracy and consistency of the extraction process, 10% of extractions were randomly double-abstracted by a second independent reviewer (AP or SK). Risk of bias (ROB) and quality assessment of retrieved studies were performed using the Joanna Briggs Institute Critical Appraisal Tools (JBI) [23]. For each article, JBI criteria questions were noted as “yes”, “no”, “unclear”, or “not applicable.” The assessment was performed by one reviewer (DRP) and none of the studies were excluded from the review. Studies that reached up to 49% of questions as “yes” were classified as high ROB; from 50 to 69% as moderate ROB; and more than 70% as low ROB [24].

Data synthesis

There was substantial heterogeneity in biofluid types, AI techniques, and study designs, and consequently a meta-analysis was not undertaken. Means and standard deviations (SD) were used to characterize the study sample(s) age(s). The study characteristics tabulated included study design, location, type of OSD, sample size, sex ratio, study aim, fluid collection methods, and a list of biofluids reported. Articles were further categorized according to OSDs or corneal disease type, statistical model, AI, or bioinformatics analyses performed. Moreover, the AI/Bioinformatics methodology purpose was noted.

Results

Study characteristics

The search strategy resulted in 10,264 articles after removal of duplicates (Fig. 1). Of the 23 articles that were found eligible for inclusion, 7 were prospective (30%), 16 were cross-sectional (70%), and one was a randomized controlled trial (Table 1). There was a global distribution in the country of origin of the included studies with China (4, 17%) and Spain (3, 13%) being the most common. There were 1058 individuals included, with 350 individuals with dry eye, 61 with keratoconus, 43 with pterygium, 179 with meibomian gland dysfunction (MGD), 59 with graft-versus-host-disease (GVHD), 51 with Sjogren, 2 with climatic droplet keratopathy (CDK), 19 with bullous keratopathy, 2 with Fuchs’ endothelial dystrophy, 18 with vernal keratoconjunctivitis (VKC), 12 with various indications for penetrating keratoplasty, and 237 healthy controls, as well as 5 myopic and 20 diabetic individuals as comparator groups.

Fig. 1. PRISMA flowchart diagram for study identification and selection.

The PRISMA flow diagram for the systematic review presenting the number of studies included and excluded at each screening step, and reasons for exclusion.

Table 1.

Summary of general characteristics of studies included in the review.

Author(s), (Publication Year)	Study design*	Country	Eye disease	Sample size	Mean age (SD)	Sex (males / females)	Study aim classification
Dry Eye, Sjögren’s Syndrome, Meibomian Gland Dysfunction
Aqrawi et al., (2017)	Cross-sectional	Norway	Sjögren’s	pSS: 27 C: 32	pSS: 52.40 (12.22) C: -	pSS: - C: -	Biomarker discovery Identification of pathophysiology
González et al., (2020)	Prospective case-controlled	Spain	DE MGD	DE: 29 MGD: 27 CT: 37	DE: 52.1 (13.5) MGD: 53.4 (15.2) CT: 40.2 (12.5)	DE: 10/19 MGD: 11/16 CT: 13/24	Biomarker discovery Identification of pathophysiology
Grus et al., (2005)	Cross-sectional	Germany	DE	DE: 88 CT: 71	–	–	Biomarker discovery Identification of pathophysiology
Huang et al., (2018)	Cross-sectional	China	DE	–	–	–	Biomarker discovery Identification of pathophysiology
Ji et al., (2019)	Prospective cohort	South Korea	DE	CsA: 9 DQ3: 9	CsA: 46.2 (1.4) DQ3: 53.3 (15.5)	CsA: 4/5 DQ3: 4/5	Disease severity Treatment outcome
Jiang et al., (2020)	Cross-sectional	China	DE	DE: 85 CT: 28	DE: 55.4 (8.8) CT: 60.8 (11.2)	DE: 50/35 CT: 15/13	Biomarker discovery Identification of pathophysiology
Piyacomm et al., (2019)	Prospective RCT	Thailand	MGD	IPL: 57 Sham: 57	IPL: 59.0 (12.7) Sham: 59.5 (11.4)	IPL: 10/47 Sham: 5/52	Treatment outcome
Sembler-Møller et al., (2020)	Cross-sectional	Denmark	Sjögren’s	pSS: 24 C: 16	pSS: 55 (11) C: 53 (16)	pSS: 2/22 C: 2/14	Biomarker discovery Identification of pathophysiology
Soria et al., (2013)	Cross-sectional	Spain	DE MGD CT	DE: 63 MGD: 38 CT: 43	DE: 55.3 (14.1) MGD: 63.4 (16.6) CT: 42.7 (14.0)	DE: 31/32 MGD: 22/16 CT: 27/16	Biomarker discovery Identification of pathophysiology
Srinivasan et al., (2012)	Cross-sectional	Canada	DE	NDE:6 MDE:6 MSDE:6 MXDE: 6	NDE: 29.8 (8.1) MDE: 59.6 (16) MSDE: 45.2 (10.5) MXDE: 36.7 (17)	NDE: 2/4 MDE:2/4 MSDE:2/4 MXDE:4/2	Biomarker discovery Identification of pathophysiology
Tong et al., (2017)	Prospective cohort	Singapore	DE	23	49.8 (14)	6/23	Treatment outcome
Zou et al., (2020)	Cross-sectional	China	Adult DM w/ DE Child DM w/ DE Adult DM Child DM	Adult DM w/ DE: 10 Child DM w/ DE:10 Adult DM:10 Child DM:10 Adult CT:10 Child CT:10	Adult DM w/ DE: 58.8 (4.3) Child DM w/ DE: 11.7 (2.8) Adult DM: 57.7 (7.2) Child DM: 12 (3.3) Adult CT: 58 (4.3) Child CT: 11.2 (1.3)	Adult DM w/ DE: 6/4 Child DM w/ DE: 6/4 Adult DM: 7/3 Child DM: 6/4 Adult CT: 7/3 Child CT: 6/4	Biomarker discovery Identification of pathophysiology
Keratoconus and other Corneal Diseases
Borges et al., (2020)	Cross-sectional	Germany	KC Pterygium GVHD	KC: 4 Pterygium: 9 GVHD: 10 CT: 6	KC: 30.5 Pterygium: 47.2 GVHD: 49.6 CT: 47.5	KC: 2/2 Pterygium: 6/3 GVHD: 3/7 CT:1/5	Biomarker discovery Identification of pathophysiology
Fodor et al., (2009)	Prospective cohort	Hungary	PKP‡	PKP: 12	PKP: 45 (14.2)	PKP: 8/4	Treatment outcome Prognosis
Fodor et al., (2021)	Prospective cohort	Hungary	KC	KC: 45	KC: 34 (12.3)	KC: 30/15	Prognosis
Kim et al., (2014)	Cross-sectional	South Korea	Pterygium	Pterygium: 24 HCC: 24	Pterygium: 49 (5.2) HCC: 49 (5.2)	Pterygium: 10/14 HCC: 10/14	Biomarker discovery Identification of pathophysiology
Leonardi et al., (2014)	Cross-sectional	Italy	VKC	VKC: 18 C: 10	VKC: 10.06 (4.76) C: -	VKC:16/2 C:-	Biomarker discovery Identification of pathophysiology Treatment outcome
Linghu et al., (2017)	Cross-sectional	China	Pterygium	Pterygium: 10	Pterygium: 52	Pterygium: 6/4	Risk factors
Menegay et al., (2008)	Cross-sectional	U.S. Germany	CDKDE	CDK: 2 C: - DE: 88 CT: 71	CDK: 69.5 (3.54) C:--	CDK:2/0 C:--	Biomarker discovery Identification of pathophysiology Biomarker discovery Identification of pathophysiology
O’Leary et al., (2020)	Cross-sectional	Switzerland	oGVHD	NIH 0: 14 NIH 1: 9 NIH 2: 16 NIH 3: 10	NIH 0: 56.1 (9.6) NIH 1: 48.4 (15.4) NIH 2: 52.6 (14.0) NIH 3: 52.6 (15.2)	NIH 0: 9/5 NIH 1:7/2 NIH 2: 9/7 NIH 3:10/1	Treatment outcome Prognosis
Soria et al., (2015)	Cross-sectional	Spain	KC	KC: 5 Myopic: 5	KC: 34.2 (9.6) Myopic: 36 (7.5)	KC: 4/1 Myopic: 3/2	Biomarker discovery Identification of pathophysiology
Wojakoswka et al., (2020)	Cross-sectional	Poland	KC	KC: 7 C: 6	KC: 42–59 C: 40–69	-	Biomarker discovery Identification of pathophysiology
Yawata et al., (2020)	Prospective cohort	Japan	BK FED	BK: 19 FED: 2	BK: 69.8 (15.1) FED: 73.2 (11.3)	BK: 9/10 FED:2/0	Treatment outcome Prognosis

CsA topical cyclosporine A, DQS diquafosol tetrasodium, NDE no symptoms and sign, MDE mildly symptomatic with aqueous deficiency, MSDE symptomatic aqueous deficiency, MXDE combination group, KC keratoconus, GVD graft-versus-host-disease, MGD meibomian gland dysfunction, DE dry eye, DM diabetes, w/ with, CDK climatic droplet keratopathy, BK bullous keratopathy, FED Fuchs’ endothelial dystrophy, PKP penetrating keratoplasty, VKC vernal keratoconjunctivitis, IPL intense pulsed light, oGVHD ocular graft versus host disease, NIH National institute of health, NIH 0 normal no symptoms, NIH 1 mild no effect on activities of daily living, hydrating drops <3 time/day, NIH 2 moderate, some effect on activities of daily living, loss of vision caused by keratopathy, pSS primary Sjögren’s syndrome (pSS), HCC healthy conjunctiva from the same patient who underwent pterygium excision.

‡PKP patients with various indications: bullous keratopathy (1), keratoconus (3), salzmann’s nodular degeneration (2), herpes keratitis, transplant rejection (2), Haab-Dimmer dystrophy, recurrence of dystrophy (2), chronic superficial keratitis (pannus) (1), bullous keratopathy, transplant rejection.

Controls refers to healthy individuals unless otherwise specified.

Validation refers to an additional analysis based on discovered candidate proteins performed on a new sample.

Treatment outcome refers to articles that analysed biofluids using AI/bioinformatics with the purpose of predicting treatment responses.

Biomarker discovery or identification of pathophysiology refers to articles that analysed biofluids using AI/bioinformatics with the purpose of identifying candidate markers for ocular surface disease pathogenesis, classify ocular surface diseases based on identified biofluids, or classify different subgroups within one ocular surface disease category based on biofluids.

Majority of studies focused on biomarker discovery and identification of pathophysiology of OSDs (15, 65%), while eight (35%) assessed treatment outcomes or prognosis, and one assessed risk factors related to OSD.

The risk of bias (ROB) assessment is presented in Appendix B. Two studies were found to have a high ROB [9, 11], nine moderate ROB [7, 13, 25–31], and twelve low ROB [12, 32–41]. The main areas of bias identified among cross-sectional studies were: criteria for inclusion in the sample were not clearly defined (n = 5, 31%), the study subjects and setting was not described in detail (n = 9, 56%), confounding factors were not identified (n = 7, 44%), and strategies to deal with confounding factors were not stated (n = 9, 56%). Among cohort studies, none of them had participants that were free of the ocular disease at the start of the study. This represents a source of bias because if the samples were taken after an OSD had occurred, it is not possible to definitively conclude if the biofluids identified are contributory to the OSD and/or a reflection of the downstream consequences of the OSD. Future studies involving long-term collection of samples prior to and following disease onset may provide more definitive evidence for the associations of biomarkers with OSD pathogenesis.

Biomarkers involved in pathogenesis of dry eye disease

Upregulation of apolipoprotein [26, 27], haptoglobin [26, 27], annexin 1 [27, 34], Glutathione S-transferase [26, 27, 32, 34], and downregulation of lipocalin-1 [7, 12, 26, 27, 31, 34], prolactin inducible protein (PIP) [26, 27, 34], lysozyme C [7, 26, 27, 31, 33, 34], lactotransferrin [7, 26, 27, 34], cystatin S [7, 26, 27, 34], and mammaglobin-b [26, 34], proline rich protein [27, 31] were associated with dry eye pathogenesis. AI analyses using bioinformatics databases implicated the upregulated proteins in biological pathways regulating lipid metabolic processes, oxidation reduction, cytokine production, while the downregulated proteins were associated with transportation, and regulation of immune response [7, 12, 26, 27, 34]. A proteomic study of contributing tear film proteins to the pathogenesis of diabetic dry eye using weighted correlation network analysis, GO and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis found three differentially expressed proteins (lysozyme C, zinc-alpha-2-glycoprotein, DNA J homolog subfamily C member 3) in adults with diabetic dry eye compared to controls, and one in children (phosphoglycerate kinase 1) with diabetic dry eye compared to controls [33]. In both adults and children, these proteins were involved in dysregulation of metabolic pathways associated with inflammation and immunity such as glycolysis, pentose phosphate pathway, and proteasomes [33]. In adults, the expression levels of these proteins significantly correlated with tear film break-up time, Schirmer I test, and corneal fluorescein staining.

There were several overlapping biomarkers between MGD [34, 40], Sjögren’s [30, 42], and dry eye [7, 26, 27, 31, 33, 34], associated with pathogenesis (e.g., lipocalin-1, lysozyme C, annexin A1, cystatin S). Additionally, Sjögren’s patients presented increased expression of TNF-a signalling, B cell survival, proteins involved in the Krebs cycle, and in oxidative stress in tear fluid [30], as well as upregulation of elastase, calreticulin, and tripartite motif-containing protein [42], proteins involved in inflammation and complement coagulation cascade [42]. A prospective randomized controlled trial investigating the efficacy of intense pulsed light (IPL) for MGD using proteomic analysis found that tear level of interleukin-1 receptor agonist was significantly lower at 3 months compared to baseline in both sham and IPL groups but there were no differences between the groups [40].

Biomarkers involved in treatment response of dry eye disease

Two articles assessed the role of biofluids in predicting response to treatments for dry eye, specifically punctual occlusion [32], and diquafosol tetrasodium or topical cyclosporine A [8]. By using tear proteomics and clustering analysis of identified proteins (i.e., measured at baseline and after 3 weeks), two distinct patient profiles of treatment response emerged, and each group presented differentially expressed tear proteins (one beneficial, one inflammatory). Patients from the group with a beneficial pattern of protein expression, a reduction in inflammatory proteins (e.g., S100A9) and an increase in lacrimal proteins protective of the ocular surface (e.g., lysozyme), also presented a lower Schirmer score at baseline than the patients from the inflammatory pattern group [32]. Thereby, allowing clinicians to identify patients with low scores who may benefit from punctual occlusion, and potentially change management in those patients less likely to benefit from this treatment. Another proteomic study found that that there were treatment specific differences in tear proteome and associated biological pathways in patients treated with diquafosol tetrasodium or topical cyclosporine A despite similar clinical outcomes, with 49 proteins showing an inverse expression pattern [8].

Biomarkers involved in pathogenesis of keratoconus

Several studies performed proteomic [7, 13, 35] and metabolomic profiling [11] of biofluids to investigate their role in keratoconus pathogenesis. However, there was little overlap in discovered biomarkers. A cross-sectional study of various OSDs, found 8 potential tear biomarkers contributing to keratoconus pathophysiology, and which allowed differentiation between keratoconus, pterygium, and graft-versus-host-disease related dry eye [7].

A proteomic study of aqueous fluid from keratoconus patients obtained during keratoplasty, identified 16 out of 137 proteins related to dysregulation of apoptosis, oxidative stress, response to vitamin D, angiogenesis, as potential markers of pathological changes in keratoconus [13]. A metabolomic analysis [11] of contributing metabolites using gas chromatography and mass spectrometry (GC/MS) and unsupervised hierarchical cluster analysis identified downregulation of 13 out of 377 metabolites related to aberrations in energy production, lipid metabolism, and amino acid metabolism in the corneal buttons of keratoconus patients compared to those of healthy donors [11].

Biomarkers involved in prognosis or treatment response of keratoconus

The release of several key inflammatory (interferon gamma, IFN-y; interleukin-13, IL-13; IL-17A, chemokine C-C motif ligand 5, CCL5; matrix metalloproteinase, MMP-13; and plasminogen activator inhibitor 1, PAI-12) factors at one year follow-up were found to predict keratoconus progression in a group of 42 patients [35]. NGF and IL-13 were found to identify progression with 100% specificity and 88% sensitivity.

Biomarkers involved in other corneal diseases

Other OSDs investigated were ocular graft-versus-host-disease [7, 39], and pterygium [7, 29, 37]. Patients with chronic ocular graft-versus-host-disease experience inflammation and fibrosis of the ocular surface, in addition to severe ocular dryness [43]. A proteomic study of 785 proteins using AI tools such as random forest and penalized logistic regression, and bioinformatics tools such as GO analysis, found that disease severity of ocular graft-versus-host-disease in patients after allogenic hematopoietic cell transplantation (AHCT) could be predicted based on the differential expression of 13 biofluids (i.e., Phosphoglycerate mutase 1, Keratin type I, cytoskeletal 9) [39]. Biochemical pathways highlighted in pathogenesis were related to complement and coagulation cascades (i.e., Clusterin, Complement factor B, Complement C3, plasminogen) [7].

A prospective cohort study of endothelial keratoplasty patients examined the kinetics of their tear profiles over the course of recovery after transplant using a clustering algorithm (i.e., principal component analysis) and found alterations in the level of expression of eleven tear fluid proteins predictive for recovery from corneal haze, with the group of patients with no corneal haze within one month after surgery having significantly lower levels at the pre-transplant baseline timepoint than the group that did develop corneal haze [28]. Several inflammatory cytokines were associated with corneal graft rejection following penetrating keratoplasty in a group of 12 patients followed for 12 – 14 months after surgery [36]. Proteomic tear profiling indicated that IL-6 and IL-8 concentrations were increased in patients with rejection, while IL-10, TNF-α, and IL-12p70 were decreased compared to patients with uncomplicated corneal grafts [36].

Climatic droplet keratopathy, a degenerative disease associated with progressive accumulation of droplets on the cornea, found to be associated with 105 proteins, mainly related to cell junction function, glycolysis, focal adhesion, regulation of cytoskeleton, fibril formation and deposits (e.g., retinal dehydrogenase, aldehyde dehydrogenase, desmoplakin, etc.) when its proteome was analysed with KEGG in a case series [9].

Biomarkers involved in pathogenesis, prognosis and treatment response of VKC

In a small sample of six VKC patients responsive to treatment with cyclosporine or corticosteroids proteomic analysis with isobaric tags for relative and absolute quantification (iTRAQ) technology showed downregulation of Hemopexin, transferrin, mammaglobin B, and secretoglobin 1D [41]. These proteins were suggested to be involved in oxidative stress regulation and inflammatory response regulation [41]. Additionally, expression of tear albumin and transferrin was found to be positively correlated with VKC disease severity, and therefore may be potential biomarkers for disease diagnosis and monitoring [41].

Applications of AI and bioinformatics

As presented in Table 2, there was prominent heterogeneity in the use and reporting of AI methodology. Seventeen articles used AI and/or bioinformatics with classification algorithms, five used predictive models, and four used both classification algorithms and predictive models.

Table 2.

Characteristics of biofluids, and artificial intelligence/bioinformatics analysis methods and aims.

Author(s) year	Biofluid sample type	Biofluids (Significant/Total)	Sample collection method	Statistical/ AI Model Type	AI Application / Bioinformatic type: Bioinformatic Purpose
Dry Eye, Sjögren’s Syndrome, Meibomian Gland Dysfunction
Aqrawi et al., (2017)	Tears Saliva	Saliva: (30/500) † Ig kappa chain C region*, Calmodulin*, Annexin A1*, Alpha-enolase*, Hemopexin*, L-lactate dehydrogenase B chain*, Granulins*, Plastin-2*, etc. Tears: (197/900) † Galectin-3*, Fibrinogen beta chain*, Copine-1*, Calpastatin*, Ig gamma-1 chain C region*, Calpastatin*, Ig gamma-3 chain C region*.	Schirmer strip	Prediction Classification	Proteomics, Scaffold, STRING, GO analysis with DAVID: protein identification, functional pathway identification, protein-protein interaction
González et al., (2020)	Tears	(2/-)†† lipocalin-1*, lysozyme C.	Tear collection with glass capillaries	Classification	Multilayer perceptron neural network, stepwise discriminant analysis, nonlinear iterative partial least squares.
					Proteomics with MASCOT: identify proteins, cluster biofluids based on similarities.
Grus et al., (2005)	Tears	(7/††)? (3700 Da),? (3916 Da), nasopharyngeal* carcinoma-associated proline-rich protein*, proline-rich protein 4*, alpha-1-antitrypsin*, c-terminal fragment*, proline rich protein 3*, calgranulin A*.	Schirmer strip	Prediction Classification	Discriminant analysis, artificial neural network MLFN with back-propagation training algorithm.
					Proteomics: classify biofluids and make predictions about their function.
Huang et al., (2018)	Tears	(18/50) † Albumin*, Lactotransferrin* Lysozyme* Transferrin*, Lipocalin 1*Zinc-alpha-2-glycoprotein*, Prolactin-induced protein*, Keratin 1*, Secretoglobin*, family 2 A, member 1,* Apolipoprotein A-I*, Serpin peptidase inhibitor 1*, Polymeric immunoglobulin receptor*, Complement component 3*, S100 calcium binding protein, A8*, Haptoglobin*,S100 calcium binding protein A9*, Lacritin*, Cystatin S*, Proline rich, lacrimal 1*, Orosomucoid 1*, Keratin 10*, Keratin 2*, Clusterin*, Annexin A1*, Secretoglobin*, family 1D*, member 1 Hemopexin*, Immunoglobulin J*, polypeptide alpha-2-HS-glycoprotein*, Apolipoprotein H (beta-2-glycoprotein I)*, Heat shock 27 kDa protein 1*, Glutathione S-transferase*,Filamin A interacting protein 1-like* POTE ankyrin domain family*, member F Alpha-2-macroglobulin Transglutaminase 3*Proline rich 4 (lacrimal)*, CCCTC-binding factor (zinc finger protein)-like*	Schirmer strip	Classification	Proteomics, Proteome Discoverer, GO analysis, STRING: identify proteins and their biological processes
Ji et al., (2019)	Tears	(0/794) †Haemoglobin subunit delta, Haemoglobin subunit alpha, Haemoglobin subunit beta, Transitional endoplasmic reticulum ATPase, Vimentin, Coronin-1A, Tubulin beta-4B chain, etc.	Schirmer strip	Classification Prediction	GO with DAVID, KEGG pathway mapping, functional annotation clustering, protein-protein interaction with STRING database: identify function of biofluids, assess similarities with existing ones, and determine biological pathways.
Jiang et al., (2020)	Tears	(48/51) † Thiodiacetic acid* Uridine* Octadecanamide* Phthalic anhydride*, 3-Acrylamidopropyl trimethylammonium*, Triglyme*, N-Heptane*, 1-Piperidinecarboxaldehyde*, 2-Methylbutyroylcarnitine*, Palmitic amide Diglyme*, N-(3-Indolylac etyl)-L-isoleucine* N,N?-Dicyclohexylurea (S)-Desoxy-D2PM* Tuckolide*, Alanyl-Alanine* Dihydroterrein* Indoline*, N-methyl corydaldine (-).	Schirmer strip	Prediction	Least absolute shrinkage and selection operator regression.
					Proteomics, KEGG and Metaboanalyst: identify biofluids associated with increased risk of disease
Piyacomm et al., (2019)	Tears	(1/2) IL-Ra*, IL-6.	Schirmer strip	Prediction	Multilevel mixed-effect linear regression: predict which biomarkers are associated with treatment response
Sembler-Møller et al., (2020)	Saliva, plasma, salivary gland tissue	Saliva (40/1013) †, Matrix Gla protein*, Basic salivary proline-rich protein 1*, Basic salivary proline-rich protein 2*, Histatin-3*, Basic salivary proline-rich protein 4*, Histatin-1*, Neutrophil elastase*, Calreticulin*, Tripartite motif-containing protein 29*, Clusterin*, Vitronectin*, Catalase*, Complement factor B*, etc.; Plasma (0/219) ††; Salivary gland tissue (0/2773) ††.	Blood draw, sialometry, labial salivary gland biopsy	Classification	Hierarchical clustering, PCA.
					Proteomics, GO analysis, KEGG analysis.
Soria et al., (2013)	Tears	(5/5)100A6*, annexin A1*, annexin A11*, cystatin-S*, phospholipase A2-activatingprotein*.	Schirmer strip Glass capillaries Merocel sponge	Classification	K-nearest neighbour, support vector machine, classification trees, random forest, naive bayes.
					Proteomics using GO and DAIVD, protein-protein interactions: identification of protein function, biological process, and classification into disease groups.
Srinivasan et al., (2012)	Tears	(33/386)†Cystatin-S, Ig lambda chain C region, Lipocalin-1, Putative lipocalin 1-like protein, Secretoglobin family, zinc-alpha-2-glycoprotein, mammaglobin-b, Polymeric immunoglobulin receptor, arin, lysozyme, Zymogen granule protein-16, etc.	Schirmer strip	Classification	Proteomics with MASCOT, GO analysis: identification or proteins and their function
Tong et al., (2017)	Tears	(8/400) †† Glutathione synthetase*, IL-1RN*, ADH1C*, AGT*, CHRNA7*, HIST1H4E*, LCP1*, H3P3A*.	Schirmer strip	Classification Prediction	Hierarchical clustering, logistic regression.
					Proteomics: classify cytokines into groups based on similarity and assess how change in cytokines predicts treatment response.
Zou et al., (2020)	Tears	(3/1922) †† Adult, lysozyme C*, zinc-alpha-2-glycoprotein*, DNA J homolog subfamily C member 3*. (1/2709) †† Child, Phosphoglycerate kinase 1*.	Schirmer strip	Classification	Weighted correlation network analysis.
					Proteomics, GO/KEGG: identify and cluster biofluids based on similarities between and within groups.
Keratoconus and other Corneal Diseases
Borges et al., (2020)	Tears	(38/208 KC), (29/322 pterygium), (79/517 GVHD)†, Keratin*, type I cytoskeletal 13*, Immunoglobulin heavy vari-able 5–10–1*, Immunoglobulin heavy variable 5–51*, Proline-rich protein 27, Immunoglobulin heavy variable 3–23*, Histone H2B type 1-A*, Apolipoprotein*, prolactin-inducible protein*, S100-A8*, annexin A2*, cystatin-C*, lipocalin-1*, lysosome C*, lysosome C*,etc.	Micropipette	Classification	Partial Least Squares analysis with Metaboanalyst, PCA
					Proteomics with KEGG: identify and cluster biomarkers in different groups based on similarities.
Fodor et al., (2009)	Tears	(5/6)/IL-8*, IL-1β, IL-6*, TNF-α*, IL-10, and IL-12p70*.	Schirmer strip	Prediction	Locally weighted regression: compare cytokines of groups over time
Fodor et al., (2021)	Tears	(8/13) IL-6, IL-10, IL-13*, IL-17A*, CXCL8*, IL-8, CCL5*, RANTES, IFN-gamma*, MMP-9, MMP-13*, TIMP-1, NGF*, t-PA, PAI-1*.	Tear collection with glass capillaries	Prediction	Logistic Regression: predict change in biomarkers at follow-up
Kim et al., (2014)	Pterygium Healthy conjunctiva	(40/230) † aldehyde dehydrogenase*, dimeric NADP- preferring*, protein disulphide-isomerase A3*, peroxiredoxin-2*, Isoform 1 of Protein-glutamine gamma-glutamyltransferase 2 TYMP*, FH Isoform Mitochondrial of Fumarate hydratase,* mitochondrial IGHV4–31*, Putative uncharacterized protein IGKC Ig kappa chain C region*, etc.	Excision	Classification	Proteomics, GO with DAVID: protein classification
Leonardi et al., (2014)	Tears Serum	(3/78) † serum albumin*, lactotransferrin, lysozyme, lacritin, secretoglobin 1D, mammoglobin B, lipocalin-1, proline-rich, protein 4, cistatine-S, hemopexin*, serotransferrin*, Ig a-1 chain	Tear collection with glass capillaries	Prediction	Stepwise linear regression.
					iTRAQ proteomics with Mascot engine: identify differences in biomarker expression and predict group differences.
Linghu et al., (2017)	Ptergyia Healthy conjunctiva	(156/156) †† Fibrinogen alpha chain*, fibrinogen gamma chain*, microfibril- associated glycoprotein 4*, fibrinogen beta chain*, fibronectin1*, collagen alpha-3*, MMP-1*, −8*, −13*, MMP-3*, −10*, −21*, −22*, CD34*, CD3*.	Cohen forceps and iridodialysis spatula	Classification	Proteomics, GO analysis with DAVID, KEGG pathway analysis: identify function of proteins, putative biological pathways.
Menegay et al., (2008)	Cornea	(105/105) 14-3-3 Protein gamma, 14-3-3 Protein sigma, 14-3-3 Protein zeta/delta, 24 kDa Protein, 60 S Ribosomal protein L3, Actin, alpha skeletal muscle, Actin, cytoplasmic 1, Actin-like protein 2, Actin-related protein 2/3 complex subunit 1B, Aldehyde dehydrogenase, Alpha 3 type VI collagen isoform 1, Alpha-actinin-4, etc	Sharp scissors and fine forceps, capturing an area of the droplets	Classification	KEGG Pathway Database/ Proteomics: identify proteins against reference healthy cornea.
O’Leary et al., (2020)	Tears Serum	(13/785) Phosphoglycerate mutase 1*, Keratin type I*, cytoskeletal 9*, Keratin type 2*, cytoskeletal 1*, Fatty acid binding protein*, epidermal Profilin-1*, Immunoglobulin κ constant*, Dermicidin*, Protein S100-A4 Lysozyme C*, Polymeric immunoglobulin receptor*, Glyceraldehyde 3 phosphate dehydrogenase*, Serum albumin*, Gelsolin*.	Schirmer strip	Prediction Classification	Random forest, logistic regression.
					Proteomics, GO analysis: classification of severity, identification of predictive biofluids, functional annotation of biological pathways.
Soria et al., (2015)	Aqueous humour	(16/137) haemoglobin subunit beta*, haptoglobin*, plasma protease C1 inhibitor*, alpha-2-HS-glycoprotein*, basement membrane-specific heparan sulphate proteoglycan core protein* haemoglobin subunit delta*, carbonic anhydrase 1*, ceruloplasmin*, hemopexin*, apolipoprotein A-II*, prostaglandin-H2 D-isomerase*, actin cytoplasmic 2, semaphoring-7A, alpha-1-acid glycoprotein 1, latent transforming growth factor beta-binding protein 2, Ig kappa chain V-I region EU.	Paracentesis KC patients during keratoplasty Controls during phakic intraocular Lens implantation	Classification	PCA, hierarchical clustering, k-nearest neighbour.
					Proteomics, APEX, MASCOT with Proteome Discover, GO analysis: determine the overlap and differences in expression of biofluids between groups.
Wojakoswka et al., (2020)	Corneal buttons	(13/377) ††Benzoic acid*, Glycolic acid*, Succinic acid*, Gluconic acid*, Linoleic acid*, Myristic acid*, Palmitic acid*, Pentadecanoic acid*, Stearic acid*, trans-13-Octadecenoic acid*, Petroselinic acid*, Cholesta-3,5-diene—isomer 1*, Cholesta-3,5-diene—isomer 2*, Cholesterol*, Cholesterol propionate*, Hexadecanol*, Phosphoric acid*.	Penetrating corneal transplantation surgery	Classification	PCA, hierarchical clustering.
					Metabolomics, MSEA: classify cytokines in multiple groups and differentiate between healthy cornea and keratoconus
Yawata et al., (2020)	Tears	(11/51) IL-1a, IL-2Ra, IL-3, IL-12 (p40), IL-16, IL-18, CTAK, GRO-a, HGF, IFN-a2, LIF, MCP-1*, M-CSF, MIF, MIG*, NGF, SCF*, SCGF-b, SDF-1a, TNF-b, TRAIL, IL-1b, IL-1Ra, IL-2, IL-4*, IL-5, IL-6*, IL-7*, IL-8, IL-9*, IL-10, IL-12p70, IL-13, IL-15, IL-17, Eotaxin, FGF basic*, G-CSF, GM-CSF, IFN-g*, IP-10, MCP-1, MIP1-a, PDGF-bb, MIP-1b, RANTES, TNF-a, VEGF*, TGF-b1, TGF-b2*, TGF-b3.	Schirmer strip	Classification	PCA.
					Proteomics; classify the cytokines into multiple groups to identify common patterns.

†All available biofluids included in the respective article.

“–“ indicates that information was not available.

*That had significant implications as determined by statistical analysis.

††Full list of proteins not available

? unknown name.

MSEA metabolite set enrichment analysis, PCA principal component analysis, MLFN multiple-layer feed-forward network, Da Dalton, GO gene ontology analysis, IL-Ra interleukin-1 receptor agonist, IL-6 interleukin receptor, DAVID Database for Annotation, Visualization and Integrated Discovery, STRING Search Tool for the Retrieval of Interacting Genes/Proteins, MASCOT Mascot Daemon by MatrixScience version 2.2.2 (Boston, MA).

Eight articles used a combination of at least two different AI classes. Most commonly, articles analysed biofluids using conventional AI ML techniques such as (1) clustering analyses, including hierarchical clustering [11, 13, 32, 42], k-nearest neighbour [13, 34], nonlinear iterative partial least squares [12], (2) discriminant analyses [31] including partial least square discriminant analysis [7], feature extraction by stepwise discriminant analysis [12], (3) decision tree algorithms including random forest [34, 39], (4) classification algorithms such as support vector machine, naive bayes [34], and (5) dimensionality reduction algorithms such as principal component analysis [11, 28, 38].

Two articles analysed biofluids using deep learning AI [12, 31]. A prospective case-controlled study of 93 patients aimed at elucidating differences between the tear proteome profile of individuals with dry eye, MGD associated dry eye, and healthy individuals, used a nonlinear iterative partial least squares algorithm to cluster the proteomic data followed up a multilayer perceptron neural network predictive model to distinguish between the three distinct tear proteome profiles. Validation of the model yielded a 89.3% correct assignment [12].

Similarly, a cross-sectional study of 88 individuals with dry eye and 71 healthy individuals, used a combination of univariate regression and multivariate discriminant analysis to identify a seven-biomarker panel of potential tear biofluids that may distinguish between the proteomic profile of individuals with dry eye and healthy individuals. These biomarkers were used to train a multiple-layer feed-forward network with back-propagation training algorithm to classify individuals into one of the two groups. Correct classification was quantified using a receiver operating characteristic curve (ROC) and area under the ROC (AUC), which was reported as 0.93, indicating high accuracy [31].

Bioinformatics methodology description largely consisted of the standard analysis protocol of established software such as GO analysis with database for annotation, visualization, and integrated discovery (DAVID), KEGG pathway analysis, iTRAQ proteomics with MASCOT engine, or STRING database searches. Overall, bioinformatic tools were used to classify biofluids into diseases subgroups [26, 33, 39], distinguish between OSD [7, 34], identify risk factors[29], or make predictions about treatment response, and/or prognosis [28, 32, 35, 36, 39].

As presented in Table 2, GO analysis was used by eleven articles, and KEGG pathway analysis was utilized by five articles. One article applied a weighted correlation network analysis (WGCNA), a data mining method, in conjunction with GO analysis and KEGG pathway analysis to identify key hub genes and proteins associated with diabetes and dry eye in adults and children. The GO and KEGG analyses pointed to differentially expressed proteins involved in various metabolic pathways in the tear proteome of adults and children with diabetes and dry eye [33]. MASCOT was used to identify proteins by four articles [12, 13, 26, 41], and STRING was used to build functional protein association networks by three articles [8, 27, 30].

Discussion

This is the first systematic review, to our knowledge, to describe the applications of AI and bioinformatics-based analyses including proteomics and metabolomics using biofluids as markers in various types of corneal and ocular surface diseases. The potential of these technologies to identify candidate biomarkers for diagnosis or potential drug targets to halt disease progression was explored. Risk factors were investigated by one cross-sectional study on pterygium using proteomics in combination with GO analysis with DAVID and KEGG pathway analysis [29]. However, most studies focused on biomarker discovery and identification of biofluids to elucidate aetiology and identify candidate markers for diagnosis, discriminate between OSDs [7, 12, 34], and even identify different subgroups within an OSD [26, 33, 39, 44].

It is increasingly recognized that inflammation of the ocular surface or cornea, specifically change in tear film cytokines, is involved in multiple OSDs including dry eye (e.g. increased expression of apolipoprotein [26, 27], haptoglobin[26, 27], annexin 1 [27, 34], S100A8, S100A9 [32], Glutathione S-transferase [26, 27, 32, 34], and decreased expression of lipocalin-1 [7, 12, 26, 27, 31, 34], prolactin inducible protein [26, 27, 34], lysozyme C [7, 26, 27, 31, 33, 34], lactotransferrin [7, 26, 27, 34], cystatin S [7, 26, 27, 34], and mammaglobin-b [26, 34], proline rich protein [27, 31], IFN-y [28, 36, 45], TNF-a [28, 36, 45]). Several of these are differentially expressed in MGD [34, 40], and Sjögren’s [30, 42] (e.g. lipocalin-1, lysozyme C, annexin A1, cystatin S), as well as keratoconus (e.g. proline rich protein). The biological function of these proteins explains the overlapping pathophysiology, as lactotransferrin and lysozyme have antibacterial properties and support the epithelium [1], while S100A8 and S100A9 are proinflammatory [32], and proline rich protein may be involved in androgen mediated lipolysis [45].

There is a clear need for advancements to the stage of direct clinical applications such as treatment response prediction or monitoring. Several studies have emphasized that tear protein profiling has the potential to provide a diagnostic signature for various OSDs and corneal diseases, specifically the tear film can differentiate between OSDs [7, 12, 46], and also provide measures of disease severity, as well as treatment effectiveness, and thereby be useful for longitudinal monitoring [28, 35, 39].

For example, in a longitudinal study of dry eye patients, hierarchical clustering revealed distinct patient profiles based on clusters of tear protein expression after 3 weeks of punctual occlusion [32]. In cluster 1 (i.e., beneficial response) patients (n = 10) showed increased expression of proteins protective of the ocular surface (e.g., prolactin-inducible protein, lactoferrin) and decreased expression of pro-inflammatory proteins (e.g., alpha enolase 1), while in cluster 2 (nonresponse), patients (n = 13), the opposite trends were observed [32]. These patient profiles correlated with baseline Schirmer scores and may be used in clinic by ophthalmologists to identify patients most at benefit from punctal plugs (i.e., low Schirmer score).

Proteomic analyses in combination with AI, may provide objective tests for evaluating treatment effectiveness for OSDs. For example, a pilot study on dry eye patients used proteomics with GO with DAVID, KEGG and functional annotation clustering, and protein-network analysis to identify 54 and 106 differential expressed biomarkers indicative of disease severity and treatment effectiveness of CsA or DQS, respectively, at 4 weeks [8]. While both treatments were found to be equally effective, tear protein expression profiles indicated distinct regulatory patterns with CsA treated tears showing upregulation of wound healing, endopeptidase activity, and protein metabolism pathways, and DQs showing upregulation of proteins involved in regulation of stress response, tissue homeostasis, and defence response [8]. Following validation in larger samples, expression levels of proteins such as phospholipase A2 group IIA, which was upregulated 2.1-fold pre-treatment in dry eye compared to control and downregulated to 0.58- and 0.78-fold after treatment with CsA and DQS, may be used as metrics indicative of treatment effectiveness with topical agents [8].

This systematic review highlighted several limitations and challenges associated with the included studies. Importantly, the quality and robustness of the AI and bioinformatics-based biofluid analyses is highly dependent on the selection of ML algorithm and the preprocessing of the data. Clustering algorithms, such as hierarchical clustering, an unsupervised ML technique, and k-nearest neighbour, a supervised technique, are useful for identifying subgroups on the basis of similarities between proteomic profiles [11, 13, 32, 34, 42]. Major disadvantages are related to data preprocessing, as clustering algorithms are sensitive to missing values, outliers, data transformation (i.e., to logarithmic scale), and selection of cluster size[47]. Although these parameters directly affect clustering results, we found that they were not consistently described, and consequently may introduce error, reduce reproducibility, and limit validation. These clustering algorithms are less accurate at datasets with more than 400 features (i.e., input variables)[47]. However, this disadvantage can be mitigated by projecting a large number of proteins or metabolites onto a smaller number of features, a procedure known as dimensionality reduction [47]. The main advantage of classifiers such as principal component analysis, random forest algorithm, partial least squares, and support vector machine in biofluid analyses with large datasets of biomarkers is that dimensionality reduction can remove irrelevant features, reduce noise or extraneous variables, and can account for highly correlated variables [47]. The major disadvantage of dimensionality reduction is that it generally requires the selection of a subset of guiding features, a step with a variable level of subjectivity. Relevant data can erroneously be labelled as noise, and this can lead to the loss of important data [47]. The deep learning algorithm, multi-layer perceptron neural network, was implemented as a “black-box model” by two articles in this review, meaning that its different layers and complex architecture was not described in sufficient detail to allow the reader to map the process from variable input to prediction [12, 31]. Therefore, despite advantages such as better handling of highly dimensional data, complex (non-linear) associations, noise, and incomplete or missing values, the interpretation and generalizability of the results are limited [48].

Small sample sizes and large datasets of proteins increase the likelihood of finding spurious associations due to the inter-day and inter-individual variability of normal tears [49]. Moreover, the reporting of large proteomic data, up to 2,733 proteins, is challenging considering reporting limits set by various scientific journals, four articles in the current review only reported them in figure format, one only reported significant proteins, and two did not report the full list of measured proteins at all. Online repositories (e.g. The Global Proteome Machine Database, PeptideAtlas, etc.) could be used for data mining in future studies [50]. Other limitations of bioinformatics-based analyses of biofluid data relate to the annotation databases (e.g., GO and KEGG) used to perform the ontological analysis necessary to map the function of input proteins and construct protein-protein interaction networks using clustering, classification, and significance analyses. These databases are manually curated by researchers and may be incomplete, imprecise, variability in the identifiers used by various research groups, and be impacted by annotation bias (i.e., well-studied biological processes are more represented) [51].

The multifactorial pathogenesis of OSDs and corneal diseases, the overlap of symptoms, and lack of concordance between clinical parameters and symptoms reported by patients [34], present a challenge to the identification of unique biomarkers for discriminating between pathologies, or monitoring treatment response. The challenge is not only compounded by small sample sizes, but also lack of healthy controls, and technical variability associated with proteomic and metabolomic studies (e.g., tear collection methods, sample preparation, pre-processing steps for mass spectrometry, lack of reporting of all investigated proteins). We found that most studies used Schirmer strips (n = 12, 52%) or micropipette/glass capillaries (n = 5, 22%) as collection methods. Generally, both collection methods, Schirmer test and capillary, are reported to produced similar results [31, 52]. However, Schirmer test was found to allow for better discrimination between dry eye and healthy samples [31, 34]. Only a handful of studies implemented a statistical validation process of the discovered proteins, with the goal of using the area under the curve (AUC) from multivariate receiver operating characteristic (ROC) analyses to calculate specificity and sensitivity, and estimate the clinical applicability of candidate biofluids [12, 42, 53]. Validation on large samples is crucial particularly considering the lack of a priori hypotheses or pre-selection of a panel of biomarkers characteristic of many proteomic and metabolomic studies, as well as the physiologic heterogeneity of OSD. Without this step, both the generalizability and the predictive specificity of candidate biomarkers remains limited. For example, dysregulation of several candidate biomarkers for pathogenesis in dry eye were also found in MGD (e.g., lipoprotein-1, lysozyme C, lactotransferrin) [7, 26, 31, 34], VKC (e.g. lactotransferrin) [41] or in pterygium and climatic droplet keratopathy (e.g., alcohol dehydrogenase) [9, 26, 37].

The combination of biofluids and imaging metrics obtained from optical coherence tomography (OCT), and analysed using AI, may compound the clinical predictive value of these techniques [54]. For example, wide corneal epithelial mapping using OCT in dry eye analysed with random forest AI, showed that superior intermediate epithelial thickness in dry eye compared to controls, was a promising marker for diagnosing dry eye (sensitivity 86.4%, specificity 91.7%) [54]. Introducing biofluids as covariates in these types of analyses would increase the robustness and validity of these analyses and bring them to clinical standards.

This systematic review appraised the use of AI or bioinformatics tools to analyse biofluid markers in OSDs and corneal disease. These tools implicated various tear film proteins in biological pathways regulating lipid metabolomic processes, oxidative stress regulation, cytokine production, vesicular transportation, and regulation of the immune response. Several studies have suggested that tear protein profiling has the potential of providing a diagnostic signature for various OSDs, may be used to identify patients most at benefit from treatments, or provide indications for treatment effectiveness and be useful for longitudinal monitoring.

Appendix A

Search strategy used to survey databases.

Appendix B

Assessment of risk of bias and quality of included studies based on criteria from Joanna Briggs Institute Critical Appraisal Tools.

Author contributions

TF, RNM: conceptualization, supervision, methodology, writing, review and editing of manuscript. DRP: conceptualization, data collection, analysis and interpretation of results, draft manuscript preparation. SHK, methodology, data collection. AP, data collection, analysis and interpretation of the results. All authors reviewed the results, and approved the final version of the manuscript.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41433-022-02307-9.