Research ArticleDevelopment of a Ferroptosis-Related lncRNA Signature toPredict the Prognosis and Immune Landscape of Bladder Cancer

Ranran Zhou,1,2 Jingjing Liang,3 Hu Tian,1,2 Qi Chen,1,2 Cheng Yang,1,2

and Cundong Liu 1,2

1Department of Urology, The Third Affiliated Hospital of Southern Medical University, Guangzhou 510000, China2The Third School of Clinical Medicine, Southern Medical University, Guangzhou 510000, China3Department of Cardiology, Shunde Hospital of Southern Medical University, Foshan 528000, China

Correspondence should be addressed to Cundong Liu; cundongliu@163.com

Received 15 April 2021; Revised 30 May 2021; Accepted 7 June 2021; Published 21 June 2021

Academic Editor: Hiroshi Miyamoto

Copyright © 2021 Ranran Zhou et al. This is an open access article distributed under the Creative Commons Attribution License,which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

The tight relationship between ferroptotic cell death and immune response demonstrated by recent studies enlightened us to detectthe underlying roles of ferroptosis-related long noncoding RNAs (frlncRNAs) in the tumor microenvironment of bladder cancer(BCa). We collected 121 ferroptosis regulators from previous studies. Based on their expression values, 408 cases with BCa wereclustered. The patients in different clusters showed diverse immune infiltration, immunotherapy response, and chemotherapyeffectiveness, revalidating the tight correlation with ferroptosis and tumor immunity. Through differential, coexpression,Kaplan-Meier, Lasso, and Cox analysis, we developed a 22-lncRNA-pair signature to predict the prognosis of BCa based ongene-pair strategy, where there is no need for definite expression values. The areas under the curves are all over 0.8. The riskmodel also helped to predict immune infiltration, immunotherapeutic outcomes, and chemotherapy sensitivity. Totally, theprognostic assessment model indicated a promising predictive value, also providing clues for the interaction between ferroptosisand BCa immunity.

1. Introduction

Bladder cancer (BCa) is one of the most common malignan-cies worldwide, with high morbidity and mortality [1]. Morethan 500,000 new BCa cases and 200,000 BCa-related deathsoccur worldwide annually [2]. BCa has two main subtypes:muscle-invasive BCa and nonmuscle-invasive BCa. Althoughthe 5-year survival rate of nonmuscle-invasive BCa is about90%, approximately 15–20% of such cases would progressto the muscle-invasive stage and even to distant metastasis,which has a dismal 5-year survival rate of 5–30% [1]. Recently,some large-scale clinical trials, such as KEYNOTE-045 andIMvigor211, demonstrated that BCa is susceptible to immunecheckpoint inhibitors (ICIs), representing an importantadvancement in the treatment of BCa [3, 4].

Ferroptosis, a form of iron-dependent and nonapoptoticcell death, is attracting increasing attention in view of the factthat apoptosis resistance is one of the hallmarks of tumors

[5]. Inducing tumor cell ferroptosis seems to be an attractiveand promising therapeutic strategy, especially for drug-resistant malignancies. Recent studies have verified that ferrop-totic tumor cells and the tumor microenvironment (TME)could influence each other [6, 7], suggesting that the combina-tion of ICIs and ferroptosis inducers is a promising treatment.

Long noncoding RNAs (lncRNAs), a type of RNA mole-cule with transcripts of >200 nucleotides, participate intumorigenesis and cancer development not only by alteringthe malignancy of cancer cells themselves but also by chang-ing the TME, as has been reported in many studies [8]. Inrecent years, the interaction between lncRNAs and ferropto-sis has also been investigated. For instance, the lncRNAP53RRA serves as a tumor suppressor by promoting p53maintenance in the nucleus, thus, facilitating ferroptosis [9].

Immune-related signatures for predicting the status ofcancer immune infiltration are receiving increasing attentionwith the clinical application of ICIs, as such signatures might

HindawiDisease MarkersVolume 2021, Article ID 1031906, 22 pageshttps://doi.org/10.1155/2021/1031906

help predict prognosis and immunotherapy outcomes. BecauselncRNAs are involved in the regulation of 70% of gene expres-sion, signatures based on lncRNA expression have been inves-tigated by many researchers. Meanwhile, some researchershave proposed the gene-pair strategy, which focuses on thedifference between two genes regardless of the exact geneexpression, to construct a prediction model, making immune-related signatures more widely applicable [10].

This study had three phases. First, unsupervised cluster-ing based on known ferroptosis regulators was performedto determine the correlation between ferroptosis and theTME in BCa. Second, using the gene-pair strategy, we con-structed a ferroptosis-related lncRNA (frlncRNA) signatureto predict overall survival (OS). Lastly, the diagnostic abilityof the risk score calculated by the model for immunotherapyresponse, chemotherapy results, and immune infiltration wasalso evaluated. From this analysis, we aimed to verify theclose association between ferroptosis and the TME and topropose an important tool for predicting the prognosis andimmune infiltration of BCa.

2. Materials and Methods

2.1. Transcriptome Data Collection and Ferroptosis-RelatedGenes. The transcription data (RNA-seq) and correspondingclinicopathological features were downloaded from The Can-cer Genome Atlas (TCGA) up to September 3, 2020 (https://portal.gdc.cancer.gov/). A total of 121 ferroptosis regulatorswere identified from a previous study [5]. To select lncRNAs,the annotation file was obtained from the Ensembl database(http://asia.ensembl.org). lncRNAs with an average expres-sion value of <0.5 were excluded from the study. Differentialanalysis was performed using the edgeR package [11], andfalse discovery rate ðFDRÞ < 0:05 and ∣logarithmic foldchange ½logFC� ∣ >1 and 2 were set as the thresholds forferroptosis regulators and frlncRNAs, respectively. ThefrlncRNAs were screened through coexpression analysis,with Pearson correlation coefficients >0.5 and p > 0:05 asfiltering criteria.

2.2. Unsupervised Clustering. The mRNA expression profileof 121 ferroptosis regulators was adopted for consensus clus-tering with the ConsensusClusterPlus R package [12] usingthe K-means method. The cumulative distribution functionwas implemented to detect the optimal K value, which repre-sented the clustering number, and the result was validatedusing principal component analysis (PCA) in R software.

2.3. Gene-Pair Strategy. The gene-pair strategy was used for0-or-1 matrix construction, as previously reported [10]. Forinstance, if the expression value of the A gene is higher thanthat of the B gene, then C, which means the gene pair com-prising A and B, is considered as 1; otherwise, C is consideredas 0. All frlncRNAs were paired, and a new 0-or-1 matrix wasconstructed. An frlncRNA pair was excluded from the studyif its value in all samples was 0 or 1.

2.4. Survival Analysis. Cases with a follow-up duration of <30days were not included in the present analysis. We imple-mented the survival R package to conduct Kaplan–Meier

(KM) survival analysis with the log-rank test for the compar-ison of OS between different cohorts. To simplify the modeland avoid overfitting as much as possible, Lasso–Cox regres-sion with 10-fold cross-validation was conducted using theglmnet package, and the threshold was set at 0.05. Univariateand multivariate Cox analyses were performed using the sur-vival package. The area under the curve (AUC) values werecalculated using the survival ROC package to estimate theeffectiveness of the established model.

2.5. Association with Clinicopathological Factors. To detectwhether the risk score, which was calculated using theestablished model, was an independent risk factor, univariateand multivariate Cox regression analyses were performedusing the survival R package. The clinicopathological featureswere all transformed into dichotomized variables, includingrisk score (high vs. low), tumor grade (high vs. low), age(≤64 vs. >64 years), sex (female vs. male), pathological Tstage (T 1–2 vs. T 3–4), N stage (N0 vs. N1–3), and M stage(M0 vs. M1).

2.6. Immune Infiltration Evaluation. Different methods wereused to evaluate the immune infiltration status. The ESTI-MATE algorithm [13] through the estimate package of Rwas used to calculate the ratio of immune and stromal com-ponents of the TME, which were quantified as ImmuneScoreand StromalScore, respectively. CIBERSORT [14] andTIMER [15] were used to evaluate the abundance profile ofimmune cells in the TME with p < 0:05 filtering.

2.7. Single-Sample GSEA (ssGSEA). The ssGSEA method wasimplemented using the Gene Set Variation Analysis (GSVA)and GSEABase packages to analyze the immune and inflam-matory infiltration profiles. The results were visualized usingthe pheatmap package.

2.8. Enrichment Analysis. Functional enrichment analysis oflncRNAs was conducted using LncSEA (http://www.licpathway.net/LncSEA/) with FDR < 0:05 and Bonferroni < 0:05. GSEAsoftware (version 4.1.0) was downloaded from the GSEAwebsite (https://www.gsea-msigdb.org/), and GSVA was per-formed using the GSVA R package. Hallmark gene sets v7.2and iron metabolism-related gene sets were collected fromthe Molecular Signatures Database (https://www.gsea-msigdb.org/gsea/msigdb/). For GSVA, ∣logFC ∣ >0:03 and adjustedp < 0:05 were considered as thresholds.

2.9. Transcription Factor Prediction. The JASPAR2020 [16]and TFBSTools [17] packages were used to predict transcrip-tion factors with sequences of 500 bp downstream to 2000 bpupstream from the transcription start point of lncRNAs.When the score, which was positively correlated with thepossibility of transcription binding, was >0.99, the transcrip-tion factors were included in further analyses.

2.10. Estimating the Effectiveness of frlncRNA Signature forClinical Treatment. The likelihood of an immunotherapyresponse was estimated using the TIDE algorithm [18]. Inaddition, the association between the risk score and someknown predictors of immunotherapeutic sensitivity, such as

2 Disease Markers

CXCL9 and CXCL13 expression, was also investigated [19].The half inhibitory concentration (IC50) values of com-monly used chemotherapeutic drugs, such as cisplatin, doxo-rubicin, gemcitabine, methotrexate, and vinblastine, wereobtained using the pRRophetic R package in patients withstage ≥ T2 BCa [20]. The survival information of patientswho had received chemotherapy and had stage ≥ T2 BCawas downloaded from TCGA.

3. Results

3.1. The Tight Association between Ferroptosis and BCaImmune Infiltration. From a recent review [5], 121 ferropto-sis regulators were identified, and the mRNA expression ofeach gene was downloaded from the TCGA BCa dataset,which included 19 paracarcinoma tissues and 408 BCa sam-ples. The edgeR package of R software (version 4.0.3) wasused to detect the differentially expressed genes, and 20ferroptosis regulatory factors were screened (Figure 1(a),Supplementary Table S1; ∣logFC ∣ >1, FDR < 0:05). Throughthe consensus clustering algorithm, by which optimalgrouping stability could be achieved, all enrolled cases wereassigned into two groups: cluster 1 and cluster 2(Figures 1(b)–1(d)). PCA indicated that our grouping methodwas reliable (Figure 1(e)). Notably, patients in cluster 1 had asignificantly poorer prognosis than those in cluster 2according to KM survival analysis (Figure 1(f)), p < 0:05).Mazdak et al. found that the serum iron ion level in 51patients with BCa was significantly lower than that in 58healthy controls, revealing the important role of iron intumor development [21]. Subsequent studies verified thatinhibition or activation of ferroptosis in BCa is a feasiblemeasure of treatment effectiveness [22]. Generally, ouranalysis results again demonstrated the significant impact offerroptosis on BCa from the perspective of big data analysis.

Some studies have reported that ferroptosis and tumor-infiltrating immune cells could influence each other [6, 7],prompting us to detect the changes in the immune landscapein different clusters. First, we estimated the infiltration ofimmune cells and the changes in immune-related pathwaysusing ssGSEA based on the reported gene sets [23].Accordingly, a different immune landscape was found in eachferroptosis cluster, which was verified by the ESTIMATE algo-rithm [13], a strategy used to predict the ratios of immune andstromal components of the TME (Figure 1(h)). GSEA showedthat cases in cluster 1 were significantly enriched in metabolicpathways (Figure 1(i)), whereas cases in cluster 2 were mostlyenriched in immune-related functions (Figure 1(j)). Asferroptosis-related clustering was immune-related, we detectedthe association between clustering and immune cell infiltrationusing the CIBERSORT algorithm [14]. We found that the dif-ferent immune cells had different infiltration proportionsbetween cluster 1 and cluster 2 according to the Wilcoxonsigned-rank test (Figure 1(g)). Generally, patients in cluster 1had a shorter survival time and relatively higher immuneinfiltration, indicating an immunosuppressive environment inthis cluster.

Considering that immune checkpoints such as PDL1,PD1, LAG3, CTLA4, TIM-3, and TIGIT could lead to an

immunosuppressive TME, we analyzed the associationbetween ferroptosis-relevant clustering and the expressionof immune checkpoints. We discovered that all of the above-mentioned genes were significantly upregulated in cluster 1(p < 0:001, Figure 1(k)). Immune checkpoints are also usuallyconsidered predictors of the immunotherapy response ofpatients with BCa, indicating that patients in cluster 1 weremore likely to benefit from ICIs, which was again validatedby the TIDE algorithm [18] (p < 0:001, Figure 1(l)). In addi-tion to immunotherapy, we also detected the difference inchemotherapy effectiveness between the two clusters usingthe pRRophetic R package [19]. Patients in cluster 1 showedsignificantly higher chemotherapy sensitivity (Figure 1(m)).In summary, this analysis revealed the potential offerroptosis-related genes to predict alterations in the immuneinfiltration of BCa.

3.2. Identification of Ferroptosis-Related lncRNAs (frlncRNAs).The transcriptome data of 408 BCa samples and 19 corre-sponding normal tissues from the TCGA website were ana-lyzed. With the annotation file downloaded from Ensembl,all lncRNAs were selected for expression differential analysisusing the edgeR package, and 927 lncRNAs were consideredas differentially expressed genes with filtering criteria of FDR< 0:05 and ∣logFC ∣ >2 (Figure 2(a)). Meanwhile, we calcu-lated the Pearson correlation coefficients between eachlncRNA and ferroptosis regulator, and the absolute values ofPearson R > 0:5 and p < 0:05 were determined to be signifi-cant (Supplementary Table S2). Ultimately, 105 lncRNAscodetermined by differential and coexpression analyses wereidentified as frlncRNAs (Figure 2(b)). Most of these 105frlncRNAs were long intervening noncoding RNAs andantisense lncRNAs (Figure 2(c)).

3.3. Construction of frlncRNA Pairs and Prognostic Signature.Patients with a follow-up period of <30 days were excludedfrom the analysis. On the basis of the gene-pair strategy,4298 frlncRNA pairs were identified. After KM survival anal-ysis with a log-rank test, 125 gene pairs were extracted withp < 0:01 filtering (Supplementary Table S3), 35 of whichwere also determined by Lasso regression (Figures 2(d)–2(f)). Ultimately, 22 frlncRNA pairs were included in theprognostic model through multivariate Cox regression witha stepwise method (Figure 2(g)). Figure 2(h) shows thefrlncRNAs in the risk model and their possible targets.

3.4. Validation of the 22-frlncRNA-Pair Prognostic Model.Various methods were implemented to determine the prog-nostic value of the established model. First, we generatedthe 1-, 3-, and 5-year receiver operating curves (ROCs) ofthe gene-pair signature and found that the AUCs were all>0.8 (Figure 3(a)). After excluding cases without definiteTNM stages, we compared the AUCs for the 5-year ROCcurves of different clinicopathological parameters and riskscores (Figure 3(b)). According to the median value of the riskscore, 197 cases were categorized into the high-risk group and199 cases were classified as the low-risk group. The high-risksubgroup had a significantly poorer clinical outcome thanthe low-risk subgroup (p < 0:001, Figure 3(c)). Higher risk

3Disease Markers

ZEB1

ATF3

FANCD2

BIDALOX12

ALOX15B

SQLE CDKN2ABBC3HILPDA

SCDCDO1 CAV1

TFRC

CSF2CBSITGB4MUC1

ALOXE3HMOX10

10

20

30

−2 0 2 4logFC

−log

10FD

R

GroupDownNotsignificantUp

(a)

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

Consensus CDF

Consensus index

CDF

2345

6789

(b)

2 3 4 5 6 7 8 9

0.1

0.2

0.3

0.4

0.5

Delta area

k

Relat

ive c

hang

e in

area

und

er C

DF

curv

e

(c)

Consensus matrix k = 2

12

(d)

Figure 1: Continued.

4 Disease Markers

−100

−50

0

50

100

−100 0 100PCA1

PCA

2

GroupCluster1Cluster2

(e)

0 1000 2000 3000 4000 5000

0.0

0.2

0.4

0.6

0.8

1.0

Survival curve (p = 4.993e−02)

Time (day)

Ove

rall

surv

ival

rate

Cluster 1Cluster 2

(f)

B ce

lls n

aive

B ce

lls m

emor

y

Plas

ma c

ells

T ce

lls C

D8

T ce

lls C

D4

naiv

e

T ce

lls C

D4

mem

ory

resti

ng

T ce

lls C

D4

mem

ory

activ

ated

T ce

lls fo

llicu

lar h

elper

T ce

lls re

gulat

ory

(treg

s)

T ce

lls g

amm

a delt

a

NK

cells

resti

ng

NK

cells

activ

ated

Mon

ocyt

es

Mac

roph

ages

M0

Mac

roph

ages

M1

Mac

roph

ages

M2

Den

driti

c cell

s res

ting

Den

driti

c cel

ls ac

tivat

ed

Mas

t cel

ls re

sting

Mas

t cel

ls ac

tivat

ed

Eosin

ophi

ls

Neu

troph

ils

ClusterCluster_1Cluster_2

⁎⁎⁎ ⁎⁎ ⁎ ⁎⁎⁎ ⁎⁎⁎ ⁎ ⁎⁎ ⁎ ⁎⁎⁎ ⁎⁎⁎ ⁎⁎ ⁎⁎⁎

0.0

0.1

0.2

0.3

0.4

Gen

e exp

ress

ion

(g)

Type_II_IFN_reponseMast_cellsB_cellsNeutrophilsNK_cellsTh2_cellsiDCsDCsAPC_co_stimulationHLAT_helper_cellsMacrophagesCCRTregMHC_class_IAPC_co_inhibitionParainflammationType_I_IFN_reponseaDCsCD8+_T_cellsCytolytic_activityInflammation−promotingTh1_cellsCheck−pointT_cell_co−inhibitionTfhpDCsT_cell_co−stimulationTIL

Tumor purity0.9

0.3

Estimate score4000

−4000

Immune score3000

−1000

Stromal score2000

−2000

SubtypeCluster_1Cluster_2

−4

−2

0

2

4

Tumor purity⁎⁎⁎Estimate score⁎⁎⁎Immune score⁎⁎⁎Stromal score⁎⁎⁎Subtype

(h)

Figure 1: Continued.

5Disease Markers

Hallmark_oxidative_phosphorylation

Hallmark_bile_acid_metabolism

Hallmark_peroxisome

Hallmark_fatty_acid_metabolism

Hallmark_adipogenesis

Hallmark_xenobiotic_metabolism

−1.65 −1.60 −1.55 −1.50 −1.45NES

Term

0.0100

0.0125

0.0150

0.0175

FDR_Q value

Size125150175

(i)

Hallmark_UV_response_up

Hallmark_TGF_beta_signaling

Hallmark_unfolded_protein_response

Hallmark_UV_response_DN

Hallmark_MTORC1_signaling

Hallmark_angiogenesis

Hallmark_hypoxia

Hallmark_myogenesis

Hallmark_MYC_targets_V1

Hallmark_apoptosis

Hallmark_mitotic_spindle

Hallmark_coagulation

Hallmark_apical_junction

Hallmark_IL2_STAT5_signaling

Hallmark_KRAS_signaling_up

Hallmark_interferon_alpha_response

Hallmark_TNFA_signaling_VIA_NFKB

Hallmark_IL6_JAK_STAT3_signaling

Hallmark_G2M_checkpoint

Hallmark_complement

Hallmark_E2F_targets

Hallmark_interferon_gamma_response

Hallmark_inflammatory_response

Hallmark_allograft_rejection

Hallmark_epithelial_mesenchymal_transition

1.5 2.0 2.5 3.0NES

Term

Size40

80

120

160

200

0.00

0.01

0.02

0.03

0.04

FDR_Q value

(j)

Figure 1: Continued.

6 Disease Markers

PDL1⁎⁎⁎ PD1⁎⁎⁎ LAG3⁎⁎⁎ CTLA4⁎⁎⁎ TIM−3⁎⁎⁎ TIGIT⁎⁎⁎

Clus

ter_

1

Clus

ter_

2

Clus

ter_

1

Clus

ter_

2

Clus

ter_

1

Clus

ter_

2

Clus

ter_

1

Clus

ter_

2

Clus

ter_

1

Clus

ter_

2

Clus

ter_

1

Clus

ter_

2

0

2

4

Cluster

Gen

e exp

ress

ion

Cluster

Cluster_1

Cluster_2

(k)

46%

17%

54%

83%

p = 2.8e−22p = 0.278

(n = 217)(n = 191)0

10

20

30

40

50

60

70

80

90

100

Cluster_1 Cluster_2

Perc

enta

ge (%

)

Response

Response

Non_response

Cluster

𝜒2Pearson = 40.27, p = 2.22e−10

(l)

Cisplatin⁎⁎⁎ Doxorubicin⁎⁎⁎ Gemcitabine⁎⁎⁎ Methotrexate⁎ Vinblastine⁎⁎⁎

Clus

ter_

1

Clus

ter_

2

Clus

ter_

1

Clus

ter_

2

Clus

ter_

1

Clus

ter_

2

Clus

ter_

1

Clus

ter_

2

Clus

ter_

1

Clus

ter_

2

−8

−4

0

4

Cluster

IC50

Cluster

Cluster_1

Cluster_2

(m)

Figure 1: Association between ferroptosis and immune infiltration. (a) Among 121 ferroptosis regulators, 14 were upregulated and 6 weredownregulated in BCa samples. (b) Consensus CDF values changed from k = 2 to k = 9. (c) Relative changes in area under CDF curve withdifferent k values. (d) Consensus clustering heat map of 408 BCa cases when k = 2. (e) Scatter diagram displaying the PCA results of theexpression of all genes when k = 2. (f) Kaplan–Meier survival analysis with log-rank test of BCa cases in different clusters (p < 0:05). (g)Box plot indicating the difference in infiltration proportion of 22 different immune cells between the two clusters. (h) Heat map ofdifferent immune-related pathways and functions. (i) GSEA for patients in cluster 1. (j) GSEA for patients in cluster 2. (k) Expressionprofiles of common immune checkpoints between different clusters. (l) Difference in immunotherapy response between the two clusters.(m) Patients in different clusters showed different chemotherapeutic sensitivity. BCa: bladder cancer; CDF: cumulative distributionfunction; PCA: principal component analysis; GSEA: gene set enrichment analysis; ∗∗∗ p < 0:001.

7Disease Markers

0

20

40

60

80

−5 0 5logFC

−log

10FD

R

GroupDownNotsignificantUp

(a)

1652 822105

Correlation Difference

(b)

Ant

isens

eln

cRN

AO

ther

0 10 20 30 40 50Frequency

(c)

−6 −5 −4 −3 −20

100

200

300

400

Log (𝜆)

Part

ial l

ikel

ihoo

d de

vian

ce330 306 282 264 257 238 155 59 18

(d)

−6 −5 −4 −3 −2Log lambda

300 268 245 86 0

7

23

4142455460

61

8188

90

93109114

116

118143145167168181

187

189195202206207

208

217227228

231

237239

243

258272278285289295

299

311

326

327332335

339

345375387388390403422424431

435

450

458

459460462477

479

483492503521526527

530

533534

555562

573

578

582

591613616625645

651

670675691699724737738739750766770779799817819831838868877

879890

893897

901

902

923940947948

965

981

986

9921017102610351047

1061

1067

1072

1075

1101

1128114011461170117111931200120112061209

1211

1219122112611262128912941299

13041310

1369

13811393142614391457146114761514152215321533156215681570

1578

15861603163316421668

1669

169517021708170917361745174917531778

1780

1783178518081815183518401864

18721883

1886

1891

1915

1941

1954

1957

1959

198320072021

2037

20422046

2075

2078

2085

20882090

2108

2113215821652175

2196

2209

2228222922602262226922702271227722782283

2287

2292

23022308

2310

2323

23352341234523592369

2376

23992413

242724302444

2447246924812495

2497

250225032521

2523

252925432575

2577

2580

2583

2602260826102621

2640

2652

2654

2660

2668

26782680268326882692271027222725273627552757277227732775

2781

2786

2787

2803

2804

28082809281628172818282728332849285328652872287828802888

2894

2900

2901

2906

2908

2909291729222926

2934

29392957296529682976298529873015304930683070

308030843099

3102

3103

310631143134

3138

3147

3159

3164

3168

3173320632123215

3225

32273241

32433244

3252

3257

3307

33203322

3338

3347

3351

3360336533673403

3406

34113428

34333434

3436

3438

3447

3448

3456345734653477349134933498

3502

3504351235193525

3526

3530

35463557

3569

35793596

3599

360436103617362236293630363136343647

3657

36593681

3700

3701370337113722

3741

3742

3771377237793788

3792

3793

3795

38023808383438453853

3854

3876

3880

38823883389239023911

3929

39313932

3934

395739633965

3989

3992400640114012401840354046

4048

4080

4081

408240894094409841064111411341164120412141284132

4140

415741594167416841694178

4194

42024211

4216

4217

421942304263427142754276428342884290

−10

0

10

20

Coe

ffici

ents

(e)

90 435

KM Lasso

(f)

Figure 2: Continued.

8 Disease Markers

scores were associated with more deaths and shorter survivaltime (Figures 3(d) and 3(e)). Furthermore, univariate andmultivariate Cox regression analyses were conducted, andthe results revealed that the risk score was an independentprognostic predictor of OS (Table 1). In addition, we foundthat the established model was superior to ImmuneScore,StromalScore, and ESTIMATEScore in predicting prognosisin BCa, which could partly reflect the level of tumor immuneinfiltration (Figures 3(g)–3(k)). We also randomly resampled30% of cases from the training set as the internal validationdataset, and the risk model also showed high efficacy in distin-guishing cases with a low risk from those with a high risk(Supplementary Figure S1). To validate the robustness of the

frlncRNA signature, we conducted KM survival estimationsin multiple subgroups according to diverse clinicopathologicalfeatures (Supplementary Figure S2). In addition, we evaluatedthe clinical associations of the risk score. We compared therisk scores between various tumor grades and TNM stagesusing the chi-square test (Figure 4(a)) and Wilcoxon signed-rank test (Figures 4(b)–4(e)).

3.5. Functional Enrichment of the 22-frlncRNA Pair. Weimplemented LncSEA [24], an online tool developed forlncRNA enrichment analysis, to primarily assess the relatedfunctions (Figure 3(f) and Supplementary Table S4). In therisk model, 36 frlncRNAs were mostly enriched in BCa-

AC135012.3_AP003071.4ADAMTS9−AS1_AC015923.1AC002398.2_AC005180.2AP001189.1_ACTA2−AS1AP001189.1_AL023755.1AP003071.4_LINC02195MIR100HG_AATBCJAZF1−AS1_AP005432.2AC099850.3_U62317.2AC099850.3_LINC01833AL161772.1_AATBCAC090673.1_FP325330.3LINC00460_PICSARAL161431.1_AC104984.6AC090825.1_MAGI2−AS3RMRP_MAGI2−AS3AC053503.4_LINC01778AC087521.1_AC079313.1AP003071.3_FENDRRLINCO0402_FENDRRRMRP_CDKN2B−AS1LINCO1615_LINC02195

0.0510.0060.0930.0020.0160.0190.1080.0100.0350.0240.0130.035

<0.0010.053

<0.001<0.001

0.1240.0210.0620.0640.0390.012

p value2.129(0.997−4.548)1.710(1.167−2.506)1.889(0.899−3.966)2.033(1.287−3.211)1.809(1.116−2.933)1.645(1.084−2.496)1.394(0.929−2.092)2.417(1.234−4.736)1.484(1.028−2.141)2.472(1.124−5.438)1.745(1.125−2.708)1.518(1.029−2.237)2.216(1.525−3.220)1.568(0.995−2.471)

8.666(3.895−19.278)4.847(2.136−10.999)

1.403(0.911−2.159)1.535(1.068−2.205)2.089(0.962−4.536)2.547(0.946−6.857)1.650(1.025−2.657)1.720(1.125−2.628)

Hazard ratio

Hazard ratio0 5 10 15

(g)

ADAMTS9-AS1AC135012.3

AC087521.1

AP003071.4

LINC01778

AC002398.2

ACTA2-AS1AC090825.1

FENDRR

MAGI2-AS3

ZEB1

AC079313.1

AP003071.3

JAZF1-AS1

FANCD2CS

CDKN2A

AC053503.4

AL023755.1

PRR4

AC015923.1

MIR100HG

AP001189.1

AL161431.1LINC01615

CAV1

AC104984.6

AATBC

CYBB

PICSAR

RMRPFP325330.3

LINC02195

U62317.2

AL161772.1

NOX4 LTF

AC090673.1 LINC00402

LINC00460

CD44

PEBP1

AP005432.2 LINC01833

AC099850.3

EIF2AK2CDKN2B-AS1

lncRNA

Ferroptosis regulator

(h)

Figure 2: Construction of the lncRNA signature. (a) A total of 927 lncRNAs were differentially expressed between the BCa samples andparacarcinoma tissues. (b) A total of 105 overlapped lncRNAs codetermined by differential and coexpression analyses. (c) Long interveningnoncoding RNAs and antisense lncRNAs accounted for most of the 105 lncRNAs. (d and e) Lasso–Cox regression with 10-fold cross-validation.(f) A total of 35 frlncRNA pairs codetermined with Kaplan–Meier survival analysis and Lasso-Cox regression. (g) A total of 22 frlncRNApairs were included in the risk model using multivariate Cox regression with a stepwise method. (h) Network displaying the frlncRNAs andtheir possible target molecules, which serve as ferroptosis regulators. frlncRNA: ferroptosis-related lncRNA.

9Disease Markers

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

1−specificity

Sens

itivi

ty

AUC at 1 years: 0.827AUC at 3 years: 0.855AUC at 5 years: 0.874

(a)

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

1−specificity

Sens

itivi

ty

Age, AUC = 0.543Gender, AUC = 0.572Grade, AUC = 0.554Stage, AUC = 0.670T, AUC = 0.645M, AUC = 0.524N, AUC = 0.650Risk score, AUC = 0.869

(b)

0 1000 2000 3000 4000 50000.0

0.2

0.4

0.6

0.8

1.0

Survival curve (p < 0.001)

Time (day)

Ove

rall

surv

ival

rate

High riskLow risk

(c)

0 100 200 300 4000

2

4

6

8

10

Patients (increasing risk score)

Risk

scor

e

Low riskHigh risk

(d)

0 100 200 300 400

02468

101214

Patients (increasing risk socre)

Surv

ival

tim

e (ye

ars)

AliveDead

(e)

lncRNADisease

PhenotypeHallmark

(f)

Figure 3: Continued.

10 Disease Markers

0 1000 2000 3000 4000 50000.0

0.4

0.8

Survival curve (p = 2.835e−02)

Time (day)O

vera

ll su

rviv

al ra

te

High ImmuneScoreLow ImmuneScore

(g)

0 1000 2000 3000 4000 50000.0

0.4

0.8

Survival curve (p = 1.929e−02)

Time (day)

Ove

rall

surv

ival

rate

High StromalScoreLow StromalScore

(h)

0 1000 2000 3000 4000 50000.0

0.4

0.8

Survival curve (p = 1.404e−01)

Time (day)

Ove

rall

surv

ival

rate

High ESTIMATEScoreLow ESTIMATEScore

(i)

StromalScore (low vs. high)

ImmuneScore (low vs. high)

ESTIMATEScore (low vs. high)

RiskScore (low vs. high)

0.020

0.030

0.142

<0.001

p value

1.551 (1.070−2.248)

0.638 (0.426−0.957)

1.272 (0.923−1.754)

4.941 (3.400−7.180)

Hazard ratio

Hazard ratio0 1 2 3 4 5 6 7

(j)

StromalScore (low vs. high)

ImmuneScore (low vs. high)

ESTIMATEScore (low vs. high)

RiskScore (low vs. high)

0.653

<0.001

0.488

<0.001

p value

1.117 (0.690−1.806)

0.403 (0.258−0.628)

1.167 (0.755−1.803)

5.311 (3.599−7.836)

Hazard ratio

Hazard ratio0 2 4 6

(k)

Figure 3: Validation of the lncRNA signature. (a) AUC of time-dependent ROC curves. (b) Comparison of 5-year ROC curves with otherclinicopathological features. (c) Kaplan–Meier survival plot for OS in cases in the high- and low-risk groups (p < 0:001). (d and e) Scatterdiagrams showing the risk scores (d) and survival status (e) of BCa cases. (f) Results of enrichment analysis using LncSEA. (g–i)Prognostic values of ImmuneScore (g, p < 0:05), StromalScore (h, p < 0:05), and ESTIMATEScore (i, p > 0:05). The optimal cutoffs weredetermined using X-tile software. (j–k) The risk score was superior to ImmuneScore, StromalScore, and ESTIMATEScore in predicting OSin univariate (j) and multivariate Cox analyses (k). AUC: area under curve; ROC: receiver operating characteristic; OS: overall survival;BCa: bladder cancer.

Table 1: Univariate and multivariate Cox analysis of the risk score.

CovariatesUnivariate analysis Multivariate analysis

HR (95% CI) p value HR (95% CI) p value

Age (≤64 vs. >64) 1.81 (0.97-3.37) 0.06 2.04 (1.07-3.89) 0.03

Gender (female vs. male) 1.62 (0.93-2.84) 0.09 1.40 (0.79-2.47) 0.24

Grade (low vs. high) 2.81e+07 (0-∞) 1.00 5.16e+06 (0-∞) 1.00

T (T 1-2 vs. T 3-4) 2.61 (1.28-5.33) 0.01 1.66 (0.79-3.50) 0.18

N (N0 vs. N1-3) 2.31 (1.37-3.89) <0.01 2.41 (1.36-4.28) <0.01M (M0 vs. M1) 2.17 (0.78-6.03) 0.14 0.87 (0.30-2.55) 0.80

Risk score (low vs. high) 6.47 (3.40-12.32) <0.01 6.41 (3.31-12.42) <0.01HR: hazard ratio; CI: confidence interval.

11Disease Markers

RiskAgeGenderGrade⁎⁎T

M

N

RiskHighLow

Age≤64>64

GenderFemaleMale

Grade⁎⁎High gradeLow grade

TT1

T2

T3

T4

MM0

M1

NN0

N1

N2

N3

(a)

9.9e−06

0

20

40

60

High grade Low gradeGrade

Risk

scor

e

GradeHigh grade

Low grade

(b)

0.410.1

0.151e−07

0.00260.57

0

25

50

75

100

T1 T2 T3 T4

T

Risk

scor

e

T

T1

T2

T3

T4

(c)

Figure 4: Continued.

12 Disease Markers

related functions, indicating that the signature had a degreeof BCa specificity.

3.6. The Immune Infiltration Landscape and ImmunotherapyResults between High- and Low-Risk Groups. Variousmethods were implemented to estimate the immune profilesof cases in the low- and high-risk subgroups. On the onehand, the Wilcoxon signed-rank test indicated that theStromalScore, ImmuneScore, and ESTIMATEScore weresignificantly higher in the high-risk cases according to theESTIMATE algorithm (Figure 5(a)), implying that high-riskpatients might have a more active TME. On the other hand,seven clusters of immune and inflammatory genes were col-lected from a previous study [25], and GSVA was imple-mented to evaluate the immune and inflammatory responseenrichment status (Figure 5(b)). Except for interferon (IFN)and MHC-I, IgG, HCK, MHC-II, LCK, and STAT1 were allpositively correlated with the risk score according to theWilcoxon test (Figure 5(c), p < 0:05). Tumor immune het-erogeneity might account for the disconnected relationshipbetween IFN and MHC-I and the risk score.

Considering that the risk score was significantly corre-lated with tumor immune activity, we analyzed the infiltra-tion proportion of immune cells. The TIMER algorithm[15], which is similar to CIBERSORT, was developed toestimate the infiltration levels of different immune cells,including B cells, CD4 T cells, CD8 T cells, neutrophils, den-dritic cells, and macrophages. TheWilcoxon signed-rank testindicated that only B cells showed no significant associationwith risk (Figure 5(d), p < 0:05), whereas the Spearman cor-relation test revealed that CD8 T cells and macrophages werepositively correlated with the risk score (Figure 5(e), p < 0:05).We also evaluated the prognostic values of the infiltration

levels of CD8+ T cells andM2macrophages, which have beenreported to be closely associated with ferroptosis in manyother malignant tumors [6, 7], via CIBERSORT algorithmand KM survival analysis with log-rank test in BCa, andX-tile was used to determine the optimal cut-off. The analy-ses results showed the patients exhibiting worse OS carriedhigh CD8+ T cell infiltration (Figure 5(f), p < 0:01) and lowM2 macrophage infiltration (Figure 5(g), p < 0:001).

The strong linkage between risk score and immune infil-tration prompted us to investigate whether the risk scorecould serve as an indicator of the response to ICIs. First, weexamined the expression difference of common immunecheckpoint molecules and discovered that PDL1, LAG3, andTIM-3 were upregulated in the high-risk cohort accordingto the Wilcoxon test (Figure 5(h), p < 0:05). We also foundthe expression level of CXCL9 and CXCL13 was significantlyassociated with the risk score (Figure 5(i), p < 0:05). Inaddition, using the TIDE algorithm, we predicted the immu-notherapy clinical outcomes of the low- and high-riskcohorts. The chi-square test revealed a significant difference(Figure 5(j), p < 0:05). Overall, we believe that the prognosismodel could also be a potential indicator of the effectivenessof immunotherapy.

3.7. The Linkage between Chemotherapy and Risk Score.Emerging evidence suggests that some types of chemothera-peutic drugs could induce malignant cancer cells to expressdamage-associated molecular patterns that bind to antigen-presenting cells, trigger a tumor-specific immune response,and enhance the immunotherapeutic sensitivity of a malig-nant tumor [26]. Considering the above findings, weanalyzed the association between risk and chemotherapysensitivity and found that the IC50 values of cisplatin and

0.0570.078

0.380.72

10.79

0

25

50

75

100

N0 N1 N2 N3

N

Risk

scor

e

N

N0

N1

N2

N3

(d)

0.055

0

20

40

M0 M1

M

Risk

scor

e

M

M0

M1

(e)

Figure 4: A strip chart (a) and box plots (b–f) showing the association between the risk score and tumor grades (b), pathological T stage (c),Nstage (d), and M stage (e).

13Disease Markers

StromalScore⁎⁎⁎ ImmuneScore⁎⁎⁎ ESTIMATEScore⁎⁎⁎

High Low High Low High Low

−2500

0

2500

5000

Risk

RiskHigh

Low

(a)

Interferon

MHC−I

STAT1

IgG

MHC−II

HCK

LCK

Risk

RiskLow

High

−2

−1

0

1

2

(b)

IgG⁎⁎ HCK⁎⁎⁎ MHC−II⁎⁎ LCK⁎⁎⁎ MHC−I STAT1⁎ Interferon

Hig

h

Low

Hig

h

Low

Hig

h

Low

Hig

h

Low

Hig

h

Low

Hig

h

Low

Hig

h

Low

−1.0

−0.5

0.0

0.5

1.0

Risk

GSV

A sc

ore

RiskHighLow

(c)

⁎⁎ ⁎⁎⁎ ⁎⁎⁎ ⁎⁎⁎ ⁎⁎⁎

0.0

0.5

1.0

1.5

2.0

B_cell CD4_T_cell CD8_T_cell Neutrophil Macrophage Dendritic

Infil

trat

ion

prop

ortio

n

Risk

High

Low

(d)

−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

1

B_ce

ll

CD4_

T_ce

ll

CD8_

T_ce

ll

Neu

troph

il

Den

driti

c

Risk

scor

e

Mac

roph

age

B_cell

CD4_T_cell

CD8_T_cell

Neutrophil

Dendritic

Risk score

Macrophage

1 −0.07

1

−0.14

0.05

1

0

0.57

0.61

1

0

0.55

0.58

0.82

1

−0.04

0.12

0.33

0.24

0.3

1

0.02

0.21

0.34

0.34

0.24

0.37

1

(e)

Figure 5: Continued.

14 Disease Markers

0 1000 2000 3000 4000 5000

0.0

0.2

0.4

0.6

0.8

1.0

Survival curve (p = 1.661e−03)

Time (day)

Ove

rall

surv

ival

rate

CD8+ T cell high infiltrationCD8+ T cell low infiltration

(f)

0 1000 2000 3000 4000 5000

0.0

0.2

0.4

0.6

0.8

1.0

Survival curve (p = 6.931e−04)

Time (day)

Ove

rall

surv

ival

rate

M2 macrophages high infiltrationM2 macrophages low infiltration

(g)

PDL1⁎⁎ LAG3⁎ TIM−3⁎⁎⁎

High Low High Low High Low

0

2

4

Risk

Gen

e exp

ress

ion

Risk

High

Low

(h)

CXCL9⁎ CXCL13⁎⁎

High Low High Low

0

3

6

9

Risk

Gen

e exp

ress

ion

RiskHigh

Low

(i)

Figure 5: Continued.

15Disease Markers

36%26%

64%74%

p = 6.15e−12p = 4.88e−05

(n = 199)(n = 197)0

10

20

30

40

50

60

70

80

90

100

High Low

Risk

Perc

enta

ge (%

)

Response

Response

Non_response

χ2Pearson = 4.58, p = 0.032

(j)

Figure 5: Evaluation of the predictive ability for immune infiltration and immunotherapy response of the risk model. (a) Cases in the high-risk group had a higher StromalScore, ImmuneScore, and ESTIMATEScore. (b) Heat map showing different immune and inflammatoryprofiles in the high- and low-risk group. (c) GSVA scores of seven clusters of immune and inflammatory genes in the high- and low-riskgroups. (d) Infiltration ratios of six immune cells in different risk groups. (e) The risk score was positively correlated with the infiltrationlevels of CD8+ T cells and macrophages. (f) Patients with BCa with a low infiltration level of CD8+ T cells had unfavorable clinicaloutcomes (p < 0:01). (g) Patients with BCa with a high infiltration level of M2 macrophages had poorer survival (p < 0:001). (h) PDL-1,LAG3, and TIM-3 were differentially expressed between the high- and low-risk groups. (i) The expression levels of CXCL9 and CXCL13 inthe high-risk group were significantly higher than those in the low-risk group according to the Wilcoxon test. (j) The chi-square testindicated that cases in the high-risk group might more likely benefit from immunotherapy. GSVA: gene set variation analysis.

IC50_cisplatin⁎⁎⁎ IC50_doxorubicin IC50_gemcitabine⁎ IC50_methotrexate⁎⁎⁎ IC50_vinblastine

High Low High Low High Low High Low High Low−8

−4

0

4

Risk

IC50

RiskHigh

Low

(a)

0 1000 2000 3000 4000 5000

0.0

0.2

0.4

0.6

0.8

1.0

Chemotherapy (p = 4.557e−05)

Time (day)

Ove

rall

surv

ival

rate

High riskLow risk

(b)

Figure 6: Evaluation of the predictive ability for chemotherapy effectiveness of the risk model. (a) Correlation analysis revealed that the riskscore was significantly correlated with the IC50 values of cisplatin, gemcitabine, and methotrexate. (b) Patients with a high-risk score showedworse prognosis among those who received chemotherapy (p < 0:001). IC50: half inhibitory concentration.

16 Disease Markers

38%

68%

62%

32%

p = 4.83e−07p = 4.81e−04

(n = 199)(n = 197)0

10

20

30

40

50

60

70

80

90

100

High Low

Risk

Perc

enta

ge (%

)

Cluster

2

1

χ2Pearson = 36.41, p = 1.6e−09

(a)

6e−11

0

20

40

60

Cluster 1 Cluster 2

Cluster

Risk

scor

e

Cluster

Cluster 1

Cluster 2

(b)

Cluster 2

Cluster 1

Low risk

High risk

(c)

Go_heme_transport

Go_4_iron_4_sulfur_cluster_binding

Go_iron_sulfur_cluster_assembly

Go_sequestering_of_iron_ion

Go_multicellular_organismal_iron_ion_homeostasis

Module_540

Go_cellular_iron_ion_homeostasis

Type

TypeHigh riskLow risk

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

(d)

Go_cellular_iron_ion_homeostasis

0.0

0.5

1.0

1.5

2.0

−0.05 0.00 0.05 0.10logFC

−log

10ad

j.p.V

al

GroupDownNotsignificantUp

(e)

Go_response_to_iron_ion

Hallmark_heme_metabolism

Go_iron_ion_homeostasis

Go_cellular_iron_ion_homeostasis

Go_iron_import_into_cell

Go_iron_ion_transmembrane_transporter_activity

Go_iron_ion_transmembrane_transport

Type

TypeCluster 2Cluster 1

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

(f)

Figure 7: Continued.

17Disease Markers

gemcitabine were negatively correlated with risk (Figure 6(a)),according to the Wilcoxon test, whereas the IC50 value ofmethotrexate had a positive association (Figure 6(a)), imply-ing that the risk model has a potential to predict the efficacyof chemotherapy for BCa. The evaluated risk score could alsoserve as a prognostic biomarker for patients who had receivedchemotherapy (p < 0:001, Figure 6(b)).

3.8. Analysis of Iron Metabolic Pathways.Under normal con-ditions, intracellular iron balance is maintained through theiron transport system. Increasing the uptake of iron ions ordecreasing their efflux can enhance the sensitivity of cancer

cells to oxidative damage and ferroptosis [27]. Hence, weanalyzed the changes in iron metabolism-related pathwaysand functions. Although the chi-square test and Wilcoxonsigned-rank test revealed that ferroptosis-related clusteringand frlncRNA-based risk grouping were highly associated(Figures 7(a)–7(c)), the changes in iron metabolism-relatedpathways were mostly different according to GSVA(Figures 7(d)–7(g), Supplementary Tables S5 and S6). It isnoteworthy that the GSVA scores of cellular iron ionhomeostasis were significantly altered in both risk groupingand ferroptosis clustering (Figure 7(h)), and the resultswere validated using GSEA (Figure 7(i)). The significantly

Go_cellular_iron_ion_homeostasis

0

3

6

9

12

−0.15 −0.10 −0.05 0.00 0.05logFC

GroupDownNotsignificant

−log

10ad

j.p.V

al

(g)

6 61

Cluster Risk

(h)

0.1

0.2

0.3

0.4

Go_cellular_iron_ion_homeostasis

Enric

hmen

t sco

re (E

S)

Nominal p−value: 0.0114FDR: 0.2413

ES: 0.4145Normalized ES: 1.5625

RiskHigh Low

0 5000 10000 15000

−0.4

−0.2

0.0

0.2

0.4

Rank in ordered gene list

Sign

al2N

oise

−0.4

−0.2

0.0

0.2

0.4

(i)

R = 0.35, p = 4.7e−13

−0.25

0.00

0.25

0 20 40 60Risk score

Go_

cellu

lar_

iron_

ion_

hom

eost

asis

(j)

Figure 7: Enrichment analysis of iron metabolic pathways. (a) Chi-square test indicated that ferroptotic clustering and risk grouping werehighly correlated. (b) Patients in cluster 1 had higher risk scores. (d) Sankey diagram showing the close association between ferroptoticclustering and risk grouping. (d and e) Heat map (d) and volcano diagram (e) indicating that seven iron metabolic pathways werechanged with risk. (f and g) Heat map (f) and volcano diagram (g) showing that seven iron metabolic pathways were differentiallyenriched in cluster 1 and cluster 2. (h) Venn plot showing overlapping of cellular iron ion homeostasis. (i) GSEA of cellular iron ionhomeostasis in the high-risk group. (j) The risk score was positively correlated with the GSVA score of cellular iron ion homeostasis.GSEA: gene set enrichment analysis; GSVA: gene set variation analysis.

18 Disease Markers

positive Spearman correlation between the GSVA score and therisk score verified the importance of maintaining cellular ironhomeostasis for the occurrence of ferroptosis (Figure 7(j)),also validating our risk model as ferroptosis related.

3.9. The Functional Prediction of the Crucial frlncRNA Pair.The gene-pair strategy focuses on the relative expressionlevels of two genes, and high prognostic and diagnostic valuesof the relative expression level imply that the two genes mightbe involved in the same molecular mechanism through amutual interaction. The present study investigated themechanisms using AC090825.1 andMAGI2-AS3, which com-prised the frlncRNA pair with the highest hazard ratio, asexamples. Despite the relatively lower expression ofAC090825.1 andMAGI2-AS3 in BCa samples (SupplementaryFigure 3a, b), the difference was significantly higher than thatin paracarcinoma tissues (Supplementary Figure 3c). Inaddition, we found that patients had poorer survivalwhen AC090825.1 was more highly expressed than MAGI2-AS3 (Supplementary Figure 3d). Moreover, AC090825.1 andMAGI2-AS3 expression had a significant association in bothparacarcinoma tissues and BCa samples (SupplementaryFigure 3e, p < 0:05). Recent studies have postulated thatlncRNAs regulate their target genes by interacting withtranscription factors. Accordingly, we predicted thetranscription factor of frlncRNAs and their target genes(Supplementary Tables S7–S10) and constructed a regulatorynetwork (Supplementary Figure 3f). HOXD4, ZNF354C,GSX2, NR2C2, MEIS1, and HOXC4 were codetermined bythe transcription factor prediction (Supplementary Figure 3g)and coexpressed with two frlncRNAs (SupplementaryFigure 3h, R > 0:3, p < 0:05), implying the possible biologicalfunctions of AC090825.1 and MAGI2-AS3.

4. Discussion

In recent years, several models have been proposed to predictthe prognosis and immune infiltration of malignancies [28].However, few studies have focused on gene-pair modelingmethods, which do not require exact gene expression levels.In addition, most studies chose to screen immune-relatedgenes to predict the immune infiltration of malignancies,whereas few studies used ferroptosis-related genes to predictthe immune infiltration status despite several reports indicat-ing that ferroptosis and the TME could affect each other. Inthis study, frlncRNAs were detected to construct a rationalmodel based on the gene-pair strategy to predict the progno-sis and tumor immune reaction of BCa for the first time.

First, we initially identified 121 ferroptosis regulatorsfrom a previous study [5] and extracted transcriptome datafrom the TCGA BCa dataset. After unsupervised clustering,all enrolled cases were divided into two subgroups. We foundthat the subgroups had different prognoses, immune infiltra-tion, immunotherapy response, and chemotherapy out-comes, implying that ferroptosis was closely correlated withthe TME in BCa, thus, supporting the relevance of construct-ing an frlncRNA signature to predict tumor immune infiltra-tion in future studies. Furthermore, all lncRNAs wereextracted, and frlncRNAs were determined through genomic

difference analysis and Pearson coefficient calculation. Afterpairing and differential analysis of frlncRNAs, we ultimatelybuilt a 0-or-1 matrix. Using KM survival analysis, the Lassoalgorithm, and multivariate Cox regression, 22 frlncRNApairs were identified. KM plot analysis, ROC curve analysis,random sampling validation, and subgroup analysis wereperformed for model verification and showed that the novelsignature is a powerful tool for BCa prognosis prediction.Moreover, it was found that the risk score calculated usingthe frlncRNA signature was significantly correlated withimmune infiltration, immunotherapeutic effectiveness, andchemotherapeutic results. GSVA showed that cellular ironhomeostasis imbalance, which is routinely described as oneof the key mechanisms of ferroptosis, was the crucial ironmetabolism-related pathway in different subgroups, validat-ing the clustering strategy and that the constructed riskmodel is ferroptosis related.

Some researchers have contributed to the construction oflncRNA signatures for predicting the immune infiltrationprofiles of BCa. Wu et al. identified eight immune-relatedlncRNAs and found that the risk score was positively corre-lated with poor clinical outcomes and high immune infiltra-tion [29]. By using the ESTIMATE algorithm and correlationanalysis, Yuan et al. constructed an immune-related lncRNAsignature to evaluate immune cell infiltration in the TME[30]. The use of lncRNAs to predict tumor immune infiltra-tion is an effective strategy because of the wide participationof lncRNAs in biological processes [31]. Most studies haveconstructed lncRNA diagnosis models based on immune-related genes or algorithms, which is a reasonable and logicalmethod. However, emerging evidence has revealed a strongassociation between ferroptosis and tumor immune reac-tions. For instance, Wang et al. found that IFNγ releasedfrom CD8+ T cells could suppress the expression ofSLC7A11, thereby promoting lipid peroxidation in cancercells and inducing ferroptosis [6] (Figure 8). In addition, itwas reported that extracellular KRAS protein released by fer-roptotic tumor cells and taken up by macrophages contrib-utes to M2 macrophage polarization, thus, resulting in thedevelopment of pancreatic tumor cells [7] (Figure 8). Simi-larly, a low infiltration level of CD8+ T cells and a high infil-tration level of M2 macrophages were found to be associatedwith poor prognosis in patients with BCa (Figures 5(f) and5(g)), and the infiltration ratio was significantly differentbetween the high- and low-risk groups (Figure 5(d)), suggest-ing that the same biological mechanisms might exist in BCa.The close relationship between ferroptotic cell death and theTME prompted us to develop a ferroptosis-linked lncRNAmodel for predicting prognosis and diagnosing immune infil-tration of BCa.

The frlncRNA signature is an effective and practical toolfor predicting the prognosis of patients with BCa. Comparedwith other clinical features, the novel model was able to dis-tinguish cases with a high or low risk with higher efficacy.Univariate and multivariate analyses revealed that the riskscore was an independent prognostic predictor. Randomresampling verification, clinical parameter association analy-sis, and subgroup analysis validated the robustness of themodel. Notably, we used a gene-pair strategy instead of

19Disease Markers

detecting the exact expression level to establish the predictivemodel, which only required examining which gene in the pairhad a higher expression, thus, extending the application ofthe risk model.

Among the lncRNAs in the frlncRNA signature, severalhave been demonstrated to be related to immunity, ferropto-sis, and malignancy. For instance, ADAMTS9-AS1, whichwas reported to suppress the malignant phenotypes of breastcancer cells [32], could also regulate colorectal cancer cellproliferation and migration [33]. The gene-pair strategy alsohelped disclose the underlying association between thepaired lncRNAs. AC090825.1 and MAGI2-AS3, comprisingthe gene pair with the highest hazard ratio in the model, werestrongly correlated based on RNA expression levels in bothparacarcinoma tissues and tumor samples (SupplementaryFigure 3e). Although it has been reported that MAGI2-AS3could inhibit the development of BCa cells [34], theinterplay between AC090825.1 and MAGI2-AS3 is stillunknown. Functional prediction analysis indicated thatAC090825.1 and MAGI2-AS3 might regulate their targets,NOX4 and ZEB1, by interacting with the transcriptionfactors (Supplementary Figure 3f). Overall, our modelhelped in identifying new biomarkers and in proposingnovel mechanisms for BCa.

To our knowledge, this is the first study to use frlncRNAsto predict the immune landscape in BCa. In addition, thisstudy is the first to use the gene-pair strategy to construct a

lncRNA signature for predicting the clinical outcomes ofBCa. From the perspective of model performance, the estab-lished model had the highest AUC values among the knownprognostic models for BCa. We also identified dozens ofnovel lncRNA biomarkers, such as AC090825.1, which maybe useful in further studies.

However, the present study had several limitations. Theanalyzed raw data were all downloaded from TCGA, andthe external validation of the established model was inade-quate. As a general rule, validation in a different cohort isrequired for a diagnosis or prognosis model because of theindividual variations in gene expression. To reduce errorscaused by gene expression differences, we innovatively usedthe gene-pair modeling method to construct an frlncRNAsignature. Random resampling, subgroup analysis, and cor-relation analysis of clinical risk parameters were conductedto validate the robustness of our model. Although externalvalidation was insufficient, these methods and the evidencesuggested that the novel model was acceptable. However,external validation in other BCa cohorts is warranted.

5. Conclusions

We developed a novel lncRNA signature to predict the prog-nosis and immune landscape of BCa based on the gene-pairstrategy and ferroptosis-related genes. This lncRNA signa-ture provides new clues for identifying the regulatory

Lipid ROS

SLC7A11SLC3A2

Cystine

KRAS G12D

KRAS G12D

Multivesicular body Autophagosome

Exosome

RAB27A

AGER

Ferroptosis

M2 macrophage polarization

Macrophage cell

Tumor cell Ferroptotic tumor cell

CD8+ T cell

IFN 𝛾

IFNR

Promote apoptosis

JAK

STAT1

STAT3

GSH

GPX4

STAT1SLC7A11

JAK

Figure 8: Dual role of ferroptosis in tumor immunity.

20 Disease Markers

relationship between ferroptosis and the TME and can helpclinicians estimate the prognosis, immunotherapy outcomes,and chemotherapy response of patients with BCa.

Data Availability

The data used to support the findings of this study are avail-able from The Cancer Genome Atlas (TCGA, https://portal.gdc.cancer.gov/), and the code would be supplied from thecorresponding author upon request.

Conflicts of Interest

The authors declared no potential conflicts of interest withrespect to the research, authorship, and/or publication ofthis article.

Acknowledgments

We thank everyone who did help in this study. This study issupported by the National Natural Science Foundation ofChina (No. 81772257), the Youth Cultivation Program ofSouthern Medical University (No. PY2018N076), and theMedical Scientific Research Foundation of Guangdong Prov-ince (No. A2019557).

Supplementary Materials

Supplementary Figure S1: validation of the risk model withresampling. Note: (a) Kaplan-Meier survival analysis of casesin high-risk group and low-risk group (p < 0:001). (b) ROCcurves of the risk score in internal validation dataset. Supple-mentary Figure S2: Kaplan-Meier subgroup survival analysisin TCGA-BLCA patients. Note: (a) age ≤ 64 (p < 0:001). (b)Age > 64 (p < 0:001). (c) Male (p < 0:001). (d) Female(p < 0:001). (e) Pathological T 1-2 stages (p < 0:001). (f) Path-ological T 3-4 stages (p < 0:001). (g) N0 stages (p < 0:001). (h)N1-N3 stages (p < 0:001). (i) M0 stage (p < 0:001). (j) M1stage (p < 0:05). Supplementary Figure S3: functional predic-tion of AC090825.1 andMAGI2-AS3. Note: (a) the expressiondifference of AC090825.1 in BCa samples and paracarcinomatissue via Wilcoxon signed-rank test. (b) The expression dif-ference of MAGI2-AS3 in BCa samples and paracarcinomatissue via Wilcoxon signed-rank test. (c) The difference valuesbetween the expression of AC090825.1 and MAGI2-AS3 inBCa samples and paracarcinoma samples. (d) The casessuffered poorer survival when the expression value ofAC090825.1 was higher than that of MAGI2-AS3 (p < 0:01).(e) The expression levels of AC090825.1 and MAGI2-AS3were positively correlated both in BCa samples and paracarci-noma samples. (f) Network of lncRNAs, the target genes andtheir possible transcription factors. (g) UpSet diagram show-ing 6 overlapped transcription factors. (h) The correlation net-work of lncRNAs, the target genes and transcription factors.BCa: bladder cancer. Supplementary Table S1: 20 ferroptosisregulators were differentially expressed between BCa andparacarinoma samples. Supplementary Table S2: the Pearsoncorrelation analysis of lncRNAs and ferroptosis regulators.Supplementary Table S3: Kaplan-Meier survival analysis of

the lncRNA pairs. Supplementary Table S4: enrichmentresults from LncSEA website. Supplementary Table S5: GSVAresults of different clusters based on ferroptosis regulators.Supplementary Table S6: GSVA results of different risk sub-groups based on frlncRNA signature. Supplementary TableS7: the prediction of the transcription factors of AC090825.1.Supplementary Table S8: the prediction of the transcriptionfactors of MAGI2-AS3. Supplementary Table S9: the predic-tion of the transcription factors of NOX4. SupplementaryTable S10: the prediction of the transcription factors ofZEB1. (Supplementary Materials)

References

[1] V. G. Patel, W. K. Oh, and M. D. Galsky, “Treatment ofmuscle-invasive and advanced bladder cancer in 2020,” CA:a Cancer Journal for Clinicians, vol. 70, no. 5, pp. 404–423,2020.

[2] A. T. Lenis, P. M. Lec, K. Chamie, and M. MSHS, “Bladdercancer: a review,” JAMA, vol. 324, no. 19, pp. 1980–1991, 2020.

[3] A. B. Apolo, J. R. Infante, A. Balmanoukian et al., “Avelumab,an anti-programmed death-ligand 1 antibody, in patients withrefractory metastatic urothelial carcinoma: results from amulticenter, phase Ib study,” J Clin Oncol, vol. 35, no. 19,pp. 2117–2124, 2017.

[4] P. Sharma, M. Retz, A. Siefker-Radtke et al., “Nivolumab inmetastatic urothelial carcinoma after platinum therapy(CheckMate 275): a multicentre, single-arm, phase 2 trial,”The Lancet Oncology, vol. 18, no. 3, pp. 312–322, 2017.

[5] D. Tang, X. Chen, R. Kang, and G. Kroemer, “Ferroptosis:molecular mechanisms and health implications,” CellResearch, vol. 31, no. 2, pp. 107–125, 2021.

[6] W. Wang, M. Green, J. E. Choi et al., “CD8+ T cells regulatetumour ferroptosis during cancer immunotherapy,” Nature,vol. 569, no. 7755, pp. 270–274, 2019.

[7] E. Dai, L. Han, J. Liu et al., “Autophagy-dependent ferroptosisdrives tumor-associated macrophage polarization via releaseand uptake of oncogenic KRAS protein,” Autophagy, vol. 16,no. 11, pp. 2069–2083, 2020.

[8] M. K. Atianand, D. R. Caffrey, and K. A. Fitzgerald, “Immuno-biology of long noncoding RNAs,” Annual Review of Immu-nology, vol. 35, no. 1, pp. 177–198, 2017.

[9] C. Mao, X. Wang, Y. Liu et al., “A G3BP1-interacting lncRNApromotes ferroptosis and apoptosis in cancer via nuclearsequestration of p53,” Cancer Research, vol. 78, no. 13,pp. 3484–3496, 2018.

[10] B. Li, Y. Cui, M. Diehn, and R. Li, “Development and valida-tion of an individualized immune prognostic signature inearly-stage nonsquamous non-small cell lung cancer,” JAMAOncology, vol. 3, no. 11, pp. 1529–1537, 2017.

[11] M. D. Robinson, D. J. McCarthy, and G. K. Smyth, “edgeR: aBioconductor package for differential expression analysis ofdigital gene expression data,” Bioinformatics, vol. 26, no. 1,pp. 139-140, 2010.

[12] M. D. Wilkerson and D. N. Hayes, “ConsensusClusterPlus: aclass discovery tool with confidence assessments and itemtracking,” Bioinformatics, vol. 26, no. 12, pp. 1572-1573, 2010.

[13] K. Yoshihara, M. Shahmoradgoli, E. Martinez et al., “Inferringtumour purity and stromal and immune cell admixture fromexpression data,” Nature Communications, vol. 4, no. 1,p. 2612, 2013.

21Disease Markers

[14] A. M. Newman, C. L. Liu, M. R. Green et al., “Robust enumer-ation of cell subsets from tissue expression profiles,” NatureMethods, vol. 12, no. 5, pp. 453–457, 2015.

[15] T. Li, J. Fu, Z. Zeng et al., “TIMER2.0 for analysis of tumor-infiltrating immune cells,” Nucleic Acids Research, vol. 48,no. W1, pp. W509–W514, 2020.

[16] O. Fornes, J. A. Castro-Mondragon, A. Khan et al., “JASPAR2020: update of the open-access database of transcription fac-tor binding profiles,” Nucleic Acids Research, vol. 48, no. D1,pp. D87–D92, 2020.

[17] G. Tan and B. Lenhard, “TFBSTools: an R/bioconductor pack-age for transcription factor binding site analysis,” Bioinformat-ics, vol. 32, no. 10, pp. 1555-1556, 2016.

[18] P. Jiang, S. Gu, D. Pan et al., “Signatures of T cell dysfunctionand exclusion predict cancer immunotherapy response,”Nature Medicine, vol. 24, no. 10, pp. 1550–1558, 2018.

[19] K. Litchfield, J. L. Reading, C. Puttick et al., “Meta-analysis oftumor- and T cell-intrinsic mechanisms of sensitization tocheckpoint inhibition,” Cell, vol. 184, no. 3, pp. 596–614.e14,2021, e514.

[20] P. Geeleher, N. Cox, and R. S. Huang, “pRRophetic: an R pack-age for prediction of clinical chemotherapeutic response fromtumor gene expression levels,” PLoS One, vol. 9, no. 9, articlee107468, 2014.

[21] H. Mazdak, F. Yazdekhasti, A. Movahedian, N. Mirkheshti,and M. Shafieian, “The comparative study of serum iron,copper, and zinc levels between bladder cancer patients and acontrol group,” International Urology and Nephrology, vol. 42,no. 1, pp. 89–93, 2010.

[22] M. Tang, Z. Chen, D. Wu, and L. Chen, “Ferritinophagy/fer-roptosis: iron-related newcomers in human diseases,” Journalof Cellular Physiology, vol. 233, no. 12, pp. 9179–9190, 2018.

[23] Y. Shen, X. Peng, and C. Shen, “Identification and validation ofimmune-related lncRNA prognostic signature for breast can-cer,” Genomics, vol. 112, no. 3, pp. 2640–2646, 2020.

[24] J. Chen, J. Zhang, Y. Gao et al., “LncSEA: a platform for longnon-coding RNA related sets and enrichment analysis,”Nucleic Acids Research, vol. 49, no. D1, pp. D969–D980, 2021.

[25] C. Zhang, Z. Zhang, G. Zhang et al., “Clinical significance andinflammatory landscapes of a novel recurrence- associatedimmune signature in early-stage lung adenocarcinoma,” Can-cer Letters, vol. 479, pp. 31–41, 2020.

[26] F. Meric-Bernstam, J. Larkin, J. Tabernero, and C. Bonini,“Enhancing anti-tumour efficacy with immunotherapy combi-nations,” Lancet, vol. 397, no. 10278, pp. 1010–1022, 2021.

[27] C. Liang, X. Zhang, M. Yang, and X. Dong, “Recent progress inferroptosis inducers for cancer therapy,” Advanced Materials,vol. 31, article e1904197, 2019.

[28] Z. Lv, C. Pang, J. Wang et al., “Identification of a prognosticsignature based on immune-related genes in bladder cancer,”Genomics, vol. 113, no. 3, pp. 1203–1218, 2021.

[29] Y. Wu, L. Zhang, S. He et al., “Identification of immune-related lncRNA for predicting prognosis and immunothera-peutic response in bladder cancer,” Aging (Albany NY),vol. 12, no. 22, pp. 23306–23325, 2020.

[30] R. Cao, L. Yuan, B. Ma, G. Wang, and Y. Tian, “Immune-related long non-coding RNA signature identified prognosisand immunotherapeutic efficiency in bladder cancer (BLCA),”Cancer Cell International, vol. 20, no. 1, p. 276, 2020.

[31] L. Statello, C. J. Guo, L. L. Chen, and M. Huarte, “Gene regu-lation by long non-coding RNAs and its biological functions,”Nature Reviews. Molecular Cell Biology, vol. 22, pp. 1–23, 2020.

[32] S. Fang, Y. Zhao, and X. Hu, “LncRNA ADAMTS9-AS1restrains the aggressive traits of breast carcinoma cells viasponging miR-513a-5p,” Cancer Management and Research,vol. Volume 12, pp. 10693–10703, 2020.

[33] N. Li, J. Li, Q. Mi et al., “Long non-coding RNAADAMTS9-AS1suppresses colorectal cancer by inhibiting the Wnt/β‐cateninsignalling pathway and is a potential diagnostic biomarker,”Journal of Cellular and Molecular Medicine, vol. 24, no. 19,pp. 11318–11329, 2020.

[34] C. Tang, Y. Cai, H. Jiang et al., “LncRNAMAGI2-AS3 inhibitsbladder cancer progression by targeting the miR-31-5p/TNS1axis,” Aging (Albany NY), vol. 12, no. 24, pp. 25547–25563,2020.

22 Disease Markers