Deep Proteomic Analysis With Machine Learning Identifies Aqueous Humor Biomarkers of ADAMTSL4-associated Congenital Ectopia Lentis

Abstract

Purpose

To systematically characterize aqueous humor (AH) proteomic alterations in ADAMTSL4-associated congenital ectopia lentis (CEL) and to identify disease-related molecular features.

Methods

Mass spectrometry–based deep data-independent acquisition (deep DIA) proteomics was employed to profile AH proteomes from pediatric ADAMTSL4-associated CEL patients. Differentially expressed proteins (DEPs) were analyzed using functional enrichment and gene set enrichment analysis. Weighted gene co-expression network analysis (WGCNA) identified disease-related protein modules. Candidate biomarkers were prioritized using machine learning, followed by technical confirmation using intelligent parallel reaction monitoring (iPRM) and clinical correlation analysis. Transcriptional changes of selected candidates were assessed by quantitative PCR in ADAMTSL4-knockdown human retinal pigment epithelial cells, human fibroblasts, and adamtsl4-knockout zebrafish.

Results

Deep DIA quantified 1865 AH proteins, among which 265 DEPs were identified and enriched in extracellular matrix (ECM) remodeling, complement–coagulation cascades, and lipid transport pathways. Expression-based stratification revealed tier-specific functional patterns. WGCNA identified modules significantly associated with ocular phenotypes. Machine learning prioritized six candidate biomarkers (ADAMTS3, APOC2, AMBP, KLKB1, SDC4, and ENPP2), of which APOC2, AMBP, KLKB1, and ENPP2 achieved targeted confirmation by iPRM; APOC2, KLKB1, and ENPP2 were correlated with axial length or choroidal thickness. In ADAMTSL4-knockdown cells, ENPP2, MYDGF, and CA2 were downregulated and LCAT was upregulated, consistent with proteomic findings. MYDGF further showed a concordant directional change in the zebrafish model.

Conclusions

This study established a high-resolution AH proteomic profile of ADAMTSL4-associated CEL, revealing coordinated molecular alterations in ECM disruption, complement–coagulation activation, and dysregulated lipid homeostasis, providing integrated molecular insights and candidate molecular features for understanding this rare ocular disorder.

Congenital ectopia lentis (CEL) is characterized by displacement of the crystalline lens from its normal anatomical position due to structural abnormalities of the zonular fibers,¹ affecting approximately six per 100,000 individuals worldwide.² Although relatively rare, CEL can exert a substantial impact on visual function.³ Despite surgical intervention being the definitive treatment modality, marked interindividual variability exists in the severity of clinical manifestations and visual prognosis,⁴ suggesting pronounced heterogeneity in the underlying molecular mechanisms. ADAMTSL4 is the second most important causative gene for CEL after FBN1.⁵ CEL associated with FBN1 variants is frequently accompanied by prominent skeletal and cardiovascular abnormalities,⁶ whereas ADAMTSL4-associated CEL caused by biallelic variants tends to present earlier, with more complex ocular phenotypes and poorer visual outcomes.^5,7 ADAMTSL4 encodes a non-enzymatic secreted glycoprotein belonging to the ADAMTS-like (a disintegrin and metalloproteinase with thrombospondin motifs-like) protein family. This protein is primarily localized to the extracellular matrix (ECM) and plays an essential role in maintaining the structural stability of microfibrils within the zonular fibers and lens capsule.^8,9 However, the molecular pathogenic mechanisms underlying ADAMTSL4-associated CEL remain poorly understood, and existing Adamtsl4-mutant mouse models fail to recapitulate the characteristic ectopia pupillae phenotype observed in human patients,¹⁰ underscoring the need for human-based studies to elucidate the authentic molecular pathogenesis.

Aqueous humor (AH) is a transparent intraocular fluid that nourishes the avascular tissues of the anterior segment and maintains ocular homeostasis.¹¹ Compared with plasma, AH is less influenced by systemic metabolic and inflammatory states and is therefore more suitable for reflecting the local molecular status of ocular tissues, particularly those of the anterior segment.¹² However, conventional proteomic analysis of AH faces substantial technical challenges.¹³ The clinically obtainable volume of AH is extremely limited, and its protein concentration is markedly lower than that of plasma.¹⁴ Moreover, the AH proteome is jointly regulated by blood–ocular barrier filtration and active secretion from the ciliary body, resulting in an exceptionally wide dynamic range of protein abundance.¹⁵ Consequently, biologically relevant signaling molecules with low abundance are frequently obscured,¹⁶ thereby imposing more stringent requirements on analytical sensitivity, dynamic range, and quantitative stability.

In this study, we employed a deep data-independent acquisition (deep DIA) strategy based on the Orbitrap Astral platform (Thermo Fisher Scientific, Waltham, MA, USA) to achieve high-depth and stable proteomic quantification of micro-volume AH samples. Doing so avoided the excessive sample consumption and batch effects associated with conventional data-dependent acquisition (DDA)-based spectral library construction.^17,18 This strategy also enabled narrow-window acquisition to reduce co-fragmentation interference and improve the quantification of low-abundance peptides compared with conventional DIA.¹⁹ For candidate protein validation, we applied high-throughput and highly reproducible intelligent parallel reaction monitoring (iPRM) to perform targeted quantitative confirmation of key differentially expressed proteins (DEPs).^20,21 Using this integrated workflow, we constructed an AH proteomic fingerprint for patients with ADAMTSL4-associated CEL that not only facilitates the identification of key molecular features of CEL caused by ADAMTSL4 variants but also highlights the strong potential of advanced proteomic technologies for elucidating the molecular mechanisms of human ocular diseases.

Methods

Clinical Cohorts and Ophthalmologic Examinations

This case–control study enrolled pediatric CEL patients with confirmed ADAMTSL4 gene mutations who underwent lens surgery at the Eye, Ear, Nose and Throat Hospital, Fudan University, between August 2021 and February 2025. The inclusion criteria were (1) under 18 years of age at the time of surgery; (2) EL confirmed by slit-lamp examination after pupillary dilation with surgical treatment; (3) presence of biallelic ADAMTSL4 variants identified through panel-based next-generation sequencing in collaboration with Amplicon Gene (Shanghai, China)⁷; and (4) availability of clinical examination records. The exclusion criteria were (1) presence of additional pathogenic variants associated with other ocular diseases; (2) previous history of ocular trauma or intraocular surgery; and (3) severe ocular comorbidities such as endophthalmitis, glaucoma, or retinal detachment. Age-matched patients with congenital cataract (CC) without lens dislocation or other ocular complications served as non-disease controls. The study adhered to the tenets of the Declaration of Helsinki and was approved by the Human Research Ethics Committee (ChiCTR2000039132). Written informed consent was obtained from the guardians of all participants.

All enrolled patients underwent detailed medical history review and comprehensive ophthalmologic examinations before and after surgery. Best-corrected distance visual acuity (BCVA) was measured by an experienced optometrist under cycloplegia. Ocular biometric measurements, including axial length (AL), keratometry (Km), astigmatism (AST), and white-to-white (WTW) distance were assessed using swept-source optical coherence tomography (SS-OCT; IOLMaster 700; Carl Zeiss Meditec, Jena, Germany). OCT imaging was performed using the SPECTRALIS OCT system (Heidelberg Engineering, Heidelberg, Germany) following pupil dilatation. Choroidal thickness (ChT) was measured at the subfoveal region (subfoveal choroidal thickness [SFCT]) and at 0.5 and 1.5 mm nasal (N0.5/N1.5) and temporal (T0.5/T1.5) to the fovea.²² All patients underwent phacoemulsification and intraocular lens implantation, during which AH samples (20–100 µL) were collected by anterior chamber paracentesis under sterile conditions and immediately stored at –80°C for subsequent proteomic analysis.

The required sample size per group (n) was prospectively calculated using the formula: $n={\left(\frac{Z\times CV}{\epsilon }\right)}^2$.²³ Assuming a 95% confidence level (Z = 1.96), an expected coefficient of variation (CV) of 0.40 to conservatively account for the technical variance of the deep DIA platform and biological heterogeneity in human biofluids, and an allowable relative error (ε) of 30% suitable for an exploratory discovery setting, the minimum sample size was determined to be 6.83 per group. Therefore, seven samples per group were adopted for the current study.

Proteomic Analysis Based on Deep DIA

Microscale Protein Extraction and Tryptic Digestion of AH Samples

AH samples were lysed in dissolution buffer (DB) comprised of 6-M urea (Sinopharm Chemical Reagent, Shanghai, China) and 100-mM triethylammonium bicarbonate (TEAB; pH 8.5; Sigma-Aldrich, St. Louis, MO, USA) and were centrifuged at 12,000g for 15 minutes at 4°C. The supernatant was reduced with 1-M dithiothreitol (Sigma-Aldrich) at 56°C for 1 hour, chilled on ice, and then alkylated with iodoacetamide (Sigma-Aldrich) in the dark at room temperature for 1 hour. Protein concentration was determined using a Bradford assay kit (Beyotime, Shanghai, China). Proteins were digested with sequencing-grade trypsin (V5280; Promega, Madison, WI, USA) in 100-mM TEAB at 37°C for 4 hours. Formic acid (Thermo Fisher Scientific) was added to adjust pH to <3, followed by centrifugation at 12,000g for 5 minutes at room temperature. Peptides were desalted on C18 columns (Thermo Fisher Scientific) and lyophilized using a freeze dryer (ScanSpeed 40; LaboGene, Lillerød, Denmark).²⁴

Deep DIA Analysis

Liquid chromatography–tandem mass spectrometry (LC-MS/MS) analysis was performed using an Orbitrap Astral mass spectrometer coupled with a Vanquish Neo ultra-high performance liquid chromatography (UHPLC) system (Thermo Fisher Scientific). Mobile phase A consisted of 99.9% H₂O with 0.1% formic acid, and mobile phase B of 80% acetonitrile with 0.1% formic acid. Lyophilized peptides were re-dissolved in 10 µL of solvent A and centrifuged at 14,000g at 4°C for 20 minutes, and 200 ng of supernatant was injected. Peptide separation was performed on a PepMap Neo C18 analytical column (150 µm × 15 cm, 2 µm; Thermo Fisher Scientific) with a trap column (300 µm × 5 mm, 5 µm) at 50°C. The segmented gradient was as follows: 0 to 0.3 minutes, 4% buffer B; 0.3 to 14.2 minutes, 8% buffer B; 14.2 to 21.1 minutes, 22.5% buffer B; 21.1 to 21.5 minutes, 35% buffer B; 21.5 to 21.9 minutes, 55% buffer B; and 21.9 to 22.6 minutes, 99% buffer B, with a flow rate of 0.8 to 2.5 µL/min.

The full mass spectrometry (MS) scan range was set to 380 to 980 m/z, with a resolution of 240,000 and an automatic gain control (AGC) target of 500%. A total of 300 DIA isolation windows were applied with a normalized collision energy of 25%. Tandem MS spectra were acquired over a range of 150 to 2000 m/z at a resolution of 80,000, with a maximum injection time of 3 ms. Raw files (.raw) were directly searched using DIA-NN 1.8.1 software²⁵ according to the “Homo_sapiens_UniProt/Swiss-Prot” database, which contained 20,436 protein sequences. Cysteine carbamidomethylation was set as a fixed modification and N-term methionine excision as a variable modification, and up to two missed tryptic cleavages were allowed. Peptides with Global.Q.Value < 0.01 and proteins with P_G.Q.Value < 0.01 were retained for quantification.

Relative Quantification and Functional Enrichment Analysis

Protein intensities were normalized across samples, and relative quantification and inter-group comparisons were performed as previously described.²⁶ DEPs were defined as those with fold change (FC) > 1.5 or < 0.67 and P < 0.05 (independent samples t-test). Functional annotation and enrichment analyses of DEPs were performed using InterProScan 5.22. Protein domain enrichment was analyzed using the Pfam database (http://pfam.xfam.org/), whereas functional categories and signaling pathways were annotated through the Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) databases. Statistical significance of enrichment was evaluated using Fisher's exact test, with all identified proteins serving as the background; multiple-testing correction was performed using the Benjamini–Hochberg procedure, and a false discovery rate (FDR) < 0.05 was considered significant. To explore functional differences among proteins with different expression trends, DEPs were divided into four clusters based on FC: Q1 (FC < 0.5), Q2 (0.5 ≤ FC < 0.67), Q3 (1.5 < FC ≤ 2), and Q4 (FC > 2), and GO, KEGG, and protein domain enrichment analyses were performed for each cluster.²⁶

Gene set enrichment analysis (GSEA) was conducted using GSEA 3.0. Gene sets corresponding to GO, KEGG, and InterPro (IPR) annotations were obtained from the GSEA database. Pathways with FDR < 0.05 and absolute normalized enrichment score (NES) > 1 were considered significantly enriched. Visualization was performed in R 4.2.3 (R Foundation for Statistical Computing, Vienna, Austria).

Identification of Clinically Relevant Co-Expression Modules

Weighted Gene Co-expression Network Analysis (WGCNA) 1.69 was used to construct a co-expression network. Proteins with excessive missing values or low variation were removed, and an appropriate soft-thresholding power (β) was selected to fit the scale-free topology. Modules were identified by dynamic tree cutting, and the correlation between module eigengenes (MEs) and clinical traits was calculated using Pearson's correlation. Modules with FDR < 0.05 were considered significantly associated with clinical features. Within each significant module, the top 10 proteins with the highest absolute gene significance (GS) values were defined as hub proteins. Protein–protein interaction (PPI) networks were predicted using STRING 11.5 (https://string-db.org/) and functionally enriched through the aforementioned databases.²⁶

To further evaluate module robustness, repeated subsampling-based module preservation analysis was performed within each cohort using the modulePreservation function in the WGCNA package. For the all-patient network, 80% of samples were randomly retained in each iteration, whereas, for the ADAMTSL4-related CEL network, five samples were randomly retained in each iteration; both procedures were repeated 20 times. Z_summary statistics were calculated with 100 permutations in each iteration, and the median Z_summary across repeated runs was used as the summary measure. Modules with Z_summary > 10 were defined as highly preserved, modules with 2 < Z_summary ≤ 10 as moderately preserved, and modules with Z_summary ≤ 2 as not preserved.²⁷

Machine Learning Strategy

For data preprocessing, proteins with more than 50% missing values were excluded, and remaining missing values were imputed using the K-nearest neighbor algorithm.^28,29 To reduce collinearity, pairwise Kendall rank correlation coefficients were calculated for all features, and one variable from each highly correlated pair (|τ| > 0.8) was retained. Feature prioritization was performed using a least absolute shrinkage and selection operator (LASSO) regression model implemented in the glmnet package of R 4.2.3.^30,31 The regularization parameter (λ) was optimized by leave-one-out cross-validation (LOOCV).³² Proteins with non-zero regression coefficients were retained as candidate biomarkers with potential mechanistic relevance for downstream analyses.

Targeted Technical Confirmation by iPRM

To provide targeted confirmation of the deep DIA findings, iPRM was performed by Novogene (Beijing, China) using the same AH cohort. The LC-MS/MS platform, chromatographic columns, and gradient conditions were identical to those used in the deep DIA analysis. Indexed retention time peptides (Biognosys, Schlieren, Switzerland) were added to each sample for retention-time calibration. The instrument operated under full mass spectrometry–parallel reaction monitoring (MS-PRM) mode. In the full-scan stage, spectra were acquired over a range of 350 to 1200 m/z with a resolution of 240,000, an AGC target of 500%, and a maximum injection time (Max IT) of 100 ms. During the PRM stage, data were collected at a resolution of 80,000 with an AGC target of 500%, a Max IT of 10 ms, and an isolation window of 1.6 m/z. Target peptides were selected based on reproducibility, signal intensity, and uniqueness. Raw iPRM data were processed in Spectronaut 17 using the same search and quantification parameters as in the DIA analysis, with a peptide mass tolerance of 10 ppm. Group differences were evaluated by comparing normalized peptide intensities, and P < 0.05 was considered statistically significant.

Exploration of Biomarkers In Vitro and In Vivo

Key candidate genes identified from the proteomic analysis were validated at the transcriptional level in human retinal pigment epithelial cells (ARPE-19, A125, RRID: CVCL_0145; QuiCell, Shanghai, China) and human fibroblasts (HFs; H452, RRID: CVCL_C9G0; QuiCell). Cells were cultured in Gibco Dulbecco's Modified Eagle Medium/Nutrient Mixture F-12 (DMEM/F-12, 11320033; Thermo Fisher Scientific) for ARPE-19 or Gibco DMEM (11965126; Thermo Fisher Scientific) for HFs, supplemented with 10% Gibco fetal bovine serum (10099141C; Thermo Fisher Scientific) at 37°C in a humidified 5% CO₂ incubator. At 30% to 50% confluence, cells were transfected with ADAMTSL4 siRNA (RiboBio, Guangzhou, China; target sequence GGACCGTCTTTCGATATAA) or negative control siRNA using Lipofectamine RNAiMAX transfection reagent (13778150; Invitrogen, Carlsbad, CA, USA). Additionally, to evaluate these molecular changes in vivo, phenotypically affected adamtsl4-knockout zebrafish (adamtsl4^–/–) were utilized; these stable mutant lines were generated via the CRISPR/Cas9 system as previously described.^33,34

After 48 hours of transfection for cell models or using enucleated whole eyes for zebrafish, total RNA was extracted using the EZBioscience RNA Isolation Kit (B0004DP; EZBioscience, Roseville, MN, USA) and reverse-transcribed with the HiFiScript cDNA Synthesis Kit (CW2020M; CoWin Biosciences, Cambridge, MA, USA). Quantitative analysis of gene expression was performed using a SYBR Green–based real-time quantitative PCR (qPCR) system (Bio-Rad Laboratories, Hercules, CA, USA). All primers are listed in Supplementary Table S1 and were synthesized by GeneWiz (Suzhou, China). Relative gene expression levels were calculated using the 2^–ΔΔCt method with GAPDH for human cells or gapdh for zebrafish as the internal control. Statistical significance was determined using an independent samples t-test, and P < 0.05 was considered significant.

Results

Study Design and Analytical Workflow

The overall design of this study is illustrated in Figure 1. AH samples from seven CEL patients with biallelic ADAMTSL4 variants and seven age-matched CC patients were collected for deep DIA proteomic analysis, followed by targeted technical confirmation via iPRM. Demographic and ophthalmic parameters of all participants including gender, age at surgery, and preoperative parameters were comparable between the two groups (Supplementary Table S2). For patients in the ADAMTSL4-associated CEL group, postoperative BCVA and ocular biometric data at the final follow-up were also recorded to evaluate surgical outcomes, with a median follow-up duration of 1.17 years (interquartile range, 0.58–1.58). All variants and their pathogenicity prediction are listed in Supplementary Table S3.

Figure 1.

Overview of the study design and analytical workflow. AH samples from ADAMTSL4-associated CEL patients and CC controls were analyzed using deep DIA-based proteomics for the discovery phase, including protein quantification, DEP identification, functional enrichment, WGCNA, and machine-learning–based feature prioritization. Subsequently, targeted technical confirmation of candidate proteins was performed by iPRM, and the corresponding transcriptional changes were examined in RPEs, HFs, and adamtsl4-knockout zebrafish models by qPCR. The correlations between selected biomarkers and clinical parameters were also evaluated. RPEs, retinal pigment epithelial cells.

AH Proteomic Profiling

In this study, AH samples were analyzed and quantified using the deep DIA strategy, leading to the identification of 1865 proteins and 14,427 peptides. The number of identified proteins per sample ranged from 1200 to 1800 (Fig. 2A), with comparable detection counts between the CEL patients with ADAMTSL4 mutations and CC controls. Peptide lengths were mainly distributed between seven and 20 amino acids (Fig. 2B), and most proteins contained multiple unique peptides (Fig. 2C). To assess the stability of MS performance, retention times of 12 representative peptides were compared across different runs and remained highly consistent (Fig. 2D). Principal component analysis results showed that the overall protein expression profiles of the two groups differed but were not completely distinct (Fig. 2E). Figure 2F presented the distribution of CV values, with medians for both groups ranging from 0.3 to 0.4. We performed a comprehensive functional annotation of the identified proteins (Supplementary Fig. S1).

Figure 2.

Summary and quality assessment of AH proteomic data obtained by deep DIA. (A) Number of proteins identified in each sample. (B) Distribution of peptide lengths. (C) Distribution of the number of unique peptides per protein. (D) RT distributions of representative peptides across different runs. (E) PCA plots of ADAMTSL4-associated CEL and CC groups based on deep DIA analysis. (F) Distribution of CV for protein quantification in two groups. PCA, principal component analysis; RT, retention time.

Functional Enrichment Analysis of DEPs

A total of 265 DEPs were identified between CEL patients with ADAMTSL4 variants and the non-disease controls, as shown in the volcano plot (Fig. 3A). Among these, 70 proteins were upregulated and 195 proteins were downregulated in the ADAMTSL4 group compared with the CC group. Functional enrichment analysis revealed that, according to GO annotation, these DEPs were mainly localized to the extracellular region and involved in biological processes such as glycosaminoglycan (GAG) metabolic process (Fig. 3B). KEGG pathway analysis showed significant enrichment in complement and coagulation cascades, ECM–receptor interaction, and cell adhesion molecules (CAMs) (Fig. 3C). IPR annotation further indicated that domains such as membrane attack complex component/perforin and low-density lipoprotein receptor class A repeat were significantly enriched (Fig. 3D).

Figure 3.

Key pathway, function, and proteins characterized in ADAMTSL4-associated CEL and CC controls. (A) Volcano plot of DEPs between the two groups. (B) GO enrichment analysis of DEPs showing the significantly enriched terms across biological process, cellular component, and molecular function categories in bar plot format. (C) Enrichment bubble plot of KEGG pathways for DEPs showing the top 20 significantly enriched pathways. (D) Enrichment bubble plot of IPR domains for DEPs showing the top 20 significantly enriched domains. (E) Hierarchical GO enrichment analysis of DEPs in Q1 to Q4 clusters, showing the top five enriched terms per cluster. (F) Hierarchical KEGG pathway enrichment analysis in Q1 to Q4 clusters. Q1 to Q4 denote DEPs stratified by FC: Q1 (FC < 0.5), Q2 (0.5 ≤ FC < 0.67), Q3 (1.5 < FC ≤ 2), and Q4 (FC > 2).

To explore whether expression magnitude was associated with distinct biological functions, DEPs were grouped into four clusters (Q1–Q4) according to FC. Q1 and Q4 represented strongly downregulated and strongly upregulated proteins, respectively, whereas Q2 and Q3 corresponded to moderately downregulated and moderately upregulated proteins, respectively. GO and KEGG enrichment analyses were subsequently performed for each cluster (Figs. 3E, 3F). The results demonstrated that the Q1 cluster was significantly enriched in pathways related to metabolism and epithelial development (GO, kidney epithelium development; GO, ureteric bud development; KEGG, nitrogen metabolism). The Q2 cluster was primarily associated with intracellular secretory transport, lysosomal degradation, and cell adhesion (GO, vesicle lumen; GO, lysosomal lumen; KEGG, lysosome; KEGG, CAMs). Proteins in the Q3 cluster were significantly enriched in immune-related functions (GO, complement activation; GO, blood microparticle; KEGG, complement and coagulation cascades). In the Q4 cluster, the immune and coagulation network exhibited further activation, accompanied by a marked increase in lipoprotein particles and circulating microparticles (GO, protein–lipid complex; KEGG, complement and coagulation cascades; KEGG, cholesterol metabolism).

Global Pathway Signatures Revealed by GSEA

To further characterize pathway-level alterations in the AH proteome of ADAMTSL4-associated CEL, GSEA was performed on 1649 comparable proteins. KEGG enrichment analysis revealed coordinated upregulation of multiple pathways in the ADAMTSL4 group (Fig. 4A), with enriched pathways further categorized into those associated with mitochondrial function and metabolic processes, immune and hemostatic regulation, and signaling and stress-response pathways (Figs. 4B–D). No KEGG pathways exhibited significant downregulation between the two groups. GO enrichment analysis revealed patterns largely consistent with KEGG results, with upregulated terms predominantly related to lipid complexes, mitochondrial energy metabolism, and transport-associated functions (Supplementary Fig. S2A). IPR enrichment analysis identified the von Willebrand factor type A (VWA) domain as significantly enriched (Supplementary Fig. S2B).

Figure 4.

Enriched KEGG pathways and domains identified by GSEA. (A) Significantly enriched KEGG pathways identified from the AH proteome of ADAMTL4-related CEL patients compared with CC controls. (B) Mitochondrial and metabolic pathways showing positive enrichment in ADAMTL4-related CEL patients. (C) Immune, inflammatory, and hemostatic pathways enriched in ADAMTL4-related CEL patients. (D) Signaling and stress-response pathways significantly enriched in ADAMTL4-related CEL patients.

Identification of Protein Modules Associated With Clinical Traits and Postoperative Outcomes

To identify protein co-expression modules related to clinical features, we first performed WGCNA on all 14 AH samples. Hierarchical clustering indicated close relationships with no outliers (Supplementary Fig. S3A). A soft-threshold power of β = 14 was chosen to approximate a scale-free topology (Supplementary Fig. S3B). A total of 11 distinct co-expression modules were detected and color labeled as yellow, black, magenta, red, brown, green, pink, blue, purple, turquoise, and gray, where the gray module was comprised of proteins not assigned to any module (Fig. 5A). The correlation heatmap summarized associations between modules and phenotypes: MEyellow, MEbrown, and MEpurple were significantly associated with diagnosis (r = 0.86, –0.80, and –0.57, respectively; FDR < 0.05). MEmagenta correlated positively with preoperative AL (r = 0.65; FDR = 0.01), whereas MEblue and MEpurple correlated negatively with preoperative AL (r = −0.72 and −0.60, respectively; FDR < 0.05) (Fig. 5B). In addition, MEpurple showed a significant positive correlation with preoperative Km (r = 0.60; FDR < 0.05). MEbrown and MEmagenta were significantly associated with ChT (FDR < 0.05).

Figure 5.

WGCNA revealed clinically and prognostically relevant co-expression protein modules. (A) Identification of 11 co-expression protein modules associated with clinical traits across all AH samples. (B) Correlation heatmap between modules and clinical traits. Positive and negative correlations are indicated in red and blue, respectively. Each cell shows the correlation coefficient and corresponding P value, and the second row denotes the statistical significance of the correlation. (C) Chord diagram illustrating the top 10 proteins with the highest absolute GS values in modules significantly correlated with each clinical phenotype. (D) GO enrichment analyses of the four core clinical modules (MEyellow, MEbrown, MEpurple, and MEmagenta). Only the top 20 enriched GO terms are displayed. (E) Identification of eight co-expression protein modules associated with postoperative outcomes in ADAMTSL4-associated CEL patients. (F) Correlation heatmap between modules and postoperative phenotypes. Red and blue represent positive and negative correlations, respectively. Each cell displays the correlation coefficient and P value, the second row indicates the significance level. (G) Chord diagram illustrating the top 10 proteins with the highest absolute GS values in modules significantly correlated with postoperative traits. (H) GO enrichment analyses of the two key postoperative modules (MEblue and MEgreen). Only the top 20 enriched GO terms are presented. BP, biological process; MF, molecular function.

To further visualize the relationships between proteins and phenotypes, we generated a chord diagram displaying the top 10 proteins with the highest absolute GS from modules significantly associated with each phenotype (Fig. 5C). Functional annotation and PPI networks were then constructed for the four key modules: MEyellow, MEbrown, MEpurple, and MEmagenta (Fig. 5D; Supplementary Figs. S3C, S4A). Module preservation analysis was performed to verify the topological robustness of these key modules, and MEyellow and MEbrown exhibited high preservation (Z_summary = 18.1 and 23.6, respectively), whereas MEpurple and MEmagenta showed moderate preservation (Z_summary = 5.79 and 3.31, respectively) (Supplementary Fig. S5A). Enrichment analyses indicated that MEyellow was dominated by plasma-derived proteins involved in complement and coagulation cascades (GO, blood microparticle; GO, high-density lipoprotein particle; KEGG, complement and coagulation cascades; KEGG, cholesterol metabolism), with representative hub proteins including fibrinogen alpha chain (FGA), alpha-1-microglobulin/bikunin precursor (AMBP), and kallikrein B1 (KLKB1). MEbrown was enriched for ECM and lysosome-related pathways (GO, collagen-containing ECM; KEGG, lysosome; KEGG, CAMs), with hub proteins including syndecan 4 (SDC4), albumin (ALB), and neural cell adhesion molecule 1 (NCAM1). MEpurple exhibited a signature of extracellular signaling and structural regulation (GO, GAG binding; GO, negative regulation of cell projection organization), with hub proteins including decorin (DCN), kinase insert domain receptor (KDR), and versican (VCAN). MEmagenta was enriched for endocrine response and energy metabolism pathways (GO, response to glucagon; GO, response to corticosteroid; KEGG, glycolysis/gluconeogenesis), with hub proteins including adenylate kinase 1 (AK1), phosphoglycerate mutase 2 (PGAM2), and glutathione S-transferase mu 1 (GSTM1).

In the ADAMTSL4-associated CEL group, we next performed WGCNA using postoperative visual outcomes and ocular biometric parameters as phenotypes to explore protein co-expression modules correlated with postoperative clinical features. A soft-threshold power of β = 7 was selected, yielding eight co-expression modules (Fig. 5E; Supplementary Figs. S3D, S3E). Among them, MEblue and MEgreen showed positive correlations with postoperative BCVA (r = 0.95 and 0.96, respectively; FDR < 0.001) (Fig. 5F). Subsequent preservation analysis indicated high preservation of MEblue (Z_summary = 18.8), while MEgreen exhibited moderate preservation (Z_summary = 5.73) (Supplementary Fig. S5B). MEblue primarily involved extracellular transport, secretion and energy metabolism (GO, secretory granule lumen; GO, cytoplasmic vesicle lumen; KEGG, carbon metabolism; KEGG, glycolysis/gluconeogenesis), with hub proteins including enolase 1 (ENO1), glucose-6-phosphate isomerase (GPI), and phosphoglycerate mutase 1 (PGAM1). MEgreen was mainly related to cell adhesion, cell junctions, and carbohydrate metabolism (GO, focal adhesion; GO, cell–substrate junction; KEGG, galactose metabolism; KEGG, fructose and mannose metabolism), with hub proteins including malate dehydrogenase 2 (MDH2), triosephosphate isomerase 1 (TPI1), and vimentin (VIM) (Figs. 5G, 5H; Supplementary Figs. S3F, S4B).

Machine Learning-Based Prioritization of Candidate Biomarkers

To further prioritize key disease-associated proteins for downstream analyses, a machine-learning analysis based on the LASSO regression model was performed. Quantified proteins passing quality control were used as input features. Using LOOCV for parameter optimization, six protein features retained non-zero regression coefficients at the minimum cross-validation error, forming the optimal expression feature subset, including A disintegrin and metalloproteinase with thrombospondin motifs 3 (ADAMTS3; O15072), apolipoprotein C2 (APOC2; P02655), AMBP (P02760), KLKB1 (P03952), SDC4 (P31431), and ectonucleotide pyrophosphatase/phosphodiesterase 2 (ENPP2; Q13822) (Figs. 6A, 6B). Kendall correlation coefficients and quantitative expression levels of the selected proteins were visualized, indicating limited redundancy among these candidate features (Figs. 6C, 6D). These candidate biomarkers captured disease-associated AH proteomic signatures of ADAMTSL4-associated CEL and were selected for subsequent targeted technical confirmation and clinical correlation analyses.

Figure 6.

Machine-learning–based identification of the optimal protein feature subset. (A) Cross-validation curve used for selection of the optimal regularization parameter λ, with the minimum cross-validation error indicated. (B) LASSO regression coefficient profiles showing changes in feature coefficients across different values of the regularization parameter λ. (C) Kendall correlation analysis among the selected protein features, showing generally low correlations and limited feature redundancy. (D) Heatmap illustrating the expression levels of the optimal protein feature subset across all samples.

Targeted Technical Confirmation of Candidate Biomarkers and Clinical Correlation Analysis

To systematically verify the quantitative reliability of the deep DIA-based proteomic, we performed targeted confirmation using large-scale iPRM in the same patient cohort, comprised of seven ADAMTSL4-associated CEL patients and seven CC controls. In total, 272 proteins were targeted, corresponding to 489 unique peptides, with one to three specific peptides selected per protein for targeted quantification. Overall, 86.03% of the peptides displayed consistent regulation directions between the iPRM and DIA analyses, confirming the reproducibility of the DIA dataset (Fig. 7A).

Figure 7.

iPRM validation, clinical correlation analysis, and transcriptional alterations in ADAMTSL4/adamtsl4–deficient ARPE-19, HFs, and zebrafish. (A) Consistency analysis between DIA and iPRM quantification of 272 targeted proteins. The x-axis represents the log₂(FC) values from iPRM validation, and the y-axis represents those from the DIA dataset. Each dot denotes a specific protein. Quadrants Q1 to Q4 indicate proteins with distinct regulation patterns between the two platforms: Q1, proteins upregulated in both DIA and iPRM; Q2, proteins downregulated in iPRM but upregulated in DIA; Q3, proteins downregulated in both DIA and iPRM; and Q4, proteins upregulated in iPRM but downregulated in DIA. (B) Relative abundance distribution of four validated proteins (APOC2, AMBP, KLKB1, and ENPP2) between CEL patients with ADAMTSL4 variants and CC controls. The iPRM validation results were consistent with the DIA findings (P < 0.05). (C) Heatmap showing Pearson correlations between iPRM-validated proteins and preoperative clinical traits. Statistical significance is denoted by asterisks (P < 0.05). (D) Pearson correlations between protein biomarkers and clinical parameters. (E) Integrated heatmap combining qPCR and deep DIA results across ARPE-19 cells, HFs, and zebrafish model for 43 candidate genes, illustrating the overall transcriptional changes following ADAMTSL4 knockdown. Color intensity represents FC (red indicates upregulation, blue indicates downregulation). Statistical significance is denoted by asterisks: *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. ACD, anterior chamber depth; LT, lens thickness; K1, flat keratometry; K2, steep keratometry; Kc, central corneal curvature; N0.5, choroidal thickness at 0.5 mm nasal to fovea; N1.5, choroidal thickness at 1.5 mm nasal to fovea; Pre., preoperative; T0.5, choroidal thickness at 0.5 mm temporal to fovea; T1.5, choroidal thickness at 1.5 mm temporal to fovea.

We next focused on the LASSO-prioritized candidate biomarkers. Four proteins—APOC2, AMBP, KLKB1, and ENPP2—were successfully quantified and showed statistically significant differences between the two groups (P < 0.05), with relative abundance trends consistent with those observed in DIA (Fig. 7B). SDC4 was excluded from the technical confirmation because its target peptides failed quality-control assessment due to low signal intensity and insufficient specificity, and ADAMTS3 did not reach statistical significance in iPRM quantification (P > 0.05).

Further correlation analysis revealed significant associations between several iPRM-confirmed proteins and preoperative clinical parameters (Fig. 7C). APOC2 (P02655) was positively correlated with preoperative AL (r = 0.63, P = 0.025) and negatively correlated with ChT across both the subfoveal and parafoveal regions (SFCT, N0.5, T0.5, N1.5, T1.5; P < 0.05). ENPP2 (Q13822) showed positive correlations with ChT in multiple regions (N0.5, T0.5, T1.5; P < 0.05), whereas KLKB1 (P03952) exhibited a strong negative correlation with T1.5 (r = −0.85; P < 0.01) (Fig. 7D).

To explore the transcriptional responses to ADAMTSL4 deficiency, qPCR analysis of 43 candidate genes was primarily performed in ADAMTSL4-deficient human RPEs, in which efficient knockdown was confirmed at both the mRNA and protein levels (Supplementary Figs. S6A–S6C). The selected genes included those with the largest FC identified by deep DIA analysis, as well as six core genes selected from the machine-learning–based biomarker screening. To further examine the reproducibility of these molecular changes in a distinct cellular background, the same candidate genes were analyzed in ADAMTSL4-knockdown HFs (Supplementary Fig. S6D). Genes that showed significant changes in either ARPE-19 or HFs were then further evaluated in vivo in the eyes of phenotypically affected adamtsl4-knockout zebrafish (Supplementary Fig. S6E).

A combined heatmap integrating qPCR and DIA results (Fig. 7E) provides an overview of the global transcriptional changes upon ADAMTSL4/adamtsl4 deficiency. Several genes exhibited transcriptional changes consistent with the proteomic data, among which lecithin cholesterol acyltransferase (LCAT) was significantly upregulated (P = 0.014), whereas ENPP2, myeloid-derived growth factor (MYDGF), and carbonic anhydrase 2 (CA2) were significantly downregulated (all P < 0.05) in ADAMTSL4-deficient RPEs. Notably, this consistent downregulation of mydgf was also observed in the in vivo zebrafish model (P = 0.027).

Discussion

As one of the most challenging pediatric ophthalmic disorders, CEL presents with highly variable surgical complexity and heterogeneous visual prognoses. In contrast to FBN1-related CEL, ADAMTSL4-associated disease is largely confined to ocular structures without systemic involvement of the cardiovascular or skeletal systems,³⁵ forcing clinical decision-making to rely predominantly on ocular phenotypes. However, these ocular features show substantial overlap among different genetic forms of CEL, thereby markedly increasing the difficulty of disease classification and evaluating postoperative visual outcomes.⁵ Although genetic studies and animal models have firmly established the pathogenicity of ADAMTSL4 variants,¹⁰ the molecular pathways driving the associated ocular anomalies, especially in humans, remain insufficiently characterized, thereby constraining a deeper understanding of disease mechanisms and phenotypic heterogeneity. Benefiting from high-depth proteomic profiling and multilayer assessment, the present study performed a systematic analysis of AH from children with ADAMTSL4-associated CEL and, for the first time to our knowledge, comprehensively delineated the molecular landscape underlying this disorder at the human level. To address the high dimensionality and limited sample size inherent to AH proteomics, we applied a machine-learning-based feature selection framework incorporating LASSO regularization with LOOCV^36–38 to prioritize exploratory mechanistic candidates, followed by targeted technical confirmation and preliminary assessment of clinical relevance.

Although proteomics has become a powerful tool for elucidating disease mechanisms and discovering candidate biomarkers,^39,40 AH proteomics has long been restricted by the extremely small sample volume, very low protein content, and wide dynamic range of analytes.^14,41 Nevertheless, as a biofluid regulated by the blood–aqueous barrier and directly reflecting the ocular microenvironment,⁴² AH possesses biological specificity that cannot be substituted by systemic samples such as plasma in studies of ocular diseases. Compared with DDA or isotope-labeling–based quantitative strategies, DIA-based analysis systematically acquires signals from all detectable ions, thereby substantially reducing stochastic sampling bias and missing values.^43,44 Based on this principle, the deep DIA strategy developed on the Orbitrap Astral platform integrates a quadrupole mass filter, Orbitrap analyzer, and Astral analyzer within a parallel high-resolution acquisition architecture, further improving scan speed, dynamic range, and quantitative robustness for low-abundance peptides, making it particularly suitable for micro-volume, low-protein samples such as AH.^17,45,46 In addition, iPRM, leveraging a high-resolution platform with automated interference removal and peak-area optimization,⁴⁷ further enhances quantification precision, reproducibility, and throughput. Using this workflow, we performed targeted confirmation of 272 proteins in AH, far exceeding previous reports,^26,40,48 underscoring the feasibility of large-scale, high-accuracy quantification in micro-volume clinical specimens.

The findings of our study suggest that ADAMTSL4-associated CEL may not be merely a disorder of a single structural protein but rather a complex disease process involving multidimensional molecular alterations related to ECM remodeling, local immune activation, and metabolic disequilibrium. Disruption of ECM- and GAG-related pathways was highly consistent with the known structural biology of zonular microfibrils. The ciliary zonule is composed of densely packed fibrillin microfibrils embedded within an inter-fibrillar matrix enriched in chondroitin sulfate and heparan sulfate proteoglycans, likely forming complexes with hyaluronic acid,^49,50 a configuration essential for viscoelasticity and force transmission. The abnormalities we observed in ECM–receptor interaction, CAMs, and GAG metabolic processes, together with reduced levels of SDC4—a key mediator of ECM-integrin adhesion and mechanotransduction⁵¹—may reflect microstructural impairment at the zonule–capsule interface. Downregulation of ADAMTS3, a protein involved in ECM processing and fibrillar assembly,⁵² further supports compromised matrix organization. Coupled with prior evidence that ADAMTSL4 promotes fibrillin microfibril formation^8,53 and exhibits its highest expression at the equatorial lens epithelium where zonules anchor to the capsule,⁵⁴ our findings suggest that ADAMTSL4 deficiency may impair GAG composition, microfibril assembly, and ECM anchoring, collectively weakening zonular integrity. These molecular alterations may help explain the clinically observed early-onset ectopia lentis and the heightened intraoperative risk.

Beyond structural abnormalities, ADAMTSL4 deficiency was accompanied by alterations consistent with possible activation of complement and coagulation pathways, along with enrichment of multiple plasma-derived proteins (e.g., FGA, AMBP, KLKB1) in AH, suggesting potential altered barrier permeability or immune response. Previous studies have demonstrated that ECM components such as cartilage oligomeric matrix protein (COMP) and decorin can bind C1q, mannose binding lectin (MBL), or C3/properdin to regulate classical, lectin, or alternative complement pathways,^55–57 and fibrinogen can form complexes with ECM proteins and growth factors to integrate tissue homeostasis and immune signaling.^57,58 These mechanisms suggest that misassembled ECM or microfibrillar rupture may expose abnormal matrix domains, potentially contributing to the activation of complement and coagulation cascades. Two of our machine-learning–prioritized candidates, KLKB1 and AMBP, are consistent with this hypothesis. KLKB1 is known to amplify intrinsic coagulation through the plasma kallikrein–factor XII (FXII) axis and has been implicated as a vascular endothelial growth factor (VEGF)-independent mediator of inflammation and vascular leakage in ocular diseases.^40,59 AMBP-derived products participate in oxidative stress regulation and inflammation,^60–62 potentially aligning with the immune-related alterations observed here. Together, these data suggest that CEL with ADAMTSL4 variants may not be a purely structural disorder, but rather a complex pathological process in which structural abnormalities are accompanied by dysregulated immune-related pathways.

The broad upregulation of cholesterol metabolism, protein–lipid complexes, and blood microparticles indicates perturbed lipid particle composition and lipid-transport networks within AH. Among the candidate biomarkers, APOC2 is a critical co-activator of lipoprotein lipase (LPL) and typically facilitates triglyceride-rich particle catabolism.⁶³ However, high concentrations of APOC2 can paradoxically inhibit LPL activity,⁶⁴ suggesting that its elevation may reflect lipid metabolic stress rather than a simple enhancement of lipid turnover. ENPP2 was consistently downregulated across DIA, iPRM, and qPCR analyses, supporting the reproducibility of this expression trend. As the principal enzyme-generating lysophosphatidic acid,⁶⁵ ENPP2 governs pathways involved in cell migration, proliferation, and epithelial–mesenchymal transition across multiple tissues.^65,66 Although direct evidence in the anterior segment is scarce, ENPP2 downregulation may reflect a potential compensatory response related to epithelial activation. Collectively, these findings indicate that ADAMTSL4 deficiency may be associated not only with altered matrix architecture but also with changes in lipid metabolism, potentially influencing local cellular activity and tissue homeostasis.

As the sole candidate demonstrating directionally consistent changes across the discovery-phase AH proteomics, in vitro cellular systems, and the in vivo zebrafish model, MYDGF merits particular attention. Previous studies have characterized MYDGF as an endoplasmic reticulum-resident secreted factor⁶⁷ that is involved in tissue repair,⁶⁸ inflammatory regulation,⁶⁹ and maintenance of glycolipid metabolic homeostasis.⁷⁰ In light of these biological properties and its consistent downregulation observed in our study, we speculate that MYDGF may serve as a candidate molecule potentially linking local tissue homeostasis and injury responses. These findings raise the possibility that ADAMTSL4 deficiency might be associated with the attenuation of protective repair-associated signaling, potentially affecting the capacity of the ocular microenvironment to buffer structural stress, immune activation, and metabolic disequilibrium.

Several limitations should be acknowledged when interpreting these findings. First, CEL with ADAMTSL4 mutations is exceedingly rare, and the sample size was constrained by patient availability. Although deep DIA, targeted iPRM confirmation, in vitro and in vivo model analyses cannot substitute for validation in larger external cohorts, these multi-tier assessments provided cross-system support for the identified molecular features. Future studies in independent cohorts will be valuable for confirming the generalizability of these findings. Second, although the present work identifies key dysregulated pathways, the full mechanistic trajectory linking ADAMTSL4 deficiency to CEL remains incompletely defined. The molecular signals uncovered here, however, provide important direction for future mechanistic exploration.

Conclusions

In conclusion, by combining deep DIA-based high-depth proteomics with targeted technical confirmation via iPRM, this study delineated the AH proteomic landscape of ADAMTSL4-associated CEL and identified candidate biomarkers associated with ECM organization, immune/coagulation regulation, and lipid metabolic pathways. Integrative network and machine-learning analyses further highlighted disease-related molecular features. Notably, while ENPP2, LCAT, and CA2 showed concordant proteomic and in vitro transcriptional changes, MYDGF further showed in vivo concordance, suggesting their potential relevance to disease-related molecular regulation. Together, these findings provide valuable molecular insights into ADAMTSL4-associated CEL and support future external independent validation and mechanism-oriented biomarker studies.