TCGA database (http://cancergenome.nih.gov) is a landmark public cancer genomics program that, as of 2021, had analyzed molecular characteristics of more than 20,000 primary cancer and normal samples covering 33 cancer types. We used The University of California Santa Cruz (UCSC) Xena website (https://xenabrowser.net/) to collect TGFBI data from the TCGA database including RNA-Seq data, clinical data, DNA methylation data, and stemness scores (Goldman et al., 2020). Strawberry Perl (http://strawberryperl.com/; version 5.32.0) was used to obtain the TGFBI expression data from the TCGA database and construct a data matrix for further analysis.

A comprehensive website TIMER2 (Tumor Immune Estimation Resource, http://timer.cistrome.org/; version 2) was used to systematic analysis the differential gene expression between different cancer types and normal tissues (Li et al., 2017). The “Gene_DE” module was used to compare TGFBI expression between various cancer types with adjacent normal tissues using TCGA data.

A comprehensive website UALCAN (http://ualcan.path.uab.edu/analysis.html) can analyze protein expression using data from the Clinical Proteomic Tumor Analysis Consortium (CPTAC) database and the TCGA database (Chandrashekar et al., 2017). The “CPTAC” module of UALCAN website was used to investigate TGFBI protein expression in various tumors and adjacent normal tissues. p < 0.05 was considered significant.

An interactive website GEPIA2 (Gene Expression Profiling Interactive Analysis, http://gepia2.cancer-pku.cn/#analysis; version 2) was used to analyze RNA-Seq expression data of cancer tissues and normal samples from the Genotype–Tissue Expression (GTEx) Project and the TCGA database (Tang et al., 2017). The “Expression Analysis-Box Plots” module of GEPIA2 was used to get box plots comparing TGFBI expression between cancer and normal tissues. Setting p value cutoff <0.01, log₂ (fold change) cutoff >1, and “Match TCGA normal and GTEx data.” In addition, we used the “Pathological Stage Plot” module to analyze TGFBI expression in different pathological stages of various tumors using TCGA data.

A publicly accessible platform Oncomine platform (www.oncomine.org) was used to analyze genome-wide expression, which, as of April 2021, contained 715 datasets and 86,733 samples (Rhodes et al., 2007). We used this platform to compare TGFBI expression between various cancer types and adjacent normal tissues using Student’s t-test. We set p value cutoff <0.05 and fold change cutoff >2, and selected the top 10% genes.

CCLE (Cancer Cell Line Encyclopedia, portals.broadinstitute.org/ccle/) database provides public access to genomic data, visualization and analysis for over 1,100 cancer cell lines.

The HPA database (https://www.proteinatlas.org) was used to map all human proteins at the cell, tissue, and organ levels by integrating various omics technologies. We used this database to explore TGFBI mRNA and protein expression data from various cancer types. We also obtained immunohistochemistry images of TGFBI protein expression in cancer tissues.

We used the “Survival Map” module in GEPIA2 to gain the overall survival (OS) and disease-free survival (DFS) data correlated with TGFBI expression across different cancer types from the TCGA database. The cases were split into high- and low-expression subgroups based on the median expression. The survival data were assessed basing on the Kaplan–Meier (KM) method, with the results being presented as hazard ratios, 95% confidence intervals, and p values of log-rank tests.

The PrognoScan database (http://dna00.bio.kyutech.ac.jp/PrognoScan/index.html) and Cox regression analysis (with p < 0.05 indicating significance) were used to verify the relationships between TGFBI expression and various survival outcomes in a pan-cancer analysis, including OS, DFS, relapse-free survival (RFS), disease-specific survival (DSS), distant metastasis-free survival (DMFS) and distant recurrence-free survival (DRFS) (Mizuno et al., 2009).

The “survival” and “survminer” R package were used to analyze the correlation of the TGFBI expression and OS, DSS, disease-free interval (DFI), and progression-free interval (PFI) across all TCGA tumors. We computed log-rank p-values and hazard ratios (HR) with 95% confidence intervals (95% CI). Data were visualized as forest plots (using the “forestplot” R package) and survival curves.

We used KM plotter (http://kmplot.com/analysis/), to assess the relationships between TGFBI expression and prognosis in various cancers from the TCGA and GEO databases. Breast, gastric, lung, ovarian, and liver cancer datasets were each split into high- and low-expression subgroups using the “autoselect best cutoff” option, and p < 0.05 was considered significant.

A comprehensive website cBioPortal (www.cbioportal.org) was used to explore, analyze, and visualize polydimensional cancer genomics data (Cerami et al., 2012). The TGFBI alteration frequencies, mutation types, and copy number alterations across the TCGA database were obtained using the “Cancer Types Summary” module of cBioPortal. To query the TGFBI genetic alteration characteristics, we set the “Quick select” field to “TCGA Pan Cancer Atlas Studies.”

The COSMIC (Catalog of Somatic Mutations in Cancer) website (https://cancer.sanger.ac.uk/cosmic/) is the largest and most comprehensive resource for exploring the impact of somatic mutations in human cancers (Tate et al., 2019). We used COSMIC to investigate the specific distribution of various TGFBI mutation types.

The Cancer Regulome Explorer (http://explorer.cancerregulome.org/) is a website that enables researchers to search, filter, and visualize analytical results obtained from the TCGA database. We used this website to investigate and visualize, at the chromosome level, the TGFBI mutation types in various cancer types, and the results were depicted in Circos plots. We set the pairwise correlation cutoff ≥0.4 and −log₁₀ (p value) cutoff ≥10.

Pearson’s correlation analysis was performed to analyze the correlations between TGFBI expression and immunoinhibitory and immunostimulatory gene subsets, TMB, and MSI. The results were displayed as heatmaps basing on the “pheatmap” package in R.

We used the “Immune_Gene” module of TIMER2 to evaluate the correlations of TGFBI expression with immune cell infiltration across diverse cancer types in the TCGA database. The immune cells comprised CD4⁺ T cells, CD8⁺ T cells, regulatory T cells (Tregs), B cells, myeloid-derived suppressor cells (MDSCs), neutrophils, dendritic cells (DCs), monocytes, macrophages, mast cells, cancer-associated fibroblasts (CAFs), natural killer (NK) cells, endothelial cells, and T follicular helper (Tfh) cells.

Tumor purity was assessed in 33 human cancers in the TCGA database basing on the “estimate” R package. Specifically, the ESTIMATE score is the sum of the immune and stromal scores, which represent the abundance of immune and stromal components, respectively (Yoshihara et al., 2013). Higher ESTIMATE scores correspond to lower tumor purity. Spearman’s correlation analysis was used to reveal the relationship between TGFBI expression and the immune, and stromal scores.

We also explored tumor RNA and DNA stemness scores (RNAss and DNAss) based on epigenetic and transcriptome data from the TCGA database. Specifically, RNAss is based on RNA-Seq data, and DNAss is based on DNA methylation data. Spearman’s correlation analysis was used to investigate the correlations of TGFBI expression with RNAss and DNAss.

The CellMiner tool (https://discover.nci.nih.gov/cellminer/home.do) was employed to obtain National Cancer Institute (NCI)-60 data on TGFBI expression (i.e., transcript data) and drug responses (i.e., drug sensitivity based on data on the drug concentration that reduces total cell growth by 50% [GI50]). The NCI-60 data comprises molecular and pharmacological data on 60 diverse human cancer cell lines. The drug response data includes data on drugs approved by the US Food and Drug Administration and those assessed in clinical trials. Pearson’s correlation analysis was employed to evaluate the correlations between TGFBI expression and drug responses.

The STRING (Search Tool for the Retrieval of Interacting Genes/Proteins) database (https://string-db.org/) contains known and predicted protein-protein interactions (PPIs) (Szklarczyk et al., 2019). We obtained the top 50 TGFBI-binding proteins using this database to construct a PPIs network. The parameters were set as follows: meaning of network edges: “evidence”; minimum required interaction score: “low confidence” (i.e., the line color represents the type of interaction evidence); and the maximum number of interactors to show: “no more than 50 interactors”. The PPIs network was visualized using STRING. Then the “Similar Gene Detection” module of GEPIA2 was used to obtain the top 100 TGFBI-correlated genes using TCGA data. Lastly, we conducted a Venn diagram analysis (http://jvenn.toulouse.inra.fr) to identify the common members of these two groups.

DAVID (Database for Annotation, Visualization, and Integrated Discovery, https://david.ncifcrf.gov/home.jsp; version 6.8) supplies a comprehensive, functional annotation tool to identify genes’ biological functions (Huang et al., 2009). It was used to conduct Kyoto Encyclopedia of Genes and Genomes (KEGG) and Gene Ontology (GO) enrichment analyses. The KEGG results were analyzed and visualized employing the “ggplot2” and “clusterProfiler” R packages. The GO biological process (BP), cellular component (CC), and molecular function (MF) results were obtained using the “cnetplots” R package.

TGFBI expression was compared between tumor and adjacent normal tissues employing the Oncomine database, with the results presented as p values, fold changes, and gene ranks. The survival results were presented as hazard ratios, 95% CI, and p values of log-rank tests. Using R version 4.0.4 (64-bit; https://www.r-project.org/) for the analyses. For all statistical analyses, p < 0.05 was considered statistically significant.

To compare TGFBI expression between tumor and normal tissues, we applied TIMER2 to analyze TGFBI expression in various cancer types of TCGA. Compared to matched normal tissues, TGFBI was upregulated in cholangiocarcinoma (CHOL; p < 0.01), colon adenocarcinoma (COAD), esophageal carcinoma (ESCA), GBM, head and neck squamous cell carcinoma (HNSC), kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), liver hepatocellular carcinoma (LIHC), rectum adenocarcinoma (READ), stomach adenocarcinoma (STAD), and thyroid carcinoma (THCA; p < 0.001). However, TGFBI was downregulated in kidney chromophobe (KICH; p < 0.001; Figure 1A).

Furthermore, we applied the Oncomine database to compare TGFBI expression between tumor and matched normal tissues. TGFBI was upregulated in colorectal cancer, esophageal cancer, gastric cancer, head and neck cancer, and lymphoma (Figure 1B). In addition, we explored TGFBI expression across different tumor cell lines in the CCLE database. As shown in Supplementary Figure S1, RNAseq showed that TGFBI was upregulated in several cell lines, including kidney cancer, chondrosarcoma, upper aerodigestive tract cancer, glioma, osteosarcoma, thyroid cancer, liver cancer and mesothelioma cell lines.

We further evaluated the difference in TGFBI protein expression between the tumor tissues and normal tissues in the CPTAC database using UALCAN tools. TGFBI protein was upregulated in colon cancer and clear cell renal cell carcinoma compared to normal tissues (p < 0.001) but downregulated in OV and uterine corpus endometrial carcinoma (UCEC; p < 0.001; Figure 1C).

We also applied GEPIA2 to analyze TGFBI expression in various pathological stages of multiple cancer types. TGFBI expression was associated with bladder urothelial carcinoma (BLCA), ESCA, KICH, KIRC and lung adenocarcinoma (LUAD; all Pr(>F) < 0.05; Figure 1D). We also compared TGFBI expression between these TCGA tumor tissues and the corresponding normal tissues in the GTEx database, which showed that TGFBI was upregulated in 19 types of cancer tissues and downregulated in three types compared to normal tissues (Supplementary Figure S2).

Using HPA data, we analyzed TGFBI protein expression. Analysis revealed that aberrant TGFBI protein expression was detected in 20 cancer types. The immunohistochemical results on TGFBI protein expression are shown in Figure 2.

We assessed the associations between TGFBI expression and cancer survival outcomes in the TCGA database using GEPIA2. High TGFBI expression was associated with poor OS for cervical squamous cell carcinoma and endocervical adenocarcinoma (CESC, p = 0.0024), GBM (p = 0.0082), HNSC (p = 0.013), KIRC (p = 0.0019), pancreatic adenocarcinoma (PAAD; p = 0.047), and uveal melanoma (UVM; p < 0.001; Figure 3A). In addition, high TGFBI expression was associated with poor DFS for CHOL (p = 0.03), COAD (p = 0.0095), KIRC (p = 0.0011), PAAD (p = 0.033), and UVM (p < 0.0029; Figure 3B).

We also applied the PrognoScan database to analyze the associations between TGFBI expression and cancer survival outcomes, mainly based on GEO datasets (GSE13507, GSE4475, GSE4271, GSE7849, GSE9893, GSE2034, GSE7390, GSE11595, GSE22138, GSE31210, GSE8894, and GSE30929). High TGFBI expression was associated in Cox regression analyses with poor DSS in bladder cancer (p = 0.031), OS in brain cancer (p = 0.047), DFS, DMFS, and RFS in breast cancer (BRCA; all p < 0.05), DMFS in eye cancer (p = 0.010), OS and RFS in lung cancer (p < 0.001), and DRFS in soft tissue cancer (p < 0.001). In contrast, high TGFBI expression was associated with better OS in blood cancer (p = 0.016), breast cancer (p = 0.045), and esophageal cancer (p = 0.047; Figure 3C).

We assessed the correlation between the expression of TGFBI and OS, DSS, PFI, and DFI in different types of cancer using a single-variate Cox regression analysis based on TCGA database. The results are summarized in forest plot (Figure 4). KM analysis showed that high TGFBI expression predicted poor prognosis of CESC, GBM, KIRC, LGG, testicular germ cell tumors (TGCT), and UVM (all p < 0.05) but good prognosis of adrenocortical carcinoma (ACC) and HNSC (all p < 0.05; Figure 5A). In addition, increased expression of TGFBI was associated with poor DSS in CESC, GBM, KIRC, LGG, PAAD, and UVM, whereas, increased TGFBI expression predicted good DSS in ACC (all p < 0.05, Figure 5B). The same method is used to analyze PFI in 33 TCGA tumors. TGFBI had a protective role in ACC. On the other hand, TGFBI had a detrimental role in BRCA, CESC, KIRC, PAAD, and UVM (all p < 0.05, Figure 5C). Finally, we also analyzed the DFI in 33 TCGA tumors. TGFBI expression had a detrimental role in BRCA, CHOL, and COAD (all p < 0.05, Figure 5D).

Then we studied the prognostic value of TGFBI expression was further evaluated using KM plotter. Notably, high TGFBI expression was significantly associated with poor DMFS, OS, post-progression survival (PPS), and RFS in breast cancer (p < 0.01; Supplementary Figure S3A). However, high TGFBI expression was associated with good DSS, OS, progression-free survival (PFS), and RFS in liver cancer (p < 0.05; Supplementary Figure S3B). High TGFBI expression was associated with poor first progression (FP), OS, and PPS in gastric cancer (p < 0.001; Supplementary Figure S3C). High TGFBI expression was associated with poor FP and OS but good PPS in lung cancer (p < 0.001; Supplementary Figure S3D). Moreover, high TGFBI expression was associated with poor OS, PFS, and PPS in OV (p < 0.05; Supplementary Figure S3E).

Using TCGA data in cBioPortal (10,967 samples from 32 studies), we assessed the TGFBI alteration frequency and mutation count in tumor samples. Melanoma had the highest frequent TGFBI alteration (>6%), with “mutation” as the primary type of alteration (Figures 6A,B). Data on the TGFBI alteration types, sites, and numbers of cases, including data on missense, nonsense (truncation), inframe, and fusion mutations, are shown in Figure 6C. The TGFBI mutation hotspot was R469C/H/S (missense mutation) in the FAS1 domain, which occurred in four cancers (UCEC, CESC, COAD, and skin cutaneous melanoma [SKCM]) in four patients.

Moreover, we explored the impact of somatic TGFBI mutations in human cancer. Using the COSMIC website, nonsense (truncation) mutations were found in lung cancer (10%); missense mutations were found in endometrial (7.14%), hematopoietic and lymphoid (40%), large intestine (15%), lung (20%), and skin (61.54%) cancer; and synonymous substitutions were found in breast (28.57%), endometrial (28.57%), and large intestine (10%) cancer. G > A was the primary mutation type in these cancer cases (Supplementary Figure S4).

Using the Cancer Regulome Explorer website, we investigated the relevant genomic correlation between TGFBI gene and certain signatures. DNA methylation, somatic copy number, somatic mutation, microRNA expression, and protein expression were showed to reveal the interrelation in different tumors using data from the TCGA database. The TGFBI was associated with above signatures in ACC, BLCA, BRCA, ESCA, STAD, GBM, HNSC, KIRC, LUAD, OV, SKCM, THCA, UCEC, lung squamous cell carcinoma (LUSC), and prostate adenocarcinoma (PRAD) (Supplementary Figure S5).

TGFBI is involved in immune cell infiltration and inflammatory responses, which play key roles in cancer initiation, progression, and metastasis. To investigate immune cell infiltration at the pan-cancer level, we used TIMER2.0 to explore the associations between TGFBI expression in human cancer and the infiltration of various types of immune cells (based on CIBERSORT, CIBERSORT-ABS, XCELL, MCPCOUNTER, QUANTISEQ, and EPIC algorithms). Overall, TGFBI expression was positively correlated with neutrophils, monocytes, macrophages, CAFs and MDSCs, and negatively correlated with B cells, T follicular helper (Tfh) cells, and CD8⁺ T cells. Thus, TGFBI expression may play vital roles in the immune cell infiltration process. The immune cell infiltration was quite different in OV, PRAD, UVM, and THCA due to the distinct TME components in the central nervous system (Supplementary Figure S6).

Tumor immunotherapy is a novel therapeutic strategy which been proven efficacious in multiple types of cancers. Thus, we explore whether TGFBI could be used as a novel target for tumor immunotherapy. As shown in Figure 7A, in most cancers, except CHOL, ESCA, sarcoma (SARC), SKCM and UCS, TGFBI expression was significantly correlated with immunoinhibitory genes. Additionally, except CESC, CHOL, mesothelioma (MESO), READ, SARC, UCEC and UCS, TGFBI expression was significantly correlated with immunostimulatory genes (Figure 7B).

We also explored the associations of TGFBI expression with TMB and MSI in various types of cancer. The TGFBI expression was negatively correlated with TMB in COAD, HNSC, LUSC and BLCA (all p < 0.05; Figure 7C). In addition, TGFBI expression was negatively correlated with MSI in LUSC, LUAD, COAD, and CHOL (all p < 0.05; Figure 7D).

Since our results have confirmed the immunoregulation role of the TGFBI in various types of cancer, it is vary needed to explore further the correlation between TGFBI expression and TME. We explored the correlations of TGFBI expression with TME components, using the ESTIMATE algorithm to calculate the ESTIMATE, stromal, and immune scores in diverse tumor types in the TCGA database. TGFBI expression was positively associated with stromal and immune scores in the pan-cancer analysis (all p < 0.05; Figures 8A,B). Furthermore, we assessed the correlations between TGFBI expression and tumor stemness in a pan-cancer analysis using TCGA data. TGFBI expression was negatively correlated with RNAss, and positively correlated with DNAss, in ACC, SARC, THCA, thymoma (THYM) and UVM (Figure 8A).

To investigate the potential relationships between TGFBI expression and drug responses in various types of human cancer, we performed a correlation analysis to identify potential drug candidates using CellMiner. TGFBI expression was positively correlated with staurosporine, dasatinib, midostaurin, and bleomycin sensitivity (all p < 0.05; Figures 8C,D,G,H), but negatively associated with lapachone, CUDC-305 byproduct, and fulvestrant sensitivity (all p < 0.05; Figures 8E,F,I).

To clarify the molecular mechanisms of TGFBI in tumorigenesis, we conducted a PPIs network analysis of TGFBI and enrichment analyses of TGFBI-binding proteins (based on STRING) that were also TGFBI-correlated genes (based on GEPIA2). The PPIs network, constructed using STRING, was in the base of experimental evidence and had 22 nodes and 34 edges (Figure 9A). We obtained 50 TGFBI-binding proteins supported by experimental evidence in STRING. Using GEPIA2, we then obtained the top 100 genes associated with TGFBI expression. Next, a Venn diagram analysis of these two groups showed one common member, namely, COL4A1 (Figure 9B). The heatmap in Figure 9C shows the positive correlation between TGFBI and COL4A1 expression.

To conduct KEGG and GO functional enrichment analyses of TGFBI, we used DAVID 6.8. The KEGG analysis showed that the “PI3K-Akt signaling pathway”, “Focal adhesion”, and “ECM–receptor interaction” might be involved in the effects of TGFBI on tumor pathogenesis (Figure 9D). The GO analysis showed similar results, as TGFBI was closely involved in a variety of terms, such as “extracellular matrix organization” in the BP category, “collagen-containing extracellular matrix” in the CC category, and “extracellular matrix structural constituent” in the MF category (Figure 9E).