Abstract

   Autoimmune uveitis (AU) is a group of autoimmune-driven diseases
   characterized by intraocular inflammation, often leading to severe
   vision loss. Ferroptosis, a recently discovered form of programmed cell
   death, has not yet been fully explored in the pathogenesis of AU. This
   study aims to investigate the role of ferroptosis-related key genes in
   AU, providing a theoretical foundation for further mechanistic studies.
   We downloaded [29]GSE198533 data set from the Gene Expression Omnibus
   (GEO). Through differential gene expression (DEG) analysis, weighted
   gene coexpression network analysis (WGCNA), and two machine learning
   models, TGFBR1 and ZFAS1 were identified as critical hub genes.
   Additionally, we validated the expression of TGFBR1 in retinal
   inflammation in a mouse model of experimental autoimmune uveitis (EAU)
   and explored its functional role. The results showed that TGFBR1
   expression was significantly downregulated in EAU. Functional
   experiments demonstrated that TGFBR1 overexpression exacerbated retinal
   inflammation by promoting lipid peroxidation, downregulating GPX4 and
   xCT, and altering immune homeostasis by suppressing FOXP3 and enhancing
   IL-17A expression. Conversely, TGFBR1 inhibition via Galunisertib
   alleviated retinal inflammation and reversed ferroptosis- and
   immunity-related protein expression. These findings suggest that TGFBR1
   contributes to AU pathogenesis by linking ferroptosis and immune
   imbalance and may serve as a potential biomarker and therapeutic
   target, particularly in BD-associated uveitis.
     __________________________________________________________________

   graphic file with name ao5c00595_0013.jpg
     __________________________________________________________________

   graphic file with name ao5c00595_0011.jpg

1. Introduction

   Autoimmune uveitis (AU) is a general name for a group of ocular
   autoimmune diseases, including conditions such as Vogt-Koyanagi-Harada
   (VKH) disease, Behçet’s disease (BD)-associated uveitis, HLA-B27
   positive uveitis, and sympathetic ophthalmia. These diseases may be
   confined to the eye or may manifest alongside systemic symptoms. This
   common intraocular inflammatory disease, which predominantly affects
   individuals of working age, frequently involves the uvea, retina,
   retinal vasculature, and vitreous body, and is one of the leading
   causes of vision impairment and blindness. Recent bibliometric analyses
   have systematically summarized global research trends in uveitis over
   the past decade, highlighting significant progress and the increasing
   attention this disease has garnered from multidisciplinary research
   fields, including immunology, ophthalmology, and bioinformatics. AU is
   characterized by a complex pathogenesis, is difficult to cure, and
   tends to recur. Evidence suggests that uveitis triggered by
   autoinflammation and autoimmunity may share common molecular and
   immunopathogenic mechanisms. Previous research has found that type 1
   helper T (Th1) cells and the primary cytokine produced by Th1,
   interferon-γ (IFN-γ), play a central role in the pathogenesis of
   uveitis. However, as research into AU has advanced, it has been found
   that various lymphocyte subpopulationsTh1, Th2, Th17, and Tregmay
   contribute significantly to the autoimmune aspects of uveitis.
   Experimental autoimmune uveitis (EAU) is a noninfectious uveitis model
   and the best animal model of AU to date. This model is induced by the
   injection of purified retinal proteins or their derived peptides. The
   immunopathogenesis of EAU is thought to be driven by aberrantly
   activated T cells, such as Th1 cells, which primarily secrete IFN-γ,
   and Th17 cells, which secrete IL-17A. These cells cross the
   blood-retinal barrier, recruiting inflammatory cells like macrophages
   and monocytes, leading to intraocular inflammation and destruction of
   the retina. Furthermore, Tregs characterized by FOXP3 expression also
   play a crucial role in maintaining immune tolerance and preventing
   excessive inflammation. Dysfunction of FOXP3^+ Tregs may contribute to
   the exacerbation of uveitis by weakening their suppressive effects on
   Th1 and Th17 cells. ,

   Ferroptosis is a form of iron-dependent programmed cell death driven by
   the overload of lipid peroxides on cellular membranes, first reported
   in 2012, and is characterized by the accumulation of lipid peroxides.
   When phospholipids containing polyunsaturated fatty acids (PUFA–PLs)
   undergo peroxidation catalyzed by iron, and if this process exceeds the
   buffering capacity of the ferroptosis defense system, lipid peroxides
   accumulate to lethal levels on the cell membrane, leading to membrane
   rupture and ultimately resulting in cell death through ferroptosis. The
   defense mechanisms against ferroptosis include the GPX4-GSH system,
   FSP1-CoQH2 system, DHODH-CoQH2 system, and GCH1-BH4 system. The
   GPX4-GSH system is regarded as the principal defense mechanism against
   ferroptosis. Glutathione peroxidase 4 (GPX4) is the only member of the
   GPX family capable of reducing lipid peroxides; due to this distinctive
   feature, it is always considered a key enzyme in this pathway and a
   critical marker of ferroptosis. , Cells utilize the transport protein
   complex known as system Xc- (composed of the transport subunit SLC7A11
   and the regulatory subunit SLC3A2) to reversibly transport cystine and
   glutamate in a 1:1 ratio across the cell membrane. Cystine imported
   into the cell is subsequently reduced to cysteine by glutathione or
   thioredoxin-2, which then participates in glutathione synthesis; the
   synergistic action of glutathione and GPX4 can eliminate lipid
   peroxides effectively.

   At present, ferroptosis has been involved in the onset and progression
   of various autoimmune diseases, including rheumatoid arthritis (RA),
   systemic lupus erythematosus (SLE), and autoimmune hepatitis (AIH). In
   autoimmune uveitis, previous studies have shown that lipid peroxidation
   levels are elevated in mouse models of EAU. , Furthermore, in
   investigations into BD, higher levels of lipid peroxides have been
   reported in patient serum compared to healthy controls. These findings
   suggest that lipid peroxidation may play a significant role in EAU, and
   that ferroptosis could contribute to its pathogenesis. Moreover, with
   the widespread application of bioinformatics and artificial
   intelligence in the medical field, machine learning algorithms have
   emerged as important tools for identifying novel retinal biomarkers and
   developing precise diagnostic and predictive models, and have already
   shown utility in Behçet’s disease-associated uveitis. Utilizing data
   from the Gene Expression Omnibus (GEO) and the Ferrdb database, we
   employed Weighted Gene Coexpression Network Analysis (WGCNA) and
   machine learning algorithms to identify two hub genes, transforming
   growth factor-β receptor 1 (TGFBR1) and ZFAS1, associated with
   ferroptosis in BD. Subsequently, we validated the expression of these
   ferroptosis-related genes in the EAU mouse model and explored the role
   of TGFBR1 in the pathogenesis of uveitis. This study provides novel
   insights and potential targets for understanding the pathological
   mechanisms of uveitis and developing therapeutic strategies.

2. Result

2.1. Weighted Gene Co-Expression Network Construction and Screening of
Differentially Expressed Ferroptosis-Related Genes

   In this study, we utilized the [30]GSE198533 data set, which comprises
   sample data from 9 patients with BD and 10 healthy controls, in
   conjunction with ferroptosis-related genes from the Ferrdb database.
   Based on the filtering criteria of |log FC| > 0.5 and an adjusted
   p-value <0.05, we identified 42 upregulated and 11 downregulated genes
   ([31]Figure A). Subsequently, we performed a heatmap analysis on the
   top 30 upregulated and downregulated genes, where red represents
   upregulated genes and blue indicates downregulated genes ([32]Figure
   B).

1.

   [33]1
   [34]Open in a new tab

   Weighted gene co-expression network construction and screening of
   cDEGs. (A) Volcano plot of DEGs after integration of the [35]GSE198533
   data set and the Ferrdb database. (B) Heatmap of the DEGs. (C) Network
   topology analysis under various soft-threshold powers. (D) Dendrogram
   of clustered genes based on topological overlap and module color
   assignment. (E) Module-trait correlation analysis. (F) Venn diagram
   showing the intersection of DEGs and turquoise module genes.

   To identify gene modules significantly associated with clinical traits,
   we constructed a gene co-expression network for the training group
   using WGCNA. First, a similarity matrix was calculated between genes,
   and the selection of a soft threshold ensured the network conformed to
   the scale-free topology. To achieve an optimal network topology, a soft
   threshold of 9 was selected, indicating high average connectivity
   within the network ([36]Figure C). Subsequently, the adjacency matrix
   was transformed into a Topological Overlap Matrix (TOM), and
   hierarchical clustering was performed using the dissimilarity matrix
   (dissTOM = 1-TOM) derived from the TOM. The Dynamic Tree Cut method was
   then applied to divide the genes into distinct modules ([37]Figure D).

   These modules were then correlated with BD, and the genes of the most
   positively correlated modules were identified. Ultimately, the
   turquoise module (MEturquoise) was found to be strongly positively
   correlated with BD (R = 0.91, p < 0.001), comprising 180 genes
   ([38]Figure E).

   To further investigate the role of ferroptosis in BD, we performed an
   intersection analysis between the differentially expressed genes (DEGs)
   and the genes within the turquoise module, resulting in the
   identification of 51 candidate differentially expressed genes (cDEGs)
   ([39]Figure F).

2.2. Functional Enrichment Analysis of Characteristic Genes and Construction
of Protein–Protein Interaction (PPI) Network

   Subsequently, we performed a functional enrichment analysis of the
   identified key genes to investigate the ferroptosis-related biological
   functions and potential pathways involved in BD. Gene Ontology (GO)
   enrichment analysis revealed that these genes were significantly
   enriched in biological processes and molecular functions, such as
   oxidative stress, TGF-β receptor binding, and ubiquitin protein ligase
   binding ([40]Figure A,B). Kyoto Encyclopedia of Genes and Genomes
   (KEGG) pathway enrichment analysis indicated that ferroptosis, human
   T-cell leukemia virus type 1 infection, arachidonic acid metabolism,
   and hepatitis B, as well as signaling pathways including the TGF-β,
   relaxin, and FoxO signaling pathways, were closely associated with
   these genes ([41]Figure C,D).

2.

   [42]2
   [43]Open in a new tab

   Functional enrichment analysis of characteristic genes. (A, B) GO
   functional annotation of characteristic genes. (C, D) KEGG signaling
   pathway annotation of characteristic genes. For all enriched GO and
   KEGG terms, p < 0.05.

   To further investigate the potential mechanisms, we constructed a PPI
   network based on the STRING database ([44]Figure D).

3.

   [45]3
   [46]Open in a new tab

   Identification of hub genes via machine learning and hub jene
   validation of expression patterns. (A) Seven hub genes identified by
   LASSO, along with the cross-validation curve of the LASSO regression.
   (B) Six hub genes identified by SVM-RFE. (C) Venn diagram showing the
   overlap of hub genes identified by LASSO and SVM-RFE. (D) PPI network
   constructed using the STRING database. (E) Expression levels of TGFBR1
   in [47]GSE198533. (F) Expression levels of ZFAS1 in [48]GSE198533.

2.3. Identification of Hub Genes

   To narrow down the set of characteristic genes, we employed two machine
   learning algorithms for feature gene identification. First, the Least
   Absolute Shrinkage and Selection Operator (LASSO) regression algorithm
   was used to extract seven genes from the feature set, including ZFAS1,
   CHMP6, OIP5-AS1, TGFBR1, FH, PIEZO1 and MMD ([49]Figure A). Next, the
   Support Vector Machine–Recursive Feature Elimination (SVM-RFE)
   algorithm identified six genes: ZFAS1, MLST8, TGFBR1, HSF1, SRC and
   OTUB1 ([50]Figure B). Finally, TGFBR1 and ZFAS1 were identified as
   overlapping hub genes by both the LASSO and SVM-RFE algorithms
   ([51]Figure C).

   To assess whether TGFBR1 and ZFAS1 could serve as diagnostic
   biomarkers, we analyzed their mRNA expression levels. Box plots were
   used to illustrate the expression levels of these two hub genes in BD
   patients compared to healthy individuals ([52]Figure E,F).

2.4. Immune Cell Infiltration and Its Correlation with Hub Genes

   We further explored the differences in immune cell infiltration between
   BD patients and healthy controls in the [53]GSE198533 data set. A total
   of 17 immune cell types were identified in the data set and are
   presented using heatmaps and boxplots. The heatmap revealed a
   significantly higher infiltration of monocytes, eosinophils, M1
   macrophages, and M2 macrophages in BD samples ([54]Figure A). The
   boxplots also displayed significant differences in the levels of γδ T
   cells, eosinophils, M1 macrophages, and M2 macrophages between healthy
   individuals and BD patients ([55]Figure B). Additionally, correlation
   analysis between immune cells and hub genes showed that, in BD samples,
   γδ T cells, NK cells, memory CD4+ T cells, and CD4+ T cells were
   positively correlated with TGFBR1 ([56]Figure C), while eosinophils, M1
   macrophages, M2 macrophages, and neutrophils were positively correlated
   with ZFAS1 ([57]Figure D).

4.

   [58]4
   [59]Open in a new tab

   Immune cell infiltration and correlation with hub genes. (A, B) Heatmap
   and boxplots of the 17 identified immune cell types in the
   [60]GSE198533 data set. (C) Correlation between immune cell
   infiltration and TGFBR1. (D) Correlation between immune cell
   infiltration and ZFAS1.

2.5. EAU Model Construction and Hub Gene Expression Verification

   The EAU model is a well-established model used to study autoimmune
   uveitis. In this study, 6–8 weeks old B10.RIII mice were selected, and
   an EAU model was induced by subcutaneously injecting a mixture of
   interphotoreceptor retinoid-binding protein (IRBP) and complete
   Freund’s adjuvant (CFA). On the 14th day postimmunization, the
   inflammatory manifestations in the anterior segment of the mice’s eyes
   were observed and evaluated using image scoring. The results showed
   that EAU mice exhibited significant inflammatory characteristics in the
   anterior segment, such as corneal edema and conjunctival hyperemia
   ([61]Figure A), with clinical scores markedly higher than those of the
   control group ([62]Figure B). Furthermore, histopathological
   examination on day 14 confirmed these anterior segment findings. HE
   staining of the retina revealed a pronounced increase in inflammatory
   cell infiltration and retinal folds in EAU mice ([63]Figure C), with
   histology scores notably elevated compared to the CON group ([64]Figure
   D). These results confirm the successful establishment of the EAU
   inflammatory model.

5.

   [65]5
   [66]Open in a new tab

   Construction of the EAU model, validation of hub genes, and retinal MDA
   expression. (A, B) Representative anterior segment images and clinical
   scores of the CON and EAU groups on day 14. Red arrows indicate
   conjunctival hyperemia; black arrows indicate posterior synechiae and
   inflammatory exudates. (C, D) Representative histological staining
   images and histopathological scores of tissue sections from the CON and
   EAU groups. Red arrows indicate vitreous cell infiltration; black
   arrows indicate retinal folds; yellow arrows indicate
   neovascularization and hemorrhage. (GCL, ganglion cell layer; INL,
   inner nuclear layer; ONL, outer nuclear layer). (E, F) qRT-PCR results
   for TGFBR1 and ZFAS1 in the retinas of the CON and EAU groups. (G)
   Measurement of MDA levels in the retinas of the CON and EAU groups. All
   qRT-PCR data were normalized to the mean value of the CON group. Data
   are presented as mean ± SD, n = 3–4. *P < 0.05; **P < 0.01; ***P <
   0.001; ****P < 0.0001; ns = not significant.

   In our previous bioinformatics analysis, we identified two
   ferroptosis-related genes, ZFAS1 and TGFBR1, exhibiting substantial
   expression differences between BD patients and healthy controls. To
   further validate these findings, we assessed their expression in the
   EAU mouse model. On the 14th day after injection, retinal samples from
   the mice were collected, and quantitative real-time polymerase chain
   reaction (qRT-PCR) was employed to observe the relative expression
   changes of ZFAS1 and TGFBR1 mRNA. The results showed that, compared to
   CON, TGFBR1 was downregulated while ZFAS1 was upregulated in the
   retinas of EAU ([67]Figure E, F). Similarly, Western blot (WB) analysis
   confirmed a significant downregulation in the relative expression of
   TGFBR1 protein ([68]Figure A,D).

6.

   [69]6
   [70]Open in a new tab

   Protein expression in the CON and EAU groups. (A, D) WB images and
   quantitative analysis of TGFBR1 in the retinas of the CON and EAU
   groups. (B, E) WB images and quantitative analysis of GPX4 in the
   retinas of the CON and EAU groups. (C, F) WB images and quantitative
   analysis of xCT in the retinas of the CON and EAU groups. All WB
   quantification values were normalized to the mean value of the CON
   group. Data are presented as mean ± SD, n = 4–5. *P < 0.05; **P < 0.01;
   ***P < 0.001; ****P < 0.0001; ns = not significant.

2.6. Measurement of Lipid Peroxidation Levels and Ferroptosis Markers in the
EAU Model

   Next, we assessed the levels of ferroptosis in the retinas of the CON
   and EAU groups. To evaluate lipid peroxidation, we measured the levels
   of malondialdehyde (MDA), a key byproduct of lipid peroxidation, in
   both groups. The results showed that MDA levels were significantly
   elevated in the retinas of EAU mice, indicating increased lipid
   peroxidation in the EAU group ([71]Figure G). Given the crucial role of
   GPX4 in scavenging lipid peroxides, it is recognized as a critical
   component in ferroptosis defense and a pivotal marker of this
   process.We analyzed GPX4 protein expression levels in the retinas of
   EAU and CON mice, finding a significant increase in GPX4 expression in
   the EAU group ([72]Figure B,E). Additionally, the expression of xCT
   (also known as SLC7A11), a key component of the cystine/glutamate
   antiporter involved in ferroptosis, exhibited a mild increase in the
   EAU group, though this result did not reach statistical significance
   ([73]Figure C,F). These findings indicate that while lipid peroxidation
   is heightened in the retinas of EAU mice, extensive ferroptosis may not
   occur.

2.7. Upregulation of TGFBR1 Exacerbates Retinal Inflammation in the EAU Model

   To investigate the role of TGFBR1 in EAU, we utilized the peptide KRFK
   derived from TSP-1 as a TGFBR1 agonist. , Initially, we assessed the
   retinal toxicity of this peptide. KRFK was diluted with PBS to a
   concentration of 5 mg/mL. Mice aged 6 to 8 weeks were randomly assigned
   to two groups: one group received intraperitoneal injections of KRFK at
   a dose of 0.08 mL per mouse per day, while the CON group was
   administered an equivalent volume of PBS daily. Both groups underwent
   continuous intraperitoneal injections for seven consecutive days. On
   the seventh day, retinal function was evaluated using
   electroretinograms (ERG) and optokinetic response (Optomotor) tests
   ([74]Figure A). ERG data indicated no significant changes in the
   amplitudes of the A-wave and B-wave between the two groups ([75]Figure
   B,C). Similarly, the Optomotor results showed comparable visual acuity
   in both groups ([76]Figure D). These findings suggest that there are no
   significant differences in visual electrophysiology or function between
   the KRFK and CON groups, indicating that KRFK does not exhibit evident
   retinal toxicity. Additionally, Additionally, WB analysis revealed that
   the protein expression level of TGFBR1 in the retinas of KRFK-treated
   EAU mice was significantly elevated compared to the untreated EAU group
   and CON group ([77]Figure A,D). Clinically and histologically, the KRFK
   treatment group exhibited exacerbated anterior segment inflammation,
   including corneal edema and neovascularization ([78]Figure E,G), along
   with more pronounced retinal detachment and other pathological changes
   ([79]Figure F,H). These results indicate that intraperitoneal injection
   of KRFK does not damage the mouse retina, while it enhances the protein
   levels of TGFBR1, thus exacerbating the inflammatory response in the
   EAU model.

7.

   [80]7
   [81]Open in a new tab

   Upregulation of TGFBR1 exacerbates anterior segment inflammation and
   increases retinal MDA levels in EAU. (A) Illustration of the optomotor
   system. (B) A-wave amplitudes recorded at 3.0 cd s/m^2 in the CON and
   KRFK groups. (C) B-wave amplitudes recorded at 3.0 cd s/m^2 in the CON
   and KRFK groups. (D) Quantification of visual acuity in the CON and
   KRFK-treated mice. (E, G) Representative anterior segment images and
   clinical scores in the CON group, EAU group, and EAU + KRFK group on
   day 14. Red arrows indicate conjunctival hyperemia; black arrows
   indicate posterior synechiae and inflammatory exudates. (F, H)
   Representative histological staining images and histopathological
   scores of tissue sections from the CON group, EAU group, and EAU + KRFK
   group. Red arrows indicate vitreous cell infiltration; black arrows
   indicate retinal folds; yellow arrows indicate neovascularization and
   hemorrhage. (GCL, ganglion cell layer; INL, inner nuclear layer; ONL,
   outer nuclear layer). (I) Measurement of MDA levels in the retinas of
   the CON group, EAU group, and EAU + KRFK group. Data are presented as
   mean ± SD, n = 3–5. *P < 0.05; **P < 0.01; ***P < 0.001; ****P <
   0.0001; ns = not significant.

8.

   [82]8
   [83]Open in a new tab

   Protein expression and serum ELISA analysis results in the CON, EAU,
   and EAU + KRFK groups. (A, D) WB images and quantitative analysis of
   TGFBR1 in the retinas of the CON, EAU, and EAU + KRFK groups. (B, E) WB
   images and quantitative analysis of GPX4 in the retinas of the CON,
   EAU, and EAU + KRFK groups. (C, F) WB images and quantitative analysis
   of xCT in the retinas of the CON, EAU, and EAU + KRFK groups. (G, I) WB
   images and quantitative analysis of FOXP3 in the retinas of the CON,
   EAU, and EAU + KRFK groups. (H, J) WB images and quantitative analysis
   of IL-17A in the retinas of the CON, EAU, and EAU + KRFK groups. (K)
   Serum IL-17A levels in the EAU and EAU + KRFK groups. All WB
   quantification values were normalized to the mean value of the CON
   group. Data are presented as mean ± SD, n = 3–4. *P < 0.05; **P < 0.01;
   ***P < 0.001; ****P < 0.0001; ns = not significant.

2.8. TGFBR1 Induces Retinal Ferroptosis in EAU Mice by Downregulating GPX4
and xCT Expression

   In a study on acute liver failure (ALF), TGFBR1 was identified as a
   driver of ferroptosis, with TGFBR1 knockout leading to a significant
   increase in xCT and GPX4 expression, thereby reducing ferroptosis in
   liver tissues. Given that excessive lipid peroxidation is a hallmark of
   ferroptosis, we assessed retinal MDA levels in CON, EAU, and
   KRFK-injected EAU mice. Our results indicated a notable elevation of
   MDA in the retinas of KRFK-treated EAU mice compared to the other two
   groups ([84]Figure I). Next, we conducted WB analysis to assess the
   expression levels of GPX4 and xCT, two critical proteins involved in
   ferroptosis defense systems. Both proteins were significantly
   downregulated in the retinas of KRFK-injected EAU mice ([85]Figure
   B,C,E,F). These findings suggest that TGFBR1 may induce retinal
   ferroptosis in EAU mice by downregulating GPX4 and xCT expression,
   thereby exacerbating retinal inflammation.

2.9. TGFBR1 Exacerbates Retinal Inflammation in EAU Mice by Modulating the
Expression Of Immune Cell-Related Cytokines

   In addition to inducing ferroptosis, TGFBR1 has been reported to
   regulate immune responses via cytokine signaling. TGFBR1 is a critical
   component of the TGF-β receptor complex and plays a pivotal role in the
   TGF-β dual receptor signaling pathway. The TGF-β pathway is crucial for
   the transcriptional regulation of FOXP3 expression and the suppressive
   function of regulatory T cells (Tregs). , To investigate the cytokine
   signaling regulation by TGFBR1 in the context of EAU, we measured the
   expression levels of key immune cell-related cytokines in the retinas
   of EAU mice. WB analysis revealed that FOXP3 expression was lower in
   the retinas of EAU mice compared to CON, and its expression was
   significantly reduced following KRFK treatment ([86]Figure G,I).
   Retinal IL-17A levels were assessed in CON, EAU, and KRFK-treated EAU
   mice, with results showing that IL-17A levels were elevated in the EAU
   group compared to CON but were significantly higher in the KRFK-treated
   group ([87]Figure H,J). The enzyme-linked immunosorbent assay (ELISA)
   results also demonstrated that the serum IL-17A levels in EAU mice were
   lower than those in the KRFK-treated EAU mice ([88]Figure K). These
   findings suggest that increased TGFBR1 expression may not enhance Treg
   cell numbers. Instead, TGFBR1 upregulation reduces FOXP3 levels and
   leads to increased IL-17A production in both the retina and serum,
   thereby exacerbating ocular inflammation.

9.

   [89]9
   [90]Open in a new tab

   Inhibition of TGFBR1 alleviates anterior segment inflammation and
   reduces retinal MDA levels in EAU mice. (A, C) Representative anterior
   segment images and clinical scores in the EAU group, Galunisertib
   group, KRFK group and KRFK + Galunisertib group on day 14. Red arrows
   indicate conjunctival hyperemia; black arrows indicate posterior
   synechiae and inflammatory exudates. (B, D) Representative histological
   staining images and histopathological scores of tissue sections from
   the EAU group, Galunisertib group, KRFK group and KRFK + Galunisertib
   group. Red arrows indicate vitreous cell infiltration; black arrows
   indicate retinal folds. (GCL, ganglion cell layer; INL, inner nuclear
   layer; ONL, outer nuclear layer). (E) Measurement of MDA levels in the
   retinas of the EAU group, Galunisertib group, KRFK group and KRFK +
   Galunisertib group. Data are presented as mean ± SD, n = 4–5. *P <
   0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001; ns = not significant.

2.10. TGFBR1 Inhibition Attenuates Retinal Inflammation in EAU Mice by
Regulating Ferroptosis and Immune-Related Protein Expression

   To further validate the role of TGFBR1, we subsequently administered
   the TGFBR1 inhibitor Galunisertib (LY2157299) via oral gavage to mice
   in the EAU group and those treated with KRFK. Each mouse received
   Galunisertib at a dose of 150 mg/kg twice daily, starting from day 7
   after EAU induction, and continued for 7 consecutive days. , Anterior
   segment photography and HE staining revealed a significant reduction in
   inflammation in both the anterior segment and retina of mice in the
   Galunisertib-treated group compared to the EAU group. Conversely, mice
   treated with KRFK alone exhibited notably aggravated ocular
   inflammation. Importantly, ocular inflammation in mice receiving
   combined treatment with KRFK and Galunisertib was significantly
   alleviated compared with that in mice treated with KRFK alone
   ([91]Figure A–D). Additionally, assessment of retinal MDA levels
   revealed that Galunisertib treatment led to a pronounced reduction in
   MDA content relative to the EAU group, whereas treatment with KRFK
   alone resulted in a clear elevation of retinal MDA levels. The combined
   treatment with KRFK and Galunisertib further attenuated retinal MDA
   levels compared to treatment with KRFK alone ([92]Figure E).

   The WB results further demonstrated that, compared to the EAU group,
   retinal GPX4 and xCT protein levels were significantly elevated in mice
   treated with Galunisertib alone, whereas these protein levels were
   markedly decreased in mice treated with KRFK alone. Combined treatment
   with KRFK and Galunisertib further increased retinal GPX4 and xCT
   protein expression levels compared to the group treated with KRFK alone
   ([93]Figure A–D). Additionally, we evaluated the expression of
   immune-related proteins FOXP3 and IL-17A in retinal tissues. Compared
   to the EAU group, Galunisertib treatment alone notably increased FOXP3
   protein levels and reduced IL-17A protein levels, whereas KRFK
   treatment alone resulted in decreased FOXP3 expression and increased
   IL-17A expression. Furthermore, compared with the group treated with
   KRFK alone, combined treatment with KRFK and Galunisertib markedly
   increased FOXP3 expression and decreased IL-17A expression in retinal
   tissues ([94]Figure E–H). These findings indicate that inhibition of
   TGFBR1 function may alleviate retinal inflammatory injury in EAU mice
   by enhancing retinal GPX4 and xCT protein levels, reducing MDA content,
   and concurrently regulating immune responses by upregulating FOXP3
   expression and suppressing IL-17A levels.

10.

   [95]10
   [96]Open in a new tab

   TGFBR1 inhibition restores ferroptosis- and immunity-related protein
   expression in the retina of EAU mice. (A, C) WB images and quantitative
   analysis of GPX4 in the retinas of the EAU group, Galunisertib group,
   KRFK group and KRFK + Galunisertib group. (B, D) WB images and
   quantitative analysis of xCT in the retinas of the EAU group,
   Galunisertib group, KRFK group and KRFK + Galunisertib group. (E, G) WB
   images and quantitative analysis of FOXP3 in the retinas of the EAU
   group, Galunisertib group, KRFK group and KRFK + Galunisertib group.
   (F, H) WB images and quantitative analysis of IL-17A in the retinas of
   the EAU group, Galunisertib group, KRFK group and KRFK + Galunisertib
   group. All WB quantification values were normalized to the mean value
   of the EAU group. Data are presented as mean ± SD, n = 4–6. *P < 0.05;
   **P < 0.01; ***P < 0.001; ****P < 0.0001; ns = not significant.

3. Discussion

   An increasing body of evidence suggests a close association between
   ferroptosis and the pathogenesis of autoimmune diseases. For instance,
   Luo et al. discovered reduced GPX4 levels and elevated MDA levels in a
   lipopolysaccharide (LPS)-induced synovitis cell model, providing
   preliminary evidence for the link between ferroptosis and synovitis. A
   recent study has demonstrated that, compared to normal mice, mice with
   SLE exhibit lower GPX4 expression alongside significantly elevated MDA
   levels and Th1/Th2 ratios. Treatment with ferroptosis inhibitors
   markedly increased GPX4 expression while decreasing MDA levels and the
   Th1/Th2 ratio, suggesting that decreased GPX4 may modulate T cell
   ferroptosis, consequently leading to an increased Th1/Th2 ratio that
   may trigger SLE. Another study reported similar findings in an
   experimental autoimmune encephalomyelitis (EAE) mouse model. Notable
   lipid peroxidation was observed in oligodendrocytes of the EAE mice,
   accompanied by significantly reduced levels of GPX4 and xCT in the
   spinal cord. Importantly, these changes could be reversed by treatment
   with ferroptosis inhibitors. These investigations indicate that
   ferroptosis significantly impacts the inflammation and organ damage
   associated with autoimmune diseases and plays a crucial role in their
   pathogenesis. These findings support our hypothesis that ferroptosis
   may be closely related to the onset and progression of autoimmune
   uveitis.

   This study initially employed bioinformatics techniques to explore the
   association between ferroptosis and autoimmune uveitis. We gained
   transcriptomic data sets for Behçet’s disease (BD) and healthy controls
   from the GEO database and utilized WGCNA, machine learning, and
   enrichment analysis to identify two hub genes related to ferroptosis:
   TGFBR1 and ZFAS1. Furthermore, we analyzed the expression differences
   of these genes in the data set, finding that TGFBR1 expression was
   significantly lower in BD patients compared to healthy individuals,
   while ZFAS1 expression was higher. Additionally, these gene expressions
   were validated in the EAU model, with qPCR results demonstrating
   downregulation of TGFBR1 in EAU mice and upregulation of ZFAS1. The
   Western blot results for TGFBR1 were consistent with those obtained
   from qPCR. These findings suggest that TGFBR1 expression is closely
   associated with disease severity and may contribute to EAU pathogenesis
   by regulating ferroptosis, supporting its potential as a therapeutic
   target in EAU and a prospective biomarker in BD.

   Given that the identification of hub genes in this study was based on
   peripheral immune transcriptomic data, we further evaluated the
   feasibility and validity of this strategy. The [97]GSE198533 data set
   was originally designed to investigate the immunopathogenesis of
   Behçet’s disease by analyzing transcriptomic differences in peripheral
   blood mononuclear cells (PBMCs) between patients and healthy controls,
   with a particular focus on the potential relationship between COVID-19
   and BD. Although PBMCs are fundamentally different in tissue origin
   from retinal tissues in EAU mice, they play a central role in systemic
   immune responses and have been widely used in studies of autoimmune
   diseases, including ocular conditions such as AU and Sjögren’s
   syndrome-related keratoconjunctivitis sicca. , In contrast to previous
   studies using this data set that primarily focused on immune cell
   infiltration and inflammatory signaling pathways, our study introduces
   ferroptosis as a novel perspective. By integrating WGCNA with multiple
   machine learning algorithms for hub gene selection and validation, we
   propose a new hypothesis that TGFBR1 may serve as a potential regulator
   of ferroptosis in AU.

   TGFBR1 has previously been implicated in the regulation of ferroptosis
   in several disease models, prompting us to further explore its
   mechanistic role in EAU. TGFBR1 has been identified as a driver of
   ferroptosis in previous studies. For instance, in an ALF mouse model,
   it was found that the deletion of TGFBR1 in the liver effectively
   alleviated ALF symptoms. Another study demonstrated that calcipotriol
   mitigated osteoarthritis (OA) by inhibiting GPX4-mediated ferroptosis
   in chondrocytes through blocking the TGF-β1 pathway. Similarly, Liu C
   and colleagues observed a comparable role of TGFBR1 in diabetic kidney
   disease (DKD). Our study demonstrated that TGFBR1 similarly acts as a
   driver of ferroptosis in the EAU mouse model. In this study, the
   upregulation of TGFBR1 significantly elevated MDA levels in the retinas
   of EAU mice, while also increasing the protein expression of GPX4 and
   xCT. This suggests that TGFBR1 upregulation may promote ferroptosis in
   retinal cells by enhancing the expression of GPX4 and xCT, thereby
   exacerbating retinal inflammation.

   In addition to its aforementioned functions, as a crucial component of
   the TGF-β receptor complex, TGFBR1 plays a pivotal role in the TGF-β
   dual receptor system. Upon ligand binding, TGFBR1 is phosphorylated by
   TGFBR2, which is a critical step for activating downstream signaling
   pathways, essential for both SMAD-dependent and SMAD-independent signal
   transduction. Recent studies have revealed that the TGF-β signaling
   pathway exerts significant regulatory effects on immune cells,
   particularly those involved in autoimmune processes. TGF-β promotes the
   differentiation of Tregs and regulates FOXP3 expression, thereby
   maintaining immune tolerance and preventing the onset of autoimmune
   diseases. Additionally, TGF-β can synergize with cytokines such as
   interleukin-6 (IL-6) to promote the development of Th17 cells. However,
   in our study, we observed that upregulating TGFBR1 did not lead to an
   increase in FOXP3 expression or a reduction in EAU-associated
   inflammation, but rather produced the opposite results. It is possible
   that the divergence in our results, compared to earlier studies, is due
   to altered expression of GPX4, which could impact the activity of
   FOXP3^+ Tregs. Xu et al. reported that the specific deletion of GPX4 in
   Tregs impaired their survival in tumor environments, increased T cell
   infiltration into tumors, and enhanced antitumor immune responses,
   indicating that GPX4 plays an important role in maintaining Treg
   survival and their immunosuppressive functions, thus facilitating tumor
   immune evasion. Moreover, our study found that IL-17A levels in the
   retina and serum of EAU mice significantly increased with the
   upregulation of TGFBR1, consistent with previous reports. These
   observations suggest that TGFBR1 may reduce FOXP3 levels by
   downregulating GPX4 expression while simultaneously enhancing Th17 cell
   activity, leading to elevated IL-17A secretion, which exacerbates
   ocular inflammation in EAU.

   The limitations of this study are mainly reflected in the following
   aspects. First, the results are primarily based on transcriptomic data
   from human PBMCs and the EAU animal model, and therefore require
   further validation in clinical samples to determine their applicability
   to patients with AU. Second, although we confirmed that TGFBR1 can
   induce ferroptosis in the EAU model, the specific mechanisms by which
   it regulates GPX4 and other related molecules remain unclear and
   warrant further investigation. In addition, this study focused on
   Behçet’s disease-associated uveitis and the EAU mouse model. While
   Behçet’s disease-associated uveitis represents a major subtype of AU,
   it does not encompass all forms of the disease. Other AU subtypes, such
   as VKH disease and HLA-B27-associated anterior uveitis, may involve
   distinct immunopathogenic mechanisms. Therefore, the findings of this
   study should not be directly extrapolated to all AU subtypes and should
   be further validated in diverse clinical contexts. Future research
   should expand to include a broader range of AU subtypes and incorporate
   patient-derived ocular tissues or retinal organoid models to
   comprehensively evaluate the clinical relevance and translational
   potential of TGFBR1 as a ferroptosis-related biomarker and therapeutic
   target.

4. Conclusions

   This study employed a combination of WGCNA, machine learning, and other
   bioinformatics approaches to identify TGFBR1 and ZFAS1 as core
   ferroptosis-related genes in AU. Subsequently, their expression
   patterns in retinal inflammation were validated using the EAU model,
   and the mechanistic role of TGFBR1 in the disease was further
   investigated. The findings indicate that TGFBR1 and ZFAS1 hold
   potential as therapeutic targets and diagnostic biomarkers for AU.
   Specifically, TGFBR1 was found to exacerbate retinal ferroptosis by
   regulating the expression of GPX4 and xCT, while also downregulating
   FOXP3 and upregulating IL-17A, thereby intensifying ocular inflammation
   in mice.

   This research provides new insights into the development of diagnostic
   and therapeutic strategies for BD. A deeper understanding of the
   mechanisms underlying TGFBR1 could facilitate the design of more
   effective interventions for BD. Future studies should continue to
   elucidate the role of TGFBR1 and its associated pathways to develop
   more targeted therapeutic strategies for BD and other immune-related
   diseases.

5. Materials and Methods

5.1. Data Download and Preprocessing

   The BD data set used in this study, [98]GSE198533, was downloaded from
   the GEO public database (GEO, [99]https://www.ncbi.nlm.nih.gov/geo/),
   which contains data from 9 BD patients and 10 healthy controls. The
   ferroptosis-related genes were sourced from the Ferrdb database
   ([100]http://www.zhounan.org/ferrdb). The data were analyzed and
   processed using R software (version 4.1.1). First, Ensembl IDs were
   converted to gene symbols and annotated for subsequent research and
   analysis. The R package “limma” (3.60.6) was then employed to normalize
   and merge the data sets, using adjusted p-values of <0.05 and |log FC|
   > 0.5 as cutoff thresholds to identify DEGs. The results were
   visualized using heatmaps and volcano plots generated by the ‘pheatmap’
   (1.0.12) and ‘ggplot2′ R packages (3.5.1), respectively.

5.2. Construction of Weighted Gene Coexpression Networks and Identification
of cDEGs

   WGCNA was performed to identify modules of highly correlated genes,
   elucidate the interconnections among these modules, and assess their
   associations with external sample traits, thereby identifying potential
   biomarkers or therapeutic targets. The R package “WGCNA” (1.73) was
   utilized to construct a weighted gene coexpression network and perform
   correlation analyses among genes. To construct a weighted gene
   coexpression network, we selected the top 5000 genes ranked by
   variance. The optimal soft-thresholding power (β) was determined using
   the pickSoftThreshold function in the WGCNA R package, which evaluates
   scale-free topology fit (R ^2) and mean connectivity across a range of
   candidate powers (1–30). According to the scale-free topology
   criterion, a power value of β = 9 was chosen, as it was the lowest
   power for which the scale-free topology model fit index exceeded 0.9
   while maintaining sufficient mean connectivity. A minimum module size
   of 30 and merge cut height of 0.25 were set. Gene modules were
   identified using hierarchical clustering based on the topological
   overlap matrix, and their correlation with the BD/control phenotype was
   evaluated. The module most strongly associated with ferroptosis was
   selected as the key module for subsequent analysis. The ‘venn’ package
   (1.12) was employed to extract intersecting feature genes between DEGs
   and WGCNA, which were visualized in a Venn diagram. Genes overlapping
   between DEGs and WGCNA were selected as cDEGs for further
   investigation.

5.3. Function Enrichment Analysis and PPI Network Construction

   The ‘clusterProfiler’ package (4.12.6) in R was implemented to perform
   GO and KEGG functional enrichment analysis on the cDEGs, assessing the
   associated biological processes (BP), molecular functions (MF),
   cellular components (CC), and gene-related signaling pathways. A
   p-value of <0.05 was considered indicative of a statistically
   significant level of enrichment analysis. Bubble plots were generated
   using the “ggplot2” R package to visualize the enrichment findings.

   To further investigate the interactions among the identified cDEGs, a
   PPI network was constructed using the STRING database (version 11.0b,
   [101]https://www.string-db.org/). The PPI network was visualized using
   Cytoscape software to better understand the interaction dynamics.

5.4. Screening Hub Genes by Machine Learning

   To identify optimal feature genes among cDEGs, we applied two
   supervised machine learning techniques: LASSO logistic regression and
   SVM-RFE.

   LASSO regression was conducted using the “glmnet” package (version 4.1)
   in R. This method applies an L1 penalty (lambda) to shrink the
   coefficients of less informative variables to zero, thereby performing
   variable selection and reducing overfitting risk. The input gene
   expression matrix was constructed from the [102]GSE198533 data set,
   filtered to include BD and control samples. Binary phenotype labels
   (“1” for BD, “0” for control) were assigned accordingly. All gene
   expression data were log-transformed and standardized prior to
   analysis. LASSO regression was performed with the following settings:
   family = “binomial” to reflect binary classification, α = 1 for LASSO
   penalty, nlambda = 1000 to explore a wide range of regularization
   parameters, and nfolds = 10 for 10-fold cross-validation. The optimal
   penalty parameter (lambda.min) was selected based on the minimum mean
   cross-validated error. Genes with nonzero coefficients at lambda.min
   were retained as the final LASSO-selected features.

   SVM-RFE analysis is a supervised machine learning technique utilized to
   identify the optimal hub genes by eliminating feature vectors generated
   by SVM. The SVM-RFE procedure was implemented using the svmRFE function
   sourced from a custom script (msvmRFE.R), along with the e1071, caret,
   and sigFeature R packages. The analysis was conducted using 10-fold
   cross-validation (k = 10) and recursive elimination was applied with
   halve.above = 100, meaning the feature set was halved when more than
   100 features remained in each iteration. The procedure was repeated
   across folds to ensure stability, and the top-ranked features were
   averaged based on cross-validated performance. The top 40 features were
   further used to construct the SVM model, and classification performance
   was assessed by plotting the error rate and accuracy curves. The
   optimal number of features corresponding to the lowest classification
   error was selected and retained as the final SVM-RFE output.

   The “venn” package was used to identify and visualize the intersecting
   hub genes shared between LASSO and SVM-RFE results. These overlapping
   genes were designated as final hub genes for subsequent validation and
   mechanistic investigation.

5.5. Differential Expression Analysis

   The R packages “limma” and “ggpubr” (0.6.0) were used to analyze and
   compare the expression levels of the final hub genes between the BD
   group and the control group, which were visualized using boxplots.

5.6. Immune Infiltration Analysis

   The CIBERSORT algorithm is based on linear support vector regression
   (SVR) to deconvolve transcriptomic expression matrices, estimating the
   composition and abundance of immune cells within mixed cellular
   populations based on transcriptomic data. Immune cell infiltration was
   evaluated using the CIBERSORT algorithm, which deconvolutes bulk tissue
   gene expression data into proportions of 22 immune cell types. The
   analysis was performed by uploading the normalized gene expression
   matrix from the [103]GSE198533 data set to the CIBERSORT web portal
   ([104]https://cibersort.stanford.edu/), using the LM22 signature
   matrix. Quantile normalization was disabled (QN = FALSE) to accommodate
   RNA-seq data, and 1000 permutations were performed to ensure robust
   statistical inference. Only samples with a deconvolution p-value <0.05
   were retained for downstream analysis. Bar plots and heatmaps of immune
   cell proportions were generated using the ggplot2 and pheatmap
   packages. Spearman correlation analysis between hub genes and immune
   cell proportions was conducted using the ggpubr and ggExtra packages.
   Immune cell types were plotted against hub gene expression values, and
   correlation strength and significance were visualized using lollipop
   plots generated from the correlation matrix.

5.7. Construction of Mouse Models

   This study was reported in accordance with ARRIVE guidelines. All
   methods were carried out following the relevant guidelines and
   regulations. The ethical and legal approval with the number
   WDRM20240702D was obtained from the Ethics Committee of the Renmin
   Hospital of Wuhan University prior to the commencement of the study.
   B10.RIII mouse parents were sourced from The Jackson Laboratory (Bar
   Harbor, ME), and all experimental mice were bred and maintained under
   specific pathogen-free (SPF) conditions with a 12-h light/dark cycle
   with free access to standard normal diet and water. Mice aged 6–8 weeks
   were selected for the experiment and anesthetized with sodium
   pentobarbital. An emulsion was prepared by mixing 50 μg of IRBP with an
   equal volume of CFA containing 1.0 mg/mL of Mycobacterium tuberculosis
   strain (MTB). Each B10.RIII mouse was subcutaneously injected with 100
   μL of the emulsion at the base of the tail and 50 μL at the
   subcutaneous tissue of each inner thigh to establish the EAU model.

5.8. Clinical and Histopathologic Assessment

   On the 14th day after injection, after anesthetizing the mice with
   sodium pentobarbital, the anterior segment inflammation of the eyes was
   observed, and images were captured for scoring. The clinical scoring of
   uveitis in EAU mice was based on five independent criteria described in
   previous studies, with scores ranging from 0 to 5. On the 14th day
   following IRBP peptide immunization, the mice’s eyes were enucleated
   and fixed in 4% buffered paraformaldehyde for 1 h at room temperature.
   The tissues were then embedded in paraffin and sectioned into slices 4
   to 6 μm thick, followed by hematoxylin and eosin (H&E) staining. The
   histopathological scoring of the EAU mouse ocular tissues was performed
   under a microscope, according to the criteria proposed by Caspi R.

5.9. Determination of MDA levels

   MDA is a metabolite of cellular lipid peroxidation. Fourteen days after
   immunization, we isolated the retinas and analyzed MDA levels. MDA
   assay kits were used to quantify MDA levels in retinal tissues
   according to the manufacturer’s instructions (Servicebio, China). The
   absorbance of samples was measured at a wavelength of 450 nm. We
   subsequently normalized the MDA levels by the total protein
   concentration, which was determined utilizing the BCA (Servicebio,
   China) assay kit.

5.10. RNA Extraction and qRT-PCR and ELISA

   Fresh mouse retinas were collected, and total RNA was extracted using
   Trizol reagent (Vazyme, China). After measuring the RNA concentration
   with a NanoDrop spectrophotometer (Thermo, America), RNA was reverse
   transcribed into cDNA using ABScript III RT Master Mix (ABclonal,
   China). Universal SYBR Green Fast qPCR Mix (ABclonal, China) was used
   for quantitative real-time PCR. The mRNA levels were normalized to
   β-actin, and the 2^–ΔΔCT method was applied to calculate relative mRNA
   expression. All primers used in this study are listed in [105]Table .

1. Gene Specific Primer.

   primers    forward (5′-3′)        reverse (5′-3′)
   TGFBR1  TGCTCCAAACCACAGAGTAGGC CCCAGAACACTAAGCCCATTGC
   ZFAS1   AGCGTTTGCTTTGTTCCC     CTCCCTCGATGCCCTTCT
   B-actin GAAGTCCCTCACCCTCCCAA   GGCATGGACGCGACCA
   [106]Open in a new tab

   On the 14th day after emulsion injection, blood was collected from the
   mouse eyeball and centrifuged to obtain serum. Serum IL-17A levels in
   the different groups were measured using the Mouse IL-17A ELISA Kit
   (Servicebio, China) according to the manufacturer’s instructions.

5.11. WB Analysis

   Mouse retinas were harvested and lysed in ice-cold lysis buffer. The
   samples were then sonicated and centrifuged to obtain the protein
   supernatant. Proteins of different molecular weights were separated
   using a 12.5% Bis-Tris polyacrylamide gel and transferred onto a
   poly(vinylidene fluoride) (PVDF) membrane. After blocking with 5% skim
   milk, the membrane was incubated overnight at 4 °C with the following
   primary antibodies: GAPDH (1:10000, HUABIO, China), TGFBR1 (1:500,
   ZENBIO, China), GPX4 (1:10000, HUABIO, China), xCT (1:500, ZENBIO,
   China), and IL-17A (1:500, HUABIO, China). Subsequently, the membrane
   was incubated with horseradish peroxidase (HRP)-conjugated secondary
   antibody for 1 h, and protein bands were detected using enhanced
   chemiluminescence (ECL) reagent (BIOPRIMACY, China). The absorbance of
   each band was measured using ImageJ 1.54 ([107]https://imagej.net/ij/).
   Relative protein expression was quantified as the ratio of the target
   band to the GAPDH band absorbance.

5.12. Optomotor

   The grating visual stimuli were scripted in MATLAB (version 2020b) and
   generated utilizing the Psychtoolbox suite. After 12 h of adaptation in
   the dark, the mice were placed on a platform surrounded by four
   screens, creating a box. These screens presented the animals with
   moving vertical sinusoidal gratings, displaying increasing spatial
   frequencies in both clockwise and counterclockwise directions
   (including 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5, and 0.6
   cycles/degree, with a contrast of 100%). The head movements of the mice
   were scored only if the angular speed of the head corresponded to that
   of the screen rotation. This data was monitored via a camera positioned
   at the top of the enclosure and assessed by an independent, blinded
   researcher. The highest spatial frequency at which a visually evoked
   response was observed was recorded as the visual acuity of the mice.

5.13. ERG

   ERG stimulation and recording of electrical responses were performed
   using the visual electrophysiological system (RetiMINER-C, China). Mice
   were subjected to dark adaptation for over 12 h. They were then
   anesthetized with intraperitoneal injection of sodium pentobarbital,
   following which the mice were secured on a mobile animal experimental
   platform. Afterward, the recording electrode, a circular platinum
   electrode, was positioned at the corneal limbus of both eyes. The
   reference electrode and ground electrode were positioned on the
   forehead and tail, respectively. ERG measurements were conducted under
   dark-adapted conditions at stimulus intensities of 3.0 cd·s/m^2, and
   all procedures were conducted under dim red light. The amplitudes of
   the a-wave and b-wave were recorded for subsequent analysis.

5.14. Statistical Analysis

   All statistical analyses, including heatmaps, volcano plots, WGCNA,
   LASSO, SVM-RFE, and immune infiltration analyses, were conducted using
   R version 4.1.1 ([108]https://www.r-project.org/). Additional
   statistical analyses were performed using GraphPad Prism version 10.0.2
   (GraphPad Software, La Jolla, California). The one-sample
   Kolmogorov–Smirnov test was used to assess the distribution of
   parametric variables. For normally distributed data, comparisons
   between two groups were made using an independent samples t test, while
   one-way analysis of variance (ANOVA) was applied for multiple group
   comparisons. For non-normally distributed data, the Kruskal–Wallis
   rank-sum test or Mann–Whitney U test was employed. All results are
   expressed as mean ± standard deviation (SD), with p < 0.05 considered
   statistically significant. Each group included at least three mice, and
   the data represent the mean ± SD from three independent experiments.
   Statistical significance levels are denoted as follows: ns = not
   significant, * p < 0.05, ** p < 0.01, *** p < 0.001, **** p < 0.0001.

Acknowledgments