Abstract

Background

   Immune checkpoint inhibitors (ICIs) have revolutionized cancer
   immunotherapy, but many patients develop resistance. While the
   immunosuppressive effects of ultraviolet (UV) light are
   well-documented, its link to ICI resistance remains unclear.

Methods

   We analyzed publicly available single-cell RNA sequencing (scRNA-seq)
   datasets from ICI-treated patients to explore the relationship between
   UV response (UVR) and treatment outcomes. A novel UVR gene signature
   (UVR.Sig) was established using 34 scRNA-seq datasets and validated in
   The Cancer Genome Atlas (TCGA) pan-cancer cohorts and 10 ICI cohorts.
   Key genes (Hub-UVR.Sig) were identified via six machine learning
   algorithms, and breast cancer (BRCA) subtypes were classified through
   consensus clustering. Biological effects of Hub-UVR.Sig genes were
   confirmed in vitro.

Results

   UVR.Sig was associated with ICI resistance and correlated with
   inhibitory immune cell infiltration and pro-tumor pathways in
   pan-cancer data. The UVR.Sig-based model achieved good predictive
   performance for ICI outcomes (AUC = 0.727). In BRCA, Hub-UVR.Sig
   stratified patients into two subtypes, with high Hub-UVR.Sig expression
   linked to stronger immune evasion and lower immunogenicity. ENO2 and
   ATP6V1F were highly expressed in BRCA tissues, and ENO2 was correlated
   with worse prognosis in BRCA patients. Knockdown of ENO2 reduced cell
   proliferation and invasion.

Conclusion

   We reveal for the first time that UVR is strongly associated with ICI
   resistance. The UVR.Sig feature offers the potential to identify
   patients who respond to immunotherapy and to tailor BRCA treatment
   strategies.

   Keywords: ultraviolet light, immune checkpoint inhibitor, single-cell
   sequencing, bulk-RNA seq, pan-cancer

1. Introduction

   Significant progress has been made in understanding the critical role
   of immune checkpoint inhibitors (ICIs) in regulating the activity of
   tumor-infiltrating T cells, leading to a revolutionary shift in cancer
   immunotherapy ([36]1, [37]2). Immunotherapeutic modalities including
   ICIs, vaccination and passive cell transfer have been extensively
   studied in the clinical setting of breast cancer (BRCA), particularly
   in patients with triple negative breast cancer (TNBC) ([38]3). However,
   only a small subset of patients benefited from immunotherapy, with the
   majority experiencing either primary or acquired resistance
   ([39]4–[40]7). Therefore, identifying appropriate biomarkers for ICI
   therapy sensitivity is crucial for optimizing treatment options and
   improving patient outcomes.

   Ultraviolet (UV) light at wavelengths ranging from 10 to 380 nm is a
   form of electromagnetic radiation. The process by which cells or
   organisms undergo changes in their activity or state (such as movement,
   secretion, enzyme production, and gene expression) in response to UV
   exposure is the UV response (UVR). Although the immunosuppressive
   effects of UVR have been well established, direct evidence linking UVR
   to immunotherapy response remains unavailable. Early studies have
   demonstrated that chronic UV exposure modulates immune and antigenic
   responses, influencing the carcinogenic process in the skin ([41]8).
   This discovery spurred further investigations into the mechanisms
   underlying UVR-induced immunosuppression. Subsequent research revealed
   that UV-induced DNA damage, reactive oxygen species generation, Treg
   induction, and the release of immunosuppressive cytokines, such as
   IL-10, are closely associated with UVR-mediated immunosuppression
   ([42]9, [43]10). Recent studies have suggested that UVR promotes
   immunosuppression by regulating the expression of immune checkpoints.
   UV exposure activates the IRF3 and NF-κB pathways via HMGB1, leading to
   the upregulation of PD-L1 expression and reduction in tumor cell
   sensitivity to CD8+ T cell-mediated cytotoxicity ([44]11). Although
   Carlos et al. identified a UVR-related gene signature that underscores
   the association between UVR and inhibitory immune cells, including
   immature dendritic cells, plasmacytoid dendritic cells, and M2
   macrophages, within the immunosuppressive microenvironment of uveal
   melanoma ([45]12), there is still a substantial gap in the literature
   regarding the role of UVR-related genes in tumors and their impact on
   immunotherapy outcomes.

   The single-cell RNA sequencing (scRNA-Seq) technique enables the
   dissection of complex interactions between tumor cells and immune cells
   at the single-cell level, facilitating a better understanding of the
   dynamic mechanisms underlying tumor-immune interactions
   ([46]13–[47]16). This study combines scRNA-seq technology with
   comprehensive bioinformatics analysis with the aim of constructing a
   predictive model of ICI efficacy based on the expression profiles of
   UVR-associated genes, laying the groundwork for improved stratification
   and personalized treatment of tumor patients. In addition, this study
   bridges a significant gap in our understanding of the role of
   UVR-associated genes in BRCA.

2. Methods and materials

2.1. Identification of UVR-related genes

   UVR-related genes were collected from the Molecular Signatures Database
   (MSigDB) ([48]17). A search using the keyword “UV Response” in MSigDB
   yielded 191 genes included in the HALLMARK_UV_RESPONSE dataset (The
   complete gene list was provided in [49]Supplementary Table S1 ).

2.2. Pan-cancer transcriptomic dataset and processing

   The pan-cancer transcriptomic dataset from The Cancer Genome Atlas
   (TCGA) was obtained from the UCSC Xena platform
   ([50]https://xenabrowser.net) ([51]18) to investigate the potential
   association between UVR-related genes and immune suppression across 30
   cancer types. To avoid interference from the dominant effects of immune
   cells, three cancer types primarily composed of immune cells were
   excluded: acute myeloid leukemia, diffuse large B-cell lymphoma, and
   thymoma. Additionally, tumor mutational burden (TMB) and microsatellite
   instability (MSI) data were obtained from the cBioPortal database
   ([52]19) for subsequent analysis. Relevant clinical and pathological
   information for the 30 cancer types were downloaded using the R package
   TCGAbiolink ([53]20). Patients included in the analysis met the
   following criteria: availability of mRNA expression and clinical data,
   completion of standard diagnosis and treatment, and a survival time
   longer than 30 d.

2.3. Acquisition and processing of ICI-related datasets

   To investigate the relationship between UVR-related genes and
   immunotherapy, the R package GEOquery ([54]21) was used to download two
   scRNA-seq datasets with well-defined efficacy for tumor immunotherapy
   from the GEO database. The Gene Set Variation Analysis (GSVA) R package
   ([55]22) was employed to assess the enrichment scores of UVR-related
   genes in these datasets and explore their association with ICI
   efficacy. These two datasets were the skin cutaneous melanoma (SKCM,
   [56]GSE115978) ([57]23) and the basal cell carcinoma (BCC,
   [58]GSE123813) datasets ([59]24). After quality control (QC), the SKCM
   dataset [60]GSE115978 included 32 patients, comprising 15
   non-responders (NRs) who did not benefit from immunotherapy, 16
   treatment-naïve patients (TN) who did not undergo immunotherapy, and 1
   responder who responded to treatment.

   Ten bulk RNA-seq datasets related to ICI treatment were systematically
   collected. These datasets included five SKCM datasets [Hugo 2016
   ([61]25), Liu 2019 ([62]26), Gide 2019 ([63]27),Riaz 2017 ([64]28) and
   Van 2015 ([65]29)], two urothelial carcinoma (UC) datasets [Mariathasan
   2018 ([66]30) and Synder 2017 ([67]31)], one GBM dataset [Zhao 2019
   ([68]32)], one renal cell carcinoma (RCC) dataset [Braun 2020
   ([69]33)], and one gastric cancer (GC) dataset [Kim 2018 ([70]34)]. The
   Hugo 2016 SKCM dataset consisted of 27 preprocessed tumor samples from
   26 patients, while the GBM dataset included 34 preprocessed tumor
   samples from 17 patients. For both datasets, one tumor sample per
   patient was randomly selected for analysis.

2.4. Collection of published signatures for comparison

   Six pan-cancer signatures [INFG.Sig ([71]35), T.Cell. Infamed.Sig
   ([72]35), PDL1.Sig ([73]36), LRRC15.CAF.Sig ([74]37), NLRP3.Sig
   ([75]38) and cytotoxic.Sig ([76]39)] were gathered along with four
   SKCM-specific signatures [CRMA.Sig ([77]40), IPRES.Sig ([78]25),
   IMS.Sig ([79]41) and TRS.Sig ([80]42)]. The codes and algorithms for
   these 10 signatures were obtained from their original studies, such as
   single-sample gene set enrichment analysis (ssGSEA) for NLRP3.Sig and
   cancer classification for ImmuneCell.Sig.

2.5. Collection and processing of scRNA-seq data

   We collected 34 scRNA-seq datasets containing stromal or immune cells
   from the TISCH database ([81]43), comprising 345 patients and 663,760
   cells across 17 cancer types. These cancer types included BCC, BRCA,
   multiple myeloma (MM), neuroendocrine tumor (NET), non-small cell lung
   cancer (NSCLC), ovarian serous cystadenocarcinoma (OV), pancreatic
   adenocarcinoma (PAAD), SKCM, stomach adenocarcinoma (STAD), uveal
   melanoma (UVM), cholangiocarcinoma (CHOL), colorectal cancer (CRC),
   GBM, head and neck cancer (HNSC), liver hepatocellular cancer (LIHC),
   medulloblastoma (MB), and Merkel cell carcinoma (MCC).

   The integration and analysis of scRNA-seq data were performed using the
   Seurat v4.0.6 R package ([82]44), with doublet QC conducted using the R
   package Scrublet v0.2 ([83]45). During QC, cells with fewer than 300
   detected genes and those with mitochondrial gene reads exceeding 20% of
   the total reads were excluded. Data normalization and standardization
   were performed using principal component analysis (PCA) ([84]46), and
   batch effects across samples were corrected using the Harmony R package
   ([85]47).

2.6. Generation of UVR.Sig

   A UVR gene signature (UVR.Sig) was generated by calculating the
   enrichment scores of UVR-related genes across various scRNA-seq
   datasets using the GSVA R package. Spearman’s correlation analysis was
   conducted between the expression levels and enrichment scores of
   UVR-related genes, marking the positively correlated genes (Spearman r
   > 0.3 and p < 0.05) as Gx. The “FindMarkers” function was used to
   identify differentially expressed genes (DEGs) in malignant tumor cells
   from each scRNA-seq dataset, where genes with |logFC| ≥ 0.30 and FDR (q
   value) < 1e-05 were considered upregulated DEGs in malignant tumor
   cells and labeled as Gy. To obtain the upregulated tumor-specific DEGs
   that were positively correlated with UVR, the intersection of Gx and Gy
   was considered for each dataset to generate the gene set, Gn. Finally,
   the Gn genes from all datasets were combined and deduplicated to form
   the UVR.Sig.

2.7. Gene Ontology and Kyoto Encyclopedia of Genes and Genomes pathway
enrichment analysis

   GO ([86]48) analysis is commonly used for functional enrichment
   studies, encompassing biological processes (BP), cellular component
   (CC), and molecular function (MF). KEGG ([87]49) is a widely used
   database that provides information on genomes, biological pathways,
   diseases, and drugs. GO and KEGG enrichment analyses were performed on
   the UVR.Sig using the ClusterProfiler R package ([88]50). The selection
   criteria for significant enrichment were set to adj.p < 0.05 and FDR (q
   value) < 0.25, with p-values adjusted using the Benjamini–Hochberg
   method.

2.8. Immune-related analysis of UVR.Sig

   To further evaluate the immune relevance of UVR.Sig, the UVR.Sig scores
   were first calculated across 30 cancer types from TCGA pan-cancer
   transcriptome dataset using the GSVA R package. A correlation analysis
   was performed between the UVR.Sig scores and 75 published
   immune-related genes ([89]51), and the results were visualized using
   correlation circle plots. Next, based on the abundance of immune cells,
   the tumor immune microenvironment was determined across different
   cancer types. Immune infiltration analysis was conducted using
   Microenvironment Cell Populations Counter (MCPcounter) ([90]52), which
   calculated absolute abundance scores for eight immune cells and two
   stromal cell types from the gene expression matrix. These cells include
   T cells, CD8+ T cells, cytotoxic lymphocytes, natural killer (NK)
   cells, B lymphocytes, monocytes, myeloid dendritic cells, neutrophils,
   endothelial cells, and fibroblasts. The results were visualized by
   generating a correlation heatmap using the “MCPcounter” function from
   the IOBR R package ([91]53). Finally, all pathways were obtained from
   the HALLMARK gene set using the MSigDB database, the correlation
   between UVR.Sig and each pathway was calculated, and the results were
   visualized using a bubble plot.

2.9. Clinical efficacy evaluation

   The primary clinical outcomes included objective response rate (ORR)
   and overall survival (OS). For all datasets, except Hugo 2016, the ORR
   was assessed using Response Evaluation Criteria in Solid Tumors
   (RECIST) version 1.1 ([92]54). The Hugo 2016 dataset utilized
   immune-related RECIST to evaluate ORR. Based on their response status,
   the participants were categorized into two groups: responders, which
   included patients with complete response or partial response, and NRs,
   which included patients with stable or progressive disease.

2.10. Construction of the ICI efficacy prediction model

   First, the five ICI RNA-seq datasets with the largest patient samples
   were combined to form a new, large dataset (n = 772), including RCC (n
   = 181), UC (n = 348), and SKCM (n = 243) datasets. The five ICI RNA-seq
   datasets were Braun 2020 RCC, Mariathasan 2018 UC, Liu 2019 SKCM, Gide
   2019 SKCM, and Riaz 2017 SKCM. The ComBat method was used to eliminate
   batch effects among different datasets. The merged dataset was randomly
   divided into two sets: a training (80%, n = 618) and validation (20%, n
   = 154) set. The remaining five ICI RNA-seq datasets were used as
   independent testing sets (n = 149): Zhao 2019 GBM, Snyder 2017 UC, Van
   Allen 2015 SKCM, Kim 2018 GC, and Hugo 2016 SKCM datasets.

   Next, the immune response prediction models were trained using UVR.Sig
   and a training set using seven common machine learning (ML) algorithms.
   The seven ML algorithms were: Naive Bayes, Random Forest (RF), Support
   Vector Machine, AdaBoost Classification Trees, Boosted Logistic
   Regressions, k-Nearest Neighbors, and the Cancerclass algorithms. For
   each ML algorithm with parameters, except Cancerclass, hyperparameter
   tuning was performed using five-fold cross-validation to optimize the
   model performance. To ensure robustness, the optimization process was
   repeated 10 times with different random seeds for each resampling. For
   the parameter-free cancer-class algorithm, the entire training set was
   directly used to train the model. Subsequently, the area under the
   curve (AUC) values of the seven models in the validation set were
   compared to identify the most effective ML algorithm for the final
   UVR-related ICI efficacy prediction model. An AUC > 0.5 indicated a
   positive association between molecular expression and event occurrence,
   and the closer the AUC was to 1, the better the diagnostic performance.
   Specifically, an AUC between 0.5 and 0.7 suggests low accuracy, between
   0.7 and 0.9 indicates moderate accuracy, and above 0.9 indicates high
   accuracy.

   To further compare the predictive performance of UVR.Sig, its AUC
   values were evaluated against six previously published ICI response
   signature-related gene sets (PDL1.Sig, LRRC15.CAF.Sig, INFG.Sig,
   T.cell.infamed.Sig, NLRP3.Sig, and Cytotoxic.Sig) across the training,
   validation, and five independent test sets. We showed AUC from the
   training, validation, and three best-performing independent test sets
   (Synder 2017 UC, Kim 2018 GC, and Hugo 2016 SKCM).

2.11. Collection and processing of CRISPR datasets

   To identify the potential therapeutic targets for UVR.Sig, data from
   seven published CRISPR/Cas9 screening studies that assessed the effects
   of individual gene knockouts on tumor immunity were gathered. These
   studies included Freeman 2019 ([93]55), Kearney 2018 ([94]56), Manguso
   2017 ([95]57), Pan 2018 ([96]58), Patel 2017 ([97]59), Vredevoogd 2019
   ([98]60) and Lawson 2020 ([99]61). Based on the model cell lines and
   the applied treatment conditions, data from the seven CRISPR studies
   were divided into 17 datasets. CRISPR analysis included cell lines from
   SKCM, GBM, CRC, and RCC. These data were used to identify genes that
   were highly likely to regulate lymphocytes and affect ICI response
   across various datasets.

   CRISPR screening was performed by performing whole-genome CRISPR-Cas9
   knockouts in various cancer cell lines. Screening was performed in two
   environments: in vitro, where different cancer cell lines were cultured
   with and without cytotoxic T lymphocytes (CTLs), and in vivo, where
   different cancer cell lines were implanted in immunodeficient or
   immunocompetent mice. Following these treatments, RNA-seq was performed
   to evaluate the abundance of the corresponding gene-specific single
   guide RNA(sgRNA). To measure the impact of gene knockouts on CTLs
   pressure or anti-tumor immunity, the logFC values between different
   groups of cell lines were calculated. Normalized z-scores were computed
   from the logFC values to eliminate batch effects and allow for
   comparisons across different CRISPR datasets. Lower z-scores indicated
   a better immune response after gene knockout. Genes were ranked based
   on the average z-scores from the 17 datasets, with lower z-scores
   indicating a higher ranking. Genes with the lowest z-scores were
   considered potential immune resistance genes.

   Additionally, to further assess the predictive value of UVR.Sig, it was
   compared with previously identified signatures used to predict ICI
   response, including five pan-cancer signatures (INFG. Sig,
   T.cell.infamed.Sig, PDL1.Sig, LRRC15.CAF.Sig, Cytotoxic.Sig) and four
   SKCM-specific features (CRMA.Sig, IPRES.Sig, IMS.Sig, and TRS.Sig).

2.12. Construction of prognostic risk models using ML algorithms

   To explore the relationship between UVR.Sig and the prognosis of
   patients with cancer, a prognostic risk model of UVR.Sig was
   constructed using six ML algorithms: Bagged Trees, Bayesian, Learning
   Vector Quantization (LQV), Wrapper (Boruta), Least Absolute Shrinkage
   and Selection Operator (LASSO), and RF. Finally, the results of the
   different algorithms were compared. Genes that appeared in at least
   four ML algorithms were selected as Hub-UVR.Sig. Kaplan–Meier (KM)
   analysis was conducted using the R package survival ([100]62), and KM
   curves for both the training and validation cohorts were generated
   based on risk scores to compare OS differences between the high- and
   low-risk groups.

2.13. Panoramic analysis of Hub-UVR.Sig

   The Hub-UVR.Sig enrichment scores were calculated using the ssGSEA
   algorithm from the R package GSVA based on TCGA pan-cancer
   transcriptomic dataset. This analysis aimed to explore the correlation
   between Hub-UVR.Sig and immune cell infiltration abundance across 30
   different cancer types. Subsequently, the relationship between
   Hub-UVR.Sig and MSI was investigated. Finally, to evaluate the
   prognostic significance of Hub-UVR.Sig in patients with cancer, the
   correlation between OS and Hub-UVR.Sig was analyzed across various
   cancer types and KM curves were constructed. Based on the results of
   the correlation and KM analyses, we focused on specific cancer types.

2.14. Construction and correlation analysis of BRCA subtypes

   Consensus Clustering ([101]63) is a resampling-based algorithm used to
   determine subgroup membership and validate clustering accuracy. Through
   multiple iterations on the subsamples of the dataset, this method
   introduces sampling variability, offering stability and metrics for
   selecting the optimal clustering parameters. Using the consensus
   clustering method from the R package ConsensusClusterPlus
   ([102]64),different subtypes of BRCA were identified based on
   Hub-UVR.Sig. During this process, the number of clusters was set
   between two and six, with 1000 resampling iterations, extracting 80% of
   the total samples each time (clusterAlg = KM, distance = Euclidean).

   To obtain Tumor Immune Dysfunction and Exclusion (TIDE) scores for
   different subtypes of BRCA, the TIDE algorithm ([103]65, [104]66) was
   applied to the expression matrix of BRCA samples. Differences in TIDE
   scores were calculated, and differences in MSI and TMB scores were
   assessed among BRCA subtypes using the Wilcoxon rank-sum test.

2.15. Somatic mutation analysis of BRCA subtypes

   The “Masked Somatic Mutation” data from TCGA website was selected as
   the somatic mutation data for BRCA samples and preprocessed using
   VarScan software. The R package maftools ([105]67) was used to
   visualize the somatic mutation landscape in the different BRCA
   subtypes.

2.16. Immune infiltration analysis of BRCA subtypes

   CIBERSORT ([106]68), utilizing linear support vector regression, was
   applied to deconvolute the transcriptome expression matrix and estimate
   the composition and abundance of immune cells in mixed-cell
   populations. By employing the CIBERSORT algorithm with the LM22
   signature gene matrix and filtering the data with immune cell
   enrichment scores greater than zero, a detailed immune cell
   infiltration matrix was obtained. The R package ggplot2 was used to
   create grouped comparison plots to illustrate the differences in immune
   cell expression among different BRCA subtypes. Additionally, the R
   package pheatmap was used to generate heatmaps displaying the
   correlation analysis results between the immune cells and Hub-UVR.Sig
   across BRCA subtypes. Finally, the correlation between Hub-UVR.Sig and
   immune cells (p < 0.05) were selected, and correlation scatter plots
   for the top two positive and top two negative correlations were
   plotted.

2.17. Cell culture, transfection and infection

   The human BRCA cell lines MDA-MB-231 and BT549 were purchased from
   ATCC. Cells were cultured in high-glucose DMEM (Procell, Wuhan, China)
   supplemented with 10% fetal bovine serum (FBS) (Gibco, USA) and 100
   U/mL penicillin/streptomycin (Procell, Wuhan, China) in a sterile
   incubator at 37 °C with 5% CO2. Tumor cells were seeded into 6-well
   plates and incubated overnight before transfection with small
   interfering RNA (siRNA). According to the manufacturer’s protocol,
   siRNA and negative control (siNC) were transfected into tumor cells
   using Lipo2000 reagent (Invitrogen, USA). The synthetic sequences for
   siRNA targeting ENO2 were as follows:
     * siENO2–1 forward: 5`- GCAACUGUCUGCUGCUCAAGG -3`
     * siENO2–1 reverse: 5`- UUGAGCAGCAGACAGUUGCAG -3`
     * siENO2–2 forward: 5`- CGAUGUGUCUGUAUUUCAUGU -3`
     * siENO2–2 reverse: 5`- AUGAAAUACAGACACAUCGUU -3`
     * siNC forward: 5`- UUCUCCGAACGUGUCACGUTT -3`
     * siNC reverse: 5`- ACGUGACACGUUCGGAGAATT -3`

2.18. RNA isolation and quantitative real-time RT-PCR

   Breast tumors and corresponding paracancerous tissues were collected
   from six patients from the Department of Breast Surgery at the Second
   Xiangya Hospital of Central South University. This study was approved
   by the Ethics Committee of the Second Xiangya Hospital of Central South
   University (Ethical approval number: K005), and all participants
   provided written informed consent. Total RNA was extracted from the
   cells using RNAex Pro reagent (AG21101, Hunan, China) according to the
   instructions. The concentration and quality of RNA were detected using
   a spectrophotometer. Reverse transcription was performed using Evo
   M-MLV kit (AG11705, Hunan, China) according to the instructions. qPCR
   was performed using 2X Universal SYBR Green Fast qPCR Mix (RK21203,
   Wuhan, China) and Gentier 96E/96R real-time PCR system (Tianlong,
   Shanxi, China) (each sample was performed in triplicate).
   Glyceraldehyde-3-phosphate dehydrogenase (GAPDH) was used as an
   internal reference gene, and the relative expression level was
   calculated by the 2^-ΔΔCt method, and three replicates were tested for
   each sample. The primers were:
     * GAPDH forward primer:5′- TGACCTGCCGTCTAGAAAAACCT -3′
     * GAPDH reverse primer:5′- GCTGTTGAAGTCAGAGGAGACCA -3′
     * ENO2 forward primer:5′- TGCCTGGTCCAAGTTCACAGC -3′
     * ENO2 reverse primer:5′- CACTGCCCGCTCAATACGTT -3′

2.19. Cell function experiment

   The proliferative capacity of tumor cells was assessed using CCK-8
   assay. Migration, invasive capacity of tumor cells was assessed by
   wound healing and Matrigel-coated transwell, respectively. In the CCK-8
   experiments, cells were seeded into 96-well plates at a density of 1.5
   × 10³ cells per well, and cell pfroliferation was evaluated daily for 4
   consecutive days using the Cell Counting Kit-8 (NCM Biotech, Suzhou,
   China). Optical density (OD) was measured at 450 nm. In wound healing
   assays, cells were seeded in 6-well plates at a density of 4 × 10^^5
   per well and grown to confluence. After removal of unadhered cells by
   washing with phosphate-buffered saline (PBS), a sterile 20 μL pipette
   tip was used to create a scratch wound in each well. Images were taken
   in the same area at 0 h and 48 h, respectively, and the distance of
   wound closure was measured to show cell migration ability. As for
   Matrigel-coated transwell, cells (1 × 10^4) were seeded in serum-free
   medium in 8 μm (Corning Incorporated, 3464, USA) upper chambers coated
   with Matrigel (Yeason, 40183ES08, Shanghai, China) and lower chambers
   with medium containing 20% FBS. After 24 hours of incubation, the cells
   were fixed with formaldehyde. Unattached cells in the upper chamber
   were carefully wiped away, and cells attached to the membrane were
   stained and counted with crystal violet.

2.20. Statistical analysis

   All data processing and analyses in this study were conducted using R
   software (Version 4.2.0) except for the cell experiments. Continuous
   variables are presented as mean ± standard deviation. Comparisons
   between the two groups were performed using the Wilcoxon rank-sum test.
   Unless otherwise specified, correlations between different molecules
   were calculated using Spearman correlation analysis. Statistical
   analyses of cellular experiments were performed using GraphPad Prism 9.
   Data are expressed as mean ± SD of at least three independent
   experiments. Statistical significance was analyzed by Student’s t-test
   (two-tailed). p-values less than 0.05 were considered statistically
   significant.

3. Results

3.1. Flow chart

   The flow of this study is shown in [107]Figure 1 .

Figure 1.

   [108]Flowchart of a UVR study process. It starts with UVR leading to
   two datasets: GSE115978 (SKCM) and GSE123813 (BCC). These form 34
   malignant single-cell RNA datasets analyzed through machine learning
   methods, producing UVR.Sig. Further analysis via Hub-UVR.Sig includes
   landscape analysis, correlation and KM analysis on TCGA-BRCA, which
   connects to consensus cluster analysis, mutation analysis, and immune
   infiltration analysis.
   [109]Open in a new tab

   Flow chart for the comprehensive analysis of an ultraviolet response
   signature (UVR.Sig).

3.2. Negative correlation between UVR-related genes and ICI efficacy

   The results from both datasets indicated a significant enrichment of
   UVR-related genes in malignant tumor cells ([110] Figures 2A–C ). In
   the SKCM dataset ([111]GSE115978), the UVR scores of the
   immunotherapy-effective group (R) were significantly lower than those
   of the immunotherapy-ineffective group (NR) (p < 0.001) and
   treatment-naïve group (TN) (p < 0.001). Furthermore, the UVR scores of
   the NR group were significantly lower than those of the TN group (p <
   0.001; [112]Figure 2D ). Similarly, in the BCC dataset
   ([113]GSE123813), the UVR scores of the R group were significantly
   lower than those of the NR group (p < 0.001; [114]Figure 2D ). These
   findings suggest a negative correlation between UVR-related genes and
   ICI efficacy.

Figure 2.

   [115]Four panels labeled A through D compare UV response data from
   datasets SKCM GSE115978 and BCC GSE123813. Panels A and B show UMAP
   plots differentiating cell types (malignant, stromal, immune) using
   color coding. Panels C and D display box plots illustrating variations
   in UV response among these cell types, with statistical significance
   marked. Panel C compares responses generally, while panel D compares
   responses across treatment types (TN, NR, R).
   [116]Open in a new tab

   Negative association between UVR-related genes and ICI efficacy in
   [117]GSE115978 (SKCM) and [118]GSE123813 (BCC). (A) Distribution of
   different cell types within the samples. (B) Enrichment scores of
   UVR-related genes in the samples, with deep red indicating higher
   scores and deep blue indicating lower scores. (C) Differences in UVR
   scores among various cell types. (D) Relationship between immunotherapy
   efficacy and UVR scores. SKCM, skin cutaneous melanoma; TN,
   treatment-naive; NR, non-responders; R, responders; ICI, immune
   checkpoint inhibitor; BCC, basal cell carcinoma. ***indicates a p value
   < 0.001, denoting highly statistically significant results.

3.3. Screening and enrichment analysis of UVR-related genes

   Through the analysis of the aforementioned two datasets, a significant
   association was identified between UVR-related genes and ICI
   resistance. Consequently, we hypothesized that the expression levels of
   UVR-related genes in patients could potentially predict the
   effectiveness of immunotherapy. We obtained 38 up-regulated
   tumor-specific DEGs in 34 scRNA-Seq datasets that were positively
   associated with UVR: ATF3, BTG3, FOS, FOSB, JUNB, NFKBIA, NR4A1, RHOB,
   SOD2, EPCAM, GGH, IGFBP2, CXCL2, BTG1, DNAJA1, DNAJB1, ALDOA, AP2S1,
   CDKN1C, ENO2, FEN1, HSPA2, OLFM1, PPIF, BTG2, ICAM1, SQSTM1, AMD1,
   ATP6V1F, BSG, CNP, CREG1, CYB5R1, MMP14, PPT1, RPN1, SELENOW, STIP1
   ([119] Figure 3A ).

Figure 3.

   [120]Circular diagram displaying cancer datasets surrounding a central
   “UVR.Sig” label, with color-coded cancer types. Bar charts compare
   ontology categories BP, CC, MF, and KEGG, showing different gene
   ontologies and pathways, each with varying significance levels.
   [121]Open in a new tab

   Development and description of UVR.Sig. (A) Venn diagram illustrating
   the intersection of genes positively correlated with UVR enrichment
   scores across various scRNA-seq datasets and DEGs upregulated in
   malignant tumor cells. Different colors represent distinct cancer
   types. (B) Bar chart depicting the results of GO and KEGG enrichment
   analyses for UVR.Sig, categorized into BP, CC, MF, and KEGG. The x-axis
   represents the GO terms. The selection criteria for GO and KEGG
   enrichment analyses were adjusted p-value (adj.p) < 0.05 and FDR value
   (q-value) < 0.25, with p-value adjustments performed using the
   Benjamini–Hochberg (BH) method. BP, biological process; CC, cellular
   component; DEG, differentially expressed gene; GO, Gene Ontology; KEGG,
   Kyoto Encyclopedia of Genes and Genomes; MF, molecular function;
   scRNA-seq, single-cell RNA sequencing; UVR, ultraviolet response;
   UVR.Sig, ultraviolet response signature.

   To analyze the BP, MF, and CC, and pathways associated with UVR.Sig, GO
   and KEGG enrichment analyses were performed, with detailed results
   provided in [122]Supplementary Table S2 . The analyses revealed that
   UVR.Sig was primarily enriched in the following biological processes:
   response to temperature stimuli, mechanical stimuli, cold, and metal
   ions; and skeletal muscle cell differentiation. In terms of CC, UVR.Sig
   was enriched in melanosomes, pigment granules, sperm heads,
   transcription regulator complexes, and membrane rafts. MF analysis
   indicated that UVR.Sig was involved in heat shock protein binding,
   ATPase activator activity, chaperone binding, glycolipid binding, and
   unfolded protein binding. Furthermore, UVR.Sig was significantly
   enriched in pathways related to TNF signaling, osteoclast
   differentiation, lipid and atherosclerosis, RNA degradation, and
   rheumatoid arthritis signaling pathways. The results of the GO and KEGG
   enrichment analyses were visualized using bar charts ([123] Figure 3B
   ).

3.4. Immune correlation analysis of UVR.Sig

   The correlation between UVR.Sig and the 75 immune-related genes was
   analyzed. The results indicated ([124] Figure 4A ) that UVR.Sig was
   significantly and positively correlated with most immune genes,
   suggesting its potential key role in the regulation of immune
   responses. Subsequently, the infiltration of immune cells across
   various cancers was assessed, which revealed that the degree of immune
   cell infiltration associated with UVR.Sig varies among different cancer
   types ([125] Figure 4B ). Specifically, in SKCM, HNSC, and mesothelioma
   (MESO), the UVR.Sig showed a predominantly negative correlation with
   immune cell infiltration, whereas a positive correlation was observed
   for CHOL, lower grade glioma (LGG), and kidney chromophobe (KICH).
   Finally, we examined the relationship between the UVR.Sig and HALLMARK
   pathways to explore whether immunosuppressive biological functions were
   upregulated in tumors with high UVR. Sig expression. The results
   indicated significant enrichment of pathways, such as angiogenesis,
   epithelial-mesenchymal transition (EMT), Hedgehog signaling, IL-2-STAT
   signaling, IL-6-JAK-STAT signaling, and inflammatory response in tumors
   with high UVR.Sig expression ([126] Figure 4C ). These findings suggest
   that UVR.Sig regulates immunosuppressive functions within the TME,
   thereby promoting tumor aggressiveness and resistance.

Figure 4.

   [127]Composite image with three panels showing correlation heatmaps and
   circular diagrams. Panel A displays a circular heatmap illustrating
   correlations between pathways and gene expressions, color-coded from
   blue to red based on correlation strength. Panel B is a dot plot with
   cancer types on the y-axis and gene sets on the x-axis; dot size and
   color represent significance and correlation. Panel C is a larger dot
   plot with cancer types and pathways, where dot size and color indicate
   p-values and correlation strength, respectively, ranging from blue to
   red.
   [128]Open in a new tab

   Immune resistance analysis of UVR.Sig. (A) Correlation circle diagram
   between UVR.Sig and immune-related genes. (B) Correlation point diagram
   of UVR.Sig with immune cell infiltration across different cancer types.
   (C) Correlation dot plot of UVR.Sig with HALLMARK pathways across
   different cancer types. Red indicates positive correlation and blue
   indicates negative correlation. UVR.Sig, ultraviolet response
   signature.

3.5. Construction of immune efficacy prediction model

   To investigate the relationship between UVR.Sig and the efficacy of
   ICIs, we collected 10 RNA-Seq datasets with clearly defined efficacy of
   immunotherapy, along with their clinical information. All cohorts were
   divided into a training set (n=618), a validation set (n=154), and five
   independent testing sets (n=149). First, we constructed immune efficacy
   prediction models using 7 ML algorithms in the training set, performing
   10 rounds of 5-fold cross-validation for parameter optimization.
   Subsequently, by comparing the AUC values of the 7 models in the
   validation set, we identified the best-performing model. The results
   indicated that the model constructed using the Cancerclass algorithm
   outperformed the others, achieving the highest AUC value of 0.727, and
   was selected as the UVR.Sig model ([129] Figures 5A, B ).

Figure 5.

   [130]Panel A shows a dot plot comparing AUC scores of seven methods,
   highlighting “Cancerclass” with the highest score at 0.727. Panel B is
   an ROC curve for Cancerclass in validation, with an AUC of 0.727. Panel
   C is a radial bar plot displaying AUC scores for different signatures
   across validation and training sets. Panel D is a heatmap showing AUC
   values for various signatures, with darker greens indicating higher
   scores.
   [131]Open in a new tab

   Prediction of immunotherapy outcomes and AUC of UVR.Sig. (A)
   Performance of predictive models for immunotherapy efficacy constructed
   using seven ML algorithms. (B) ROC curve of the prediction model
   constructed by the Cancerclass algorithm. (C) Circle plot comparing the
   performance of UVR.Sig and other signatures in the training,
   validation, and testing sets (Hugo 2016 SKCM, Synder 2017 UC, Kim 2018
   GC). (D) Heatmap comparing the predictive value of UVR.Sig and other
   signatures. Mean AUC represents the average AUC value. AUC, area under
   the curve; ML, machine learning; UVR.Sig, ultraviolet response
   signature.

   Then, we compared the predictive capability of UVR.Sig with 6 other
   signatures (INFG.Sig, T.cell.infamed.Sig, PDL1.Sig, LRRC15.CAF.Sig,
   NLRP3.Sig, and Cytotoxic.Sig) across the training, validation, and
   testing sets. The results demonstrated that UVR.Sig exhibited the best
   predictive performance in the training set, validation set, and Hugo
   2016 SKCM, Snyder 2017 UC, and Kim 2018 GC of the testing sets ([132]
   Figure 5C ). In contrast, most previously published signatures only
   achieved high stability in one or two datasets, while their performance
   in other external cohorts was considerably unsatisfactory, likely due
   to poor generalizability. This underscores the potential of UVR.Sig as
   a pan-cancer predictive model for ICI response. Finally, the heatmap
   ([133] Figure 5D ) shows the performance of individual signatures in
   predicting ICI efficacy across different datasets, further confirming
   the stability of UVR.Sig.

3.6. CRISPR analysis of UVR.Sig

   Gene-knockout immune response data were systematically collected from
   seven CRISPR datasets and further categorized into 17 distinct datasets
   based on the model cells and treatment conditions utilized in each
   cohort. The 17 CRISPR datasets contained 22,505 genes. Initially, the
   genes were ranked according to their average Z-scores, with the
   top-ranking genes identified as potential immune resistance genes,
   indicating that their knockouts may enhance anti-tumor immunity.
   Conversely, genes that ranked lower were classified as immune
   sensitive, which may inhibit the biological functions of anti-tumor
   immunity upon knockout ([134] Figure 6A ). Specifically, the top 1%,
   2%, and 3% corresponded to 225, 450, and 675 genes, respectively.
   Subsequently, the percentage of UVR.Sig among the top genes was
   compared to that of nine other ICI response signatures (INFG.Sig,
   T.cell.infamed.Sig, PDL1.Sig, LRRC15.CAF.Sig, Cytotoxic.Sig, CRMA.Sig,
   IPRES.Sig, IMS.Sig, and TRS.Sig). The results indicated that UVR.Sig
   occupied a higher percentage of the top genes than the other signatures
   ([135] Figure 6B ). Notably, among the top 20% of genes, six genes from
   the UVR.Sig were identified: DNAJA1, STIP1, JUNB, EPCAM, OLFM1, and
   NR4A1. Validation of the ICI resistance characteristics of these six
   UVR.Sig genes across 17 CRISPR datasets demonstrated their potential as
   predictive targets for immunotherapy ([136] Figure 6C ).

Figure 6.

   [137]Panel A shows a heatmap displaying z-scores for various genes
   across different cohorts. Genes are listed on the right, cohorts at the
   top, with a color-coded z-score key from -2 to 2. Panel B is a radar
   chart illustrating the percentage of top-ranked genes across various
   signatures like IPRES.Sig and PD1.Sig in different colors. Panel C is
   another heatmap focusing on specific genes with a similar z-score color
   key, linked to cohorts listed on the right.
   [138]Open in a new tab

   CRISPR analysis of UVR.Sig. (A) Gene ranking based on z-scores across
   17 CRISPR datasets. Green indicates genes with immune resistance
   characteristics, where knockout enhances anti-tumor immune function;
   yellow indicates immune-sensitive genes, where knockout suppresses
   anti-tumor immunity; white represents missing values for genes in the
   core data. (B) Radar plot comparing the percentage of top-ranking genes
   from UVR.Sig and nine other signatures. (C) Heatmap visualization of
   the z-scores for six UVR.Sig genes across 17 CRISPR datasets. UVR.Sig,
   ultraviolet response signature.

3.7. ML-based selection of Hub-UVR.Sig

   To further refine the UVR.Sig, six ML algorithms were calculated to
   analyze the relationship between UVR.Sig and ICI efficacy based on the
   training set: the Wrapper (Boruta) ([139] Figure 7A ), Bayesian ([140]
   Figure 7B ), Bagged Trees ([141] Figure 7C ), RF ([142] Figure 7D ),
   LASSO ([143] Figure 7E ), and LQV ([144] Figure 7F ). The intersection
   of results from different ML algorithms was obtained ([145] Figure 7G
   ), selecting genes that appeared in at least four ML algorithms as
   Hub-UVR.Sig, including ATF3, ATP6V1F, BTG1, BTG3, ENO2, FOS, and ICAM1.

Figure 7.

   [146]A series of seven graphs showing different model analyses: A)
   Boruta showing variable importance with a boxplot. B) Bayesian model
   showing cross-validation accuracy decreasing as variables increase. C)
   Bagged Trees model showing fluctuating accuracy with variable changes.
   D) Random Forest model illustrating increasing accuracy with more
   variables. E) LASSO frequency of models with a bar chart. F) LQV model
   showing feature importance ranking. G) An intersection size chart
   comparing overlap between different models, featuring a bar graph of
   set sizes with intersecting dots.
   [147]Open in a new tab

   ML and Hub-UVR.Sig analysis. (A) Graph of wrapper (Boruta) algorithm
   prognostic risk model of UVR.Sig. (B) Bayesian algorithm prognostic
   risk model map of UVR.Sig. (C) Bagged Trees algorithm prognostic risk
   model diagram of UVR.Sig. (D) Random Forest algorithm prognostic risk
   model map of UVR.Sig. (E) LASSO prognostic risk model map of UVR.Sig.
   (F) Learning vector quantification prognostic risk model plot of
   UVR.Sig. (G) Intersection plot of results from six different ML
   algorithms. ML, machine learning; UVR.Sig, ultraviolet response
   signature.

   To assess the prognostic value of Hub-UVR.Sig, a multivariate Cox
   regression analysis was performed on the training and validation sets.
   Risk score was calculated for each sample based on the coefficients of
   the Cox regression model, and samples were classified into high- and
   low-risk groups using the “surv_categorize” function, according to the
   optimal cutoff-value. Risk factor plots were used to visualize the
   relationships among samples ([148] Figures 8A, C ), and Kaplan–Meier
   curve analysis was performed to evaluate the OS ([149] Figures 8B, D ).
   The results showed that the low-risk group had significantly better OS
   in both the training and validation sets compared with the high-risk
   group (log-rank p < 0.01). Finally, the chromosomal locations of the
   seven Hub-UVR.Sig genes were analyzed using the RCircos R package, and
   a chromosome location map was generated ([150] Figure 8E ). The map
   revealed that BTG1 and ENO2 were located on chromosome 12.

Figure 8.

   [151]Figure comprising five panels. Panel A features a scatter plot and
   heat map of gene expressions associated with survival time and risk
   group. Panel B shows a Kaplan-Meier survival curve comparing high and
   low-risk groups with a p-value. Panel C presents another scatter plot
   and heat map with distinct gene expressions. Panel D contains a
   Kaplan-Meier curve with survival data and significance. Panel E depicts
   a circular ideogram of human chromosomes with gene locations marked.
   [152]Open in a new tab

   Hub-UVR.Sig and Kaplan–Meier curve analysis. (A) Risk factor plot of
   the Hub-UVR.Sig associated with patient prognosis in the training set.
   (B) Kaplan–Meier curve for high-risk and low-risk patients in the
   training set. (C) Risk factor map of the Hub-UVR.Sig and the prognosis
   of patients with cancer in the validation set. (D) Kaplan–Meier curve
   for high-risk and low-risk patients in the validation set. (E)
   Chromosomal mapping of the Hub-UVR.Sig. Blue represents the low-risk
   group, while yellow represents the high-risk group. UVR.Sig,
   ultraviolet response signature.

3.8. Panoramic analysis of Hub-UVR.Sig

   The enrichment scores of Hub-UVR.Sig in the TCGA dataset of 30
   different cancer types were evaluated using ssGSEA ([153] Figure 9A ).
   The correlation between Hub-UVR.Sig and immune cell infiltration was
   examined in various cancer types. Hub-UVR.Sig showed a significant
   positive correlation with infiltration of activated dendritic cells,
   resting dendritic cells, M1 macrophages, M2 macrophages, activated mast
   cells, and neutrophils. Conversely, it exhibited a significant negative
   correlation with the infiltration of resting mast cells and naïve CD4+
   T cells ([154] Figure 9B ).

Figure 9.

   [155]Panel A displays a scatter plot comparing HBZ-UVR signature across
   various cancer types, shown in distinct colors. Panel B is a
   correlation matrix with colored circles representing correlation
   strength, redness indicating a positive, and blue a negative
   correlation. Panel C shows a heat map with shades indicating log10p
   values, linking cancer types to specific molecular features. Panel D
   presents another heat map for gene expression values and significance.
   Panel E is a Kaplan-Meier survival plot comparing high-risk and
   low-risk groups, illustrating differing survival probabilities over
   time, with statistical significance p = 0.0037.
   [156]Open in a new tab

   Landscape analysis of Hub-UVR.Sig. (A) Scores of the Hub-UVR.Sig across
   different cancer types. (B) Correlation between Hub-UVR.Sig and the
   abundance of immune cell infiltration in various cancer types. (C)
   Correlation between Hub-UVR.Sig and MSI across 30 cancer types. (D)
   Relationship between the expression levels of Hub-UVR.Sig and MSI in 30
   different cancer types. (E) Kaplan–Meier curves of the risk scores of
   the Hub-UVR.Sig and OS in BRCA. The absolute value of the correlation
   coefficient (r value) is interpreted as follows: values below 0.3
   indicate weak or no correlation, values between 0.3 and 0.5 indicate
   weak correlation, values between 0.5 and 0.8 indicate moderate
   correlation, and values above 0.8 indicate strong correlation. Red
   indicates positive correlation, while blue indicates negative
   correlation. Blue represents the low-risk score group, while yellow
   represents the high-risk score group. BRCA, breast cancer; MSI,
   microsatellite instability; OS, overall survival; UVR.Sig, ultraviolet
   response signature. *p<0.05, **p < 0.01.

   Next, the correlation between Hub-UVR.Sig and MSI was assessed across
   the 30 cancer types. Hub-UVR.Sig showed the strongest positive
   correlation with MSI in BRCA and COAD and the strongest negative
   correlation with ESCA ([157] Figure 9C ). Furthermore, the relationship
   between the expression levels of Hub-UVR.Sig and MSI was determined in
   different cancer types. The results revealed that ICAM1 exhibited the
   strongest negative correlation with MSI in CHOL, whereas ENO2 showed
   the strongest positive correlation with MSI in COAD ([158] Figure 9D ).

   Finally, the prognostic ability of Hub-UVR.Sig was analyzed across 30
   different cancer types in TCGA cohort. The results of multivariate Cox
   regression and survival analyses indicated that in BRCA, patients with
   high-risk scores had significantly worse OS than those with low-risk
   scores (p < 0.01; [159]Figure 9E ), suggesting that UVR.Sig may serve
   as a potential marker for poor prognosis in patients with BRCA.

3.9. Construction and correlation analysis of BRCA subtypes

   Given the excellent predictive capability of Hub-UVR.Sig for OS in
   patients with BRCA, we focused our investigation on its relationship
   with BRCA. Based on the Hub-UVR.Sig expression levels in TCGA-BRCA
   samples, a consistency clustering analysis was performed using the R
   package ConsensusClusterPlus, ultimately identifying two subtypes: BRCA
   subtypes A (Cluster 1) and B (Cluster 2) ([160] Figures 10A, B ).
   Subtype A comprised 591 samples, whereas subtype B comprised 484
   samples. The PCA results revealed significant differences between the
   two subtypes ([161] Figure 10C ). Further analysis of TIDE ([162]
   Figure 10D ), TMB ([163] Figure 10E ), and MSI scores ([164] Figure 10F
   ) in the different BRCA subtypes indicated that there were
   statistically significant differences in TIDE and TMB scores for the
   different subtypes, with subtype A showing higher TIDE scores and lower
   TMB scores compared with subtype B (p < 0.05).

Figure 10.

   [165]Composite image of six graphs: (A) Heatmap shows a consensus
   matrix for k=2 clusters. (B) Cumulative distribution function plot with
   curves for k=2 to k=6. (C) PCA scatter plot displaying two clusters,
   Cluster1 in pink, Cluster2 in green. (D) Box plot for TIDE scores
   between Cluster1 and Cluster2, significant difference noted (p =
   4.2e-08). (E) Box plot for TMB between clusters, significant difference
   (p = 0.03). (F) Box plot for MSI between clusters, no significant
   difference (p = 0.23).
   [166]Open in a new tab

   Consensus cluster analysis. (A) Consensus cluster analysis results for
   BRCA samples. (B) Consistency cumulative distribution function plot
   from the consensus clustering analysis. (C) PCA plot of two BRCA
   subtypes. (D–F) Comparison of TIDE (D), MSI (E), and TMB (F) across
   different BRCA subtypes. Pink represents subtype A (Cluster 1), while
   green represents subtype B (Cluster 2). BRCA, breast cancer; MSI,
   microsatellite instability; PCA, principal component analysis; TIDE,
   Tumor Immune Dysfunction and Exclusion; TMB, tumor mutational burden.

3.10. Somatic mutation analysis of BRCA subtypes

   Subsequently, the mutation frequencies were analyzed in the different
   BRCA subtypes. The results indicated that the TP53 gene exhibited the
   highest mutation frequency in subtype A, reaching 37%, while in subtype
   B, the PIK3CA gene had the highest mutation frequency at 40% ([167]
   Figure 11A ). Furthermore, the biological and functional changes
   induced by mutations were assessed in different subtypes. In subtype A,
   mutations primarily affected the functions of the TGF-β and PI3K
   signaling pathways ([168] Figure 11B ). In contrast, in subtype B, the
   mutations mainly affected the RTK-RAS and PI3K signaling pathways
   ([169] Figure 11C ).

Figure 11.

   [170]Panel A presents a bar graph comparing genetic mutations across
   two clusters, highlighting gene-specific variations like TP53 and
   PIK3CA. Panel B illustrates pathways affected by mutations, with a
   focus on fractions affected in samples. Panel C continues this analysis
   for another set of pathways. Panels D and E list druggable categories,
   specifying gene and protein targets, marked by color codes for various
   categories like tyrosine kinases and DNA repair mechanisms.
   [171]Open in a new tab

   SNP analysis. (A) Mutation landscape of BRCA subtype A and subtype (B)
   Each color represents a different type of mutation, with the vertical
   axis listing the genes with the highest mutation frequencies, including
   TP53, PIK3CA, TTN, CDH1, GATA3, and MUC16. (B, C) Biological function
   analysis of mutations affecting patients in subtype A (B) and B (C).
   The vertical axis lists various signaling pathways, with the horizontal
   axis indicating the proportion of samples and pathways affected. (D, E)
   Classification of potentially druggable genes in subtype A (D) and B
   (E). Each classification includes the top five genes in parentheses,
   and the horizontal axis indicates the number of genes within each
   druggable gene classification. BRCA, breast cancer.

   Finally, based on mutation data and the Drug–Gene Interaction database
   (DGIdb), we explored the gene druggability and drug–gene interactions
   in patients from different subtypes. As shown in [172]Figures 11D, E ,
   in subtype A (Cluster 1), the predicted drugs were potentially targeted
   at CLINICALLY ACTIONABLE genes including ARID1A, CDH1, GATA3, KMT2C,
   and MAP3K1; In subtype B (Cluster 2), the predicted drugs were mainly
   associated with the DRUGGABLE GENOME, involving CDH1, DMD, DST, HMCN1,
   and MAP3K1.

3.11. Immune analysis of BRCA subtypes

   The differences in immune cell infiltration were examined across BRCA
   subtypes. Using the CIBERSORT algorithm, the correlation between 22
   immune cell types was assessed within various subtypes and a bar chart
   illustrating the proportion of immune cells in each BRCA subtype was
   generated ([173] Figure 12A ). The results indicated that the
   enrichment scores for all 22 immune cell types were greater than zero.
   A grouped comparison chart ([174] Figure 12B ) was used to highlight
   the differences in immune cell infiltration abundance among BRCA
   subtypes. Thirteen immune cell types displayed statistically
   significant differences in expression levels across subtypes (p <
   0.05), including naïve B cells, plasma cells, CD8+ T cells, naïve CD4+
   T cells, resting memory CD4+ T cells, follicular helper T cells, Tregs,
   monocytes, M1 macrophages, M2 macrophages, resting dendritic cells,
   activated mast cells, and neutrophils. [175]Figure 12C presents a
   heatmap of the infiltration abundance of various immune cells across
   different BRCA subtypes.

Figure 12.

   [176]Panel A shows a stacked bar chart of immune cell composition
   across two clusters with a legend indicating cell types. Panel B
   displays box plots comparing immune infiltration scores between Cluster
   1 and Cluster 2 for various cell types. Panel C provides a heatmap and
   dendrogram illustrating the clustering of immune cell annotations.
   Panels D to G contain scatter plots showing correlations between
   specific genes and immune cells, with correlation coefficients and
   p-values indicated.
   [177]Open in a new tab

   Immune analysis by CIBERSORT algorithm. (A) Bar chart showing the
   proportion of immune cells in different BRCA subtypes. (B) Comparison
   of immune cell infiltration abundance across BRCA subtypes. (C) Heatmap
   illustrating the abundance of various immune cell infiltrations in
   different BRCA subtypes. (D–G) Scatter plots displaying the highest
   positive and negative correlations between Hub-UVR.Sig and immune cell
   infiltrations. BRCA, breast cancer; UVR.Sig, ultraviolet response
   signature. *p<0.05, **p < 0.01, ***p < 0.001, ns: not significant.

   Finally, the correlation between Hub-UVR.Sig and various immune cell
   infiltrations in BRCA samples were analyzed, focusing on the two pairs
   with the highest positive and negative correlations ([178]
   Figures 12D–G ). The results indicated a positive correlation between
   ICAM1 and M1 macrophage infiltration scores (r = 0.36; [179]Figure 12D
   ). Additionally, BTG1 was positively correlated with resting memory
   CD4+ T cell infiltration scores (r = 0.24) ([180] Figure 12E ).
   Conversely, ICAM1 exhibited a negative correlation with plasma cell
   infiltration scores (r = -0.24; [181]Figure 12F ), while BTG3
   demonstrated a negative correlation with resting mast cell infiltration
   scores (r = -0.23; [182]Figure 12G ).

3.12. Expression and function of Hub-UVR.Sig in BRCA

   We examined the mRNA expression levels of Hub-UVR.Sig genes in tumor
   tissues and paired adjacent normal tissues from six breast cancer
   patients. The results demonstrated that ENO2 and ATP6V1F were
   significantly upregulated in tumor tissues ([183] Figure 13A ). Further
   analysis of TCGA database revealed that ENO2 expression was markedly
   higher in breast cancer tissues compared to normal breast tissues
   ([184] Figure 13B ). Kaplan-Meier Plotter database ([185]69) indicated
   that high ENO2 expression was significantly associated with worse OS in
   breast cancer patients (HR = 1.27, 95% CI = 1.05-1.55, [186]Figure 13C
   ).

Figure 13.

   [187]Charts and graphs depict various experimental results related to
   ENO2 expression in breast cancer. Panels include bar graphs comparing
   expression levels, survival probability, transwell assays, wound
   healing assays, and scatter plots analyzing infiltration levels. Panel
   A shows expression in normal versus tumor tissues. Panel B involves box
   plots, and Panel C is a survival curve. Panel D and E depict bar and
   line graphs on cell proliferation. Panel F shows images from a
   transwell assay. Panel G presents wound healing results. Panel H
   features scatter plots for immune cell infiltration. Panel I shows dot
   plots comparing ENO2 expression with treatment controls.
   [188]Open in a new tab

   Expression and function of Hub-UVR.Sig in BRCA. (A) Hub-UVR.Sig mRNA
   expression levels in cancer tissues and corresponding paracancerous
   tissues of six BRCA patients. (B) Expression of ENO2 in BRCA tissues
   and normal tissues. (C) ENO2 was associated with worse OS in BRCA
   patient. (D) Knockdown of ENO2 in TNBC cells. (E) Knockdown of ENO2
   reduced proliferation of TNBC cells. (F, G) Knockdown of ENO2 decreased
   migration (F) and invasion (G) of TNBC cells. (H) Relationship between
   ENO2 expression and immune infiltrating cells in BRCA. (I) Expression
   of ENO2 increased in mice after treatment with ICIs. BRCA, breast
   cancer; TNBC, triple-negative breast cancer; UVR.Sig, ultraviolet
   response signature; *p<0.05, **p < 0.01 ***p < 0.001 and ****p<0.0001.
   Error bars indicate SD. All results were representative of or combined
   from at least three independent experiments. ns: not significant.

   To investigate the biological functions of ENO2, we conducted
   experiments using TNBC cell lines MDA-MB-231 and BT549. RT-qPCR
   confirmed that ENO2-specific siRNA effectively knocked down its
   expression ([189] Figure 13D ). CCK8 assays demonstrated that silencing
   ENO2 reduced the proliferative capacity of tumor cells ([190]
   Figure 13E ). Furthermore, ENO2 knockdown inhibited both the migration
   and invasion abilities of the cells ([191] Figures 13F, G ). Immune
   infiltration analysis based on the Timer database ([192]70)showed that
   ENO2 expression levels were negatively correlated with the infiltration
   of CD8+ T cells, naive B cells, activated NK cells, and M1 macrophages
   ([193] Figure 13H ). Analysis of two GEO datasets ([194]GSE268752 and
   [195]GSE229422) demonstrated that ENO2 expression was significantly
   elevated in ICI-treated mice with BRCA compared to control groups
   ([196] Figure 13I ).

4. Discussion

   In this study, we employed the GSVA method to evaluate UVR enrichment
   scores in malignant cells and found a negative correlation between UVR
   scores and ICI responses in two scRNA-seq cohorts (SKCM and BCC). Based
   on these findings, we hypothesized that this negative correlation may
   be a common phenomenon across various cancer types. To test this
   hypothesis, we conducted a large-scale comprehensive analysis,
   identifying a set of genes, termed UVR.Sig, that were highly expressed
   in malignant cells and significantly positively correlated with UVR
   scores in 34 scRNA-seq datasets. Rigorous validation of UVR.Sig
   demonstrated that it outperformed multiple previously used models in
   predicting ICI responses across bulk RNA-seq independent immunotherapy
   cohorts. Moreover, the Hub-UVR.Sig, further refined through multiple
   machine learning methods, showed strong prognostic value for
   ICI-treated patients, especially in BRCA patients. Based on
   Hub-UVR.Sig, two distinct BRCA subtypes were identified, each
   exhibiting unique molecular mutation profiles and immune
   characteristics, providing a critical foundation for the development of
   personalized treatment strategies. Finally, by analyzing clinical
   tissue samples from BRCA patients, we found that ENO2 and ATP6V1F, two
   genes from the Hub-UVR.Sig, were highly expressed in tumor tissues.
   High ENO2 expression was associated with worse OS in BRCA patients and
   negatively correlated with the infiltration of cytotoxic immune cells.
   Moreover, knockdown of ENO2 significantly suppressed the proliferation,
   migration, and invasion of TNBC cells.

   UV light has been shown to alter the expression of cytokines,
   chemokines, and cell surface receptors in cells ([197]71–[198]73),
   thereby modulating various interactions between tumors and the immune
   system. These interactions may result in either immune activation or
   immune suppression. For example, UV light can stimulate BRCA cells to
   secrete pro-inflammatory chemokines, leading to the recruitment of
   antitumor effector T cells ([199]73). In a CT26 colon cancer mouse
   model, a single high-dose radiation exposure induced a durable complete
   remission mediated by CD8^+ T cell infiltration ([200]74). Conversely,
   studies have also shown that UV exposure can cause a variety of DNA
   lesions—including strand breaks, base damage, and cross-linking—which
   subsequently induce apoptosis in radiation-sensitive tissues such as
   lymphocytes and result in systemic immunosuppression ([201]75,
   [202]76). Furthermore, tumor-associated macrophages exposed to
   radiation express higher levels of iNOS, arginase-I, and COX-2, thereby
   promoting tumor growth ([203]77). Therefore, UVR may affect the immune
   microenvironment in a complex manner, and elucidating the molecular
   mechanisms by which UV light influences tumor biology is of critical
   importance.

   Our study showed that UVR.Sig, consisting of 38 genes, was mainly
   enriched in the TNF signaling pathway. In the TME, activated immune
   cells, such as macrophages and T lymphocytes, fibroblasts, and cancer
   cells secrete large amounts of TNF-α. This cytokine accumulates within
   the tumor, triggering and maintaining inflammatory responses that
   promote tumor growth and progression ([204]78, [205]79). The TNF-α
   receptor, TNFR2, is highly expressed in several tumor types, including
   BRCA, cutaneous T-cell lymphoma, and colorectal cancer
   ([206]80–[207]82). Overexpression of TNFR2 not only enhances tumor cell
   resistance to apoptosis, but also inhibits anti-tumor immune responses,
   thereby facilitating immune evasion ([208]83). Further analysis
   revealed that UVR.Sig was positively correlated with pathways, such as
   EMT, hypoxia, IL-6-JAK-STAT3, and inflammatory responses. Hypoxia is a
   hallmark of the TME ([209]84), that suppresses the activity of T and NK
   cells, enhances Treg function, and upregulates the expression of immune
   checkpoint molecules, such as PD-L1 ([210]85, [211]86). Additionally,
   STAT3 activation not only upregulates PD-L1 expression ([212]87), but
   is also associated with the expansion and functional enhancement of
   various immunosuppressive cells, which secrete factors, such as IL-10,
   that inhibit effector T cell activity, thereby weakening anti-tumor
   immune responses ([213]88, [214]89). Collectively, the association
   between UVR.Sig and these pathways suggested its potential role in
   promoting immunosuppression, immune evasion, and the deterioration of
   the TME.

   UVR.Sig was negatively correlated with CD8+ T cells, NK cells, CTLs,
   and B cells in almost all tumor types, while showing a positive
   correlation with neutrophils, endothelial cells, and fibroblasts. As
   key CTLs, CD8+ T and NK cells play critical roles in the immune system
   by directly killing infected or cancerous cells, thus protecting the
   body from pathogens and tumors ([215]90–[216]94). Similarly, B cells in
   the TME recognize tumor-specific antigens and produce antibodies that
   neutralize tumor cells, thereby preventing their growth and spread of
   tumor cells, which is closely related to improving the effectiveness of
   immunotherapy ([217]95, [218]96). In contrast, a high infiltration of
   neutrophils, endothelial cells, and fibroblasts in the TME typically
   indicates immunosuppression ([219]97–[220]102). Tumors with high
   UVR.Sig exhibited pronounced immunosuppressive characteristics, making
   them less responsive to ICI therapy. This highlights the potential of
   UVR.Sig as a valuable predictive marker of immunotherapy response.

   Using the optimal Cancerclass algorithm, UVR.Sig was identified as a
   novel signature that effectively predicted ICI response across various
   cancer types, including RCC, UC, SKCM, GC, and GBM. To validate its
   predictive power, UVR.Sig was systematically compared with six widely
   used pan-cancer signatures. The results demonstrated that UVR.Sig
   exhibited superior performance (AUC = 0.727) and consistently
   outperformed other pan-cancer signatures across multiple cancer types
   and independent cohorts, likely due to its stronger generalizability.
   In addition, compared with other molecular markers, UVR.Sig may provide
   a more comprehensive reflection of the overall immune status of the
   tumor microenvironment. In contrast, PD-L1 primarily reflects surface
   molecule expression, TMB indicates mutation burden, and MSI reflects
   genomic stability ([221]103); these markers do not fully capture the
   complexity of immune cells and signaling pathways within the
   microenvironment. Therefore, UVR.Sig could represent a complementary
   tool with potential clinical value in assessing the immunological
   landscape.

   Given the outstanding predictive ability of UVR.Sig for immunotherapy
   outcomes, the CRISPR dataset was used to identify potential drug
   targets. This strategy not only revealed novel therapeutic targets but
   also supported personalized medicine, enabling more precise treatments
   to enhance efficacy. Based on the correlation between genes and immune
   response, the genes were ranked and UVR.Sig genes closely associated
   with immune resistance, including DNAJA1, STIP1, JUNB, EPCAM, OLFM1,
   and NR4A1, were identified. DNAJA1, a member of the heat shock protein
   40 (Hsp40) family, plays a critical role in regulating B cell function
   by enhancing the expression and activity of activation-induced cytidine
   deaminase in mice ([222]104). Additionally, DNAJA1 prevents proteasomal
   degradation of unfolded mutant p53, thereby promoting tumor metastasis
   ([223]105). JunB, a member of the activator protein 1 (AP-1)
   transcription factor family, regulates Treg differentiation, and
   promotes CD25 expression and IL-2 production ([224]106, [225]107). JunB
   plays a pivotal role in immunosuppression and may be a critical factor
   in predicting adverse reactions to immunotherapy ([226]108). EpCAM, a
   marker of circulating tumor and cancer stem cells, is expressed in
   various cancer types ([227]109, [228]110). It inhibits the activity of
   CD8+ T cells and upregulates PD-L1 expression, making it a potential
   immunotherapy target in cancers such as BRCA, COAD, and oral squamous
   cell carcinoma ([229]111–[230]113). In conclusion, these genes
   represent potential therapeutic targets across various cancer types,
   and further investigation of these core UVR.Sig genes will contribute
   to the development of more effective combination strategies for
   immunotherapy.

   To enhance the prognostic efficacy of UVR.Sig, ML algorithms were
   employed to identify seven hub genes, termed as Hub-UVR.Sig, including
   ATF3, ATP6V1F, BTG1, BTG3, ENO2, FOS, and ICAM1. FOS and JUNB can be
   directly regulated by p53—p53 binds to the response elements in the
   promoter region of FOS and promotes its expression. FOS then forms a
   heterodimer with JUNB, which activates the transcription of downstream
   immunosuppression-related genes ([231]114). This mechanism may, at
   least in part, explain the immunosuppressive effect of Hub-UVR.Sig. We
   observed that Hub-UVR.Sig was positively correlated with activated mast
   cells and negatively correlated with resting mast cells. Although mast
   cells are traditionally linked to allergic responses, recent studies
   have shown that activated mast cells play a critical role in tumor
   progression and are often associated with poor prognosis
   ([232]115–[233]117). The mechanisms involved include immune
   suppression, angiogenesis promotion, and extracellular matrix
   degradation ([234]118, [235]119). This suggests that Hub-UVR.Sig may
   also be involved in the regulation of inflammatory and immune responses
   by affecting mast cell activation. Risk scores generated based on
   UVR.Sig are effective in identifying patients with BRCA, and patients
   with higher risk scores typically exhibit worse OS. This led us to
   shift our focus to BRCA to further explore the potential value of
   Hub-UVR.Sig in this context.

   Using consensus clustering analysis, two BRCA subtypes that exhibited
   significant differences in their molecular mutation characteristics and
   immune infiltration profiles were identified. Subtype A (Cluster 1) had
   a high mutation frequency in TP53 (37%), whereas Subtype B (Cluster 2)
   had a high mutation frequency in PIK3CA (40%). TP53 mutations are the
   most common in BRCA, occurring in 30–35% of all BRCA cases and
   approximately 80% of triple-negative breast cancer cases ([236]120,
   [237]121). Strong evidence has linked TP53 mutations to poor
   disease-free survival and OS in BRCA ([238]122). PIK3CA mutations are
   present in 25–46% of BRCA cases and are associated with chemotherapy
   resistance, poor prognosis, and reduced OS ([239]123). Additionally,
   Subtype A has a high TIDE score and low TMB score, suggesting a strong
   immune escape potential and low immunogenicity, potentially leading to
   a poor response to immunotherapy ([240]65, [241]124). It should be
   noted that the consensus clustering in this study was performed at the
   molecular level and does not fully correspond to pathological BRCA
   subtypes. Cluster 1 and Cluster 2 were identified through a systematic
   molecular-level exploration of Hub-UVR.Sig using consensus clustering,
   whereas pathological classification is mainly based on tumor morphology
   and conventional biomarkers. Our study therefore represents an initial
   exploration, and future work will aim to integrate molecular and
   pathological features to achieve a more comprehensive breast cancer
   classification.

   Finally, we validated the expression of the Hub-UVR.Sig genes in BRCA
   patients and found that ENO2 and ATP6V1F were highly expressed in tumor
   tissues. ENO2, a glycolytic enzyme, has been reported to promote
   stem-like properties, tumorigenesis, and metastatic progression in BRCA
   cells by activating the glycolytic pathway ([242]125, [243]126).
   Consistently, knockdown of ENO2 in TNBC cells resulted in a significant
   reduction in tumor cell proliferation, migration, and invasion.
   Moreover, high ENO2 expression was negatively correlated with the
   infiltration of cytotoxic immune cells and was upregulated following
   anti-PD-L1 treatment, suggesting that ENO2 may contribute to adaptive
   ICI resistance.

   This study has several limitations. First, the currently available
   immunotherapy cohorts (GC, SKCM, RCC, UC, and GBM) offer limited tumor
   type coverage, which may affect the generalizability of our findings.
   Further validation of UVR.Sig in other cancer types lacking
   immunotherapy data is warranted, and its broader application across
   pan-cancer contexts requires support from additional clinical cohorts
   to ensure robustness and universality. Nevertheless, our immune
   correlation analysis of UVR.Sig across 30 cancer types in the TCGA
   partially compensates for this limitation. Second, our study primarily
   focused on the association between gene expression and response to
   ICIs, while other key prognostic factors—such as genomic mutations, DNA
   methylation, histone modifications, and non-coding RNAs—were not
   considered. The heterogeneity among datasets may also introduce batch
   effects. Given the robust predictive performance of Hub-UVR.Sig for
   ICIs response and its prognostic value in BRCA, future studies should
   prioritize validation using real-world clinical data from BRCA
   patients, including the prospective collection of clinicopathological
   information and integration of additional variables that may influence
   tumor prognosis. We are also aware of the importance of in vitro
   experiments. In the future, we will add functional data verification of
   Hub-UVR.Sig genes such as ATP6V1F, ICAM1 and BTG family members to
   further support their biological and clinical relevance and improve the
   completeness and credibility of this article.

5. Conclusion

   To our knowledge, this is the first study to reveal a strong
   association between UVR mechanisms and ICIs resistance in cancer.
   Through pan-cancer single-cell transcriptomic analysis, we developed a
   UVR-related gene signature (UVR.Sig) that outperformed existing
   biomarkers in predicting ICIs response and showed significant
   prognostic value in breast cancer. While our findings offer a promising
   tool for refining immunotherapy patient selection, further validation
   in additional tumor types and incorporation of other prognostic factors
   are needed to strengthen its clinical applicability.

Acknowledgments