The expression landscape and clinical significance of cancer-specific RNA transcripts across human cancers

Abstract

Background

Transcript-level analyses allow for the precise characterization of gene expression and its functional role in cancer. However, most of these studies rely on reanalyses of next-generation sequencing data, whose incomplete or inaccurate assemblies limit the comprehensive and faithful characterization of transcripts. To systematically elucidate transcriptomic expression in tumors and define broadly applicable therapeutic strategies, we investigated cancer-specific RNA transcripts (cancer-SRTs) expressed across multiple cancer types based on long-read sequencing data.

Methods

We characterized the expression profiles of 44,405 cancer-SRTs across multiple cancer types using t-SNE and correlation analyses. Transcripts expressed in more than 10 cancer types were further investigated through enrichment, survival, and correlation analyses to elucidate their functions and clinical relevance. To explore the mechanisms driving cancer-SRT generation, we analyzed alternative splicing events within these transcripts and integrated copy number variation, DNA methylation, and ATAC-seq data from matched TCGA tumor samples. Using the expression of 131 transcripts strongly associated with tumor hallmarks, we developed a risk-score model to evaluate associations with patient survival, tumor stage, immune characteristics, and responses to immune checkpoint blockade. Finally, the in vitro anti-tumor effects of siRNAs targeting two cancer-SRTs were evaluated using CCK-8 assay, colony formation, and transwell assays.

Results

Cancer-SRTs exhibit substantial structural diversity and are enriched in malignancy-associated pathways. The expression of these transcripts is associated with multiple genomic and epigenetic processes. We identify 131 transcripts that are strongly associated with tumor hallmarks and develop a risk-score model for evaluating patient prognosis and tumor progression. The model also exhibited strong associations with features of immune evasion.

Conclusions

Cancer-SRTs are widely expressed yet highly heterogeneous across tumor types, and are subject to multiple regulatory mechanisms underlying their functional and clinical significance. These findings advance our understanding of tumor biology and lay the groundwork for developing diagnostic, prognostic, and therapeutic strategies based on these transcripts. Future studies investigating their underlying mechanisms and applications in immunotherapy will be critical for precision cancer treatment.

Supplementary information

The online version contains supplementary material available at 10.1186/s12967-026-08498-z.

Background

More than 95% of human genes generate multiple transcripts, with an average of 8.6 per gene, adding great complexity to functional studies [1]. These transcripts display distinct, tissue-specific expression, and selective transcription is widespread across tissues [2, 3]. In tumors, genomic rearrangements, aberrant alternative splicing, trans-splicing, and RNA editing further expand isoform diversity, resulting in the generation of many tumor-specific variants [4]. Importantly, different isoforms of the same gene can perform distinct or even opposing functions [5]. For example, one study demonstrated that isoforms of MYC box-dependent interacting protein 1 (also known as bridging integrator 1 [BIN1] or amphiphysin 2) play distinct roles in normal liver tissue and hepatocellular carcinoma (HCC) [6]. In the liver, BIN1-S suppresses tumorigenesis by blocking c-Myc, whereas BIN1-L in HCC promotes oncogenesis by stabilizing polo-like kinase 1. These findings highlight the need for systematic transcript-level analyses that go beyond gene-level studies. Extensive efforts have been made to characterize isoform landscapes across cancers at the transcriptional level. Databases such as GEPIA2 [7], SRTdb [8], TAiC [9], and OncoDB [10] facilitate the exploration of transcript features in diverse tumors and tissues. Zatzman et al. [11] developed RNAmp, an algorithm that quantifies transcriptional overexpression in tumors, thereby revealing gene dysregulation and supporting patient selection for innovative treatments. Despite these advancements, most of these studies rely on reanalyses of next-generation sequencing data, which often fail to span full-length transcripts, leading to incomplete or inaccurate assemblies that limit the comprehensive and faithful characterization of transcripts.

To overcome these inherent limitations of next-generation sequencing, long-read sequencing (LR-seq) technologies such as PacBio and Nanopore have emerged in recent years, generating reads ranging from 10 kb to several megabases. By employing single-molecule real-time sequencing, LR-seq avoids PCR amplification and preserves native epigenetic modifications, thereby enabling direct capture of full-length transcripts without reference-based assembly. This approach enhances accuracy and supports more complete transcriptome annotation [12, 13]. Building on LR-seq data, our previous work [14] established FLIBase, the first comprehensive full-length transcriptome database spanning both human cancers and normal tissues, which revealed numerous unannotated cancer- and tissue-specific RNA transcripts (SRTs) and provided new opportunities to investigate their roles in cancer initiation and progression.

To systematically elucidate transcriptomic expression in tumors and define broadly applicable therapeutic strategies, we investigated cancer-SRTs expressed across multiple cancer types. We comprehensively analyzed the expression patterns, functions, and clinical relevance of these transcripts across cancer types, revealing their regulation by diverse genomic and epigenetic mechanisms. On the basis of the expression of 131 cancer-SRTs strongly associated with cancer hallmarks, we developed a risk-score model and constructed machine learning models, highlighting the translational potential of cancer-SRTs as biomarkers for cancer diagnosis, therapeutic targeting, and evaluation of immunotherapy response.

Materials and Methods

Data collection

Transcript expression data and transcript annotation information for 30 cancer types matched to The Cancer Genome Atlas (TCGA; https://portal.gdc.cancer.gov) were obtained from the FLIBase database (http://www.flibase.org/) [14]. Clinical phenotype data, copy number variation (CNV) data, and methylation 450K array data corresponding to TCGA samples were sourced from the Xena database (https://xenabrowser.net/datapages/). Immune infiltration levels of TCGA samples were acquired from the TIMER2.0 database (http://timer.comp-genomics.org/) [15]. We retrieved the following H3K27ac ChIP-seq datasets from Gene Expression Omnibus (GEO): 7 bladder urothelial carcinoma (BLCA) cell lines (5637, KK47, L1207, MGHU3, RT112, SCaBER, and SD48) [16] (GSE193886); 5 cervical squamous cell carcinoma and endocervical adenocarcinoma (CESC) cell lines (C33A, CaSki, HeLa, S12, and SiHa) [17] (GSE195631); and 9 lung squamous cell carcinoma (LUSC) cell lines (CALU1, EBC1, HARA, HCC2279, HCC2450, HCC2814, KNS62, LK2, and SQ1) [18] (GSE137459). Additionally, we acquired ChIP-seq data for YY1 [19] (GSE32465), ELK1 [19] (GSE91923), and H3K27ac [20] (GSE175131), along with ATAC-seq data [20] (GSE169955) for the LUAD cell line A549 from the ENCODE database. Global methylation levels of TCGA tumor samples, estimated from average methylation at LINE-1 sites, were obtained from a published study [21]. ATAC-seq data matched to TCGA samples were downloaded from the Genomic Data Commons (GDC; https://gdc.cancer.gov/about-data/publications/ATACseq-AWG) [22].

Definition of cancer-SRTs

Cancer-SRTs were defined by applying stringent expression-based criteria to guarantee tumor specificity, leveraging our previous study [14]. Cancer-SRTs detected by LR-seq were restricted to transcripts observed exclusively in tumor samples, with no detectable expression in normal tissues. In contrast, cancer-SRTs derived from RNA-seq were identified using the criteria described below:

1. For each transcript, the median expression level in tumor samples had to be at least 10-fold greater than the maximum expression level detected across normal tissues; normal tissue expression was evaluated against a comprehensive reference set encompassing all GTEx tissues except testis, plus adjacent normal tissues from The Cancer Genome Atlas (TCGA).

2. To ensure consistent tumor-associated expression, transcripts were required to be detectably expressed (TPM > 0.5) in more than 5% of tumor samples from at least one TCGA cancer type.

Following application of these criteria, 44,405 cancer-SRTs were identified. This approach ensured that transcripts classified as cancer-SRTs were markedly enriched in tumor tissues while being entirely absent from normal tissue profiles, minimizing the potential misclassification of tissue-restricted transcripts as cancer-specific.

T-distributed stochastic neighbor embedding (t-SNE) analysis

To perform t-SNE analysis in R (version 4.4.0) using the “Rtsne” package (version 0.17), we applied a multi-step preprocessing procedure to the expression data of 44,405 cancer-SRTs. First, expression values were log-transformed to stabilize variance and then standardized to a mean of 0 and a standard deviation of 1. These preprocessing steps improved data quality and comparability. Based on the processed data, t-SNE analysis was conducted to visualize sample relationships in a reduced-dimensional space.

Enrichment analysis

Pathway enrichment analysis and Gene Set Variation Analysis (GSVA) were applied to genes corresponding to cancer-SRTs expressed in more than 10 cancer types. These analyses were carried out using the R package “clusterProfiler” (version 4.12.0) [23]. Gene sets from Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), and Hallmark gene sets were sourced from the Molecular Signatures Database (MSigDB, https://www.gsea-msigdb.org/gsea/msigdb/index.jsp), while immune-related gene sets were obtained from the Immunology Database and Analysis Portal (ImmPort, https://immport.org/shared/home) [24]. Estimates of the proliferation index of TCGA samples [25] and the epithelial-mesenchymal transition (EMT) gene list [26] were derived from previously published studies.

Calculation of cancer-SRT burden and survival analysis

Cancer-SRT burden for each tumor sample was defined as the number of cancer-SRTs with TPM values exceeding 0.5. Using the R packages “survival” (version 3.6.4) and “survminer” (version 0.4.9), we generated Kaplan-Meier survival curves for patients across 30 cancer types. The “surv_cutpoint()” function from “survminer” identified the optimal cutoff for patient risk scores, enabling stratification into risk groups. We evaluated survival differences between these groups via the log-rank test. Additionally, univariate and multivariate Cox regression analyses were conducted using the “ezcox” package (version 1.0.4) to estimate hazard ratios (HRs) and their corresponding p-values within each cancer type.

Alternative splicing event analysis

Alternative splicing events (ASEs) were identified with SUPPA2 (version 2.3), a computational tool for transcriptome-wide splicing quantification. SUPPA2 classified ASEs into seven categories: exon skipping (SE), mutually exclusive exons (MX), alternative 5’ splice site (A5SS), alternative 3’ splice site (A3SS), retained intron (RI), alternative first exon (AFE), and alternative last exon (ALE). For each ASE type, the software inferred potential splice isoforms using transcript structures from GENCODE v35. PSI values were derived by comparing the relative abundances of isoforms that include or exclude specific splicing events.

Processing of CNV data, ATAC-seq, and methylation array data

The R package “biomaRt” (version 2.44.4) converted gene copy number loci from GISTIC2 thresholded copy number data into chromosomal positions. We then used BEDtools (version 2.31.1) [27] to extract overlapping sites among the gene copy number loci, processed ATAC-seq log2CPM data, and methylation array data within ±1000 bp of cancer-SRTs’ transcription start sites (TSSs). When transcripts corresponded to multiple copy number regions, the locus with the largest overlap was selected. For transcripts associated with multiple methylation probes, the probe exhibiting the smallest Spearman correlation coefficient was retained. Only autosomal overlapping sites were included in the analysis.

ChIP-seq data processing and motif enrichment analysis

Adapter-trimmed FASTQ files were aligned to the reference genome with BWA (version 0.7.18) [28], and the resulting SAM files were coordinate-sorted using Picard (version 3.2.0). These SAM files were then converted to BAM format via SAMtools (version 1.20). Potential PCR duplicates were removed using the MarkDuplicates tool in Picard. MACS2 (version 2.2.9.1) [29] was subsequently employed to identify genomic regions enriched for H3K27ac.

Gene set enrichment analysis (GSEA)

For each of the 30 cancer types, patients were stratified into high and low risk-score groups. Differential expression analysis of gene expression data for each cancer type was performed using the R package “easyTCGA” (version 0.0.4.2000), and genes were ranked according to their log fold change (logFC) values from high to low. Based on this ranking, GSEA was conducted using the R package “clusterProfiler”, and the normalized enrichment scores (NESs) for each pathway were extracted.

Selection of characteristic cancer-SRTs and construction of risk-score

Correlation analysis was conducted between the TPM values of cancer-SRTs expressed in more than 10 cancer types and the GSVA enrichment scores of their corresponding Hallmark gene sets. We applied screening thresholds of p < 0.05 and a correlation coefficient |R| > 0.6, which identified 131 significantly correlated characteristic cancer-SRTs.

To further explore how these 131 characteristic cancer-SRTs (S1, S2, S3, …, Sn) relate to tumor hallmark features, cancer subtypes, and patient survival, we performed univariate Cox regression using their TPM values as independent variables to derive regression coefficients (β). A risk-score was then constructed according to the following formula:

The gene-level risk-score was constructed using the same analytical framework and an analogous formula, derived from the host genes of the 131 characteristic cancer-SRTs (G1, G2, G3, …, Gn):

Concordance indices (C-indices) analysis

To quantitatively compare the prognostic performance of transcript-level and gene-level risk-score models, we assessed model discrimination using concordance indices (C-indices). For each cancer type, patients were stratified into high- and low-risk groups based on the optimal risk-score cut-point derived from survival analysis. This binary classification was applied consistently to both models to ensure comparability.

C-indices were calculated for each cancer type using the “concordance()” function from the “survival” R package, with overall survival time and status as outcome variables. The binary risk group was treated as an ordered factor, where the high-risk group was designated as the higher-risk category to ensure that larger risk values corresponded to a poorer prognosis. This explicit ordering prevented ambiguity in the directionality of the concordance estimation.

Calculation of EMT score

The EMT gene list was obtained from a previous study [26]. EMT scores for each tumor sample were calculated based on the TPM expression values of these EMT genes, and their correlation with the risk-score was subsequently assessed. The EMT score was calculated using the following formula:

Here, M and E represent the TPM expression values of mesenchymal and epithelial genes, respectively, while m and e denote the numbers of mesenchymal and epithelial genes, respectively.

Calculation of TIDE score

We calculated TIDE scores for 30 cancer types using the TIDE database (http://tide.dfci.harvard.edu/) [30]. The tumor expression matrices were standardized by normalizing each sample against the average expression across all samples within the respective study.

Random forest-based machine learning

TPM values for the 131 signature cancer-SRTs, along with corresponding cancer progression information, were obtained from patients representing 30 cancer types. These patients were randomly divided into training and independent test sets at a 3:1 ratio, with comparable distributions of cancer types in both groups. Random forest models were implemented with the R package “randomForest” (version 4.7.1.2). A fixed random seed was used during data partitioning to ensure reproducibility. The number of trees (ntree) was set to 500, a commonly used value that provides stable estimates of variable importance and prediction performance while maintaining computational efficiency. The remaining parameters were kept at the package defaults, including mtry and node size. No additional hyperparameter optimization was performed. This fixed hyperparameter configuration was intentionally adopted to minimize overfitting and to ensure consistency and comparability of model construction across different cancer types. We evaluated the predictive performance of this machine learning model, which was based on the 131 cancer-SRTs, by conducting receiver operating characteristic (ROC) curve analysis.

Cell lines and cell culture

All human liver cancer cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM; Gibco, Grand Island, NY, USA) supplemented with 10% fetal bovine serum (FBS; Gibco), 100 µg/mL penicillin (Gibco), and 100 µg/mL streptomycin (Gibco) in a humidified incubator at 37 °C with 5% CO₂. A549 cells were cultured in RPMI-1640 medium containing 10% FBS. HCT-116 cells were cultured in McCoy’s 5A Modified medium (Shanghai Basal Media Technologies Company) supplemented with 10% fetal bovine serum (FBS), 100 U/mL penicillin, and 100 μg/mL streptomycin. Prior to experiments, cells were treated with Mycoplasma-OUT (Genechem, Shanghai, China) for 1 week, followed by treatment with a Mycoplasma Test Kit (Genechem) to confirm the absence of mycoplasma contamination. All cells were authenticated via STR analysis.

RNA extraction, reverse transcription, and qRT-PCR analysis

Total RNA was isolated from cells using TRIzol reagent (Invitrogen, CA, USA). cDNA was synthesized using the PrimeScript^TM RT Reagent Kit (TaKaRa, Tokyo, Japan). Quantitative real-time polymerase chain reaction (qRT-PCR) was performed using SYBR Premix Ex Taq II (TaKaRa, Tokyo, Japan). The relative expression level of each target RNA was calculated using the 2^^(-ΔΔCt) method, with β-actin as the internal reference gene. Primer sequences are provided in Table S1.

RNA interference

Small interfering RNA (siRNA) oligonucleotides targeting the candidate genes were designed and synthesized by RiboBio (Guangzhou, China). Cells were transfected with the indicated siRNAs (final concentration: 20 nM) using Lipofectamine RNAiMAX reagent (Invitrogen, CA, USA). After 48 hours of transfection, cells were harvested for subsequent experiments, including RNA isolation, cell proliferation assays, transwell migration assays, and Western blotting. Sequences of the gene-specific siRNAs are provided in Table S1.

Cell proliferation, colony formation, and migration assays

Cell proliferation was assessed using the Cell Counting Kit-8 (CCK8; Dojindo, Kumamoto, Japan). Cells were seeded into 96-well plates at a density of 1,000 cells per well. After 1, 3, or 5 days of culture, the medium was replaced with 90 µl of fresh medium containing 10 µl of CCK-8 solution, followed by incubation at 37 °C for 2 hours. Absorbance was measured at 450 nm using a microplate reader.

For the colony formation assay, 1.0 × 10³ MHCCLM3, 1.5 × 10³ Huh-7, 1.0 × 10³ HCT116 and 1.0 × 10³ A549 cells were seeded into 6-well plates and incubated at 37 °C for approximately 1–2 weeks. Colonies were fixed with absolute methanol for 10 minutes and stained with 0.5% crystal violet solution (Sigma-Aldrich, St. Louis, MO, USA) for 30 minutes. After washing with distilled water, colonies were counted and quantified.

Transwell migration assays were conducted using 24-well plates with 8-µm pore size chambers (Corning, NY, USA). A total of 5 × 10⁴ cells were resuspended in 200 µL of serum-free medium and added to the upper chamber, while 800 µL of complete medium was added to the lower chamber. After incubation at 37 °C for a specified time, the chambers were fixed with absolute methanol and stained with 0.5% crystal violet solution for 30 minutes. Non-migrated cells on the upper surface of the membrane were wiped off with a cotton swab. Migrated cells on the lower surface were imaged and counted under an inverted light microscope (Olympus, Tokyo, Japan).

Western blotting

Proteins were separated by 10% SDS-PAGE and transferred onto a 0.45 μm nitrocellulose membrane (GE, CT, USA). Membranes were blocked with 5% non-fat milk in TBST for 30 minutes at room temperature, then incubated with primary antibodies overnight at 4 °C. After three washes with TBST, membranes were incubated with HRP-conjugated secondary antibodies for 1 hour at room temperature. Protein bands were visualized using the LumiBest ECL Reagent Solution kit (Share-Bio, Shanghai, China). The primary antibodies, including anti-YY1 (66281–1-lg, Proteintech, China) and anti-β-actin (66009–1-lg, Shanghai, China), were used.

Statistical analysis

Differences between the two groups were assessed with the Wilcoxon rank-sum test, and multiple comparisons were adjusted using the Benjamini–Hochberg method. Kaplan–Meier survival curves were generated with the R package “survminer”, and overall survival differences between groups were compared via the log-rank test implemented in the R package “survival”. Immune response rates between high and low risk-score groups were compared using the chi-square (χ²) test. Associations among different groups were evaluated by Spearman’s rank correlation analysis. All statistical analyses were performed using SPSS (IBM, NY, USA) and R, with a significance threshold of p < 0.05.

Results

The expression landscape of cancer-SRTs across human cancers

First, we classified 44,405 cancer-SRTs from the FLIBase database according to the coding potential of their corresponding genes (Fig. S1A). The majority (84.57%) originated from protein-coding genes, whereas 13.04% originated from lncRNAs, and 1.41% and 0.98% originated from pseudogenes and other categories, respectively. Next, we examined transcript boundary features (Fig. S1B), including CAGE peaks at transcription start sites (TSSs) and polyadenylation (polyA) sites at the 3’ ends. Overall, 58.49% of the cancer-SRTs displayed detectable CAGE peaks, and 80.23% contained polyA sites. These findings indicate that cancer-SRTs generally exhibit high transcript integrity and maturation.

On the basis of optimal alignment with reference-annotated transcripts, we quantified both the proportion (Fig. S1C) and number (Fig. 1A) of different classes of cancer-SRTs across cancer types. Novel not in catalog (NNC) and novel in catalog (NIC) isoforms were the most abundant, underscoring their greater structural novelty relative to reference annotations [31] and their potential involvement in diverse splicing events. On average, approximately 1,480 cancer-SRTs were expressed in each cancer type, with stomach adenocarcinoma (STAD) exhibiting the highest cancer-SRT count (11,099) and thyroid carcinoma (THCA) exhibiting the lowest (974).

Fig. 1.

The expression landscape of cancer-SRTs across human cancer types. A. The number of cancer-SRTs identified per cancer type. Transcripts were categorized into five groups according to their best-matching reference transcript: novel in catalog (NIC), novel not in catalog (NNC), full-splice match (FSM), antisense, and intergenic. The cancer types are abbreviated as follows: ACC, adrenocortical carcinoma; BLCA, bladder urothelial carcinoma; BRCA, breast invasive carcinoma; CESC, cervical squamous cell carcinoma and endocervical adenocarcinoma; CHOL, cholangiocarcinoma; COAD, colon adenocarcinoma; ESCA, esophageal carcinoma; GBM, glioblastoma multiforme; HNSC, head and neck squamous cell carcinoma; KICH, kidney chromophobe; KIRC, kidney renal clear cell carcinoma; KIRP, kidney renal papillary cell carcinoma; LGG, brain lower grade glioma; LIHC, liver hepatocellular carcinoma; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; MESO, mesothelioma; OV, ovarian serous cystadenocarcinoma; PAAD, pancreatic adenocarcinoma; PCPG, pheochromocytoma and paraganglioma; PRAD, prostate adenocarcinoma; READ, rectum adenocarcinoma; SARC, sarcoma; SKCM, skin cutaneous melanoma; STAD, stomach adenocarcinoma; THCA, thyroid carcinoma; THYM, thymoma; UCEC, uterine corpus endometrial carcinoma; UCS, uterine carcinosarcoma; UVM, uveal melanoma. B. Global expression patterns of 44,405 cancer-SRTs visualized via t-distributed stochastic neighbor embedding (t-SNE). Sample positions indicate transcriptional similarity, and samples are colored according to their assigned analysis cohort. C. Counts of cancer-SRTs and log₂-normalized TPM expression levels across cancer types. In the heatmap, the color intensity (red gradient) corresponds to the TPM value, and cancer types are arranged according to the number of expressed cancer-SRTs. D. Expression proportions of cancer-SRTs detected in more than 10 cancer types. Proportions represent the percentage of expressed cancer-SRTs relative to the total set of 1,232 cancer-SRTs expressed in at least 10 cancer types, with cancer types ordered by expression proportion. E. Bubble plots displaying the results of GO (left), KEGG (middle), and Hallmark gene set (right) enrichment analyses of genes corresponding to cancer-SRTs expressed in more than 10 cancer types. Bubble size reflects the number of enriched genes, and color intensity represents −log₁₀(FDR), with darker red indicating smaller FDR values

To investigate the expression patterns of cancer-SRTs across different cancer types, we performed t-distributed stochastic neighbor embedding (t-SNE) analysis (Fig. 1B). Cancer-SRT expression showed both cancer-type specificity and broad expression across multiple cancers (Fig. 1C), with highly consistent patterns within each cancer type (Fig. S1D). To identify broadly expressed cancer-SRTs as candidate biomarkers or therapeutic targets across multiple cancers, we selected transcripts expressed in more than 10 cancer types. This yielded 1,232 transcripts from 937 genes (Fig. S1E, Table S2). The percentage of tumors expressing these transcripts varied substantially, averaging 45.63%, with bladder urothelial carcinoma (BLCA) showing the highest expression (89.45%) and kidney chromophobe (KICH) the lowest (4.87%) (Fig. 1D). Functional annotation of the corresponding genes revealed significant enrichment in cancer-related pathways, including cell cycle regulation, E2F targets, the p53 signaling pathway, and epithelial–mesenchymal transition (EMT) (Fig. 1E).

Clinical significance of the cancer-SRT burden across human cancers

To investigate the effects of cancer-SRTs on tumorigenesis and progression, we performed gene set variation analysis (GSVA) to examine the associations between cancer-SRT burden (defined as the number of cancer-SRTs with TPM values > 0.5 in each tumor sample) and Hallmark pathways (Fig. 2A). The cancer-SRT burden showed significant positive correlations with cell cycle–related pathway activity in most cancer types, suggesting that an elevated burden promotes tumor cell proliferation and accelerates disease progression. The tumor immune microenvironment plays a critical role in tumorigenesis, tumor progression, and the response to immunotherapy [15, 24]. We next explored the links between the cancer-SRT burden and the tumor immune microenvironment. Correlation analysis revealed that high cancer-SRT burden was significantly negatively associated with most immune-related pathways (Fig. 2B). Immune infiltration analysis further revealed positive correlations with regulatory T (Treg) cells and M0 macrophages but negative correlations with M1 and M2 macrophages (Fig. 2C). To evaluate the clinical relevance of these cancer-SRTs, patients were stratified by cancer-SRT burden across various cancer types. A high cancer-SRT burden was significantly associated with poor survival in patients with 12 cancer types (Fig. 2D, Fig. S2A). In addition, the cancer-SRT burden correlated with tumor stage in 9 cancer types, with advanced-stage tumors having a greater burden (Fig. 2E, Fig. S2B). Collectively, these findings highlight cancer-SRT burden as a potential indicator of aggressive disease and adverse prognosis.

Fig. 2.

Clinical relevance of cancer-SRT burden across human cancer types. A. Correlations between cancer-SRT burden (defined as the number of cancer-SRTs with TPM values > 0.5 in each tumor sample) and hallmark gene set GSVA scores across cancer types are shown. Red and blue indicate positive and negative correlations, respectively, and color intensity corresponds to the magnitude of the correlation coefficient. B-C. Correlations between the cancer-SRT burden and the GSVA scores of representative immune pathways (B) and the infiltration levels of 22 immune cell types (C). Dot size represents the correlation coefficient; red and blue denote positive and negative correlations, respectively. The color intensity reflects the p value, with darker shades indicating smaller values. D. Kaplan–Meier curves comparing overall survival in patients with high (red) and low (blue) cancer-SRT burdens. E. The cancer-SRT burden distribution across different cancer stages

Genomic and epigenetic regulation of cancer-SRTs in human cancer

To explore the mechanisms underlying cancer-SRT generation, we first examined genomic alterations. Copy number variation (CNV), a major class of structural variation with a higher mutation rate than single nucleotide polymorphisms (SNPs), is strongly linked to cancer initiation and progression [32–34]. Analysis revealed that the expression of most cancer-SRTs was significantly positively correlated with gene copy number at their genomic loci (Fig. 3A). We next investigated epigenetic regulation. DNA methylation, which is typically associated with transcriptional repression [35], was higher in samples lacking cancer-SRT expression than in those expressing them (Fig. 3B). A subset of cancer-SRTs also showed negative correlations with CpG methylation levels (Fig. S4A–C). Furthermore, global tumor methylation, estimated using LINE-1 sites [21], was inversely correlated with cancer-SRT burden across most cancer types (Fig. 3C, Fig. S4D), supporting a role for DNA hypomethylation in inducing cancer-SRT expression.

Fig. 3.

Genomic and epigenetic regulation of cancer-SRTs in human cancers. A. Hierarchical clustering heatmap depicting the correlations between cancer-SRT expression and copy number variations (CNVs). Red and blue represent positive and negative correlations, respectively. B. Differences in methylation levels between samples expressing and not expressing cancer-SRTs. A TPM value > 0.5 is considered the expression threshold. *p < 0.05, **p < 0.01, ***p < 0.001. C. Correlations between the cancer-SRT burden and the global methylation level in different cancer types according to TCGA data, estimated by the average methylation level at LINE-1 loci. Each point represents a tumor sample and is colored according to its cancer type. D. Correlations between cancer-SRT expression levels and ATAC-seq signal values in matched tumor samples. Red indicates a positive correlation, blue indicates a negative correlation, and the color intensity reflects the magnitude of the correlation coefficient. E. Proportion of expressed (TPM > 0) versus nonexpressed (TPM = 0) cancer-SRTs possessing H3K27ac peaks in BLCA, CESC, HCC, and LUSC. F. The top five transcription factor motifs enriched for cancer-SRTs in LUSC. Sequence logos illustrate nucleotide conservation, with taller stacks and larger letters indicating higher conservation. Bar plots compare motif occurrence frequencies in target versus background sequences, with p values indicating the statistical significance of enrichment. G. Heatmaps and average profiles of YY1 binding, chromatin accessibility (ATAC-seq), and H3K27ac signals around transcription start sites (TSSs). Genome-wide distributions of YY1 ChIP-seq (red), ATAC-seq (blue), and H3K27ac ChIP-seq (purple) signals are shown across ±3 kb regions centered on cancer-SRT TSSs in LUSC. The heatmaps display the locus-specific signal intensities, whereas the line plots show the aggregated mean enrichment. H. GSEA of YY1-bound cancer-SRTs in LUSC. Transcripts are ranked by the log₂fold change in TPM values between si-NC and si-YY1 samples

Cancer cells frequently regulate chromatin accessibility through chromatin remodeling, thereby promoting tumor initiation and progression [36]. To determine whether the generation of cancer-SRTs is associated with chromatin accessibility, we analyzed chromatin accessibility profiles using ATAC-seq data from TCGA. Compared with samples not expressing cancer-SRTs, samples expressing cancer-SRTs displayed significantly higher ATAC-seq signals (Fig. S3A). Furthermore, a substantial number of transcripts were positively correlated with local chromatin accessibility (Fig. 3D), suggesting that chromatin openness plays a regulatory role in the generation of cancer-SRTs. Analysis of H3K27ac ChIP-seq data for BLCA, CESC, LUSC, and HCC revealed that transcripts with H3K27ac peaks at transcription start sites (TSSs) were significantly more likely to be expressed (Fig. 3E), suggesting that histone modifications contribute to the transcriptional activation of cancer-SRTs. In LUSC, where overlap was most pronounced, motif enrichment revealed the top 5 motifs enriched in LUSC H3K27ac peaks (Fig. 3F). YY1, as the top transcription factor, bound the promoters of many cancer-SRTs, which also exhibited elevated H3K27ac and increased chromatin accessibility (Fig. 3G). GSEA further revealed that cancer-SRT expression was correlated with YY1 levels (Fig. 3H). Given the dual role of YY1 in transcriptional activation and repression [37], its regulatory impact likely varies across targets.

Finally, we examined alternative splicing, a key mechanism that generates transcriptomic diversity [38]. Among these cancer-SRTs, we identified 38,368 alternative splicing events (ASEs), with an average of 31 events per transcript (Fig. S3B). Among the seven splicing types, alternative 3’ splice site (A3) and alternative 5’ splice site (A5) events were the most frequent, which is consistent with their known involvement in multiple malignancies, including acute myeloid leukemia (AML) [39], uveal melanoma (UVM) [40], chronic lymphocytic leukemia (CLL) [41], and medulloblastoma [42]. These findings highlight extensive splicing plasticity as a critical feature of cancer-SRTs.

Construction of the cancer-SRT signature for risk evaluation

To identify cancer-SRTs associated with fundamental tumor processes across cancers, we performed GSVA to assess correlations between transcript expression and Hallmark pathways across cancer types (Fig. 4A). This analysis yielded 131 cancer-SRTs with strong associations (|R| > 0.6, p < 0.05), most of which correlated positively with proliferation-related pathways (Fig. 4B, Table S3). These findings suggest that these 131 cancer-SRTs may actively promote tumor cell growth and cancer progression. We next examined their clinical relevance. A risk-score was calculated for each patient on the basis of these featured cancer-SRT expression data and univariate Cox regression results. Patients were then stratified into high- and low-risk groups, revealing that a high risk-score was significantly associated with poor survival across 27 cancer types (Fig. 4C, Fig. S5). The results of the multivariate Cox regression analysis revealed that among age, gender, and tumor stage, risk-score remained an independent prognostic factor for nearly all cancer types (Fig. 4D, Fig. S6).

Fig. 4.

Construction of a cancer-SRT signature for risk evaluation. A–B. Correlations between TPM values of the 131 signature cancer-SRTs and GSVA scores of Hallmark gene sets across cancer types. Hallmark gene sets were categorized into five groups: development, DNA damage, metabolism, pathway, proliferation, and signaling. Red and blue denote positive and negative correlations, respectively. In the chord diagram (A), cancer-SRTs are ranked by the number of associated pathways and cancer types in which they are expressed. In the heatmap (B), the color intensity corresponds to the magnitude of the correlation coefficient. C. Kaplan–Meier survival curves comparing high-risk-score (red) and low-risk-score (blue) groups across cancer types. D. Forest plot displaying hazard ratios (HRs) with 95% confidence intervals (CIs) and p values for risk-score, age, sex, and stage across cancer types. HR and p values were not computed for CESC, as all patients were female, or for GBM, because stage information was missing. The gray dashed line indicates HR = 1. *p < 0.05, **p < 0.01, ***p < 0.001

To determine whether the transcript-level cancer-SRT signature offers prognostic value beyond conventional gene-level expression, we compared the concordance indices (C-indices) of risk models based on the 131 cancer-SRTs with models constructed from their corresponding host genes. The transcript-level models consistently achieved higher C-indices in most cancer types (Fig. S7), indicating that the cancer-SRT signature reveals clinical heterogeneity obscured by gene-level quantification.

Clinical implications associated with the cancer-SRT risk-score in human cancers

To explore the biological basis of the risk-score and its link to prognosis, we analyzed its distribution across tumor stages. Significant differences were observed across the 11 cancer types, with higher risk-scores frequently associated with advanced-stage disease (Fig. 5A, Fig. S8). GSEA of Hallmark pathways revealed significant enrichment of cancer-related pathways, especially proliferation- and EMT-related pathways, in the high risk-score group (Fig. 5B), indicating that elevated risk-scores reflect aggressive tumor behavior.

Fig. 5.

Clinical implications associated with the cancer-SRT risk-score in human cancers. A. Distribution of risk-score across different cancer stages. B. Normalized enrichment scores (NESs) from GSEA using Hallmark gene sets and correlations of risk-score with the EMT score and proliferation index across cancer types. Boxes highlight pathways with p < 0.05 per cancer type, and the bubble size corresponds to the correlation coefficient. The color reflects the p value, with darker hues indicating greater significance. C. Heatmap showing disparities in immune cell infiltration abundance between the high and low risk-score groups across cancer types. “+” indicates significantly greater infiltration in the high risk-score group (FC > 1.5, p < 0.05), and “–” indicates significantly reduced infiltration (FC < 0.67, p < 0.05). Only immune cell types showing significant differences in at least one cancer type are included. D. Divergence in immunotherapy response between the high and low risk-score groups across cancer types. “Responder” refers to patients who responded to immunotherapy, and “Nonresponder” refers to those who did not respond. Only cancer types exhibiting significant p values and reduced response rates in the high risk-score group are shown. E. Differences in TIDE scores between high and low risk-score groups within the same cancer types in (D). Elevated TIDE scores suggest a higher probability of immune evasion following immunotherapy. *p < 0.05, **p < 0.01, ***p < 0.001. F. Receiver operating characteristic (ROC) curves for cancers with area under the curve (AUC) values of 0.7–0.8 (left), 0.8–0.9 (middle), and above 0.9 (right). Cancers with AUC values between 0.7 and 0.8 include GBM, KICH, READ, THYM, and UVM; those with AUC values between 0.8 and 0.9 include ACC and KIRC; and those with AUC values exceeding 0.9 include CHOL

Given the role of tumor proliferation in modulating immune responses, we compared immune cell infiltration between the high and low risk-score groups. Although the fold change in immune cell infiltration between the two groups varied across cancer types, patients with high risk-scores generally exhibited increased infiltration of immunosuppressive cells, such as Treg cells and M0 macrophages (Fig. 5C), suggesting an association with a potentially suppressed effector immune state. Using the TIDE database [30], TIDE analysis further revealed that these patients had higher TIDE scores and lower response rates to immune checkpoint blockade (ICB) therapy across eight cancer types (Figs. 5D, 5E), which aligns with an immune-evasive phenotype.

We next evaluated the impact of the risk-score on tumor progression. Patients with high risk-scores had significantly shorter progression-free survival (PFS) in 29 cancer types (Fig. S9), suggesting that an elevated risk-score is linked to an increased likelihood of disease progression. To further explore the predictive capacity of these biomarkers, we developed a machine learning model using the expression levels of 131 signature cancer-SRTs to estimate the probability of cancer progression. The model demonstrated robust predictive performance across multiple cancers, with an AUC value > 0.7 in eight cancer types. Notably, the AUC value for cholangiocarcinoma (CHOL) was > 0.9; those for adrenocortical carcinoma (ACC) and kidney renal clear cell carcinoma (KIRC) ranged from 0.8–0.9; and those for glioblastoma multiforme (GBM), kidney chromophobe (KICH), rectum adenocarcinoma (READ), thymoma (THYM), and uveal melanoma (UVM) ranged from 0.7–0.8 (Fig. 5F), underscoring the promise of transcript-level features in risk stratification within the TCGA cohort.

Functional characterization of cancer-SRTs in cancers

To investigate the functional role of cancer-SRTs, we employed hepatocellular carcinoma (HCC) as a primary experimental model and applied a stepwise prioritization strategy. Among the 131 featured cancer-SRTs, 74 were detected in HCC and were considered as initial candidates. We first evaluated their expression patterns and clinical associations in HCC, including transcript abundance and prognostic relevance, which yielded 40 candidate transcripts with potential clinical significance. These candidates were subsequently aligned with their corresponding reference annotations. Transcripts exhibiting clear sequence consistency or intron-retention characteristics were prioritized to facilitate transcript-specific siRNA design and minimize potential off-target effects. We further examined their expression in RNA-seq datasets from HCC cell lines using the Integrative Genomics Viewer (IGV) and selected 10 transcripts with relatively high endogenous expression for experimental evaluation. Baseline expression levels of these candidates were then assessed by qPCR across multiple HCC cell lines. Five transcripts with consistently detectable expression were subjected to transcript-specific knockdown experiments. Among them, TCONS_00892847 (locus: LINC02542) and TCONS_00392001 (locus: METTL26) exhibited favorable expression characteristics, showed associations with patient prognosis, demonstrated robust endogenous expression in HCC cell lines, and achieved efficient transcript-specific knockdown (Fig. 6A–B, Fig. S10). Therefore, these two transcripts were selected for subsequent functional characterization.

Fig. 6.

Functional characterization of cancer-SRTs in cancer. A. Genomic structures and loci of cancer-SRTs TCONS_00892847 and TCONS_00392001 and their respective host genes. B. Survival analysis (left) and expression profiles (right) of TCONS_00892847 and TCONS_00392001 in the TCGA-LIHC cohort and 50 paired tumor and adjacent normal tissue samples. C. Relative mRNA expression levels of TCONS_00892847 and TCONS_00392001 in Huh-7 (left) and MHCCLM3 (right) cells after siRNA-mediated silencing, as determined by qRT‒PCR. D. CCK-8 assays in Huh-7 and MHCCLM3 cells transfected with siRNA targeting the indicated transcripts or with siNC (negative control). E. Colony formation assays (with representative images) in Huh-7 and MHCCLM3 cells transfected with siRNAs targeting the indicated transcripts or with siNC. The bar graphs show the quantification of the results of the colony formation assay. F. Transwell migration assays of Huh-7 and MHCCLM3 cells were performed after the cells were transfected with siRNAs targeting the indicated transcripts. Representative migration images and statistics in the bar graphs are shown. The data are presented as the mean ± SEM; n = 3 in (D-F). One-way ANOVA, * p < 0.05; ** p < 0.01; ***p < 0.001

TCONS_00892847 and TCONS_00392001 were silenced in HCC cells using siRNA-mediated knockdown (Fig. 6C), and their biological functions were evaluated through a series of in vitro assays, including CCK-8 proliferation assays, colony formation assays, and Transwell migration assays. Consistent results across these experiments demonstrated that silencing these cancer-SRTs markedly suppressed HCC cell proliferation, clonogenic capacity, and migratory ability (Fig. 6D–F), indicating their critical roles in HCC development and progression. We also evaluated the functions of these transcripts in other cancer types. siRNA-mediated knockdown of the selected transcripts was performed in the lung cancer cell line A549 and the colorectal cancer cell line HCT116. Consistent with the results observed in HCC cells, silencing these cancer-SRTs markedly impaired cell viability, colony formation, and migration (Fig. S11). These findings suggest that the tumor-promoting functions of these cancer-SRTs may be conserved across multiple cancer types, although context-dependent variability among tumor types cannot be excluded.

Among the two validated cancer-SRTs, TCONS_00392001 was selected for further investigation because it is predicted to encode a protein and displayed reproducible oncogenic phenotypes across multiple assays. To gain preliminary insight into the downstream effects of TCONS_00392001, we performed RNA sequencing in Huh-7 cells following its knockdown. Gene set enrichment analysis (GSEA) revealed that, among all downregulated pathways, the Wnt/β-catenin signaling pathway exhibited the most significant enrichment (Fig. S12A–B). Multiple genes associated with the Wnt/β-catenin signaling pathway were markedly downregulated (Fig. S12C). The seven genes showing the largest fold changes were further validated in an independent HCC cell line (MHCCLM3), in which their reduced expression upon TCONS_00392001 knockdown was similarly observed (Fig. S12D). Given the established role of Wnt/β-catenin signaling in promoting HCC growth and metastasis [43, 44], the coordinated downregulation of multiple pathway-associated genes following TCONS_00392001 knockdown suggests that this transcript may contribute to HCC progression, at least in part, through modulation of Wnt/β-catenin-related transcriptional programs. Additional mechanistic studies will be required to determine the association between TCONS_00392001 and Wnt/β‑catenin signaling and whether this transcript directly regulates pathway activity.

Discussion

Compared with genomic-level analyses, transcript-level analyses offer a more precise delineation of gene expression specificity. Building on our previous work integrating LR-seq and RNA-seq, we constructed a comprehensive full-length transcriptome database and identified cancer-SRTs in tumors [14]. In this study, by integrating expression, regulatory, and clinical analyses, we revealed that cancer-SRTs constitute a highly dynamic and functionally relevant layer of the cancer transcriptome. These findings extend prior transcript-level investigations that were largely limited by short-read sequencing and underscore the importance of full-length transcript analyses for understanding tumor biology. Recent LR-seq studies have revealed numerous previously undetectable transcriptomic alterations in cancer. For instance, LR-seq enabled the discovery of 18 novel neoantigens in glioma cell lines that were not detectable by next-generation sequencing [45], underscoring the utility of this approach in the development of targeted therapies. In another study, Chen et al. [46] performed single-molecule real-time long-read RNA sequencing on patient-derived HCC cells and discovered 362 previously unannotated alternative splicing variants. In this study, we provide a systematic delineation of cancer-SRTs across cancer types and elucidate their structural and functional diversity. Cancer-SRTs are widely expressed yet highly heterogeneous across tumor types. While many arise from protein-coding genes, a substantial proportion originate from lncRNAs and pseudogenes, with most classified as NIC or NNC, revealing a hidden layer of transcriptomic complexity.

Mechanistically, we identified multiple genomic and epigenetic processes that contribute to cancer-SRT expression. Copy number variation, DNA hypomethylation, chromatin accessibility, and histone acetylation were strongly associated with transcript generation. In parallel, the high prevalence of alternative splicing events highlights splicing plasticity as a driver of transcriptomic diversity in cancer, which are consistent with prior observations in both solid and hematologic malignancies. Importantly, functional enrichment analyses revealed that broadly expressed cancer-SRTs participate in key oncogenic pathways, including cell cycle regulation, p53 signaling, and epithelial–mesenchymal transition, which is consistent with active tumor cell states [47]. These results suggest that cancer-SRTs are not transcriptional noise but are integral to malignant progression.

Clinically, the cancer-SRT burden correlated positively with proliferation-related pathway activity, immunosuppressive features in the tumor microenvironment, advanced tumor stages, and poor overall survival across diverse cancer types. These results indicate that the cancer-SRT burden may serve as a global indicator of tumor progression and adverse prognosis. Studies have indicated that Treg cells (regulatory T cells) can suppress the maturation and differentiation of macrophages, resulting in a higher proportion of macrophages remaining in the M0 stage rather than differentiating into M1 or M2 subsets, thereby contributing to an immunosuppressive effect [48, 49]. In addition, Tregs can inhibit the function of effector immune cells through multiple mechanisms, representing key factors in tumor immune evasion [50]. The enrichment of Treg cells and depletion of M1/M2 macrophages in tumors with high cancer-SRT burdens highlight a close link between cancer-SRTs and the tumor immune landscape, reflecting a microenvironment characteristic of immune evasion and limiting the efficacy of immunotherapy.

Furthermore, accumulating evidence indicates that the parental genes of certain cancer-SRTs contribute to tumor immune evasion and therapy resistance. Zeng et al. [51] demonstrated that EP300 facilitates immune escape by modulating the SOCS1–STAT1 signaling pathway, thereby impairing antigen presentation and suppressing CD8⁺ T-cell activity. Consistent with this notion, LINC00460 fosters an immunosuppressive tumor microenvironment through the upregulation of MYC, CD47, and PD-L1 [52]. In the context of therapy resistance, ADAR1 promotes chemoresistance by stabilizing SCD1 mRNA via RNA editing, leading to activation of JAK/STAT signaling and enhances tumor self-renewal capacity [53]. Collectively, these findings indicate that cancer-SRTs may serve as potential biomarkers that reflect the complex and dynamic interplay between tumors and the immune system.

To obtain these insights, we constructed a transcript-based risk-score from 131 signature cancer-SRTs associated with malignancy. This score served as an independent prognostic factor across multiple tumor types, stratifying patients by survival, tumor stage, and immune features and predicting reduced responses to immune checkpoint blockade. Importantly, a machine learning-based model further achieved promising preliminary performance in predicting disease progression, which provides a conceptual framework for using transcript-level biomarkers in future precision oncology, although its clinical translation requires further validation. In multiple cancer types, functional studies of these signature cancer-SRTs remain limited. For instance, in liver cancer, only a few cancer-SRTs, such as PRAME and DLEU1, have been linked to EMT and invasion [54, 55], providing indirect support for the functional importance of the other cancer-SRTs identified in this study. For another subset of cancer-SRTs, their corresponding genes are known cancer-associated genes. HMMR has been shown to promote immune evasion by forming a complex with FAK/SRC, thereby activating NF-κB signaling, which sustains CD47-mediated “don’t eat me” signaling and suppresses antitumor immunosurveillance [56]. Similarly, ASPM drives tumor growth by interacting with FOXM1 through liquid-liquid phase separation, enhancing FOXM1 stability and establishing a double-positive feedback loop that co-activates oncogenic transcription [57]. However, this study reveals novel transcript variants of these genes. The specific functions of these novel transcripts may differ from those of known canonical transcripts. Finally, the functions of approximately half of these cancer-SRTs in liver cancer remain unknown. Functional assays of two intron-retained cancer-SRTs revealed that their silencing suppressed cell proliferation, colony formation, and migration in multiple cancer types. These findings provide direct experimental evidence that cancer-SRTs can exert tumor-promoting effects in multiple tumor contexts. However, whether such effects are broadly conserved across all cancer types remains to be determined. The functional mechanisms of cancer-SRTs require further investigations to elucidate their potential functional involvement in tumor progression and immune evasion.

Despite these advances, limitations remain. First, functional mechanisms of cancer-SRTs require further experimental validation to elucidate their specific roles in tumor progression and immune evasion. Second, the lack of validation in independent external cohorts represents a major limitation of this study. While our clinical prediction model demonstrated strong internal consistency, its generalizability requires further validation. Future studies utilizing emerging long-read datasets or evaluating these signatures in independent patient cohorts will be necessary to confirm their stability and clinical utility. Furthermore, limited immunotherapy response data precluded direct modeling for immunotherapy responsiveness. As more clinical datasets focused on immunotherapy become available, the cancer-SRT risk-score could be evaluated as a predictive or stratification biomarker for immunotherapy response. It may be assessed either alone or in combination with established immune indicators such as TIDE, tumor mutational burden, and PD-L1 expression.

Conclusion

In summary, this study characterizes cancer-SRT expression patterns across cancer types and explores their regulatory mechanisms and functional and clinical relevance. These findings advance our understanding of tumor biology and lay the groundwork for developing cancer-SRT-based diagnostic, prognostic, and therapeutic strategies. Future studies investigating their underlying mechanisms and applications in immunotherapy will be critical for precision cancer treatment.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

A3SS

Alternative 3’ splice site

A5SS

Alternative 5’ splice site

ACC

Adrenocortical carcinoma

AFE

Alternative first exon

ALE

Alternative last exon

ASEs

Alternative splicing events

BLCA

Bladder Urothelial Carcinoma;

BRCA

Breast invasive carcinoma

Cancer-SRTs

Cancer-specific RNA transcripts

CESC

Cervical squamous cell carcinoma and endocervical adenocarcinoma

CHOL

Cholangiocarcinoma

CNV

Copy number variation

COAD

Colon adenocarcinoma

EMT

Epithelial-mesenchymal transition

ESCA

Esophageal carcinoma

GBM

Glioblastoma multiforme

GDC

Genomic Data Commons

GEO

Gene Expression Omnibus

GO

Gene Ontology

GSVA

Gene Set Variation Analysis

HCC

Hepatocellular carcinoma

HNSC

Head and Neck squamous cell carcinoma

HRs

Hazard ratios

ICB

Immune checkpoint blockade

ImmPort

Immunology Database and Analysis Portal

KEGG

Kyoto Encyclopedia of Genes and Genomes

KICH

Kidney Chromophobe

KIRC

Kidney renal clear cell carcinoma

KIRP

Kidney renal papillary cell carcinoma

LGG

Brain Lower Grade Glioma

LIHC

Liver hepatocellular carcinoma

logFC

Log fold change

LR-seq

Long-read sequencing

LUAD

Lung adenocarcinoma

LUSC

Lung squamous cell carcinoma

MESO

Mesothelioma

MX

Mutually exclusive exons

NESs

Normalized enrichment scores

NIC

Novel in catalog

NNC

Novel not in catalog

OV

Ovarian serous cystadenocarcinoma

PAAD

Pancreatic adenocarcinoma

PCPG

Pheochromocytoma and paraganglioma

PFS

Progression-free survival

PRAD

Prostate adenocarcinoma

READ

Rectum adenocarcinoma

RI

Retained intron

SARC

Sarcoma

SE

Exon skipping

SKCM

Skin cutaneous melanoma

SNPs

Single nucleotide polymorphisms

STAD

Stomach adenocarcinoma

TCGA

The Cancer Genome Atlas

THCA

Thyroid carcinoma

THYM

Thymoma

TSSs

Transcription start sites

t-SNE

t-distributed stochastic neighbor embedding

UCEC

Uterine corpus endometrial carcinoma

UCS

Uterine carcinosarcoma

UVM

Uveal melanoma

Author contributions

Conceptualization: Z.C., X.H. Methodology: H.L, J.D. Visualization: H.L, J.D. Funding acquisition: Z.C., X.H. Project administration: Z.C., X.H. Supervision: Z.C., X.H. Writing – original draft: H.L, Z.C., J.D. Writing – review & editing: Z.C., X.H. All authors read and approved the final manuscript.

Funding

National Natural Science Foundation of China 82472641 (Z.C.). National Key R&D Program of China 2024YFC3405900 (X.H.).

Data availability

The data generated in this study are available within the article and its supplementary data files.

Declarations

Ethics approval and consent to participate

Not applicable

Consent for publication

Not applicable

Competing interests

The authors declare that they have no competing interests.