SEARCHCROSS-SECTIONALophthalmology, dry eye

cross-sectional·ophthalmology, dry eye, machine learning, proteomics, biomarkers·PMC10746717

Using machine learning model explanations to identify proteins related to severity of meibomian gland dysfunction

Scientific Reports · 11 authors, 13 centres

AI SUMMARY

FIDELITY 80%

POPULATION233 patients with MGD (levels 2-4) from the Norwegian Dry Eye Clinic

INTERVENTIONMachine learning models (LGBMClassifier) trained on relative tear protein quantifications to predict MGD severity

COMPARISONComparison of protein importance across MGD levels 2, 3, and 4 using SHAP values (unweighted and weighted by model uncertainty) and PEAKS X Pro software

This summary was generated by AI from a single paper. It has not been reviewed by a clinician and is not clinical advice. Verify against the source before acting on it.

AGENT 01 · STRUCTURED SUMMARY

TRACE TO SOURCE

“

TL;DR

This study used machine learning (ML) models and SHAP feature importance to identify tear proteins associated with meibomian gland dysfunction (MGD) severity in 233 patients. Several proteins previously linked to dry eye disease (DED), such as S100-A8, PRP4, and lysozyme C, were among the top-ranked features, and the ML approach also highlighted novel candidates like glutathione peroxidase 1 and dynactin subunit 2. These findings suggest that ML-driven proteomic analysis can uncover potential biomarkers and therapeutic targets for MGD.

Full summary

3,460 CHARS

**Background:** Meibomian gland dysfunction (MGD) is a leading cause of dry eye disease (DED), affecting hundreds of millions globally. While the tear film lipid layer is primarily affected, tear proteins are thought to play key roles in MGD etiology. Machine learning (ML) can identify patterns in complex proteomic data and highlight features important for disease classification. This study aimed to use ML to identify tear proteins associated with MGD severity, focusing on differences across severity levels rather than comparing healthy vs. diseased.

**Methods:** The dataset included 233 patients with MGD (levels 2-4) from the Norwegian Dry Eye Clinic (2017-2019). Tear samples were collected via Schirmer strips, and proteins were quantified using untargeted LC-MS and PEAKS X Pro software (FDR ≤1%, fold change ≥2, ANOVA). After preprocessing (removing contaminants, replacing missing values, log transformation, scaling), 2188 proteins were analyzed. An LGBMClassifier (boosting tree-based) was trained to predict MGD level (2, 3, or 4) from protein quantifications, with class imbalance handled via the 'is_unbalanced' parameter. Three additional binary models predicted each level vs. others. Model performance was evaluated on the training data using balanced accuracy, F1 score, and Matthews correlation coefficient (MCC). Feature importance was assessed using Shapley Additive Explanations (SHAP), with the top 15 proteins examined. A novel weighting technique adjusted SHAP values by model uncertainty (based on predicted probabilities). PEAKS X Pro identified proteins significantly different between MGD levels.

**Key Results:** The multiclass model achieved balanced accuracy 72%, F1 0.82, MCC 0.72. Binary models: MGD level 2 (balanced accuracy 95%, F1 0.80, MCC 0.78), level 3 (85%, 0.86, 0.69), level 4 (87%, 0.82, 0.73). Unweighted SHAP identified top proteins including S100-A8 (upregulated in DED), proline-rich protein 4 (PRP4, downregulated in DED), lysozyme C (downregulated in DED), polymeric immunoglobulin receptor, prostaglandin reductase 1, and immunoglobulins (IGKV2-24, IGLV8-61, IGLV6-57, IGA2, IGG1). Weighted SHAP added glutathione peroxidase 1, dynactin subunit 2, psoriasin, and ADH7 to the top 15 for the multiclass model. PEAKS analysis found 13 proteins significantly different between MGD levels, including thymidine phosphorylase; only mannose-1-phosphate guanyltransferase alpha overlapped with SHAP-identified proteins. S100-A8 showed statistically significant differences between MGD levels 2 and 3 (p<0.05) and borderline between levels 2 and 4 (p=0.052). PRP4 and lysozyme C showed trends but not significant.

**Clinical Implications:** The study demonstrates that ML combined with SHAP can identify tear proteins associated with MGD severity, many of which are known DED biomarkers (e.g., S100-A8, PRP4, lysozyme C) and others potentially novel (e.g., glutathione peroxidase 1, dynactin subunit 2). These proteins are involved in inflammation, immune response, oxidative stress, and normal ocular function, suggesting MGD is a complex condition with multiple pathways. The findings could lead to new biomarkers for MGD grading and potential therapeutic targets. However, the models were not externally validated, and the study did not include a healthy control group, limiting generalizability. Future work should validate these proteins in independent cohorts and explore their roles in MGD pathogenesis.

PICO

PPOPULATION

233 patients with MGD (levels 2-4) from the Norwegian Dry Eye Clinic

IINTERVENTION

Machine learning models (LGBMClassifier) trained on relative tear protein quantifications to predict MGD severity

OOUTCOME

Identification of proteins most important for predicting MGD severity; model performance metrics (balanced accuracy, F1 score, MCC)

STUDY TYPE

cross-sectional

SPECIALTY

ophthalmology

SUMMARISED BY

AI pipeline

FIDELITY CHECK

80% · B

ORIGINAL PAPER

PMC10746717

Using machine learning model explanations to identify proteins related to severity of meibomian gland dysfunction

Scientific Reports

READ THE FULL PAPER

JUMP TO SECTION

ABSTRACT

Abstract Meibomian gland dysfunction is the most common cause of dry eye disease and leads to significantly reduced quality of life and social burdens. Because meibomian gland dysfunction results in impaired function of the tear film lipid layer, studying the expression of tear proteins might increase the understanding of the etiology of the condition. Machine learning is able to detect patterns in complex data. This study applied machine learning to classify levels of meibomian gland dysfunction from tear proteins. The aim was to investigate proteomic changes between groups with different severity levels of meibomian gland dysfunction, as opposed to only separating patients with and without this condition. An established feature importance method was used to identify the most important proteins for the resulting models. Moreover, a new method that can take the uncertainty of the models into account when creating explanations was proposed. By examining the identified proteins, potential biomarkers for meibomian gland dysfunction were discovered. The overall findings are largely confirmatory, indicating that the presented machine learning approaches are promising for detecting clinically relevant proteins. While this study provides valuable insights into proteomic changes associated with varying severity levels of meibomian gland dysfunction, it should be noted that it was conducted without a healthy control group. Future research could benefit from including such a comparison to further validate and extend the findings presented here. Subject terms: Machine learning, Protein analysis, Diagnostic markers, Computer science, Biomarkers, Corneal diseases

CITATION EXPORT

Andrea M Storås, Fredrik Fineide, Morten Magnø et al.. Using machine learning model explanations to identify proteins related to severity of meibomian gland dysfunction. Scientific Reports. (2023). https://doi.org/10.1038/s41598-023-50342-7

AUTHORS

11 · 13 CENTRES

Andrea M Storås, Fredrik Fineide, Morten Magnø, Bernd Thiede, Xiangjun Chen, Inga Strümke, Pål Halvorsen, Hilde Galtung, Janicke L Jensen, Tor P Utheim, Michael A Riegler

1Department of Holistic Systems, Simula Metropolitan Center for Digital Engineering, Oslo, Norway
2Department of Computer Science, OsloMet - Oslo Metropolitan University, Oslo, Norway
3The Norwegian Dry Eye Clinic, Oslo, Bergen, Norway
4Department of Ophthalmology, Sørlandet Hospital Arendal, Arendal, Norway
5Department of Plastic and Reconstructive Surgery, Oslo University Hospital, Oslo, Norway
6Department of Biosciences, University of Oslo, Oslo, Norway
7Department of Medical Biochemistry, Oslo University Hospital, Oslo, Norway
8Department of Ophthalmology, Vestre Viken Hospital Trust, Drammen, Norway
9Department of Computer Science, Norwegian University of Science and Technology, Trondheim, Norway
10Institute of Oral Biology, University of Oslo, Oslo, Norway
11Department of Oral Surgery and Oral Medicine, University of Oslo, Oslo, Norway
12Department of Ophthalmology, Oslo University Hospital, Oslo, Norway
13Department of Computer Science, UiT The Arctic University of Norway, Tromsø, Norway