Abstract

Background:

   Age-related Macular Degeneration (AMD) poses a growing global health
   concern as the leading cause of central vision loss in elderly people.

Objection:

   This study focuses on unraveling the intricate involvement of Natural
   Killer (NK) cells in AMD, shedding light on their immune responses and
   cytokine regulatory roles.

Methods:

   Transcriptomic data from the Gene Expression Omnibus database were
   utilized, employing single-cell RNA-seq analysis. High-dimensional
   weighted gene co-expression network analysis (hdWGCNA) and single-cell
   regulatory network inference and clustering (SCENIC) analysis were
   applied to reveal the regulatory mechanisms of NK cells in early-stage
   AMD patients. Machine learning models, such as random forests and
   decision trees, were employed to screen hub genes and key transcription
   factors (TFs) associated with AMD.

Results:

   Distinct cell clusters were identified in the present study, especially
   the T/NK cluster, with a notable increase in NK cell abundance observed
   in AMD. Cell-cell communication analyses revealed altered interactions,
   particularly in NK cells, indicating their potential role in AMD
   pathogenesis. HdWGCNA highlighted the turquoise module, enriched in
   inflammation-related pathways, as significantly associated with AMD in
   NK cells. The SCENIC analysis identified key TFs in NK cell regulatory
   networks. The integration of hub genes and TFs identified CREM, FOXP1,
   IRF1, NFKB2, and USF2 as potential predictors for AMD through machine
   learning.

Conclusion:

   This comprehensive approach enhances our understanding of NK cell
   dynamics, signaling alterations, and potential predictive models for
   AMD. The identified TFs provide new avenues for molecular interventions
   and highlight the intricate relationship between NK cells and AMD
   pathogenesis. Overall, this study contributes valuable insights for
   advancing our understanding and management of AMD.

   Keywords: Age-related macular degeneration, NK cell, transcription
   factors, machine learning, single-cell RNA sequencing

Introduction

   As the global population ages, projections indicate a rise in the
   number of age-related macular degeneration (AMD) patients, reaching
   196 million by 2020 and escalating to 288 million by 2040.^ [35]1 AMD,
   characterized by the degeneration of the macula, results in damage to
   photoreceptors and is the leading cause of central vision loss in
   elderly people.^[36]2,[37]3 The pathogenesis of AMD involves various
   factors, including genetics, environment, and immunology.^[38]4 -[39]6
   Environmental and lifestyle variables, such as smoking, hypertension,
   and body mass index, exert a significant influence on AMD risk.^ [40]7
   Similarly, genetic factors play a crucial role, with individuals having
   a family history of AMD facing an elevated risk compared to those
   without such familial predisposition.^ [41]8

   There is an increasing consensus suggesting that immunology also plays
   an important role. Despite the immune privilege status of the eye,
   abnormal immune and inflammatory responses occur, intensifying during
   the development of AMD.^[42]9,[43]10 Complement C3a, C3d, Ba, Bb, C5a,
   and CFH levels are reportedly higher in AMD patients than healthy
   controls.^ [44]11 Besides, the abnormal recruitment of
   macrophages/microglia to tissue lesions reportedly contributes to the
   development of AMD.^ [45]12 It is now understood that Natural Killer
   (NK) cells, integral to the immune system, impact cytokine production
   and cytotoxicity. They exert antiviral and antitumoral effects by
   killing cells with low expression of the MHC antigen presentation
   system.^ [46]13 Goverdhan et al^ [47]14 found that the NK receptor AA
   haplotype, coupled with the human leukocyte antigen (HLA)-Cw*0701
   allele, was associated with AMD. In an additional study of gene
   expression profiles related to AMD, the hub genes, including C1S, ADM,
   and 1ER5L, exhibited a positive correlation with the activation of NK
   cells during the progression of AMD.^ [48]15

   Transcription factors (TFs), as key regulators of gene expression, play
   a vital role in NK cell maturation, development, and differentiation.
   Ets family members play a regulatory role in the development of NK
   cells. Depletion of ETS proto-oncogene 1 (Ets-1) in mice results in a
   significantly diminished NK cell compartment.^ [49]16 TFs of the notch
   family operate as the common lymphoid progenitor (CLP) stage,
   contributing to lymphocyte development. Their role involves, at least
   in part, inhibiting the commitment to a myeloid lineage fate. Notably,
   the Notch ligands Jagged1/2 and Delta-like protein-1 demonstrate a
   preference for fostering the in vitro development of NK
   cells.^[50]17,[51]18

   The advancement of single-cell sequencing technology presents an
   opportunity to explore NK cell subpopulations and examine gene
   expression patterns related to transcription factors. Employing
   high-dimensional weighted gene coexpression network analysis (hdWGCNA)
   and single-cell regulatory network inference and clustering analysis
   (SCENIC), we identified hub transcription factor genes in NK cells
   within the choroid of early-stage AMD patients. By revealing the
   expression profile, our objective is to comprehensively grasp the
   distinctive regulatory mechanisms of these transcription factors in the
   pathogenesis of AMD. This approach provides a new perspective for a
   more comprehensive understanding of AMD’s underlying mechanisms.

Materials and Methods

Data acquisition and processing

   Transcriptomic data were collected from the Gene Expression Omnibus
   (GEO) database. Single-cell analyses were conducted using dataset
   [52]GSE203499.The [53]GSE203499 dataset includes samples from 20
   healthy controls, 24 individuals with early AMD, and 6 donors with
   geographic atrophy. The majority of samples were collected from four
   distinct anatomical regions of the eye: the macular choroid, macular
   retina, peripheral choroid, and peripheral retina. We only included
   samples from the macular choroid for the healthy control and early AMD
   groups. Machine learning models were trained with datasets
   [54]GSE135092 and [55]GSE29801, and accuracy assessments were conducted
   with dataset [56]GSE99248 ([57]sFigure 1). The [58]GSE135092 dataset
   comprises samples from 106 donors with no history of ocular diseases
   and 23 donors previously diagnosed with AMD, collected from the macular
   or peripheral regions of donor eyes. The [59]GSE29801 dataset includes
   177 samples from the macular or extramacular regions of RPE-choroids
   and 118 samples from the macular or extramacular regions of retinas,
   encompassing cases with no reported ocular disease, possible
   preclinical AMD, or AMD. The [60]GSE99248 dataset contains retinal and
   RPE-choroid-scleral tissue punches from 8 normal and 8 AMD donor eye
   samples. For all dataset, we only enrolled samples from the macular
   choroid for the healthy control and AMD groups

Single-cell RNA-seq analysis

   The data from dataset [61]GSE203499 were further analyzed using Seurat
   in R.^ [62]19 Potential doublets and low-quality cells with both small
   and large library sizes within individual datasets were systematically
   excluded. The application of Sctransform facilitated the normalization
   and variance stabilization of molecular count data. The identification
   of differentially expressed marker genes for each cluster was achieved
   using the FindAllMarkers function. To visualize the distribution of
   cells, Uniform Manifold Approximation and Projection (UMAP) analysis
   was conducted.

Cell-cell communication analysis

   To infer and analyze cell-cell communication, we employed CellChat, a
   comprehensive repository encompassing ligands, receptors, cofactors,
   and their interactions.^ [63]20 Facilitated by the versatile and
   user-friendly toolkit CellChat, we explored novel intercellular
   communications and constructed cell-cell communication atlases. For the
   analysis of cell interactions, expression levels were computed relative
   to the total count mapped to the identical set of coding genes across
   all transcriptomes. The resulting expression values were then averaged
   within each single-cell cluster.

Differential expression analysis and gene set enrichment analysis (GSEA)

   In comparing NK cells between the AMD and Normal groups, we conducted
   differential expression analysis. Differentially expressed genes (DEGs)
   were identified through pseudobulk RNA analysis, where the RNA count
   values for each gene were aggregated across all cells in each sample.^
   [64]21 DESeq2 was then employed for the differential expression
   analysis.^ [65]22 Additionally, GSEA was utilized to investigate the
   functional pathways enriched by the differentially expressed genes.

Pseudo-time analysis

   Monocle2 was utilized to sequence the 2 CD8+ T cell clusters along a
   “pseudotime” trajectory.^ [66]23 Following successful quality control,
   genes were incorporated into Monocle’s Reversed Graph Embedding
   algorithm to define the trajectory. Subsequently, Monocle executed
   dimensionality reduction on the data, resulting in the ordered
   arrangement of cells based on pseudotime.

hdWGCNA analysis

   hdWGCNA is a tool designed for conducting weighted gene co-expression
   network analysis (WGCNA) on high-dimensional data, including
   single-cell RNA-seq.^ [67]24 It serves to construct cell type-specific
   co-expression networks, identify resilient modules of interconnected
   genes and offer the biological context for these modules. In this
   study, hdWGCNA was employed to identify the key module associated with
   NK/T clusters. Subsequently, hub genes within this key module were
   identified through screening.

SCENIC analysis

   SCENIC represents a gene regulatory networks (GRNs) algorithm tailored
   for the analysis of single-cell data.^ [68]25 SCENIC introduces a gene
   co-expression network inferred through TF motifs, subsequently
   delineating high-reliability GRNs primarily governed by TFs. In this
   study, the SCENIC algorithm was employed to unravel the specific
   network of transcription factors associated with NK cells.

Machine learning

   To filter the variables with high accuracy and stable performance, we
   used the method of nested resampling in the mlr3 package.^ [69]26 The
   best model was selected among 4 machine learning models, including
   logistic regression (LR), random forest (ranger), decision trees
   (rpart), and adaptive best subset selection (abess). For comparing the
   performance of different learners, we used the benchmarking function of
   mlr3 parkage in the training set ([70]GSE135092 and [71]GSE29801).
   Finally, we used the data from the testing set ([72]GSE99248) to test
   the performance of each learner and screen the hub gene.

Statistical analyses

   Statistical analyses were performed using R software (Version 4.3).
   Analyses were conducted using two-tailed unpaired t-test. Statistical
   significance was defined as ns = P > .05; *P < .05; **P < .01;
   ***P < .001; and ****P < .0001.

Results

Single-cell RNA-seq profiling of different cell clusters in AMD patients and
healthy controls

   To depict the landscape and dynamics of choroid in the macula,
   single-cell transcriptomes were assessed with subcluster analysis and
   visualized by the UMAP approach ([73]Figure 1A). A total of 55 543
   cells from AMD and normal eye tissue were classified into 29 clusters
   ([74]Figure 1B). Based on the gene signature of distinct subclusters,
   these cells were mainly composed of 5 immune cell clusters, including T
   lymphocyte and NK cells (T/NK: PTPRC, CD2), fibroblasts (Fibro: DCN,
   CFH), pericyte (Peri: ACTA2, RGS), melanocyte (Melano: MITF, MLANA),
   schwann cell (Schwa: CDH19, PLP1), macrophyte (Macro: AIF1, CD68), B
   lymphocyte (B lym: CD79A, MS4A1), erythrocytes (Eryth: HBB, HBA2),
   endothelium cells (Endo: VWF, PECAM1), mast cells (Mast: KIT, CPA3),
   retinal pigment epithelium cells (RPE: RPE65, RLBP1), Plasma cells
   (Plasma: MZB1, SDC1), Rod cells (ROD: NRL, PDE6A) ([75]Figure 1B and
   [76]C). The T/NK subcluster comprised 16 144 cells, the highest among
   all cell clusters ([77]Figure 1D). To elucidate the origin of all
   clusters, a total of 30 479 cells derived from AMD samples and 25 064
   cells derived from normal samples were independently subjected to
   clustering ([78]Figure 1E). We further quantified the absolute numbers
   and relative contribution of all clusters in different disease stages
   ([79]Figure 1F and [80]G). Given that the NK/T subcluster constitutes
   the highest proportion among all subtypes, it is significant to note
   that NK cells and T cells play pivotal roles in immune responses and
   inflammation. Consequently, the NK/T clusters were subjected to further
   exploration and decoding at the single-cell level.

Figure 1.

   [81]Figure 1.
   [82]Open in a new tab

   Single-cell RNA-seq profiling of different cell clusters in age-related
   macular degeneration (AMD) patients and health control: (A) eye
   anatomy; (B) uniform manifold approximation and projection (UMAP)
   visualization of all the single cells, (C) dot plot displaying the key
   marker gene of each cell cluster, (D) bar plot representing the cell
   number of each cell, (E) UMAP visualization of different originated
   from AMD and normal patients, (F) bar plot representing the relative
   contribution of cells originated from AMD and normal for each cell
   type, and (G) bar plot representing the number of cells originated from
   AMD and normal for each cell type.

Characterization of the identities of sub-cluster in NK/T cell

   In the overall cluster analysis, considering the close spatial
   distribution of T cells and NK cells in the UMAP plot, we initially
   grouped T cells and NK cells together as a single cluster for analysis.
   However, for more in-depth exploration, the NK/T cluster was
   re-clustered, resulting in the identification of 5 distinct cell
   subclusters ([83]Figure 2A). The absolute numbers and relative
   contributions of all clusters at different disease stages revealed
   significant differences between AMD and normal control, particularly
   for the T/NK_3 and T/NK_4 subclusters ([84]Figure 2B and [85]C). Based
   on marker genes, we further characterized these 5 clusters ([86]Figure
   2D). CD3D and CD3E were commonly expressed in T cells, whereas the
   T/NK_3 subcluster was the only one that did not express CD3D and CD3E.
   Simultaneously, it exhibited high expression of NK cell-related genes
   such as KLRB1, NKG7, and GZMB. Consequently, we defined T/NK_3 as NK
   cells. Specific marker genes CD4 and CD40LG were expressed in T/NK_2,
   indicating that these clusters represented CD4 T cells ([87]Figure 2D).
   Additionally, T/NK_1, T/NK_4, and T/NK_5 expressed CD8A and CD8B,
   specific markers of CD8 T cells. However, the T/NK_5 subcluster also
   showed an enrichment of NK cell marker genes, suggesting that T/NK_5
   might be NKT cells ([88]Figure 2D). Finally, T/NK_1 and T/NK_4 were
   confirmed as CD8 T cells ([89]Figure 2D). The results indicated a
   marked increase in the abundance of NK cells in AMD. To elucidate a
   more refined landscape of T cell and NK cell subsets in AMD, the 2 CD8+
   T cell subclusters were further characterized. FindMarkers differential
   expression analysis between T/NK_1 and T/NK_4 performed in Seurat
   revealed T/NK_1 exhibited significant upregulation of pro-inflammatory
   genes, including CXCL13, IGLC1, IFNG, and LIME1 ([90]Figure 2E).

Figure 2.

   [91]Figure 2.
   [92]Open in a new tab

   Characterization of the identities of sub-cluster in NK/T cell: (A)
   UMAP visualization of each NK/T cluster, (B) violin plot displaying the
   expression of key marker genes in each NK/T cluster, (C) bar plot
   representing the relative contribution of cells from AMD and normal for
   each, (D) bar plot representing the relative proportions of each
   cluster in AMD and normal, (E) volcano plot displaying the differential
   gene expression analysis between NK/T_1 and NK/T_4 cell, (F)
   differentiation trajectory of NK/T_1 and NK/T_4, with a color code for
   pseudo-time, state, disease original, and cell type, and (G) heatmap
   plot representing the expression of T cell specified marker genes along
   pseudo-time.

   To explore potential disparities between the 2 CD8 T subclusters, we
   conducted pseudo-time and trajectory analysis. After mapping all cell
   labels (pseudotime, state, disease, and cell type) onto the pseudotime
   trajectory ([93]Figure 2F), we identified 3 branch points and 7 states
   for the 2 CD8 T subclusters ([94]Figure 2F). Gene expression changes
   along pseudotime revealed 3 distinct patterns associated with the
   differentiation trajectory of CD8 T cells ([95]Figure 2G). All cells
   originated from the Naïve T-related pattern, characterized by high
   expression of LEF1, SELL, and CCR7. During the mid-to-late timeline,
   there was an enrichment of the memory T cell-related pattern,
   characterized by increased expression of CPR183, PRF1, CST7, CD69, and
   IL15RA. The last stage was associated with the expression of NKG7,
   GZMA, IFNG, and CXCR3, representing the cytotoxic T cell-related
   pattern ([96]Figure 2F and [97]G). Given the T/NK_4 subcluster was
   mostly enriched in the latest timeline, we defined it as effector CD8 T
   cells, while the T/NK_1 cluster was labeled as memory/naïve CD8 T
   cells.

Differential disease states of NK cells reveal the heterogeneity of cellular
communication levels and function characteristics

   After identifying all subclusters of the T/NK cluster, we ultimately
   identified four clusters, including CD4T, CD8T, NK, and NKT cells. To
   further elucidate the underlying impact of the T/NK cluster in AMD, we
   conducted an analysis of cellular communication within all cell
   clusters. The number of cell interactions and the interaction strength
   among all cell clusters indicated that CD8T, NK, Macro, Endo, and NKT
   cells exhibited higher interaction strength among clusters ([98]Figure
   3A). Additionally, we performed a ligand-receptor pairs interaction
   analysis to gain insight into the differences between AMD and Normal
   groups. By comparing the information flow in the AMD condition with
   Normal, we identified some conserved and context-specific signaling
   pathways ([99]sFigure 2A). Interestingly, the complement system was
   found to increase in AMD, identified as a key factor in the
   pathogenesis of AMD in a previous study.^ [100]27 Conversely, growth
   factor signaling, such as HGF, VEGF, IGF, and FGF, decreased in the AMD
   group compared with the Normal group. Subsequently, we visualized the
   differential number of interactions and interaction strength in the
   cell-cell communication network. The results showed a significant
   decrease in incoming and outgoing interaction strength in NK cells in
   the AMD group compared to the Normal group ([101]Figure 3B and
   [102]sFigure 2B).

Figure 3.

   [103]Figure 3.
   [104]Open in a new tab

   Differential disease states of NK cells reveal the heterogeneity of
   cellular communication levels and function characteristics: (A)
   analysis of cell-cell interaction in all cells, (B) dot plot
   representing the cell with significant changes in sending or receiving
   signals between AMD and normal, (C) dot plot representing the specific
   signaling changes of NK cell, (D) volcano plot of differential gene
   expression analysis of NK cell between AMD and normal, (E) gene set
   enrichment analysis (GSEA) for differential gene expression, and (F)
   wordcloud plot of down-regulated signaling ligand-receptor pairs
   between AMD and normal.

   Given the significant alterations in the quantity of NK cells in AMD
   and the notable differences in cell interactions, we decided to focus
   on NK cells for the next phase of our research. We further identified
   the specific signaling changes of NK cells in AMD compared to normal
   samples, with SEMA4 specifically present in AMD ([105]Figure 3C). In
   addition, we performed differential expression analysis (DEA) and GSEA
   analysis of NK cells between AMD and Normal. A volcano plot was
   generated to visualize the top 10 upregulated and downregulated genes
   ([106]Figure 3D). IL-related genes were significantly upregulated,
   while the HLA family was notably decreased in AMD. The differential
   genes underwent KEGG enrichment analysis, revealing a highly activated
   immune response in AMD. Conversely, MHC-related antigen processing
   signaling was suppressed in AMD compared to Normal ([107]Figure 3E and
   [108]sFigure 2C). Cellchat, with its capability to identify upregulated
   and downregulated signaling ligand-receptor pairs based on DEA,
   produced results consistent with GSEA, indicating an increased immune
   response in AMD ([109]Figure 3F).

hdWGCNA revealed that NK cell is characterized by the turquoise module

   We conducted hdWGCNA to investigate gene modules in NK cells and
   identified 9 modules ([110]Figure 4A). Subsequently, we constructed the
   co-expression network for each gene module, highlighting the 3 hub
   genes ([111]Figure 4B). The co-expression modules were visualized using
   the hub gene signature in the original UMAP distribution. Notably, the
   green, yellow, and turquoise modules were enriched in the T/NK cluster,
   with the turquoise module showing the most specificity for NK cells
   ([112]Figure 4C). To assess the relationship with AMD, module trait
   correlation was performed for each cell cluster, revealing a
   significant positive correlation between the turquoise module and AMD
   in NK cells ([113]Figure 4D). Differential module eigengene analysis
   corroborated this finding, indicating a significant positive
   association between the turquoise module and AMD ([114]Figure 4E). In
   light of these findings, we conducted a more in-depth exploration of
   the co-expression network within the turquoise module. The display of
   the top 25 hub genes and pathway enrichment highlighted a significant
   focus on inflammation-related pathways, as indicated by KEGG pathway
   enrichment ([115]Figure 4F and [116]G).

Figure 4.

   [117]Figure 4.
   [118]Open in a new tab

   High dimensional weighted gene co-expression network analysis (hdWGCNA)
   revealed that the NK cell cluster is characterized by the turquoise
   module: (A) dendrogram was utilized to visualize all modules in the
   scale-free network, (B) UMAP visualization of top five genes from each
   module in co-expression networks, (C) UMAP plot showing the expression
   distribution of hub genes for each module across the NK cells, (D)
   heatmap plot representing correlations between different modules and
   different stage of age, race, gender or disease in different cell type,
   (E) lollipop plot showing the differential module eigengene analysis
   result of different modules, an “X” is placed over each point that does
   not reach statistical significance, (F) ModuleNetworkPlot showing the
   top 25 hub genes for the turquoise module, and (G) bar plot
   representing the KEGG enrichment results of genes in the turquoise
   module.

SCENIC analysis revealed the key transcription factor in NK cells

   We next utilized SCENIC to unravel the gene regulatory network (GRN)
   across all clusters, followed by hierarchical clustering based on
   regulon activity. The results revealed a striking similarity in the GRN
   between NK and NKT cells ([119]Figure 5A). Additionally, we employed
   the RSS to assess regulons and ranked them within each cluster
   ([120]Figure 5B). NK cells were thus selected for further study.
   Notably, TBX21, EOMES, and SNAI3 were identified as distinctive
   transcription factors in the GRN of NK cells ([121]Figure 5C). The
   regulatory networks of TBX21, EOMES, and SNAI3 were further validated
   using iRegulon ([122]sFigure 3A-C). Consistent with the above regulon
   results, TBX21, EOMES, and SNAI3 were identified as the master
   regulators in AMD.

Figure 5.

   [123]Figure 5.
   [124]Open in a new tab

   Single-cell regulatory network inference and clustering (SCENIC)
   analysis revealed the key transcription factor in NK cells: (A)
   t-Distributed Stochastic Neighbor Embedding (tSNE) plot showing the
   different cells based on raw data (top) and regulon activity (AUcell),
   (B) dot plot showing regulon specificity scores (RSS) in each cell
   type, and (C) heatmap plot showing cell type-specific regulators.

Identification and validation of a predictive model for AMD based on hub TF
of NK cell

   We conducted further investigations to investigate the relationships
   between NK cells and AMD, aiming to gain insights into the underlying
   mechanisms driving AMD. Our focus was on predicting AMD, and we
   constructed a predictive model using the hub TFs of NK cells by
   integrating the hub genes of the turquoise module from hdWGCNA and the
   key transcription factors in NK cells from SCENIC. A Venn plot was
   generated to illustrate the overlap between the key genes from hdWGCNA
   and SCENIC, leading to the selection of 13 genes (CEBPB, CREM, ETV6,
   FOSB, FOXP1, IRF1, IRF8, JUNB, JUND, KLF3, NFKB2, RUNX3, USF2)
   ([125]Figure 6A). Subsequently, we utilized the function of nested
   resampling in the mlr3 package to select a subset of TFs for unbiased
   performance estimates. The training set was obtained by combining data
   from [126]GSE135092 and [127]GSE29801 after correcting for batch
   effects ([128]sFigure 4A), with 220 samples enrolled in the training
   set. The distribution of the 13 hub genes in the training set also
   indicated a certain degree of correlation among some hub genes
   ([129]sFigure 4B). Therefore, we employed three machine learning
   algorithms—random forest (RF), decision trees (DT), and ABESS—for
   feature selection to reduce bias. Twenty-two-fold cross-validation was
   applied for resampling, and ROC and precision-recall curves were
   plotted ([130]Figure 6B). The area under the curve (AUC) and accuracy
   (ACC) of all resampling results for each learner was calculated,
   revealing that the RF learner exhibited the best performance among all
   learners ([131]Table 1). Further validation of the 22 RF learners in
   the testing set from [132]GSE136661 indicated that learner 16 yielded
   the best performance (AUC = 0.66), identifying CREM, FOXP1, IRF1,
   NFKB2, and USF2 as hub genes ([133]Figure 6C and [134]D). In summary,
   our findings suggest that the expression of the identified 5 hub TF
   genes may serve as a potential predictor for AMD.

Figure 6.

   [135]Figure 6.
   [136]Open in a new tab

   Identification and validation of a predictive model for AMD based on
   hub genes of NK cells: (A) Venn plot displaying the thirteen
   overlapping key regulators identified by merging the genes in hdWGCNA
   turquoise module and genes in SCENIC in NK cell, (B) receiver operating
   characteristic curve (ROC) and precision-recall curve of four feature
   selection in the training data set ([137]GSE135092 and [138]GSE29801),
   (C) area under curve (AUC) of total 22 random forest learners performed
   in the testing set ([139]GSE136661), and (D) ROC of learners 16 in the
   testing set.

Table 1.

   The AUC and ACC of each learner.
   Learner_id                 Resampling_id Iters AUC   ACC
   Random forest              CV            22    0.857 0.784
   Decision trees             CV            22    0.725 0.798
   Fast best-subset selection CV            22    0.460 0.491
   Logistic regression        CV            22    0.328 0.370
   [140]Open in a new tab

   Abbreviations: ACC, accuracy; AUC, area under curve; CV, cross
   validation.

Discussion

   In the present study, single-cell analysis revealed different cell
   types in the choroid of macula from AMD and Normal samples, especially
   the T/NK cluster. Subsequently, we analyzed the five subclusters of
   T/NK cells, identifying CD4T, effector CD8 T cells, naïve/memory CD8T,
   NK, and NKT cell populations. Notably, there was a significant
   upregulation in the proportion and number of NK cells in AMD. To gain
   further insights, we employed CellChat to analyze cell-cell
   communication pathways between AMD and Normal. Our findings indicated a
   significant alteration in the interaction strength of NK cells.
   Building upon these results, we established a link between NK cells and
   the occurrence and development of AMD. Moreover, we conducted hdWGCNA
   on NK cells in AMD, revealing that the turquoise module was
   significantly associated with AMD in NK cells. The hub genes within
   this module were enriched in pathways related to immune regulation and
   inflammatory response. Additionally, SCENIC analysis on NK cells in AMD
   unveiled master transcription factors crucial in regulating NK cell
   function, subsequently playing a key role in AMD. Combining the hub
   genes from the turquoise module and the transcription factors from the
   regulatory network of NK cells, we identified 13 candidate
   transcription factors with a crucial regulatory effect on the
   occurrence of AMD. Subsequently, employing 4 machine learning
   techniques, we filtered these transcription factors to identify those
   with the potential to predict the prevalence of AMD. Following
   validation in an independent dataset, we ultimately identified 5 key
   transcription factors. These findings offer novel insights into the
   genesis process and biological importance of NK cells in AMD.

   In AMD, NK cells exhibited a notable increase in both proportion and
   number. Concurrently, we observed a significant decrease in HLA class
   II molecules, including HLA-DR and HLA-DQ, in NK cells. HLA-DR-positive
   NK cells demonstrated functional activity and the ability to present
   antigens to T cells, indicating a protective effect in infectious
   diseases or chronic inflammation.^ [141]28 Interestingly, some
   interleukin (IL) receptors related genes, such as ILDR2, IL18RAP, and
   IL12RB2, showed a significant increase in NK cells in AMD. IL-18, known
   to enhance NK activity by binding to IL-18R,^ [142]29 and the increased
   expression of IL-12R on NK cells were associated with elevated levels
   of NK cell-derived cytokines.^ [143]30 Pathway enrichment analysis
   further supported these findings. In summary, this study highlights the
   crucial role of NK cells in AMD, showcasing highly functional activity
   implicated in the pathologic process through excessive cytokine
   secretion and incorrectly directed cytotoxicity.

   In the present study, based on machine learning approaches, we
   identified CREM, FOXP1, IRF1, NFKB2, and USF2 as key genes. Previous
   research has explored the roles of FOXP1, IRF1, and NFKB2 in the
   development and progression of AMD.^[144]31 -[145]33 FOXP1 exhibited
   expression in the retinal pigment epithelium (RPE), and its knockout in
   RPE led to reduced vascular endothelial growth factor (VEGF),
   contributing to choroidal neovascularization in AMD.^ [146]31 IRF1 has
   been implicated in the etiology of AMD, with IL-27 increasing the
   expression of complement factor H, a key player in AMD development,
   through the activation of the STAT1-IRF-1 pathway in RPE.^ [147]32
   Furthermore, the H haplogroup of Mitochondrial DNA (mtDNA) was
   protective against AMD, and NFKB2 showed differential expression
   between the H and J haplogroups.^ [148]33 Collectively, FOXP1, IRF1,
   and NFKB2 emerged as crucial players in AMD development. Moreover, our
   study has uncovered two hitherto undocumented TFs associated with AMD.
   CREM emerged as the top differentially expressed TF in
   melphalan-induced vascular toxicity.^ [149]34 Notably, CREM-1
   demonstrated protective effects against retinal cell apoptosis induced
   by light exposure.^ [150]35 Considering that vascular injury and cell
   apoptosis are pivotal pathological stages in AMD, the implication is
   that CREM may exert influence in the context of AMD.^ [151]36
   Additionally, USF2 exhibited high expression during embryonic eye
   development, suggesting a potential role in eye development.^ [152]37
   Overall, the identification of these predictive genes might represent a
   breakthrough in the pathogenesis and development of AMD.

   This study acknowledges certain limitations that warrant attention in
   future research. Firstly, the limited sample size in the machine
   learning analysis may impact the accuracy and reliability of the
   results. Expanding the sample size could enhance the robustness of the
   findings. Secondly, the selected TFs of NK cells in AMD should be
   subject to further experimental exploration to elucidate their
   molecular mechanisms and assess their therapeutic potential.

Conclusion

   In conclusion, our study provides valuable insights into the intricate
   relationship between NK cells and AMD. The substantial increase in NK
   cell proportions among AMD patients, along with altered cell
   interactions and communication strength, underscores the pivotal role
   of NK cells in the pathogenesis of AMD. Utilizing hdWGCNA and SCENIC,
   we unveiled hub module genes and identified key TFs associated with NK
   cells in AMD. The predictive model, integrating hub genes and key TFs,
   revealed 5 potential TFs that could serve as predictors for the
   development of AMD. This comprehensive approach enhances our
   understanding of the regulatory mechanisms within NK cells during AMD.
   The identified TFs, particularly CREM, FOXP1, IRF1, NFKB2, and USF2,
   open new avenues for exploring molecular interventions in the context
   of AMD.

Supplemental Material

   sj-docx-1-evb-10.1177_11769343241272413 – Supplemental material for
   Single-cell RNA Sequencing Identifies Natural Kill Cell-Related
   Transcription Factors Associated With Age-Related Macular Degeneration
   [153]sj-docx-1-evb-10.1177_11769343241272413.docx^ (2.9MB, docx)

   Supplemental material, sj-docx-1-evb-10.1177_11769343241272413 for
   Single-cell RNA Sequencing Identifies Natural Kill Cell-Related
   Transcription Factors Associated With Age-Related Macular Degeneration
   by Yili Luo, Jianpeng Liu, Wangqiang Feng, Da Lin, Mengji Chen and
   Haihua Zheng in Evolutionary Bioinformatics

Acknowledgments