Paired longitudinal tear proteomics reveals persistent ocular surface dysregulation after teprotumumab therapy in thyroid eye disease

Abstract

Background

Thyroid eye disease (TED) is a complex autoimmune disorder that is frequently disfiguring and sight-threatening. Recent advances have shifted treatment from surgery toward targeted biologic therapies like teprotumumab (TMB), an IGF-1R inhibitor that improves proptosis, diplopia, and inflammation. However, ocular surface symptoms often persist.

Methods

We conducted a longitudinal pre-post comparative study of patients with active, moderate-to-severe TED treated with TMB at a tertiary academic medical center, with tear sampling at two prespecified time points: baseline (collected within the 3 weeks strictly antecedent to the first infusion) and post-treatment (within 6 months of the final infusion). Adults with active, moderate-to-severe TED who received ≥4 TMB infusions were eligible. All patients contributed paired pre- and post-treatment tear samples (22 eyes). Controls were adults without TED or ocular surface disease (32 eyes). Tear fluid was collected by Schirmer strips and analyzed by liquid chromatography-tandem mass spectrometry. Differential expression was assessed using a linear mixed-effects model to explicitly account for the intrinsic correlation between paired eyes; correlations between LFQ intensities and clinical metrics used Spearman rank correlation.

Results

A total of 2974 proteins were identified across 76 tear samples. Unsupervised analyses demonstrated separation between control, pre-treatment, and post-treatment TED tear proteomes. Although TMB induced substantial proteomic changes, post-treatment profiles did not revert to a control-like state. Proteins involved in inflammation, oxidative stress, and cytoskeletal organization remained persistently dysregulated, while a subset of tear film stability–associated proteins normalized. Several persistently dysregulated protein targets overlapped with FDA-approved drugs.

Conclusions

These observational findings support tear proteomics as a platform for biomarker discovery and therapeutic prioritization and warrant validation in larger, independent cohorts.

Plain Language Summary

Thyroid eye disease is an autoimmune condition that can cause eye bulging (proptosis), double vision, and chronic eye discomfort. Although the drug teprotumumab improves many clinical signs, patients often continue to experience dry eye symptoms, and the reasons for this are not well understood. In this study, we analyzed proteins in tear fluid collected from patients with thyroid eye disease before and after teprotumumab treatment and compared them with tears from people without the disease. Tears are easy to collect and reflect the health of the ocular surface. We found that while treatment caused major changes in tear composition, many proteins linked to inflammation and cellular stress remained abnormal, even after therapy. Some protective tear proteins improved, but overall the tear profile did not fully return to normal. These findings suggest that eye surface disease may persist despite treatment and that tear protein measurements could help guide future therapies.

Velez et al., conduct a longitudinal pre-post study of tear proteomics in patients with active moderate-to-severe thyroid eye disease treated with teprotumumab, comparing paired baseline and post-treatment samples with controls. Teprotumumab induced broad proteomic changes but did not restore a control-like state, with persistent dysregulation of inflammatory, oxidative stress, and cytoskeletal proteins alongside partial normalisation of tear-film proteins.

Introduction

Thyroid eye disease (TED), a debilitating complex autoimmune disease with a yearly incidence of 19 cases per 100,000 people¹, affects approximately 40% of patients with Graves’ disease (GD)². While the pathophysiology of TED is not completely understood, it is believed to result from an interplay of genetic and environmental factors. A primary mechanism involves autoantibodies directed against the thyrotropin receptor (TSHR), and possibly the insulin-like growth factor-1 receptor (IGF-1R), on orbital fibroblasts^3,4. The resulting autoantibody activation of orbital fibroblasts leads to downstream signaling pathways that drive the recruitment and production of proinflammatory cytokines, which activate orbital fibroblasts to proliferate and differentiate into both myofibroblasts and adipocytes. Signaling through the TSHR/IGF-1R complex stimulates excessive synthesis of glycosaminoglycans and hyaluronan. These changes, along with recruitment of T and B lymphocytes, lead to periorbital edema, tissue-remodeling, and soft tissue expansion characteristic of the active inflammatory phase of disease. Characteristic clinical manifestations of TED include eyelid retraction, proptosis, and restrictive strabismus, with 3–8% of patients progressing to sight-threatening disease due to exposure keratopathy or compressive optic neuropathy². However, dry eye symptoms may be the most common cause of discomfort in patients with TED, occurring in up to 85% of patients⁵. As the active phase subsides, typically over 6–24 months, the disease enters a chronic inactive phase where inflammation recedes, yet expansive soft tissue changes and fibrosis persist.

Anti-inflammatory therapies like corticosteroids and radiotherapy have historically been used to treat active inflammatory TED; however, surgical interventions such as orbital decompression, strabismus surgery, and eyelid reconstruction are frequently necessary to treat chronic persistent disease⁶. Teprotumumab (TMB), a monoclonal antibody that inhibits the insulin-like IGF-1R, became the first drug approved by the US Food and Drug Administration for the treatment of TED in January 2020⁷. The IGF-1R normally functions to regulate tissue growth and metabolism and is overexpressed by orbital fibroblasts and T and B cells in patients with TED^8–10. In randomized controlled trials, TMB was found to significantly reduce orbital inflammation and proptosis, as well as improve diplopia and quality of life in patients with active, moderate-to-severe TED^11,12. There is also limited evidence that TMB decreases lacrimal gland volume and increases tear volume in TED patients, although its efficacy in reducing ocular surface inflammation is yet to be established¹³. Despite its clinical efficacy, TMB carries a high financial cost and is associated with potentially severe adverse events such as hearing loss, hyperglycemia, encephalopathy, and gastrointestinal disturbance¹⁴. While TMB is typically reserved for patients with active, moderate to severe TED, patients with mild disease, particularly those with prominent ocular surface symptoms, often go untreated. Additionally, approximately 65.38% of TED patients treated with TMB display regression of proptosis within one year of initial treatment¹⁵. These limitations raise concern about TMB’s durability and the need for new targeted therapies that address relapsing forms of TED and ocular surface symptoms.

Studies suggest direct lacrimal gland involvement in TED. Lacrimal gland enlargement is a common finding in TED and has been positively correlated to the activity and severity of disease^16–19. Eckstein et al demonstrated that lacrimal acinar cells physiologically express the TSHR and suggests autoantibodies may bind to lacrimal TSHRs and contribute to lacrimal impairment and symptoms of dry eye disease (DED). Other contributing factors for DED include mechanical impairment of ocular surface lubrication (i.e., lagophthalmos and widening of the palpebral fissure), meibomian gland dysfunction, and ocular surface inflammation^3,20–25. Additionally, changes to the molecular composition of the tear film (e.g., increased tear film osmolarity, disruption of the tear lipid layer, etc.) further exacerbate the debilitating DED symptoms commonly experienced by TED patients²⁵. Thus, recent efforts have sought to characterize the tear film proteome of patients with TED^26–35. Tear samples are not only easily accessible but also serve as a reliable proxy for proteomic changes of the ocular surface³⁴; this is a method that has been used to evaluate inflammatory proteins as biomarkers in dry eyes³⁶. While the aforementioned proteomics studies have consistently found upregulation of inflammatory proteins and downregulation of protective proteins in TED, none assessed the tear proteome serially to evaluate changes across time, correlation with disease severity, and response to treatment. To our knowledge, this is the first study to evaluate tear proteome changes in TED patients following treatment with TMB.

In this study, we present a paired longitudinal analysis of human tear film proteomic changes in patients with TED before and after TMB therapy. Using high-resolution discovery proteomics, we demonstrate that TED is characterized by a distinct aqueous-phase molecular signature that tracks with clinical disease activity. Additionally, we show that while systemic IGF-1R inhibition drives profound clinical and structural improvements in the orbit, it does not fully revert the tear proteome to a healthy control-like state. Widespread abnormalities in pathways governing innate immunity, neutrophil degranulation, and cellular stress responses persist after therapy, providing a molecular basis for ongoing ocular surface dysfunction. Finally, we provide a curated in silico drug-repositioning framework that identifies 10 clinically viable, FDA-approved therapeutic candidates tailored to reverse these persistent ocular surface cascades, offering a rational blueprint for targeted adjunctive therapies.

Methods

Ethics statement

This study was conducted in strict accordance with the tenets of the Declaration of Helsinki for Human Research and the Health Insurance Portability and Accountability Act. The study protocol, patient selection registries, and biofluid collection procedures received formal approval and oversight from the Stanford Institutional Review Board (protocol approval number #42754). Prior to participation, written informed consent was obtained from all active patients and control individuals for biofluid sampling, evaluation of clinical parameters, and the open-access publication of de-identified clinical metrics, baseline demographics, and corresponding figure panel images.

Patient selection and outcome definition

Participants were recruited at the Stanford Byers Eye Institute in Palo Alto, CA, USA. Recruitment occurred from March 2020 to December 2020. Follow-up and post-treatment sampling occurred through August 2021. Eligible TED participants were identified from patients receiving TMB in routine clinical care and were enrolled consecutively. Control participants were recruited from general ophthalmology clinics at Byers Eye Institute and were excluded if they had TED, GD, DED, meibomian gland dysfunction, or other ocular surface disease. Subjects were included in the study if they were over 18 years of age, had active, moderate-to-severe TED at baseline, and received 4 or more infusions of TMB. A standard full therapeutic course of TMB consists of 8 intravenous infusions. To reflect real-world clinical tolerability and pragmatic patient management, our inclusion criteria predefined a minimum threshold of ≥4 infusions. Within our cohort, one participant discontinued therapy after 4 infusions due to the development of treatment-related ototoxicity. To maintain clinical generalizability and prevent the artificial sanitation of the real-world biological landscape, this participant was retained in the primary analytical cohort. To verify that this lower-exposure individual did not introduce statistical instability or skew our data topologies, a leave-one-out sensitivity validation test was conducted as described below. TMB was administered intravenously at 10 mg/kg for the first infusion and given every 3 weeks at 20 mg/kg for a total of 8 infusions. Subjects underwent tear sample collection and examinations that included slit-lamp examination, retinal biomicroscopy, indirect ophthalmoscopy, clinical activity score (CAS), orbital imaging, retinal nerve fiber optical coherence tomography, and visual field testing. Examinations and data collection were performed at two time points: (1) baseline measurements were taken within the 3 weeks strictly antecedent to the first clinical TMB infusion, ensuring a completely drug-naïve state, and (2) post-treatment measurements were taken after the 8th infusion, but no more than 6 months after the last infusion. Our post-treatment endpoint was explicitly selected to capture the completed, mature biological and clinical remodeling induced by TMB. In clinical practice, the therapeutic response of orbital tissue decompression and localized inflammatory reduction matures fully upon the completion of the full eight-infusion protocol. While sampling mid-therapy captures highly volatile, transitional pharmacokinetic states, evaluating true baseline coordinates against the completed post-treatment milestone successfully isolates stable, long-term biological signatures and definitive therapeutic outcomes⁷.

Disease activity was measured using CAS, a validated scoring system that assesses the presence of 7 clinical symptoms, including spontaneous orbital pain, gaze-evoked orbital pain, eyelid swelling, eyelid erythema, conjunctival redness, chemosis, and caruncle/plica inflammation³⁷. Disease severity was graded using the European Group on Graves’ Orbitopathy criteria, which classifies TED severity as mild, moderate-to-severe, or sight-threatening³⁸. To reduce measurement bias, TED activity/severity assessments were performed by a single oculoplastic surgeon (ALK) using standardized criteria. Our inclusion criterion of moderate-to-severe TED is defined as having two or more of the following: greater than or equal to 2 mm of lid retraction, moderate or severe soft-tissue involvement, proptosis that is greater than or equal to 3 mm above the normal range for a given race and gender, and inconstant or constant diplopia. In these patients, the disease typically sufficiently impacts daily life to warrant immunosuppression or surgical intervention. All patients underwent a thyroid laboratory workup at baseline that included thyroid-stimulating immunoglobulin (TSI), free T4, total T3, and thyroid-stimulating hormone (TSH).

Sample size was determined by the number of eligible TMB-treated TED participants with serial tear samples during the study period; no formal a priori power calculation was performed due to the exploratory nature of high-dimensional proteomics. The primary exposure was TMB treatment time point (baseline and post-treatment). Primary molecular outcomes were tear protein abundance by LC–MS/MS (spectral counts) and label-free quantification (LFQ) intensity (i.e., MaxQuant). Clinical outcomes included CAS, proptosis, Gorman diplopia score (GDS), Schirmer values, and thyroid markers (TSH, TSI). CAS response was defined as a ≥ 2 reduction in CAS score from baseline. Proptosis response was defined as a ≥ 2 mm reduction in proptosis from baseline in the study eye without a corresponding increase of ≥2 mm in the fellow eye. The study eye is the eye with greater proptosis at baseline. Overall response was defined as having both CAS and proptosis response, each as described above.

Sample collection

Tear collection, processing, and LC–MS/MS acquisition were performed using the same protocol across all groups and time points. All participants provided written consent to undergo Schirmer’s tests for the purpose of the study. Schirmer strips without printed mm scales were used (Haag-Streit, Köniz, Switzerland). While in the sterile package, the Schirmer strip was bent at a 90-degree angle along the 0 mm line. Jewelers’ forceps were used to grasp the tip of the Schirmer’s strip and place the bent portion of the strip behind the lower lid margin, approximately 5 mm medial to the lateral canthus and just lateral to the lateral limbus. The investigator ensured that the strip rested at a 90-degree angle with respect to the lid and did not touch the skin. The patient was instructed to close their eyes for 5 min. The strips were then removed and placed in a sample collection tube and stored at −80 °C until processed.

Protein extraction and digestion

Following thawing from 80 °C storage, each individual Schirmer strip was placed flat inside a sterile 60 × 15 mm petri dish. While holding the dry anchoring end of the strip with sterile forceps, the wetted portion of the filter paper was sectioned using sterilized micro-scissors into five compact 5 mm transverse segments, starting at the zero-mark notch. These individual 5 mm paper pieces were then transferred via sterile forceps into the internal insert basket of a 0.1 μm PVDF centrifugal filter unit (Ultrafree-MC Centrifugal Filter, MilliporeSigma, Cat #UFC30VV), where they layer flat and fit loosely³⁹. Next, 450 µL of 100 mM ammonium bicarbonate in 50 mM NaCl was added and filter units were incubated at 25 °C for 4 h at 300 rpm. Samples were then centrifuged at 5500 rcf for 5 min and flow through was collected and dried to completion in speed vac. Extracted proteins were re-suspended in 50 mM ammonium bicarbonate in the presence of ProteaseMAX surfactant (Promega), followed by reduction with 10 mM DTT at 55 °C for 30 min. Following reduction, proteins were alkylated with 30 mM acrylamide for 30 min at room temperature. Digestion was performed with Trypsin/LysC (Promega) in a standard overnight digest at 37 °C. After digestion, the reaction was quenched using 1% formic acid, and peptides were de-salted on C18 Monospin reversed-phase columns (GL Sciences). The de-salted peptides were dried in a speed vac before reconstitution in 100 µL of reconstitution buffer (2% acetonitrile with 0.1% Formic acid); 10 µL of this solution was used to estimate the amount of peptide material using Pierce quantitative fluorometric peptide assay. Based on the material recovered, 1 µg was injected onto the mass spectrometer.

Liquid chromatography-tandem mass spectrometry

Mass spectrometry experiments were performed on a Q Exactive HF-X Hybrid Quadrupole - Orbitrap mass spectrometer (Thermo Scientific, San Jose, CA) with liquid chromatography using a Nanoacquity UPLC (Waters Corporation, Milford, MA). For a typical LC/MS experiment, a flow rate of 450 nL/min is used, where mobile phase A was 0.2% formic acid in water and mobile phase B was 0.2% formic acid in acetonitrile. For this experiment 50 cm µPAC column (Pharmafluidics) was used. Peptides were directly injected onto the analytical column using a gradient (2–45% B, followed by a high-B wash) of 80 min. The mass spectrometer was operated in a data-dependent fashion using HCD fragmentation for MS/MS spectra generation. Raw LC–MS/MS data files were processed using two distinct, specialized bioinformatic pipelines optimized for their respective quantitative frameworks. High-confidence peptide spectral matches (PSMs) and discrete spectral count matrices were extracted utilizing Byonic software (Protein Metrics, San Carlos, CA) to leverage its specialized fragmentation search algorithms⁴⁰. Proteolysis with Trypsin/LysC was assumed to be semi-specific, allowing for N-ragged cleavage with up to two missed cleavage sites. Precursor and fragment mass accuracies were held within 12 ppm. Proteins were held to a false discovery rate (FDR) of 1%, using standard approaches. In parallel, to generate continuous, high-resolution LFQ tracking, raw files were analyzed using the MaxQuant computational platform (version 2.7.5.0), with sample abundances calculated via the native MaxLFQ algorithm⁴¹. Protein intensities were calculated using the MaxLFQ algorithm with default parameters, and LFQ values of zero, representing missing quantification, were treated as missing values in downstream analyses.

Statistical and bioinformatics analysis

Differential protein expression

For longitudinal differential expression analysis, discrete spectral counts (PSMs) derived from the Byonic informatics pipeline were utilized. Normalized proteomics data were imported to R Studio (version 2025.09.1 + 401, R version 4.5.1). To adjust for variations in total sample-wide spectral sampling depth across runs, raw count matrices were normalized using the Weighted Trimmed Mean of M-values algorithm via the calcNormFactors function in edgeR. While this function has been originally implemented for RNA-Seq count data, it applies equally well to proteomic (e.g., spectral count) data⁴². Since highly abundant proteins contribute more peptide/spectra than lowly abundant proteins, the calcNormFactors function adjusts for the total number of spectral counts per sample (analogous to original library size in RNA-seq experiments) and normalizes for protein composition⁴³. Normalized count allocations were subsequently log2-transformed with an added fractional prior count of 1 (i.e., log2[normalized counts]+1). This systematic transformation stabilized the lower boundary of the matrix, satisfied linear model normality assumptions, and mitigated the distorting impact of zero-count entries without requiring artificial data imputation. To prevent pseudoreplication and the inflation of degrees of freedom inherent to evaluating paired left and right eyes from the same individual, the nested intra-subject structure was explicitly modeled. The consensus correlation between paired eyes across the entire dataset was computed using the duplicateCorrelation function within the limma framework⁴⁴. This consensus correlation parameter was integrated into a generalized least squares linear mixed model, treating individual patient batches as a random effect to ensure conservative and statistically valid p-values. Differentially expressed proteins (DEPs) were determined using the updated mixed-model matrix. To ensure the distributional safety of this missing value pipeline, a parallel left-censored sensitivity stress test was executed (described below), demonstrating stability of our primary protein modules. P-values were corrected for multiple testing using Benjamini–Hochberg (BH) adjustment implemented in the limma package. DEPs were defined using an adjusted p-value < 0.05 and an absolute log2 fold-change (log2FC) threshold of > 0.5. Because tear fluid is a highly dynamic extracellular matrix where important homeostatic and inflammatory pathways operate via tightly regulated, lower-magnitude shifts, applying a conservative |log2FC| > 0.5 (rather than an arbitrary twofold boundary) prevents the high false-negative rates that obscure clinically vital downstream trajectories^45–47. False positives are mathematically restricted by the stringent 5% FDR control across all selected features. Furthermore, to maximize feature stringency and suppress variance-driven false positives, this differential testing was coupled to a linear mixed-effects model via duplicateCorrelation in limma, treating the individual patient as a random effect block to account for paired-eye dependencies. DEPs identified from each group comparison (i.e., pre-TMB vs. control, post-TMB vs. control, and post- vs. pre-TMB) were classified into expression trajectories based on their paired longitudinal expression patterns. Proteins with consistent up or downregulation across pre- and post-TMB groups were assigned to “progressive up” or “progressive down” trajectories. Proteins with changes reversed in the post-TMB group were labeled “reversed by TMB,” whereas trajectories with expression changes unique to the post-TMB group or persistent across time points were labeled “treatment-specific” and “persistent,” respectively.

Clinical correlation analysis

To evaluate whether tear protein abundance reflected clinical disease severity and treatment response, a dual-architecture correlation framework was deployed using unaveraged, eye-specific data. Cross-sectional baseline associations were evaluated between log2-transformed LFQ protein intensities and pre-treatment clinical metrics, including CAS, proptosis (mm), GDS, Schirmer test values (mm), and systemic thyroid markers (TSH and TSI). Non-normal clinical distributions (specifically systemic TSI titers) were handled by executing non-parametric Spearman rank correlation (ρ). To evaluate molecular changes tracking with the clinical resolution of active inflammation, longitudinal delta modeling was performed for patients with paired pre- and post-treatment eye measurements. Quantifiable protein shifts were calculated as ∆LFQ = Log2(Post)–Log2(Pre) and correlated against clinical improvement, defined as ∆CAS = Pre–Post (where positive values denote clinical reduction in inflammation). For all models, a minimum dataset filter was applied, requiring a protein to be quantified in at least 30% of samples and possess at least 8 complete pairs. Consistent with MaxQuant output conventions, LFQ MS1 intensity values of zero were treated as missing observations^41,48. Continuous correlation mapping was performed strictly on log2-transformed LFQ intensities utilizing pairwise deletion to preserve true missingness and prevent the creation of artificial trend lines^41,48. To ensure the distributional safety of this missing value pipeline, a parallel left-censored sensitivity stress test was executed, demonstrating stability of our primary protein modules. Multiple-testing corrections were executed globally using the BH FDR method, with significance defined at an FDR < 0.05. Proteins meeting nominal significance thresholds (p < 0.05) but not FDR correction were considered exploratory. To compare the relative strength of associations across clinical metrics, the number of proteins meeting FDR < 0.05 and nominal p < 0.05 thresholds was summarized for each metric. For visual clarity in regression plots, highly skewed clinical variables were log-transformed, and semi-log lines with 95% confidence intervals were fitted^41,48.

Imputation sensitivity stress testing

To verify that the primary differential expression signatures were robust and independent of low-intensity matrix boundary handling, a parallel validation sensitivity analysis was performed on the primary expression dataset. Missing values were simulated using a standard left-censored Missing Not at Random imputation strategy, drawing random variables from a down-shifted normal distribution meant to mimic the instrument’s technical limit of detection (width: 0.3 SD, down-shift: 1.8 SD from the dataset mean). Re-running the entire longitudinal linear mixed-effects model on this imputed matrix demonstrated structural preservation of the primary therapeutic signatures; all 375 DEPs identified in the main analysis remained stable. This mathematical resilience confirms that the key biological findings represent robust, frequently sampled signaling features positioned well above the instrument noise floor and are entirely independent of alternative low-intensity boundary configurations.

Pathway and network analysis

Pathway enrichment, network analyses, and drug repositioning were conducted as exploratory secondary analyses to prioritize mechanisms and therapeutic hypotheses. Functional annotation of DEPs was performed using the Reactome database⁴⁹. Pathway enrichment analysis of each DEP trajectory was performed via the clusterProfiler R package. Enrichment was assessed using the hypergeometric test, and p-values were adjusted for multiple testing using the BH method. Metascape and STRING (Search Tool for the Retrieval of Interacting Genes) v11 were used to identify protein-protein interactions and illustrate interconnectivity among proteins^50,51. Data were analyzed at a FDR of 5% and the protein-protein interaction networks were analyzed.

Drug repositioning analysis

To translate our longitudinal proteomic findings into therapeutically actionable candidates, an in silico drug repositioning pipeline was constructed using the Drug–Gene Interaction Database (DGIdb)⁵². DGIdb contains data mined from several resources (e.g., DrugBank, HGNC, etc.) and provides a comprehensive overview of the druggability of identified protein targets^53,54. To ensure this exploratory screening went beyond unweighted database matching, a rigorous multi-tier prioritization and curation framework was implemented: (1) candidates were restricted to FDA-approved agents whose primary mechanisms of action are directionally congruent with reversing the specific longitudinal inflammatory and tissue-remodeling pathways dysregulated during active TED pathogenesis; (2) drug-target interactions were ranked using a composite priority score weighting the empirical protein log2FC, FDA-approval status, and quantitative DGIdb interaction confidence parameters. Finally, to filter out computational noise and isolate clinically viable candidates, the algorithmic outputs were subjected to manual clinical curation. Systemic compounds lacking a plausible therapeutic role for orbital remodeling or ocular surface protection (e.g., antipsychotics, central nervous system agents, and heavy systemic cytotoxics) were excluded. Conversely, priority was assigned to: (1) targeted biologics and systemic immunomodulators with documented second-line efficacy in refractory TED; (2) small-molecule inhibitors targeting complement cascades or intracellular signaling pathways active in orbital fibroblasts; and (3) commercially available topical ophthalmic therapeutics capable of mitigating exposure-driven ocular surface pathology.

Results

TMB induces proteomic changes in the tear film

A cohort of eleven patients (22 eyes) were diagnosed with active moderate to severe TED and used for the analysis (Supplementary Table 1). The average patient age was 46.6 ± 11.2 years; 91% were female. The median duration of disease from TED diagnosis to first infusion was 21 months (IQR: 8.0–29.1; Fig. 1b). Previous GD treatments included radioactive iodine (n = 4), methimazole or propylthiouracil (n = 8), and thyroidectomy (n = 3). Previous TED treatments included pulsed-dosed corticosteroids (n = 8), orbital radiation therapy (n = 3), orbital decompression (n = 1), and tocilizumab (n = 1). No patients had concurrent steroid use, and three patients were treatment naïve. Mean baseline CAS was 4.5 ± 0.8 (post-treatment CAS score was 0.8 ± 0.7), mean baseline proptosis was 20.6 ± 3.2 mm in the study eye. In patients that had diplopia, the mean GDS was 2.0 ± 0.7. Patients received a mean of 8 infusions. One patient discontinued TMB during treatment due to hearing loss after four doses. Mean CAS and GDS reductions were 90 and 67%, respectively. Mean reduction of proptosis was 1.8 ± 1.1 mm (16%; Fig. 1c). Three patients had regression of proptosis 1 year following initial TMB treatment. Fifteen patients (30 eyes) with no history of TED, GD, ocular surface disease, or meibomian gland dysfunction were used for comparative controls. The mean age of the control group was 48.4 ± 13.2 years (p = 0.61), and patients were 81.8% female. All control patients had baseline TSH and TSI within normal limits. Tear fluid specimens were safely obtained using Schirmer strips without complication.

Fig. 1. Mass spectrometry workflow.

a Clinical presentation of severe thyroid eye disease with pronounced conjunctival injection, proptosis, and eyelid retraction, complemented by corresponding computed tomography images. After TMB, the patient’s inflammatory signs have improved. b Heatmap detailing the demographic characteristics of the study cohort, including age, sex, smoking status, TED duration, baseline clinical activity score, diplopia score, and the number of TMB infusions administered. Most patients were female, predominantly in the 40–59 age group, non-smokers, experienced TED for 1–2 years, and received 8 infusion treatments. c Graph correlating the clinical activity score pre- and post-TMB treatment with factors such as the duration of having thyroid eye disease, gender, and baseline ocular protrusion, measured using Hertel ophthalmometer. d Proteomics workflow for acquiring tear fluid samples and subsequent analyzing via LC–MS/MS. Graphics in panel D were generated using BioRender.

Proteins from TED and control samples were then extracted from Schirmer strips, trypsin digested, and separated by liquid chromatography before analysis by tandem mass spectrometry (LC–MS/MS; Supplementary Data 2 and Fig. 1d). We identified a total of 2974 proteins across 76 tear samples (32 control eyes, 22 TED pre-treatment, 22 TED post-treatment). Global protein abundance trends derived from spectral counting (semi-quantitative) were concordant with MS1 intensity-based LFQ measurements (Pearson r = 0.67, 95% CI 0.63–0.72; Spearman ρ = 0.43; p < 2.2 × 10⁻¹⁶; n = 532 proteins; n = 76 eyes; Supplementary Fig. 1). There were 256 proteins with a mean spectral count greater than two that were commonly expressed in all samples. These results show considerable variability among low-abundant proteins, similar to previous studies reporting a wide dynamic range and variability of protein expression in human tears^55–59. To avoid the noise and poor reproducibility associated with low-abundant proteins, spectral counts were normalized across all samples prior to differential expression analysis using the calcNormFactors function from the R package edgeR⁴³. Unsupervised principal component analysis demonstrated separation between the TED cases and control tears based on normalized protein intensities (Fig. 2a). Intriguingly, while the post-treatment cohort demonstrated a global directional coordinate shift away from the pre-treatment baseline on PCA, a subset of two post-TMB samples exhibited an intermediate layout, clustering more closely with the pre-TMB baseline space (Fig. 2a). To determine whether this visual proximity reflected an incomplete therapeutic response, a dedicated retrospective chart review was conducted for these two specific individuals. Both patients met the standardized criteria for clinical success, showing a ≥ 2—point reduction in CAS and an objective reduction in proptosis, with no evidence of disease reactivation, atypical adnexal features, or distinct adverse events. Consequently, this intermediate visual positioning on the global PCA matrix does not correlate with clinical non-response. Instead, it likely reflects individual baseline physiological variance—such as minor baseline fluctuations in tear film volume or historical ocular surface homeostasis—rather than a failure of therapeutic molecular normalization. Another important observation was that TMB therapy did not shift the TED tear proteome back to the normal tear cluster but instead remained independent. This suggests that while anti-IGF-1R therapy alters the tear proteome, it does not fully restore the molecular composition of the tear film, something that correlates with clinical observations¹³.

Fig. 2. Protein expression changes in the TED tear film induced by TMB.

a The score plot of PC1 and PC2 between TED cases and normal tears based on normalized protein expression. Spectral counts were normalized across all samples prior to differential expression analysis using the calcNormFactors function from edgeR. b Differentially expressed proteins between control and TED tears (pre-TMB) detected by LC–MS/MS represented as a volcano plot. The horizontal axis (x-axis) displays the log2 fold-change value (pre-TMB vs. controls) and the vertical axis (y-axis) displays the noise-adjusted signal as the −log10 (p-value). c Differentially expressed proteins between control and TED tears following treatment with TMB. d Differentially expressed proteins between pre-TMB and post-TMB tears in TED patients. Statistical significance across group abundance trends is determined using a two-sided linear mixed-effects model framework via the limma package to account for paired inter-eye correlations, with adjustments for multiple comparisons executed using the Benjamini–Hochberg false discovery rate (FDR) procedure (threshold set at FDR < 5%). e Overlap of differentially expressed proteins across contrasts. UpSet plots show the intersections of differentially expressed proteins (DEPs) between the three pairwise contrasts: Pre-TMB vs. control, post-TMB vs. control, and post- vs. pre-TMB. The top panel shows the number of DE proteins shared among contrasts, with each column representing a unique combination of contrasts. The connected black dots below indicate which contrasts are included in each intersection. Separate plots are shown for upregulated and downregulated proteins to highlight direction-specific changes.

To ensure the structural integrity of our longitudinal signals and confirm that group homogeneity was not compromised by the inclusion of a single participant who discontinued TMB after 4 infusions, we executed a parallel leave-one-out validation stress test. Omitting this patient’s complete paired longitudinal sample set, we re-ran our entire linear mixed-effects differential expression pipeline across all detected features. The baseline proteomic topology was stable: of the 375 primary DEPs identified between pre- and post-treatment samples, 92.8% (348 proteins) were preserved as statistically significant DEPs in the sensitivity model. Furthermore, evaluating the global correlation of longitudinal abundance changes between the full and sensitivity frameworks yielded a Pearson correlation coefficient of r = 0.9885. This confirms that our primary statistical signatures and directional trajectories represent genuine collective therapeutic behaviors rather than outlier effects driven by a single individual. We identified 361 proteins that were differentially expressed among control and pre-TMB samples (133 upregulated proteins and 228 downregulated proteins; adjusted p-value < 0.05; |log2FC| >0.5; Fig. 2b and Supplementary Data 3). Among the upregulated proteins were ryanodine receptor 3 (RYR3), complement factor B (CFB), CDK5 and ABL1 enzyme substrate 1 (CABL1), interleukin-1 receptor antagonist (IL1RN), mesothelin (MSLN), S100-A8, and S100-A9. Among the downregulated proteins were glutathione S-transferase A2 (GSTA2), superoxide dismutase C, annexin A5, and tissue metalloproteinase inhibitor 1 (TIMP1).

The tear proteome was significantly altered after TMB therapy. A total of 393 proteins were differentially-expressed between control and post-TMB tears (261 upregulated, 132 downregulated; adjusted p-value < 0.05; |log2FC| >0.5; Fig. 2c and Supplementary Data 4). Additionally, there were 375 proteins that were differentially expressed between pre-TMB and post-TMB tears (318 upregulated, 57 downregulated; adjusted p-value < 0.05; |log2FC| >0.5; Fig. 2d and Supplementary Data 5). To ensure that our chosen feature-selection boundaries captured genuine biological signals rather than background noise, we applied direction-specific multi-set intersection analyses using the UpSetR package to cleanly segregate upward and downward regulation patterns across our cohorts. This directional partitioning isolated four highly distinct kinetic sub-trajectories, revealing that lower-magnitude regulatory features maintained high coordinate stability and functional consensus rather than behaving as sporadic false positives (Fig. 2e). DEPs proteins were classified based on their expression trajectories across control, pre-TMB, and post-TMB groups (Fig. 3a). The largest subset of proteins (n = 167) exhibited treatment-specific changes, characterized by minimal alteration in pre-TMB samples followed by significant modulation after TMB, indicating a dominant TMB-driven proteomic response (Fig. 3b and Supplementary Data 6). Among these proteins were those involved in immune regulation [e.g., polymeric immunoglobulin receptor, lymphocyte-specific protein 1, and high mobility group protein B1] and actin cytoskeletal organization [e.g., adenylyl cyclase-associated protein 1, cofilin-1, destrin, and WD repeat-containing protein 1].

Fig. 3. Trajectory analysis of differentially expressed proteins identifies protein alterations reversed by TMB.

a DEPs identified from each group comparison (i.e., pre-TMB vs. control, post-TMB vs. control, and post- vs. pre-TMB) were classified into trajectories based on their paired longitudinal expression patterns. DEPs identified from each group comparison were classified into distinct sub-trajectories. Proteins were explicitly partitioned into distinct sub-trajectories: (1) persistent upregulation (elevated at baseline and remaining elevated post-treatment); (2) persistent downregulation (depressed at baseline and remaining depressed post-treatment); (3) TMB-reversed (up to down; significantly elevated at baseline but restored back toward healthy control coordinates following therapy); and (4) TMB-reversed (down to up; significantly depressed at baseline but rescued back toward healthy control levels post-treatment).. Results are represented as a heatmap and display protein expression levels on a logarithmic scale. Orange indicates high expression, while dark green/black indicates low or no expression. b Mean z-scores of protein expression changes across the six protein trajectories. Each graph represents the standardized change in expression of proteins within a specific trajectory, with the z-score quantifying how much a protein’s expression deviates from the overall mean of all proteins. Results are displayed as mean ± SEM.

A substantial proportion of proteins (n = 153) showed persistent alterations in both pre- and post-TMB tears relative to controls, suggesting sustained pathway dysregulation that was not normalized by treatment. Eighty-nine of these proteins were persistently downregulated, while 64 were persistently upregulated following TMB treatment. Among the persistently downregulated proteins were those related to antimicrobial responses [e.g., lactoperoxidase and neutrophil defensin 1] and cellular detoxification [e.g., superoxide dismutase-1, superoxide dismutase-2, and GSTA2; Fig. 3b and Supplementary Data 6]. Among the persistently upregulated proteins were CFB, annexin A2, ERO1-like protein alpha, NADPH dehydrogenase [quinone] 1, and NADPH oxidase organizer 1. A smaller subset of proteins (n = 42) demonstrated complex or mixed expression patterns, reflecting heterogeneous responses that did not conform to monotonic or reversible trends across groups (Fig. 3b and Supplementary Data 6). In contrast, relatively few proteins exhibited progressive monotonic changes, with 53 proteins showing a progressive increase and only 12 proteins displaying a progressive decrease across pre- and post-TMB groups. Among the 53 proteins displaying a progressive increase in expression were IL1RN, ezrin (EZR), sciellin, involucrin (IVL), and small proline-rich protein 3 (Fig. 3b and Supplementary Data 6). Among the 12 proteins displaying a progressive decrease across pre- and post-TMB groups was albumin (Fig. 3b and Supplementary Data 6).

Notably, 143 proteins displayed significant alterations prior to treatment that were partially or fully normalized following TMB, consistent with molecular reversal of baseline perturbations. Sixteen of these proteins [e.g., CABL1, granulocyte macrophage colony-stimulating factor receptor subunit α, and TBC1 domain family member 8B] were significantly upregulated in pre-TMB tears and were normalized to control levels following TMB treatment (Fig. 3b and Supplementary Data 6). GM-CSF has been previously implicated in DED pathogenesis through the promotion of CD11b+ cell activation and migration to the ocular surface⁶⁰. The remaining 127 proteins were significantly downregulated in pre-TMB tears (compared to controls) and were normalized to control levels following TMB treatment. Among these proteins were mucin-5AC (MUC5A), TIMP1, gelsolin (GSN), perlecan (HSPG2), complement factor H, apolipoprotein D, and several immunoglobulin fragments (e.g., IGLV3-21, IGLV1-51, IGLV6-57, IGKV2-30, IGKV1-39, and IGLV1-40; Fig. 3b and Supplementary Data 6). Decreased MUC5A, GSN, and perlecan levels have been previously linked to DED^61–64. These results suggest that TMB restores MUC5A and perlecan levels in tears, promoting tear film stabilization and ocular surface health.

Correlation of proteomic shifts with clinical phenotypes

To evaluate whether tear protein abundance reflected clinical disease features, we performed correlation analyses between LFQ-based tear protein expression and clinical metrics in pre-treatment TED samples, including CAS score, proptosis, GDS, Schirmer test values, and systemic thyroid markers (i.e., TSH and TSI levels). Correlations were assessed using Spearman rank correlation to accommodate non-normality and potential outliers, and p-values were adjusted for multiple testing using the BH FDR. Among all parameters evaluated at baseline, Schirmer tear volume demonstrated the most robust and profound molecular footprint, yielding the highest number of proteins surviving stringent multiple-testing correction (Fig. 4a and Supplementary Data 7). Volcano plot profiling of the Schirmer axis identified several distinct, FDR-significant protein clusters (Fig. 4b). Specifically, alpha-1-antitrypsin (SERPINA1; ρ = −0.7148, FDR = 0.011; Fig. 4c), Transferrin (TF; ρ = −0.6218, FDR = 0.041; Fig. 4d), and glutathione S-transferase P (GSTP1; ρ = −0.7934, FDR = 0.003; Fig. 4e) all exhibited strong inverse correlations with tear production. The robust upregulation of these plasma-associated leakage markers and cellular detoxification enzymes in eyes with low tear volume indicates that severe aqueous deficiency in TED directly co-occurs with severe local oxidative stress and blood-tears barrier breakdown, a pattern previously described in dry eye and ocular surface disease^65,66. Conversely, metrics assessing distal orbital volume expansion (proptosis, diplopia) or systemic immune signaling (e.g., TSH, TSI) yielded few or no proteins capable of surviving FDR correction (Fig. 4a), underscoring that the cross-sectional tear proteome primarily reflects immediate, localized ocular surface physiology. However, at nominal significance thresholds (i.e., p < 0.05), key pathways emerged that highlighted indirect mechanical tissue stress. Most notably, the lacrimal-secreted cysteine protease inhibitor cystatin-SN (CST1) demonstrated a powerful, log-linear positive correlation with systemic, pathogenic TSI titers (ρ = 0.9326, FDR < 0.001; Fig. 4f). Because high circulating TSI antibodies drive proptosis and subsequent eyelid lag, this prominent upregulation of CST1 likely serves as a protective localized feedback mechanism to counteract exposure-driven proteolytic breakdown of the corneal epithelium^3,67. Finally, to identify molecular markers that act as direct kinetic indicators of TMB efficacy, we matched longitudinal proteomic fluctuations against therapeutic shifts in inflammation (i.e., CAS response; Fig. 4g). Strikingly, haptoglobin (HP) emerged as the sole protein to survive rigorous FDR correction across the longitudinal delta model (ρ = −0.7858, FDR = 0.012; Fig. 4g). Plotting the kinetic velocity (ΔLFQ vs. ΔCAS) revealed a highly symmetrical relationship: patients who experienced the most dramatic clinical reductions in active orbital inflammation (ΔCAS) exhibited the most profound clearance of local tear haptoglobin. Given haptoglobin’s established role as an acute-phase responder to hyperosmolar epithelial friction and macrophage infiltration, these data indicate that haptoglobin may serve as a highly responsive, tear-specific “molecular thermometer” capable of tracking the resolution of localized ocular surface pathology during TED therapy^67,68.

Fig. 4. Summary of associations between tear protein abundance and clinical metrics in pre-treatment TED.

a For each clinical variable (CAS, proptosis, GDS Gorman diplopia score, TSH thyroid-stimulating hormone, TSI thyroid-stimulating immunoglobulin, and Schirmer test values), the number of tear proteins showing significant associations with LFQ-based protein abundance is shown. Filled circles indicate the number of proteins remaining significant after multiple-testing correction (false discovery rate [FDR] <0.05), whereas open circles denote proteins meeting nominal significance (Spearman p < 0.05). Schirmer test values yield the largest molecular footprint, indicating that the cross-sectional tear proteome primarily reflects immediate, localized ocular surface physiology rather than distal orbital or systemic disease metrics. b Volcano plot showing Spearman correlations between LFQ-based tear protein abundance and Schirmer test values in pre-treatment TED samples. Each point represents a protein; the x-axis denotes the Spearman correlation coefficient (ρ), and the y-axis indicates −log₁₀(FDR-adjusted p-value). The dashed horizontal line represents the threshold for significance (FDR = 0.05). Proteins remaining significant after multiple-testing correction (FDR < 0.05) are highlighted, with representative proteins involved in inflammatory leakage and epithelial oxidative stress annotated. c–e Linear scatter plots displaying robust inverse baseline correlations between Schirmer test values (mm) and the log2-transformed intensities of c alpha-1-antitrypsin (SERPINA1), d transferrin, and e glutathione S-transferase P (GSTP1). Green shaded regions denote 95% confidence intervals; goodness-of-fit (R²), Spearman rank coefficient (ρ), and Benjamini–Hochberg FDR values are reported. f Log-linear scatter plot demonstrating a baseline positive correlation between local tear cystatin-SN (CST1) abundance and systemic, pathogenic thyroid-stimulating immunoglobulin (TSI) titers. The x-axis is plotted on a logarithmic scale to account for clinical skew; the orange shaded region represents the 95% confidence interval. g Longitudinal delta model tracking molecular shifts against clinical treatment response. Left panel: Spaghetti plot demonstrating individual patient eye trajectories for Clinical Activity Score (CAS) across Pre-, During-, and Post-treatment clinical intervals. Right panel: Kinetic velocity scatter plot pairing the longitudinal change in local tear haptoglobin [∆LFQ = Log2(Post)–Log2(Pre)] against the magnitude of clinical anti-inflammatory resolution (∆CAS = Pre–Post). The strong, symmetrical negative correlation indicates that localized haptoglobin clearance closely tracks physical disease improvement, acting as a responsive reporter of homeostatic ocular surface restoration. The coral shaded region denotes the 95% confidence interval.

Functional characterization of the TED tear proteome

To further classify the dynamics of the TED tear proteome and resolve coordinate tissue responses, functional enrichment analysis was performed across the defined kinetic trajectories using ReactomeDB and Metascape, complemented by STRING-db network mapping of key pathogenic hubs^49–51. Functional classification revealed that a significant fraction of DEPs govern the regulation of cell survival, necrosis, and tissue motility. The molecular footprint of active orbital remodeling was clearly reflected on the ocular surface; pathways associated with cellular responses to mechanical stimuli (p = 1.1 × 10⁻⁴) and endothelial response to fluid shear stress (p = 1.1 × 10⁻⁴) were highly enriched within the complex/mixed trajectory, capturing the downstream physical strain, tissue friction, and severe barrier distress imposed by proptosis and eyelid lag (Fig. 5a and Supplementary Data 8). Within the complex/mixed and TMB-specific trajectories, neutrophil degranulation emerged as a dominant, highly significant signature (p = 4.2 × 10⁻¹⁶; Fig. 5a), highlighting sustained innate immune activation on the ocular surface. While normal, tightly regulated neutrophil degranulation restricts host tissue damage through localized phagosomal release, chronic hyperosmolar or mechanical friction can drive runaway extracellular release, promoting localized tissue breakdown⁶⁹. This inflammatory footprint was further mirrored by an enrichment in complement regulatory signaling (p = 6.9 × 10⁻⁵) within both the complex and reversed trajectories. Pathways associated with carbohydrate metabolism, specifically glycolysis and gluconeogenesis, were enriched among progressively downregulated DEPs (p = 1.1 × 10⁻⁴), indicating a localized metabolic disruption in the TED tear film that escapes TMB corrective action. Conversely, the clotting cascade and pathways regulating fibrin clot formation were the most highly enriched terms among proteins exhibiting progressive upregulation (p = 4.2 × 10⁻⁵), likely tracking microvascular permeability and plasma leakage across the ocular barrier (Fig. 5a and Supplementary Data 8). Importantly, our analysis captured direct, localized evidence of successful disease-specific therapeutic targeting. In healthy control tears, the pathway governing the regulation of insulin-like growth factor (IGF) transport and uptake by IGF binding proteins was highly enriched, but was severely depleted in the baseline, pre-TMB tear film. Because TED pathogenesis is driven fundamentally by autoantibody-mediated crosstalk at the IGF-1 receptor (IGF-1R) on orbital fibroblasts, this baseline tear depletion highlights a major local loss of protective IGF regulators. Following treatment, this cascade demonstrated a robust therapeutic reversal (p = 2.9 × 10⁻⁷; Fig. 5a and Supplementary Data 8). To dissect this recovery at a resolution of individual effectors, a high-confidence protein-protein interaction network was constructed (Fig. 5b). Indeed, 24 coordinated network proteins (including local protective barrier components and systemic acute-phase modulators like haptoglobin, SERPINA1, CST1, and transferrin) exhibited significant recovery post-treatment, confirming that systemic TMB therapy successfully restores the local homeostatic balance of IGF regulatory effectors within the ocular surface environment.

Fig. 5. Functional characterization of the TED tear proteome.

a Top enriched pathways across protein expression trajectories. Dot plots display the most significantly enriched pathways for each trajectory of protein expression changes from ReactomeDB. Each facet corresponds to a distinct trajectory group (e.g., progressive up, progressive down, reversed by TMB, TMB-specific, persistent). Dot size represents the number of proteins contributing to the pathway, and dot color corresponds to the significance level of enrichment (−log₁₀ P-value). Pathways are ordered by significance within each trajectory facet. Statistical significance for individual protein trajectory longitudinal normalizations is evaluated using a two-sided empirical Bayes moderated t-test, with multiple testing corrections applied via the Benjamini–Hochberg method. b Protein-protein interaction network of IGF regulators identified in TED tear samples. Nodes (proteins) are sized according to their respective log2FC (post- vs pre-TMB) and colored by their respective −log(p-value) from 1-way ANOVA.

Network analysis was performed to identify functional protein-protein interactions (PPIs; verified by experimental data or databases) among upregulated proteins in each group (control, pre- and post-TMB; Supplementary Fig. 2)⁷⁰. Among the DEPs between the three groups, 268 nodes (proteins) formed a complex network with 2158 edges (interactions; Supplementary Fig. 2a). The Molecular Complex Detection (MCODE) algorithm was used to extract protein complexes embedded in the complex network for functional annotation through pathway and process enrichment (Supplementary Fig. 2b)⁷⁰. A majority of the DEPs clustered in protein complexes representing MAPK signaling, acute-phase response, VEGF signaling, keratinization, and hemostasis (Supplementary Fig. 2b). When separated by group, control tear proteins clustered in complexes representing gene ontology (GO) terms reflecting metabolic pathways and cellular detoxification processes. Pre-TMB tear proteins clustered in complexes representing GO terms such as “signaling by Rho GTPases” and “regulation of toll-like receptor by endogenous ligands.” Conversely, MCODE identified proteins clustered in complexes reflecting metabolic and protein folding pathways in post-TMB tears. This overall suggests that there is a loss of proteins mediating metabolism and protein folding in the TED tear film that may be partially restored after treatment with TMB. Such alterations may reflect downstream consequences of chronic inflammation and oxidative stress rather than direct IGF-1R–dependent effects, providing a potential explanation for the incomplete molecular recovery observed following TMB therapy.

Drug repositioning

A total of 64 proteins remained persistently upregulated following TMB treatment, while 53 proteins demonstrated a progressive increase in expression across the pre- and post-TMB groups (Fig. 3b and Supplementary Data 6), suggesting sustained or treatment-associated dysregulation and potential relevance to disease biology. Given their persistence in the TED tear film proteome, these proteins could be prioritized as candidate targets for drug repositioning. To identify pharmacologic agents with potential therapeutic relevance, the upregulated protein set was queried against the Drug–Gene Interaction database (DGIdb)⁵². A composite priority score was calculated for each predicted protein-drug interaction as the sum of three components (i.e., DGIdb interaction score, fold-change contribution, and drug approval status; see “Methods”). This approach allowed systematic prioritization of interactions supported by multiple lines of evidence while favoring compounds with established regulatory approval. To bridge these bioinformatic predictions with clinical utility, candidate interactions were subjected to secondary clinical curation to exclude high-toxicity systemic agents and cross-reference candidates against known ocular toxicities. This refined profiling yielded a highly focused, translational matrix consisting of 9 distinct protein targets overlapping with 10 FDA-approved therapeutic candidates (Fig. 6a and Supplementary Data 9). Hierarchical clustering of the final curated matrix revealed distinct patterns of drug–protein interaction profiles and highlighted potential hubs of therapeutic relevance (Fig. 6a). For example, the complement cascade was represented by the prominent upregulation of CFB, which was targeted with high priority by the small-molecule inhibitor iptacopan⁷¹.

Fig. 6. Partial molecular normalization of the tear film proteome by TMB.

a Drug repositioning candidates were identified among the differentially expressed proteins in TED tears by searching for the implicated proteins in DGIdb. A composite priority score was calculated for each protein-drug interaction as the sum of three components: (1) DGIdb interaction score, (2) fold-change contribution, and (3) drug approval status (see “Methods”). Results are displayed as a heatmap highlighting proteins targeted by multiple high-priority drugs, as well as drugs acting on multiple dysregulated proteins. Hierarchical clustering groups proteins and drugs with similar interaction profiles. b Schematic representation of TED pathophysiology and pathways mediating orbital fibroblast activation, inflammation, and fat proliferation. c Schematic representation of the TED tear film proteome and representative proteins from the following trajectories: Persistent (Up and Downregulated), Progressive Up, and Reversed by TMB. Graphics in (b and c) were generated using BioRender.

Localized ocular surface stress and mechanical friction hubs were captured by the altered expression of COX8A. This metabolic distress axis mapped to a pathway-level cluster of commercially available topical ophthalmic non-steroidal anti-inflammatory drugs (e.g., nepafenac, ketorolac, and diclofenac), which are standardly implemented to suppress downstream prostaglandin cascades and mitigate exposure-driven hyperosmolar tissue injury⁷². Chronic epithelial barrier disruption was highlighted by the upregulation of IVL, matching downstream targeted anti-TNFα biologics (e.g., infliximab and adalimumab)^73,74. Furthermore, systemic immunomodulatory pathways capable of targeting multiple dysregulated targets were proteins; most notably, methotrexate exhibited pleiotropic interactions by simultaneously targeting both IL1RN and S100A8^75–77. Finally, cellular structural remodeling and intracellular calcium handling were captured by tubulin alpha 1B (TUBA1B) and RYR3, mapping to colchicine and dantrolene, respectively^78,79. Taken together, this curated drug-repositioning landscape isolates highly targeted, clinically rational therapeutic pathways tailored to the persistent molecular signatures of the TED tear film.

Discussion

The pathophysiology underlying TED is complex and incompletely understood. Exploration of the various molecular mediators of TED pathogenesis has identified several molecular processes contributing to TED development, including cytokine signaling, cellular immunity, adipogenesis, and glycosaminoglycan synthesis⁸⁰. These recent findings have shifted the treatment paradigm towards targeted biologic agents, such as TMB. In this study, we present the first paired longitudinal analysis of tear film proteomic changes in patients with TED treated with TMB, using a noninvasive approach to identify novel disease mechanisms and therapeutic targets associated with ocular surface involvement. Using shotgun mass-spectrometry, we identified 133 significantly expressed proteins in pre-TMB TED tears, including proteins that have not been previously implicated in TED pathogenesis, such as RYR3, CABL1, and MSLN. Notably, several tear proteins correlated with CASs, including S100A9, thioredoxin, and CABL1, suggesting that tear proteomics may provide a noninvasive means of assessing disease activity and treatment response. S100A8 and S100A9 have been previously detected in the TED tear film. S100A8 is an innate immune response activator that is secreted by neutrophils and monocytes and forms a complex with S100A9 (calprotectin) to mediate the recruitment of immune cells to sites of inflammation⁸¹. In addition, elevated serum levels of S100A8 and calprotectin have been previously associated with autoimmune thyroid diseases and Graves’ orbitopathy and increase expression of IL-6 and IL-8 in orbital fibroblasts⁸². More broadly, these results build upon established frameworks demonstrating that high-throughput tear proteomics can successfully dissect the complex subclinical mechanisms driving active orbital inflammatory disorders⁸³.

The tear film plays roles in immune surveillance, antimicrobial defense, and providing oxygen and nutrition to the avascular corneal epithelium⁸⁴. Thus, it was interesting to find that proteins implicated in innate immunity (e.g., complement cascade, neutrophil degranulation, and humoral immune response), metabolism (e.g., nucleotide, selenium, and aerobic metabolism), and maintenance of cell survival were depleted in the TED tear film compared to healthy control tears, suggesting that disruption of these processes also contribute to the ocular surface phenotype of TED. While we did not detect IGF-1R in the TED tear film, downregulation of IGF regulatory effectors and binding proteins indicated dysregulation of this pathway in our patient cohort. Because it is possible to perform serial tear collections, future studies could investigate patient-specific trends in protein expression in response to different therapies⁸⁵. More importantly, our findings demonstrate that TED is associated with a distinct tear proteome that remains molecularly abnormal despite substantial clinical improvement following IGF-1R inhibition. While TMB induces broad proteomic alterations, the tear film does not revert to a normal, control-like state, highlighting a dissociation between orbital disease control and ocular surface homeostasis. This observation provides molecular context for the persistent dry eye symptoms commonly reported by patients with TED, even after improvement in inflammatory signs and proptosis¹³. This independent compartmentalization is strongly corroborated by recent cross-sectional profiling of the TED tear film by ref. ⁸⁶. While their study design relied on pooled tear samples, which can inadvertently smooth over individual biological extremes, it serves as an invaluable macro-level reference verifying that core tissue-remodeling cascades remain significantly altered at the ocular surface interface. By utilizing an unpooled, individualized patient architecture, our longitudinal data extend these macro-level findings to track how these specific tissue-remodeling cascades unfold within individual patients across therapy. Furthermore, our exploration of individual proteomic trajectories highlighted the expected heterogeneity inherent to human tear fluid^45,47. Because these intermediate PCA outliers demonstrated objective clinical success (i.e., ≥2-point CAS reductions and proptosis decompressions), designating them as a distinct “incomplete responder” subgroup based solely on exploratory visual coordinates would risk type-I selection bias. Instead, these profiles demonstrate that patients can achieve identical clinical outcomes while maintaining subtle variations in their underlying proteomic recovery pathways.

TED-associated ocular surface disease is a highly complex, mixed-mechanism pathology driven by a combination of evaporative loss, mechanical exposure keratopathy, lagophthalmos, and meibomian gland dysfunction. Classic parameters such as tear break-up time (TBUT), meibography, and corneal fluorescein staining were not uniformly available for this retrospective cohort due to its clinical focus on deep orbital remodeling. Thus, our tear proteomic data must be interpreted as a high-resolution snapshot primarily capturing the aqueous-phase alterations and underlying subclinical inflammatory pathways mediated by the lacrimal gland axis, rather than an exhaustive representation of the evaporative or mechanical spectrum of dry eye pathology. Future prospective clinical studies matching high-throughput tear omics with uniform TBUT and Ocular Surface Disease Index testing are required to establish granular mechanistic links between individual protein trajectories and dry eye symptom profiles.

Our protein trajectory analysis revealed that TMB treatment resulted in significant changes in tear protein composition, consistent with its known effects on immune signaling and fibroblast activation. However, only a limited subset of proteins exhibited partial or complete normalization to control levels following treatment, indicating that IGF-1R blockade alone is insufficient to fully reverse ocular surface dysregulation. Several proteins that were normalized after TMB, including MUC5A and perlecan, are known to play roles in tear film stability and epithelial barrier function and have been previously implicated in DED^61–63. Restoration of these protective components suggests that TMB may confer some benefit to the ocular surface, potentially through indirect effects on lacrimal gland function or local inflammatory signaling. Nonetheless, the persistence of widespread molecular abnormalities underscores the complexity of TED-associated ocular surface disease and suggests that additional pathways contribute to ongoing pathology.

A prominent feature of the post-treatment tear proteome was the persistence of inflammatory, oxidative stress–related, and epithelial stress proteins, including components of neutrophil degranulation and ROS detoxification pathways. Dysregulation of these processes has been linked to chronic ocular surface inflammation and epithelial damage in other dry eye and inflammatory eye diseases⁶⁹. For example, the sustained upregulation of proteins such as EZR and HS90A suggests ongoing cellular stress and immune activation despite clinical quiescence of orbital disease. This disruption aligns closely with the comprehensive systematic review by Zhang et al., which synthesized multi-omic evidence demonstrating widespread, coordinated dysregulation across interconnected immune activation, apoptosis, cell cycle regulation, and protein processing pathways during active disease phases⁸⁷. These findings support a model in which the ocular surface represents a partially independent disease compartment in TED, with molecular drivers that may persist after systemic or orbital inflammation has improved.

Finally, our proteomic data point to actionable targets for drug repositioning in TED. By leveraging persistent and progressively upregulated proteins as candidate therapeutic targets, we demonstrate the feasibility of integrating tear proteomics with drug–protein interaction databases to identify rational drug repositioning targets. Our manual curation strategy highlighted a spectrum of FDA-approved drugs capable of targeting pathological cascades not addressed by standard TMB therapy, including complement and IL-1 signaling. Complement activation is well documented in ocular surface inflammation and previous proteomic studies of TED tears have identified increased complement expression⁸⁸. Similarly, tear film levels of IL-1R1 levels are implicated in inflammatory dry eye and Sjögren’s syndrome⁸⁹. Our analysis also pointed to systemic immunomodulators and biologics, including methotrexate (which exhibits pleiotropic interactions across both IL1RN and S100A8 macrophage recruitment pathways) and anti-TNFα monoclonal antibodies (adalimumab and infliximab)^75–77. Notably, these anti-TNF biologics map downstream to the marked upregulation of IVL, capturing a known axis where anti-TNF therapy successfully attenuates localized tissue stress and suppresses downstream chronic epithelial barrier breakdown^73,74. Collectively, our results support a multi-pathway model of TED in which IGF-1R–mediated orbital inflammation represents only one component of a broader disease process affecting the ocular surface. The persistence of tear proteomic abnormalities after TMB treatment underscores the need for combination or sequential therapeutic strategies that address residual inflammation, oxidative stress, and epithelial dysfunction. Findings may generalize best to adults with active moderate-to-severe TED receiving TMB in tertiary referral settings; applicability to mild or inactive TED requires validation. It is important to note that our computational drug repositioning workflow serves as an exploratory, hypothesis-generating discovery platform rather than a definitive translational clinical protocol. Physical target validation within localized ocular surface tissue models, detailed pharmacokinetic profiling, and explicit evaluations of delivery routes and local bioavailability are absolute prerequisites. Future studies with larger cohorts and longer follow-up will be essential to validate candidate biomarkers, assess their predictive value, and determine whether targeted adjunctive therapies can meaningfully improve ocular surface disease and long-term outcomes in patients with TED.

This study has limitations. As an observational cohort study without randomization, residual confounding and selection bias are possible, and causal inferences regarding TMB’s effects on the tear proteome should be made cautiously. We acknowledge that our participant cohort is compact (n = 11 patients, 16 controls), which precludes traditional, independent a priori sample size calculations. However, high-dimensional discovery proteomics workflows utilizing paired longitudinal sampling operate under vastly different statistical mechanics than standard cross-sectional clinical studies^90–93. Because each patient serves as their own direct internal baseline control, confounding inter-individual genetic, environmental, and physiological noise is minimized. This self-controlled architecture substantially amplifies statistical power, enabling the detection of true biological signals within small cohorts. Furthermore, the 375 identified DEPs did not distribute randomly across the proteome, as would be expected with widespread false discovery under BH thresholding. Instead, they demonstrated tight functional convergence into highly coordinated, disease-relevant biological pathways—including IGF transport networks, neutrophil degranulation cascades, and extracellular matrix wound-healing pathways. While these findings represent a high-confidence, internally validated discovery blueprint, prospective multi-center studies with larger independent sample sizes remain essential to confirm these biomarker signatures externally. Additionally, while our small patient cohort naturally presents constraints regarding absolute group homogeneity, our inclusion threshold (≥4 infusions) was intentionally structured to model real-world clinical pragmatism, where unexpected treatment cessations due to toxicities (such as ototoxicity) occur. In addition, technical factors related to tear collection using Schirmer strips, variability in protein recovery, and the dynamic range limitations of mass spectrometry may influence protein detection and quantification. Consequently, validation in larger, independent cohorts is warranted. Although serial sampling was performed, additional longitudinal sampling within individual patients could further elucidate patient-specific proteomic trajectories and treatment responses. Future studies should include prospective validation and integrate complementary biospecimens, such as orbital fat and serum, to determine whether protein expression changes observed in tears reflect broader molecular alterations across disease-relevant tissues.

Author contributions

Study concept and design: A.L.K. and V.B.M. Acquisition of data: G.V., L.A., A.L.K., V.B.M. Analysis and interpretation of data: G.V., J.W., G.N., D.Y., V.B.M. Drafting of the manuscript: G.V., G.N., A.L.K., V.B.M. Critical revision of the manuscript for important intellectual content: G.V., J.W., A.D., A.L.K., V.B.M. Statistical analysis: G.V., J.W., and V.B.M. Obtained funding: A.L.K. and V.B.M. Administrative, technical, and material support: A.L.K. and V.B.M. Study supervision: A.L.K. and V.B.M. Dr. Mahajan had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Peer review

Peer review information

Communications Medicine thanks David Toro Tobon and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

The mass spectrometry proteomics data and search outputs generated during the current study were deposited to the ProteomeXchange Consortium via the PRIDE partner repository with the dataset identifier PXD066511 and are publicly accessible at https://www.ebi.ac.uk/pride/archive/projects/PXD066511. Source data underlying all figure panels, including individual-level quantification profiles and summary statistics, are provided as a multi-tab workbook in Supplementary Data 1. The source data for Fig. 1 are in Supplementary Data 1 (Tab: Fig. 1); the source data for Fig. 2 are in Supplementary Data 1 (Tab: Fig. 2); the source data for Fig. 3 are in Supplementary Data 1 (Tab: Fig. 3); the source data for Fig. 4 are in Supplementary Data 1 (Tab: Fig. 4); the source data for Fig. 5 are in Supplementary Data 1 (Tab: Fig. 5); and the source data for Fig. 6 are in Supplementary Data 1 (Tab: Fig. 6). Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Vinit B. Mahajan (vinit.mahajan@stanford.edu).

Competing interests

A.L.K. is a consultant for Amgen (consultant, independent contractor), Immunovant (consultant/advisor), Viridian (consultant, independent contractor), Sling Therapeutics (independent contractor), Genentech (consultant), Kriya (consultant), Aceylrin (consultant, independent contractor), Argenx (consultant), Lassen (consultant, independent contractor). The funding organizations had no role in design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s43856-026-01757-6.