The transcriptome data and clinical information of BLCA samples were collected from The Cancer Genome Atlas (TCGA). The FPKM values of the raw matrix were transformed into TPM values for follow-up studies. The TCGA BLCA dataset included 403 BLCA samples and 19 normal samples. Four immunotherapy cohorts were collected, including IMvigor210 (248 samples), Gide (32 samples), Van Allen (42 samples) and Lauss (25 samples). Besides, the copy number variation (CNV) data of BLCA samples in the TCGA BLCA dataset, processed with GISTIC 2.0, were downloaded from the UCSC Xena data portal (http://xena.ucsc.edu/). Drug information, including 184 common anticancer drugs and the corresponding target genes were collected from the DrugBank database (www.drugbank.ca).

According to our previous studies, the Xiangya real-world cohort was based on BLCA samples after surgical resections in the Xiangya Hospital, Central South University. The Xinagya real-world cohort, already uploaded with the number as GSE188715, included 57 BLCA samples sequenced by the BGISEQ-500 platform (BGI-Shenzhen, China) (Liu et al., 2021a; Hu et al., 2021). The TPM values of the raw matrix were used for the follow-up analysis.

IFN-γ related genes have never been systematically summarized as a list. Originally, we collected IFN-γ-related genes from previous studies as comprehensively as possible (Gao et al., 2016; Hu et al., 2020). The least absolute shrinkage and selector operation (LASSO) regression analysis was first performed on these genes for dimension reduction. Each gene can be regarded as a factor, and factors which contribute relatively less to final outcomes of the analysis were assigned the value zero. Remaining factors, genes with non-zero coefficients, were ultimately selected for multivariable Cox regression analysis and used as variables to construct the IFN-γ-related signature. The formula is as follows: Risk score = ∑ I F N i * R N A i , where IFNi is the coefficient of the genes and RNAi is the expression value of the genes in multivariable Cox regression analysis. To evaluate the performance of the IFN-γ-related signature, we simultaneously perform the time-dependent receiver operating characteristic (ROC) and calibration curves using the R package “timeROC” in the TCGA BLCA cohort, Xiangya real-world cohort, and IMvigor210 cohort.

The univariate and multivariate Cox regression analyses were adopted to filtrate independent prognostic factors from clinicopathologic characteristics and the IFN-γ-related signature. After analyzing results, prognostic factors in univariate Cox analysis were screened and integrated to construct a nomogram. The performance of the nomogram was evaluated by the time-dependent ROC and calibration curves using the R package “timeROC” in the TCGA BLCA dataset and the Xiangya real-world cohort.

The mutation landscape of the IFN-γ-related signature was visualized based on CNV data using the R package “maftools.” Somatic mutation data of BLCA samples in the TCGA BLCA dataset was used to calculate the tumor mutation burden (TMB).

In our previous studies, relevant immunological characteristics and algorithms in the TME were described in detail (Liu et al., 2021a; Hu et al., 2021). The cancer immunity cycle is composed of seven key steps: the release and presentation of cancer cell antigens (Steps 1 and 2), the priming and activation of the immune system (Step 3), then the trafficking and infiltration of immune cells into tumors (Steps 4 and 5), finally recognizing and killing cancer cells by T cells (Steps 6 and 7) (26). The cancer immunity cycle and immune infiltrating cells involved were quantified using the single-sample gene-set enrichment analysis (ssGSEA) in BLCA samples.

We calculated the sensitivity of anticancer drugs using data downloaded from the DrugBank. The predictive value of the IFN-γ-related signature for the immunotherapy was validated in three immunotherapy cohorts, namely the Gide, Van Allen, and Lauss.

Correlation coefficients were determined by the Spearman and distance correlation analyses. Normally distributed continuous variables between the two groups were compared using the t-test, while the non-normally ones between the two groups were compared using the Mann-Whitney U test instead. Chi-square or Fisher exact tests were used for comparison between categorical variables. The “survcutpoint” function from the R package “survminer” for the maximum rank statistic was used to determine the optimal cutoff value of the IFN-γ-related signature. The survival curves were generated using the Kaplan-Meier method, while the statistical significance was determined using the log-rank test. The threshold of significance was set at p < .05, and all statistical tests were two-sided. R project (version 3.6.3, http://www.r-project.org) was used for all analyses.

Differential analysis performed on the IFN-γ-related genes between BLCA samples and normal samples was shown in Figures 1A, B. Both sample sets (the BLCA set and normal one) could be clearly separated by IFN-γ-related genes (Figure 1C). Correlations between every two IFN-γ-related genes were shown in Figure 1D, in which 22 IFN-γ-related genes were favorable factors, while the remaining 11 were risk factors. The CNV frequency of IFN-γ-related genes was shown in Figure 1E. Roughly a half of the genes showed CNV loss, while the other half reversed. The mutation frequency of the IFN-γ-related genes was shown in Figure 1F, in which PARP14, OAS2, OAS3, C1S, and TNFAIP3 were the top five mutated IFN-γ-related genes.

We successfully reduced the dimension of IFN-γ-related genes using LASSO regression analysis (Figure 2A). The coefficients of the IFN-γ-related genes in LASSO regression analysis were shown in Figure 2B. TNFAIP2, CXCL10, and TAP1 were finally included for the multivariable Cox regression analysis. Figure 2C displayed the coefficients of these three genes. The distribution of the IFN-γ-related signature in BLCA samples was shown in Figure 2D. BLCA patients with high IFN-γ-related signature scores were associated with decreased survival time (Figure 2E). The accuracy of the signature in predicting 1-year, 3-year, and 5-year OS was .609, .614, and .639 respectively (Figure 2F).

Univariate Cox regression analysis was performed on clinical characters including the IFN-γ-related signature in the TCGA BLCA cohort. Results revealed that the age, tumor stage, TN grading system, and the IFN-γ-related signature were independent prognostic factors (Figure 3A). Multivariate Cox regression analysis was ultimately performed likewise in the TCGA BLCA dataset, in which age and the IFN-γ-related signature were independent prognostic factors (Figure 3B). Nomogram was constructed based on clinical factors including the IFN-γ-related signature in the TCGA BLCA dataset (Figure 3C).1-year, 3-year, and 5-year ROC of the nomogram in the TCGA BLCA dataset had respective values of .72, .71, and .74 (Figure 4A). 1-year, 3-year, and 5-year calibration curves of the nomogram in the TCGA BLCA dataset were shown in Figure 4B. 1-year, 3-year, and 5-year ROC of the nomogram in the Xiangya real-world cohort had values of .82, .87, and .86 (Figure 4C). 1-year, 3-year, and 5-year calibration curves of the nomogram in the Xiangya real-world cohort were shown in Figure 4D.

TP53, TTN, KMT2D, MUC16, and KDM6A were the top five mutated genes in the high IFN-γ-related signature score group (Figure 5A). TTN, TP53, MUC16, KMT16, and ARID1A were the top five mutated genes in the low IFN-γ-related signature score group (Figure 5B). The high IFN-γ-related signature score group was significantly associated with a lower TMB level (Figure 5C). However, there was no significant correlation between the MANTIS score and the IFN-γ-related signature (Figure 5D).

As was shown in Figure 6A, low IFN-γ-related signature scores significantly indicated some cancer immunity cycles including T cell recruiting, Th 1 cell recruiting, and macrophage recruiting. Besides, the low IFN-γ-related signature scores were generally associated with immune infiltrating cells including activated CD4 cells, activated CD8 cells, and natural killer T cells (Figure 6B). Correlations between the IFN-γ-related signature and each stroma-activated pathway were shown in Figure 6C. Immunotherapy-predicted pathways were relatively more active in the low IFN-γ-related signature score group (Figure 6D). We divided TCGA samples into different binary groups according to the sex and stage. Results of validating our IFN-γ signature in female and male groups, high stage and low stage groups proved the conclusion as expected (Supplementary Figure S1–S4).

BLCA patients with a high IFN-γ-related signature score were associated with shorter survival time (Figure 7A). 1-year, 3-year, and 5-year ROC of the IFN-γ-related signature had values of .84, .66, and .66, respectively (Figure 7B). The signature also showed negative correlation with multiple immune checkpoint molecules, including CD274, LAG3, CTLA4, PDCD1, and HAVCR2 (Figure 7C). As expected, the signature was negatively associated with cancer immunity cycles, including T cell recruiting, Th 1 cell recruiting, and macrophage recruiting (Figure 7D). The IFN-γ-related signature was negatively associated with immunotherapy-predicted pathways, including APM signal, microRNAs in cancer, mismatch repair, cell cycle, and p53 signaling pathway (Figure 7D). In addition, the low signature score group was generally significantly associated with immune infiltrating cells, including activated CD4 cells, activated CD8 cells, and natural killer T cells (Figure 7E).

BLCA patients with high IFN-γ-related signature scores were associated with decreased survival time (Figure 8A). 1-year, 3-year, and 5-year ROC of the IFN-γ-related signature had values of .74, .59, and .6 (Figure 8B). The IFN-γ-related signature was negatively associated with multiple immune checkpoint molecules, including CD274, LAG3, CTLA4, PDCD1, and HAVCR2 (Figure 8C). The IFN-γ-related signature was negatively associated with cancer immunity cycles, including T cell recruiting, Th 1 cell recruiting, and macrophage recruiting (Figure 8D). The IFN-γ-related signature was negatively associated with immunotherapy-predicted pathways, including APM signal, microRNAs in cancer, mismatch repair, cell cycle, and p53 signaling pathway (Figure 8D). In addition, the low IFN-γ-related signature scores were generally significantly associated with immune infiltrating cells, including activated CD4 cells, activated CD8 cells, and natural killer T cells (Figure 8E).

We predicted the sensitivity of BLCA patients to common drugs. Atezolizumab_CD274, Cetuximab_FCGR2A, Cetuximab_FCGR3A, Cetuximab_C1QC, Cetuximab_C1QB, Cetuximab_FCGR1A, Cetuximab_G1QA, Trastuzumab_ERBB2, and Bevacizumab_VEGFA had significantly lower drug sensitivity in the high IFN-γ-related signature score group (Figure 9A). Likewise, melanoma patients with high IFN-γ-related signature scores were less likely to respond to ICB (Figures 9B–D). Notably, melanoma patients with high IFN-γ-related signature scores were associated with decreased survival time (Figure 9C).