https://orcid.org/0000-0003-3605-9373
Figures
Abstract
Intronic PolyAdenylation (IPA) is an important post-transcriptional mechanism that can alter transcript coding potential by truncating translation regions, thereby increasing transcriptome and proteome diversity. This process generates novel protein isoforms with altered peptide sequences, some of which are implicated in disease progression, including cancer. Truncated proteins may lose tumor-suppressive functions, contributing to oncogenesis. Despite advancements in Alternative PolyAdenylation (APA) analysis using RNA-seq, detecting and quantifying novel IPA events remains challenging. To address this, we developed IPScan, a computational pipeline for precise IPA event identification, quantification, and visualization. IPScan has been benchmarked against existing methods using simulated data, different human and mouse cell lines, and TCGA (The Cancer Genome Atlas) breast cancer datasets. Differential IPA events under different biological conditions were quantified and validated via qPCR.
Author summary
APA affects transcript coding potential by truncating translation regions, thereby increasing transcriptome and proteome diversity. A specific type, IPA, generates alternative protein isoforms. In this study, we introduce IPScan, a newly developed framework for detecting novel IPA events from RNA-seq data, as well as quantifying and visualizing these events. Our evaluation on simulated RNA-seq data, human and mouse cell lines, and real cancer patient data demonstrates that IPScan significantly outperforms state-of-the-art models. Furthermore, we confirm that the detected IPA events are associated with cancer outcomes, offering an additional layer of molecular signatures for downstream analysis.
Citation: Fahmi NA, Cheng S, Overstreet J, Song Q, Yong J, Zhang W (2025) IPScan: Detecting novel intronic PolyAdenylation events with RNA-seq data. PLoS Comput Biol 21(11): e1013668. https://doi.org/10.1371/journal.pcbi.1013668
Editor: Wei Li, University of Maryland School of Medicine, UNITED STATES OF AMERICA
Received: May 19, 2025; Accepted: October 28, 2025; Published: November 11, 2025
Copyright: © 2025 Fahmi et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant data are within the manuscript and its Supporting Information files. The dbGaP accession number to the specific version of the TCGA dataset is phs000178.v11.p8. Detailed implementation and execution instructions of IPScan can be found at https://github.com/compbiolabucf/IPScan.
Funding: This work was supported by by the National Science Foundation [NSF-III2246796 to J.O. and W.Z.; NSF-III2152030 to N.A.F. and and W.Z.] and the National Institutes of Health [NIH NIGMS 5R01GM113952-08 to S.C. and J.Y.]. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Alternative PolyAdenylation (APA) is a key post-transcriptional mechanism that regulates gene expression by generating mRNA isoforms with distinct 3′ untranslated regions (3′-UTRs). This process influences mRNA stability, localization, and translation, adding complexity to transcriptome regulation [1,2]. A specialized form of APA, Intronic PolyAdenylation (IPA), occurs within introns and can lead to early transcription termination, producing either truncated coding mRNAs or non-coding transcripts. This mechanism alters protein function by either eliminating essential domains or introducing novel peptide sequences. Dysregulated IPA has been implicated in multiple diseases, including leukemia, multiple myeloma, diabetes, and non-alcoholic fatty liver disease [3–6]. Moreover, IPA-driven premature termination can disrupt tumor suppressors such as DICER and FOXN3 while generating oncogenic isoforms of genes like CARD11, MGA, and CHST11 [3]. Additionally, IPA contributes to drug resistance by altering the expression of key genes, such as TOP2α, which affects the response to inhibitors in leukemia cell lines (HL-60, CEM, and K562) [7]. The widespread prevalence of IPA, with more than 12,500 genes in the human genome exhibiting annotated intronic APA events [8], underscores its significance in transcriptome regulation and disease pathogenesis. Understanding IPA dynamics is crucial for refining genome annotations and developing predictive models for disease phenotypes.
Leveraging high-throughput RNA-seq technology enables accurate quantification of the transcriptome, providing deeper insights into previously unexplored molecular mechanisms [9,10]. While standard RNA-seq protocols generate sequencing libraries from the entire transcript, 3′-end-seq technology specifically amplifies only the 3′ end of transcripts [11]. This method offers quantitative, genome-wide profiling of the 3′ ends of polyadenylated coding and non-coding transcripts. Typically, peaks in 3′-end-seq data occur upstream of a transcript’s termination point, indicating potential isoform truncation sites [12]. In humans, most polyadenylation sites (polyA sites) are preceded by upstream hexamers (ATTAAA or AATAAA), known as polyadenylation signals (PAS) [13,14]. The presence of 3′-end-seq peaks or PAS within a gene’s coding region can reveal mRNA truncation and potentially lead to the discovery of previously unannotated isoforms and novel protein products.
Several studies have employed RNA-seq, alongside other sequencing protocols, to identify and quantify APA events across various genomic regions. Computational tools such as DaPars [15], APAtrap [16], TAPAS [17], APA-Scan [18], and QAPA [19] are comprehensive and well-established for detecting APA events, though they primarily focus on 3′-UTR APA events. Among more recent studies, IPAFinder [20] targets IPA events by analyzing RNA-seq read coverage fluctuations within intronic regions, while APAlyzer [21] relies on the previously annotated PolyA_DB [22] database to identify and quantify IPA isoforms. APAIQ [23] combines DNA sequence information with RNA-seq read coverage to detect APA sites but is not specifically designed for IPA. InPACT [24], a newer IPA detection method, uses a convolutional neural network module to identify potential polyA sites and a read module to identify unannotated terminal exons.
In this study, we introduce IPScan, a tool that integrates RNA-seq read coverage data with 3′-end-seq peaks to identify novel IPA events and validate detected sites using wet-lab qPCR experiments and PacBio Iso-Seq long-read sequencing data. Tests on both simulated datasets and real human and mouse cell lines demonstrate IPScan’s ability to detect differential IPA events across diverse biological conditions. Furthermore, analysis of The Cancer Genome Atlas (TCGA) breast cancer (BRCA) dataset [25] reveals associations between IPScan-identified IPA events and potential biomarkers relevant to tumorigenesis.
Results
IPScan identifies two types of IPA events: Type 1 and Type 2, as shown in Fig 1. In Type 1 events, polyadenylation occurs within a downstream intron, resulting in a hybrid exon-intron structure that serves as the 3′-end exon. In Type 2 events, a cryptic exon within the downstream intron is recognized and used as the 3′-end exon.
a) Type 1: Occurs when the first step of splicing is inhibited, resulting in polyadenylation within the downstream intron. b) Type 2: Involves inclusion of a cryptic exon within the downstream intron, which serves as the new 3’-end exon. Introns are shown as black solid lines, and polyA sites are indicated by red vertical lines.
Moreover, IPScan is designed to handle two key scenarios. First, it detects novel IPA events present in the input sample (Fig 2a). Second, it quantifies differential IPA site usage by comparing two distinct biological conditions (Fig 2b). To evaluate its performance, we applied IPScan to multiple simulated and real datasets, including human and mouse cell lines as well as TCGA breast cancer patient samples. Our analysis uncovered several unannotated IPA sites, underscoring their potential for inclusion in protein annotation databases.
(a) Schematic overview of IPScan’s process for detecting novel IPA events and generating peptide sequences from newly identified truncated isoforms. (b) Diagram illustrating the detection and quantification of differential IPA events between two conditions, with coverage plots depicting read coverage flanking the IPA sites in both conditions.
Experimental results with simulated RNA-seq data
In our simulation experiment, we generated synthetic RNA-seq samples with sequencing depths ranging from 5 million (M) to 50M short reads using Flux-Simulator [26]. To model IPA events, we introduced 1,000 synthetic events per sample, consisting of 500 Type 1 and 500 Type 2 events. Further details of this process are provided in the section “Materials and Methods”. To simulate dynamic changes in IPA events under different biological conditions, we generated paired samples (control vs. case), each containing 50M reads and designed to include 1,000 synthetic differential IPA events to assess IPScan’s ability to detect differential IPA usage. For each condition, three technical replicates with varied parameter settings were created using Flux-Simulator.
IPScan is robust across varying sequencing depths.
IPScan demonstrates high sensitivity, detecting IPA events even at low read coverage. Among the 1,000 simulated IPA events, it successfully identified 609 (61%), 699 (70%), 852 (85%), and 918 (92%) at sequencing depths of 5M, 10M, 30M, and 50M reads, respectively. As shown in Fig 3a, while detection performance improves with increasing sequencing depth, IPScan maintained a comparatively high AUC score of 0.72 even at the lowest tested coverage (5M reads), underscoring its robustness in detecting IPA events across a wide range of sequencing depths.
(a) Simulation experiment evaluating IPScan at different sequencing depths. ROC curves show detection accuracy across RNA-seq read depths ranging from 5M to 50M. (b–c) Comparison of IPScan with four baseline methods: (b) identification of potential IPA sites and (c) detection of differential IPA events between conditions. In both analyses, performance was benchmarked against 1,000 simulated ground-truth events.
IPScan outperforms the baselines.
To evaluate the accuracy of IPA event detection and the estimation of dynamic usage across conditions, we compared IPScan with four state-of-the-art computational tools: IPAFinder, APAlyzer, InPACT, and APAIQ. On the simulated RNA-seq dataset with 50M reads, IPScan demonstrated superior precision and accuracy in IPA site detection (Fig 3b). Although InPACT and APAIQ performed comparably, IPScan achieved the highest AUC score of 0.94, outperforming all other baselines. At varying read depths (30M, 10M, and 5M), IPScan achieved top or near-top performance, with the corresponding AUROC plots shown in Figs A–C in S1 Appendix.
Among the 1,000 synthetic differential IPA events between control and case samples, IPScan correctly identified 897 as significantly different between conditions, surpassing InPACT (870), IPAFinder (801), and APAlyzer (648). APAIQ was not included in this comparison because it is not designed to evaluate condition-specific changes in IPA site usage. As shown in Fig 3c, AUC scores were calculated for paired samples with 50M reads per condition. IPScan achieved the highest AUC (0.93), exceeding InPACT (0.88), while IPAFinder and APAlyzer exhibited lower values. For each condition, three technical replicates were included, and significance was determined using a Wilcoxon rank-sum test (p-value <0.05) together with truncation ratio (TR) difference between conditions ().
Beyond accuracy, false positive rates varied among methods. APAlyzer reported the largest number of significant events, but only 70% were true IPA events, indicating a high false positive rate. IPAFinder performed better, with a 19% false positive rate, whereas IPScan achieved the lowest rate at 13.6%. Together, these results highlight IPScan’s superior detection accuracy and robustness for identifying differential IPA events across diverse biological conditions.
IPA profiling across human and mouse cell lines
To further evaluate IPScan’s performance, we applied it to various mammalian cell types, including Tsc1−/− mouse embryonic fibroblasts (MEFs) and wild-type (WT) MEFs, as well as the human breast cancer cell lines MCF7 and BT549. In both breast cancer cell lines, we analyzed samples treated with DMSO (mock) or Torin 1, a potent mTOR inhibitor.
IPScan detects novel IPA events.
IPScan successfully identified novel IPA sites across both human and mouse cell lines. As shown in Fig 4 (left panels), the number of detected IPA events varied across samples. In human cell lines, Torin-treated samples exhibited a higher number of IPA events compared to mock-treated ones. Modulations in mTOR activity within cellular systems result in a bifurcated expression profile of IPA. Therefore, the results explain well the identification of novel IPAs in cells treated with Torin 1 [8].
The vertical bar plots (left) depict the number of novel IPA events identified by IPScan, while the horizontal bar plots (right) present KEGG pathway enrichment results for (a) mouse MEF cell lines and (b, c) breast cancer cell lines. The x-axis indicates the p-value (-log10 normalized), and the y-axis lists the corresponding KEGG pathways. Pathways enriched by genes with higher expression in the first group are shown at the top, whereas those enriched in the second group appear at the bottom.
The detected IPA events were further classified into two types: Type 1, where splicing inhibition leads to polyadenylation within a downstream intron, and Type 2, where a cryptic exon within the downstream intron becomes the new 3’-end exon (Fig 1). As summarized in Table 1, IPScan detected a significant number of novel Type 1 events in all samples, while Type 2 events were less frequent. This suggests that mTOR pathway perturbation primarily affects splicing dynamics and promotes premature polyadenylation within introns (Type 1) rather than cryptic exon inclusion (Type 2).
Events are classified into Type 1 (polyadenylation within a downstream intron due to splicing inhibition) and Type 2 (cryptic exon inclusion forming a new 3’-end exon).
To validate the predicted IPA sites, we analyzed the nucleotide composition surrounding these regions. Polyadenylation is typically driven by PAS, such as ATTAAA, AATAAA, or single nucleotide variants, located upstream of polyA sites to signal transcript termination. In the Mouse Tsc1−/− MEF sample, the nucleotide profile around IPA sites identified by IPScan showed a significant enrichment of A and T nucleotides in upstream regions, supporting the potential for isoform truncation (Fig 5). This AT-rich signature aligns with known PASs, reinforcing the accuracy of IPA site detection. Similarly, nucleotide distribution analysis in human cell lines revealed a higher frequency of A and T bases surrounding the detected IPA sites (Figs D–H in S1 Appendix).
The x-axis indicates positions relative to the IPA sites (±50 bp), and the y-axis shows the frequency of each nucleotide at each position.
Pathway analysis is essential to understand gene function and the broader biological significance of detected IPA events. After identifying genes affected by IPA, we performed KEGG pathway enrichment analysis to examine their impact on biological processes. As shown in Fig 4 (right panels), the enriched KEGG pathways for IPA events detected by IPScan vary across the three cell lines analyzed in this study. The results indicate that mouse and human samples with different mTOR activation levels are involved in key cellular processes associated with cancer hallmarks [27]. Notable pathways include the cell cycle, focal adhesion, Ras signaling, and PD-L1 expression and PD-1 checkpoint pathways, all of which play critical roles in cancer progression [27]. Interestingly, the enriched pathways exhibit species- and cell line-specific patterns, suggesting that the biological functions of these novel IPA genes may depend on cellular context and influence breast cancer aggressiveness. Additionally, pathways such as mRNA surveillance, axon guidance, and actin cytoskeleton regulation were consistently observed among both upregulated and downregulated IPA genes when comparing samples with high and low mTOR activation. These findings underscore the significance of mTOR-regulated pathways in tumorigenesis, particularly through IPA regulation.
To independently validate the novel IPA isoforms identified in our study, we leveraged publicly available long-read transcriptomic data. Long-read sequencing technologies, such as PacBio single-molecule real-time (SMRT) Iso-Seq [28], sequence entire transcripts end to end without requiring computational assembly or inference [29]. By capturing complete cDNA molecules without fragmentation, Iso-Seq provides an orthogonal layer of evidence free from read-coverage biases and assembly heuristics, making it particularly well suited for validating candidate IPA isoforms predicted from short-read data [30,31].
Using Iso-Seq as an additional benchmark, we found that approximately 18.6% of the novel IPA isoforms predicted by IPScan were corroborated by full-length long-read transcripts. By comparison, IPAFinder, APAlyzer, InPACT, and APAIQ achieved lower validation rates of 10.19%, 6 %, 16.2%, and 12.5%, respectively (Fig 6a). These results highlight IPScan’s superior accuracy in capturing bona fide IPA isoforms, particularly those overlooked by methods that rely more heavily on pre-annotated polyA sites. Notably, among the events exclusively detected by IPScan, two isoforms derived from FLOT1 and GALNT10 were supported by Iso-Seq data (Fig 6b–6c). Both of these isoforms are novel with respect to current RefSeq annotations. The ability of Iso-Seq to recover these unannotated transcripts provides strong, independent confirmation of IPScan’s predictions.
(a) Ratio of IPA events detected by each method that were validated by Iso-Seq data. (b–c) Two RefSeq unannotated isoforms from (b) FLOT1 and (c) GALNT10, identified only by IPScan, are supported by long-read Iso-Seq annotations and corresponding read-coverage plots.
IPScan measures the differential usage of IPA events.
Using IPScan, we identified widespread differential usage of novel IPA sites across various conditions in human and mouse cell lines. Detection was based on a significance threshold defined by a chi-squared test (p-value <0.05) and a TR difference between conditions (). In the absence of ground truth data for real experimental samples, we evaluated IPScan’s performance by assessing its concordance with baseline methods. Unlike IPAFinder, which detects both annotated and novel events by leveraging read-coverage fluctuations, APAlyzer primarily captures expression changes between regions upstream and downstream of known IPA sites and depends on a pre-annotated database (polyA_DB) [22]. To quantify the overlap of differential IPA events identified by IPScan and these tools, we constructed a four-set Venn diagram comparing IPScan, IPAFinder, APAlyzer, and InPACT in mouse WT vs. Tsc1−/− MEF samples (Fig 7). APAIQ was not included in this comparison, as it is not designed to evaluate condition-specific changes in IPA site usage. Among the methods, APAlyzer reported the largest number of differential events, largely due to its reliance on pre-annotated polyA sites. Notably, the limited overlap among the four methods underscores the novelty of IPScan’s findings.
The diagram highlights shared and unique IPA events identified by each method.
Among the 231 differential IPA events detected by IPScan, 56 events were exclusively detected only by IPScan. We further validated two of these unique events, MAP3K10 and TROAP, through qPCR and alignment plot analyses. The qPCR results confirmed increased expression of the truncated isoforms in Tsc1−/− compared to WT, consistent with IPScan’s predictions (Fig 8). Details of the qPCR experiments and primer design are provided in the S1 Appendix.
Quantitative PCR (qPCR) analysis shows significantly elevated expression of MAP3K10 and TROAP transcripts in Tsc1−/− MEF cells compared with WT controls, supporting the activation of novel IPA sites in these genes.
IPA analysis in TCGA breast cancer data
IPScan was applied to 1,219 TCGA breast cancer (BRCA) samples [25]. To assess differential IPA usage between tumor and normal tissues, we analyzed 1,106 tumor samples and 113 normal samples. Clinical data for 1,108 BRCA patients were obtained from cBioPortal [32].
IPScan identifies differential IPA usage in breast cancer tumor-normal samples.
Using IPScan, we identified a substantial number of IPA events that exhibited significant differences in site usage between breast cancer tumor and normal samples. These findings suggest that differential IPA usage is linked to cancer-specific regulatory mechanisms, as IPA events influence mRNA stability, translation, and protein production. Such dynamic alterations in IPA usage highlight their potential as biomarkers for cancer diagnosis or prognosis, offering valuable insights into the molecular mechanisms driving tumor progression. As illustrated in Fig 9a, an alignment plot demonstrates the read coverage of a novel IPA event across multiple samples, showing a clear separation between tumor and normal samples based on TR values. The statistical significance of this separation was validated using an unpaired t-test (p-value < 0.05).
(a) Validation of an IPA event detected by IPScan (CSNK1E: Chr22:38296209) using read-coverage plots from five normal and five tumor samples. The x-axis indicates genomic position, and the y-axis shows read coverage; gene annotation is provided in the bottom panel. (b–c) Kaplan–Meier plots showing associations between IPScan truncation ratio (TR) values and (b) overall survival or (c) disease-free rate for two representative IPA events. (d) Significant IPA events across tumor subtypes (ER, PR, HER2, and triple-negative [TN]); the significant event for each subtype is shown separately for receptor-positive and receptor-negative samples. (e) Boxplots of IPScan TR values for an IPA event in FTSJ1 across four tumor stages. (f) Boxplots of IPScan TR values for an IPA event in RBBP6 across four tumor stages.
IPScan-detected IPA events associated with clinical variables.
We analyzed the associations between IPScan-derived TR values and various clinical variables. Specifically, we evaluated the relationship between IPA site usage and both overall survival and disease-free rates. The analysis focused on a selected set of IPA events from tumor samples, prioritizing genes with high expression levels to reduce ambiguity. Kaplan–Meier survival plots revealed significant correlations between IPA usage and survival outcomes (Fig 9b) as well as disease-free rates (Fig 9c). These findings were further supported by log-rank tests, which demonstrated statistical significance based on the corresponding p-values.
We also examined IPA patterns, measured as truncation ratio (TR) values, across hormone receptor subtypes in breast cancer tissues. These subtypes—Estrogen Receptor positive/negative (ER+/ER-), Progesterone Receptor positive/negative (PR+/PR-), Human Epidermal Growth Factor Receptor 2 positive/negative (HER2+/HER2-), and Triple-Negative (TN/non-TN)—play critical roles in cancer progression. TR values from IPScan analysis revealed clear distinctions between receptor-positive and receptor-negative subgroups (Fig 9d), providing insights that may inform treatment strategies. In addition, TR values showed significant differences across the four stages of cancer progression and tumor development (Fig 9e, 9f). These patterns highlight the potential of IPScan-derived IPA events as robust indicators for disease staging and prognosis.
IPScan identifies novel peptide sequences in breast cancer patient samples.
From 1,219 TCGA BRCA samples, IPScan identified 3,019 IPA events. Among these, 1,315 unique unannotated peptide sequences were predicted based on the open reading frames (ORFs) of upstream annotated exons. To validate these peptides, we examined their overlap with Pfam domains affected by IPA events. Our analysis revealed that 535 IPA events resulted in the loss of functional Pfam domains due to missing peptides within these regions. For example, an IPA event in CASP10 (caspase-10) leads to loss of the peptidase_C14 domain while retaining the DED (death effector domain), potentially generating truncated CASP10 isoforms that function independently of canonical death-signal activation [33,34]. These findings suggest that protein isoforms arising from IPA events may gain deregulated or novel functions compared with their full-length counterparts, thereby reshaping the functional proteome.
Additionally, we applied PepQuery [35], a tool designed to identify known and novel peptides in local or publicly available mass spectrometry-based proteomics datasets. Cross-referencing two proteomics datasets, “TCGA Breast Cancer Proteome PDC000173” and “TCGA Breast Cancer Phospho-Proteome PDC000173”, PepQuery found no overlap between the novel peptide sequences identified by IPScan and annotated databases. This indicates that these sequences were previously undocumented in existing proteomic resources. Detailed results of the PepQuery analysis are provided in the S1 Appendix.
Discussion
In this study, we introduced IPScan, a computational tool designed to identify novel IPA sites from RNA-seq data and to quantify dynamic IPA changes under different conditions. IPScan’s performance was validated through multiple layers of experimental evaluation, including comparisons with four state-of-the-art methods—IPAFinder, APAlyzer, InPACT, and APAIQ. Our results demonstrate that IPScan outperforms these methods in accurately and precisely identifying IPA events. Notably, two novel IPA isoforms in FLOT1 and GALNT10, detected exclusively by IPScan, were independently validated using PacBio Iso-Seq long-read sequencing, providing strong orthogonal evidence for their authenticity. Evaluations on simulated RNA-seq data showed that IPScan more effectively predicts IPA isoforms and quantifies dynamic IPA site usage across conditions compared with baseline methods. Importantly, IPScan maintained robust performance even in low-coverage samples, achieving moderate to high accuracy across varying sequencing depths. Furthermore, it exhibited the highest precision and the lowest false positive rates among all tested methods. To further assess its practical utility, we applied IPScan and the baseline tools to human and mouse cell lines. While IPScan consistently reported the lowest false positive rates in simulated datasets, it also identified differential IPA events that were undetected by the baseline methods in cell line data. Two such events, involving MAP3K10 and TROAP, were validated through wet-lab qPCR experiments, confirming their differential expression and supporting the accuracy of IPScan’s predictions. These findings establish IPScan as a robust and reliable tool for IPA detection, offering significant improvements over existing methods.
IPA events are frequently observed in breast cancer [36]. Using IPScan, we analyzed 1,219 TCGA breast cancer samples and identified significant alterations in IPA events between tumor and normal samples, including those in key cancer driver genes. Novel IPA events were detected in oncogenes such as GNB1, SMARCE1, STAT6, CUL4A, and IQGAP1 [37]. Additionally, dynamic IPA site usage was observed in well-known tumor suppressor genes, including TSC1, HSD17B2, CD58, RUNX1 and INPP4B [38]. In the case of the transcription factor RUNX1, the IPA event results in truncation within the Runt domain (an evolutionarily conserved protein domain), rendering this domain non-functional and potentially disrupting the critical dimerization activity required for RUNX1 function. Under normal conditions, RUNX1 is robustly expressed in both luminal and basal cells of breast tissue [39,40]. In contrast, many breast cancers—particularly aggressive subtypes—exhibit reduced or absent RUNX1 expression [39,41]. Functionally, RUNX1 helps preserve the epithelial phenotype by promoting E-cadherin expression, thereby supporting cell–cell adhesion and suppressing migratory behavior. When epithelial–mesenchymal transition (EMT) is induced (e.g., by TGF-β signaling or serum deprivation), RUNX1 expression declines, leading to increased mesenchymal marker expression and enhanced invasiveness [39]. Together, these observations suggest that IPA-mediated disruption of RUNX1 introduces previously unrecognized pathogenic mechanisms that may contribute to breast cancer progression.
Further investigation into the biological functions of the differentially expressed IPA genes in TCGA breast cancer samples, particularly those regulated by mTOR signaling, revealed their involvement in critical oncogenic pathways. For instance, dysregulation in axon guidance pathways, which are known to promote cell proliferation and adhesion in cancer, was identified [42]. Additionally, we observed disruptions in the PD-1/PD-L1 immune checkpoint pathway, a mechanism frequently exploited by cancer cells to evade immune surveillance [43]. As IPA events often produce truncated protein products with altered functions, the newly identified unannotated IPA proteins regulated by mTOR signaling may contribute to breast cancer progression. These proteins likely dysregulate axon guidance and immune surveillance pathways, potentially promoting tumor growth and reducing the clinical efficacy of PD-1/PD-L1 inhibitors in breast cancer patients. These findings highlight the importance of IPA in understanding breast cancer biology and improving therapeutic strategies.
IPScan offers broader and more comprehensive annotations than existing methods. Despite these strengths, several areas remain for future improvement. For example, the optional use of matched 3’-end-seq data, while clearly indicated as an “optional” input, may need further consideration given the limited availability of datasets that include both matched 3’-end-seq and RNA-seq. In practice, however, 3’-end-seq data do not have to be perfectly matched to every RNA-seq sample. 3’-end-seq data derived from the same tissue type, disease context, or well-characterized cell lines can provide valuable reference information for annotating and validating polyA sites. This flexibility helps extend IPScan’s applicability even when fully matched datasets are not available.
In addition, IPScan’s current implementation relies on read-coverage-based quantification of truncation ratios and is therefore not directly applicable to single-cell RNA-seq analysis. Unlike bulk RNA-seq, single-cell RNA-seq typically suffers from lower sequencing depth, higher technical noise, and substantial dropout rates, all of which hinder the accurate detection of fine-scale read-coverage patterns needed to identify IPA events. As part of our future work, we plan to extend IPScan to single-cell resolution while addressing these challenges. One promising direction is to integrate PacBio long-read sequencing with single-cell RNA-seq to improve isoform-level resolution and enhance the accuracy of IPA event detection. In addition to detection accuracy, the computational cost of IPA analysis tools is an important consideration for large-scale studies. Our benchmarking shows that runtimes vary widely across methods, from less than an hour for APAlyzer to several hours on CPU-based systems for IPAFinder and InPACT, with deep learning-based approaches such as APAIQ requiring even longer CPU runtimes but benefiting from significant GPU acceleration. Relative to these tools, IPScan exhibits competitive runtime performance. On a standard compute node equipped with 16 CPU cores (Intel(R) Xeon(R) CPU E5-2620 v4 @ 2.10GHz, 64-bit) and 125 GB RAM, the complete analysis of an RNA-seq dataset containing ∼50 million paired-end aligned reads required about 80 minutes using IPScan. Notably, extraction of peaks from the 3’-end-seq data and generation of the merged database accounted for nearly 30 % of the total runtime. Once a pre-generated list of potential IPA events is available, IPScan runs substantially faster when applied to large-scale datasets such as TCGA BRCA, requiring only ∼40 minutes per sample. These results indicate that, while not the fastest among all benchmarks, IPScan provides a practical balance between computational cost and analytical depth, making it well suited for large-scale transcriptomic analyses.
Materials and methods
Data preparation
Simulation data. In our experiments, we used Flux Simulator [26] to generate paired-end short reads and simulate in silico RNA-seq experiments based on a ground truth transcript expression profile and IPA events. Specifically, we generated four sets of synthetic RNA-seq samples with sequencing depths of 5M, 10M, 30M, and 50M reads, with 500 Type-1 and 500 Type-2 events on each sample. To assess dynamic IPA usage under different conditions, we prepared paired control and case samples with 50M reads each. For all simulated samples, gene expression levels were assigned to follow a Poisson distribution to mimic real RNA-seq data. To ensure accurate IPA detection and minimize ambiguity from low-expression genes, IPA events were simulated only in highly expressed genes. Each IPA site was modeled with a short and long isoform pair, where the IPA site represents the endpoint of the unannotated short isoform, while the original full-length isoform serves as the long isoform. Genomic locations, including intron-exon structures, were extracted from RefSeq annotations [44]. The specific parameter configurations are provided in Table A in S1 Appendix. All simulated RNA-seq samples were aligned to the reference genome using HISAT2 [45] and converted into read-coverage files using SAMtools [46]. This pipeline ensured high-quality simulation and alignment, allowing for a robust evaluation of dynamic IPA usage.
Cell-line data. To conduct our experiments on real tissue samples, we evaluated different sets of human and mouse cell lines [12,47,48] using IPScan. Analyses were performed on both RNA-seq and matched 3′-end-seq datasets for each sample. For mouse embryonic fibroblast (MEF) RNA-seq samples, we analyzed poly(A + ) RNAs isolated from Tsc1−/− and wild-type (WT) MEFs. The RNA-seq analysis generated 63,742,790 paired-end reads for WT and 74,251,891 paired-end reads for Tsc1−/− MEFs using the HiSeq platform, with each read being 50 base pairs long. Short reads were aligned to the mm10 reference genome using HISAT2. For the 3′-end-seq analysis, reads from WT and Tsc1−/− MEFs were preprocessed by trimming polyA tails from the 3′-ends and filtering out low-quality reads (Phred score < 30) or reads shorter than 25 base pairs. The remaining reads were aligned to the mouse mm10 reference genome using Bowtie2 [49], with no mismatches allowed. In total, 6,186,893 reads were aligned for WT samples, and 5,382,111 reads were aligned for Tsc1−/− samples. For MEF data analysis, 24,511 IPA sites were annotated by overlapping 27,091 UCSC polyA sites from the mouse mm10 genome with 65,488 3′-end peaks identified from the 3′-end-seq MEF samples.
The human cell line data included two sets of Mock vs. Torin-treated samples: MCF7 Mock vs. Torin and BT549 Mock vs. Torin. For MCF7, the RNA-seq data generated 115,450,097 paired-end reads for Mock and 108,176,924 paired-end reads for Torin-treated samples. In the 3′-end-seq analysis, 4,355,286 reads were mapped to the human hg38 reference genome for MCF7 Mock, and 3,343,638 reads were mapped for MCF7 Torin. Similarly, for BT549, the RNA-seq data produced 131,955,082 paired-end reads for Mock and 138,127,113 paired-end reads for Torin-treated samples. The mapped 3′-end reads totaled 3,348,578 for BT549 Mock and 4,386,406 for BT549 Torin. The same pipeline used for the MEF data was applied to generate read coverage files for the human cell line data. To annotate IPA locations in human breast cancer samples, we identified 20,868 potential IPA sites by intersecting 42,615 UCSC polyA sites from the human hg38 genome with 66,617 3’-end peaks derived from the combined MCF7 and BT549 cell line data.
Human MCF7 Iso-Seq data (SRA: SRX7505753; GEO accession: GSM4251592) were downloaded from the GEO database and consist of single-end PacBio RS II reads from three biological replicates. Raw FASTQ files were aligned to the human reference genome using minimap2 [50] with default Iso-Seq settings, achieving an alignment rate of approximately 91%. The aligned transcripts were then processed with SQANTI3 [51], which integrates long-read isoform evidence with reference annotations to generate an updated transcriptome annotation (GTF) file. SQANTI3 classified the transcripts as either known isoforms (matching reference annotations) or novel isoforms (containing unannotated splice junctions, alternative transcription start or end sites, or intronic polyadenylation events).
TCGA breast cancer data. A total of 1,219 breast cancer samples were downloaded from TCGA to investigate the association of IPA events with cancer patient data. The dataset included 1,106 TCGA BRCA tumor samples and 113 normal tissue samples. All samples were aligned to the reference genome, and read coverage files were generated using HISAT2 and SAMtools. Clinical data for the 1,106 BRCA patients were obtained from cBioPortal [32]. While TCGA provides extensive RNA-seq data across multiple cancer types, it does not include 3′-end-seq data. To address this limitation, we utilized the 20,868 potential IPA sites previously identified in human breast cancer cell line data for the analysis of TCGA BRCA samples.
IPScan workflow
IPScan is designed to discover novel IPA events through two complementary approaches. First, it identifies novel IPA sites and reconstructs the resulting unannotated peptide sequences. Second, it quantifies changes in IPA usage across biological conditions and evaluates their statistical significance. To broaden the search space, IPScan integrates UCSC-annotated polyA sites (PolyA_DB 3) [22] with 3’-end-seq peaks (optional input), generating an extensive catalog of candidate sites. When matched 3’-end-seq data are not available, the method instead leverages existing polyA site databases and performs genome-wide scans for canonical PASs, such as AATAAA and its variants, to predict sites. Candidate IPA events are then refined by focusing on independent introns, introns that do not overlap with any exon from any isoform of the gene, ensuring that only true intronic sites are retained for downstream analyses. This restriction reduces ambiguity arising from alternative splicing and annotated transcript ends, thereby improving both the specificity and reliability of IPA detection.
If a polyA site is present in the nearby region, the read coverage distribution is likely to exhibit a significant drop near the site. Based on this principle, IPScan calculates the TR for each novel event as TR = where C2 represents the average read coverage in the intron region upstream of the IPA site, and C1 denotes the average read coverage of all the exons of the gene. The detected events are categorized into Type 1 and Type 2 based on their structural formation (Fig 1). If no Type 1 event is observed in the designated region, IPScan searches for Type 2 events, where a cryptic exon serves as the new 3′-end exon, which is entirely composed of intronic sequences (Fig 1). The cryptic exon boundary is defined when the average read coverage of that region reaches at least 80% of the average read coverage of all exons of the gene. This allows for a refined estimation of both the average read depth across all exons (C1) and the read coverage of the intronic region within the IPA site and the 3′-end exon margin (C2).
The process involves parameter tuning to enhance the identification of IPA events. For TR = , events with TR < 0.2 are filtered out. Additionally, a threshold of C2 > 10 is applied to remove a large portion of false events. For each identified position, peptide sequences are generated based on the Open Reading Frames (ORFs) of the upstream annotated exons. Codons are translated up to the first available stop codon within the given range. For Type 2 events occurring in the intronic region, all three possible ORFs from the tentative start position of the cryptic exon are considered when generating the peptide sequences.
To assess dynamic IPA changes between conditions, IPScan identifies differential events based on differences in TR values. A change in TR for a gene between samples or groups is reported as a differential IPA event. The significance of these events is evaluated using a chi-squared test with a threshold of p-value < 0.05 and a TR difference greater than 0.2. For samples with replicates, dynamic IPA usage is assessed using the Wilcoxon rank-sum test, with significance defined as p-value < 0.05 and an average TR difference greater than 0.2.
Implementation
IPScan is implemented as a command-line tool using Python3 scripts. It requires RNA-seq data in BAM format as input. To identify IPA events, IPScan operates in two modules: ‘single’ and ‘differential’. The ‘single’ module detects potential novel IPA sites along with de novo peptide sequences, while the ‘differential’ module compares IPA events between two biological conditions. The current implementation supports and has been tested against both human (hg38) and mouse (mm10) genomes using RefSeq annotation. IPScan is easily adaptable to other species when provided with proper annotations and required input files. For graphical visualization, IPScan generates read coverage and annotation plots for user-specified regions. Detailed implementation and execution instructions can be found at github/IPScan.
Baselines and evaluation methods
In this study, four widely used IPA identification approaches, InPACT, APAIQ, IPAFinder and APAlyzer, were applied to evaluate the performance of IPScan. InPACT characterizes IPA events by reporting the parameters ‘PolyAsite’ and ‘IPA usage’. For APAIQ, candidate IPA events were derived from the set of detected PASs. IPAFinder identifies potential IPA sites through the annotation of ‘IPA_terminal_region’, whereas APAlyzer directly provides the PAS identifier together with its genomic position. To ensure a consistent and unbiased comparison, all detected events from these methods were evaluated against the ground truth within ±50 bp from the reference site.
To evaluate changes in IPA site usage between simulated case and control samples (three replicates per condition), we combined InPACT with DRIMSeq [52] to conduct differential transcript usage analysis, as recommended in the InPACT manuscript [24]. While InPACT itself is not inherently designed for differential comparisons, its output can be coupled with DRIMSeq to enable such analyses. DRIMSeq uses a Dirichlet-multinomial model to assess isoform usage and enables detection of condition-dependent changes in the IPA sites with the criteria of > 1 and FDR < 0.05. Because APAIQ does not support case–control designs or provide a framework for differential usage testing, it was excluded from this comparison. IPAFinder used a DEXSeq-based statistical approach to determine the dynamic usage of IPA sites. The selection criteria were set to an FDR-adjusted p-value < 0.05 and a difference in IPA site usage of
IPUI| > 0.1. Significant differential IPA events from APAlyzer were selected using an FDR-adjusted p-value < 0.05. Detailed information on the execution of these baseline methods is available in the S1 Appendix. As an evaluation metric, the area under the ROC curve (AUC) was used to estimate the IPA detection capacity of all methods on the simulated datasets.
Supporting information
Table A in S1 Appendix. Parameters for running Flux Simulator to generate 50 million reads.
Fig A in S1 Appendix. Comparison of IPScan with four baseline methods for the identification of potential IPA sites with 30M reads. Performance was benchmarked against 1,000 simulated ground-truth events.
Fig B in S1 Appendix. Comparison of IPScan with four baseline methods for the identification of potential IPA sites with 10M reads. Performance was benchmarked against 1,000 simulated ground-truth events.
Fig C in S1 Appendix. Comparison of IPScan with four baseline methods for the identification of potential IPA sites with 5M reads. Performance was benchmarked against 1,000 simulated ground-truth events.
Fig D in S1 Appendix. Nucleotide composition around the IPA sites detected by IPScan in the MEF WT sample. The x-axis represents the position relative to the IPA sites (±50 bp), while the y-axis indicates the proportion of each nucleotide at each position.
Fig E in S1 Appendix. Nucleotide composition around the IPA sites detected by IPScan in the MCF7 Mock sample. The x-axis represents the position relative to the IPA sites (±50 bp), while the y-axis indicates the proportion of each nucleotide at each position.
Fig F in S1 Appendix. Nucleotide composition around the IPA sites detected by IPScan in the MCF7 Torin sample. The x-axis represents the position relative to the IPA sites (±50 bp), while the y-axis indicates the proportion of each nucleotide at each position.
Fig G in S1 Appendix. Nucleotide composition around the IPA sites detected by IPScan in the BT549 Mock sample. The x-axis represents the position relative to the IPA sites (±50 bp), while the y-axis indicates the proportion of each nucleotide at each position.
Fig H in S1 Appendix. Nucleotide composition around the IPA sites detected by IPScan in the BT549 Torin sample. The x-axis represents the position relative to the IPA sites (±50 bp), while the y-axis indicates the proportion of each nucleotide at each position.
Realtime quantitative PCR (RT-qPCR) analysis and primer sequences. PepQuery run parameters. Running the baselines.
https://doi.org/10.1371/journal.pcbi.1013668.s001
(PDF)
Acknowledgments
The results are based upon data generated by The Cancer Genome Atlas (TCGA) established by the NCI and NHGRI. Information about TCGA and the investigators and institutions who constitute the TCGA research network can be found at http://cancergenome.nih.gov.
References
- 1. Tian B, Manley JL. Alternative polyadenylation of mRNA precursors. Nat Rev Mol Cell Biol. 2017;18(1):18–30. pmid:27677860
- 2. Fahmi NA, Saha S, Song Q, Lou Q, Yong J, Zhang W. Computational methods for alternative polyadenylation and splicing in post-transcriptional gene regulation. Exp Mol Med. 2025;57(8):1631–40. pmid:40804481
- 3. Lee S-H, Singh I, Tisdale S, Abdel-Wahab O, Leslie CS, Mayr C. Widespread intronic polyadenylation inactivates tumour suppressor genes in leukaemia. Nature. 2018;561(7721):127–31. pmid:30150773
- 4. Singh I, Lee S-H, Sperling AS, Samur MK, Tai Y-T, Fulciniti M, et al. Widespread intronic polyadenylation diversifies immune cell transcriptomes. Nat Commun. 2018;9(1):1716. pmid:29712909
- 5. Jobbins AM, Haberman N, Artigas N, Amourda C, Paterson HAB, Yu S, et al. Dysregulated RNA polyadenylation contributes to metabolic impairment in non-alcoholic fatty liver disease. Nucleic Acids Res. 2022;50(6):3379–93. pmid:35293570
- 6. Locke JM, Da Silva Xavier G, Rutter GA, Harries LW. An alternative polyadenylation signal in TCF7L2 generates isoforms that inhibit T cell factor/lymphoid-enhancer factor (TCF/LEF)-dependent target genes. Diabetologia. 2011;54(12):3078–82. pmid:21913056
- 7. Elton TS, Hernandez VA, Carvajal-Moreno J, Wang X, Ipinmoroti D, Yalowich JC. Intronic polyadenylation in acquired cancer drug resistance circumvented by utilizing CRISPR/Cas9 with homology-directed repair: the tale of human DNA Topoisomerase IIα. Cancers (Basel). 2022;14(13):3148. pmid:35804920
- 8. Sun J, Kim J-Y, Jun S, Park M, de Jong E, Chang J-W, et al. Dichotomous intronic polyadenylation profiles reveal multifaceted gene functions in the pan-cancer transcriptome. Exp Mol Med. 2024;56(10):2145–61. pmid:39349823
- 9. Wang Z, Gerstein M, Snyder M. RNA-Seq: a revolutionary tool for transcriptomics. Nat Rev Genet. 2009;10(1):57–63. pmid:19015660
- 10. Yeh H-S, Zhang W, Yong J. Analyses of alternative polyadenylation: from old school biochemistry to high-throughput technologies. BMB Rep. 2017;50(4):201–7. pmid:28148393
- 11. Beck AH, Weng Z, Witten DM, Zhu S, Foley JW, Lacroute P, et al. 3’-end sequencing for expression quantification (3SEQ) from archival tumor samples. PLoS One. 2010;5(1):e8768. pmid:20098735
- 12. Chang J-W, Zhang W, Yeh H-S, Park M, Yao C, Shi Y, et al. An integrative model for alternative polyadenylation, IntMAP, delineates mTOR-modulated endoplasmic reticulum stress response. Nucleic Acids Res. 2018;46(12):5996–6008. pmid:29733382
- 13. Proudfoot NJ. Ending the message: poly(A) signals then and now. Genes Dev. 2011;25(17):1770–82. pmid:21896654
- 14. Tian B, Graber JH. Signals for pre-mRNA cleavage and polyadenylation. Wiley Interdiscip Rev RNA. 2012;3(3):385–96. pmid:22012871
- 15. Xia Z, Donehower LA, Cooper TA, Neilson JR, Wheeler DA, Wagner EJ, et al. Dynamic analyses of alternative polyadenylation from RNA-seq reveal a 3’-UTR landscape across seven tumour types. Nat Commun. 2014;5:5274. pmid:25409906
- 16. Ye C, Long Y, Ji G, Li QQ, Wu X. APAtrap: identification and quantification of alternative polyadenylation sites from RNA-seq data. Bioinformatics. 2018;34(11):1841–9. pmid:29360928
- 17. Arefeen A, Liu J, Xiao X, Jiang T. TAPAS: tool for alternative polyadenylation site analysis. Bioinformatics. 2018;34(15):2521–9. pmid:30052912
- 18. Fahmi NA, Ahmed KT, Chang J-W, Nassereddeen H, Fan D, Yong J, et al. APA-Scan: detection and visualization of 3’-UTR alternative polyadenylation with RNA-seq and 3’-end-seq data. BMC Bioinformatics. 2022;23(Suppl 3):396. pmid:36171568
- 19. Ha KCH, Blencowe BJ, Morris Q. QAPA: a new method for the systematic analysis of alternative polyadenylation from RNA-seq data. Genome Biol. 2018;19(1):45. pmid:29592814
- 20. Zhao Z, Xu Q, Wei R, Wang W, Ding D, Yang Y, et al. Cancer-associated dynamics and potential regulators of intronic polyadenylation revealed by IPAFinder using standard RNA-seq data. Genome Res. 2021;31(11):2095–106. pmid:34475268
- 21. Wang R, Tian B. APAlyzer: a bioinformatics package for analysis of alternative polyadenylation isoforms. Bioinformatics. 2020;36(12):3907–9. pmid:32321166
- 22. Wang R, Nambiar R, Zheng D, Tian B. PolyA_DB 3 catalogs cleavage and polyadenylation sites identified by deep sequencing in multiple genomes. Nucleic Acids Res. 2018;46(D1):D315–9. pmid:29069441
- 23. Long Y, Zhang B, Tian S, Chan JJ, Zhou J, Li Z, et al. Accurate transcriptome-wide identification and quantification of alternative polyadenylation from RNA-seq data with APAIQ. Genome Res. 2023;33(4):644–57. pmid:37117035
- 24. Liu X, Chen H, Li Z, Yang X, Jin W, Wang Y, et al. InPACT: a computational method for accurate characterization of intronic polyadenylation from RNA sequencing data. Nat Commun. 2024;15(1):2583. pmid:38519498
- 25. Cancer Genome Atlas Network. Comprehensive molecular portraits of human breast tumours. Nature. 2012;490(7418):61–70. pmid:23000897
- 26. Griebel T, Zacher B, Ribeca P, Raineri E, Lacroix V, Guigó R, et al. Modelling and simulating generic RNA-Seq experiments with the flux simulator. Nucleic Acids Res. 2012;40(20):10073–83. pmid:22962361
- 27. Hanahan D, Weinberg RA. Hallmarks of cancer: the next generation. Cell. 2011;144(5):646–74. pmid:21376230
- 28. Rhoads A, Au KF. PacBio sequencing and its applications. Genomics Proteomics Bioinformatics. 2015;13(5):278–89. pmid:26542840
- 29.
Gonzalez-Garay ML. Introduction to isoform sequencing using pacific biosciences technology (Iso-Seq). Transcriptomics and gene regulation. Springer; 2015. p. 141–60.
- 30. Abdel-Ghany SE, Hamilton M, Jacobi JL, Ngam P, Devitt N, Schilkey F, et al. A survey of the sorghum transcriptome using single-molecule long reads. Nat Commun. 2016;7:11706. pmid:27339290
- 31. Wang T, Wang H, Cai D, Gao Y, Zhang H, Wang Y, et al. Comprehensive profiling of rhizome-associated alternative splicing and alternative polyadenylation in moso bamboo (Phyllostachys edulis). Plant J. 2017;91(4):684–99. pmid:28493303
- 32. Gao J, Aksoy BA, Dogrusoz U, Dresdner G, Gross B, Sumer SO, et al. Integrative analysis of complex cancer genomics and clinical profiles using the cBioPortal. Sci Signal. 2013;6(269):pl1. pmid:23550210
- 33. Wachmann K, Pop C, van Raam BJ, Drag M, Mace PD, Snipas SJ. Activation and specificity of human caspase-10. Biochemistry. 2010;49(38):8307–15.
- 34. Sprick MR, Rieser E, Stahl H, Grosse-Wilde A, Weigand MA, Walczak H. Caspase-10 is recruited to and activated at the native TRAIL and CD95 death-inducing signalling complexes in a FADD-dependent manner but can not functionally substitute caspase-8. EMBO Journal. 2002.
- 35. Wen B, Wang X, Zhang B. PepQuery enables fast, accurate, and convenient proteomic validation of novel genomic alterations. Genome Res. 2019;29(3):485–93. pmid:30610011
- 36. Qiao P, Zhang C, Shi Y, Du H. The role of alternative polyadenylation in breast cancer. Front Genet. 2024;15:1377275. pmid:38939531
- 37. Chakravarty D, Gao J, Phillips SM, Kundra R, Zhang H, Wang J, et al. OncoKB: a precision oncology knowledge base. JCO Precis Oncol. 2017;2017:PO.17.00011. pmid:28890946
- 38. Zhao M, Sun J, Zhao Z. TSGene: a web resource for tumor suppressor genes. Nucleic Acids Res. 2013;41(Database issue):D970-6. pmid:23066107
- 39. Hong D, Messier TL, Tye CE, Dobson JR, Fritz AJ, Sikora KR, et al. Runx1 stabilizes the mammary epithelial cell phenotype and prevents epithelial to mesenchymal transition. Oncotarget. 2017;8(11):17610–27. pmid:28407681
- 40. van Bragt MPA, Hu X, Xie Y, Li Z. RUNX1, a transcription factor mutated in breast cancer, controls the fate of ER-positive mammary luminal cells. Elife. 2014;3:e03881. pmid:25415051
- 41. Mercado-Matos J, Matthew-Onabanjo AN, Shaw LM. RUNX1 and breast cancer. Oncotarget. 2017;8(23):36934–5. pmid:28455962
- 42. Harburg GC, Hinck L. Navigating breast cancer: axon guidance molecules as breast cancer tumor suppressors and oncogenes. J Mammary Gland Biol Neoplasia. 2011;16(3):257–70. pmid:21818544
- 43. Yang T, Li W, Huang T, Zhou J. Immunotherapy targeting PD-1/PD-L1 in early-stage triple-negative breast cancer. J Pers Med. 2023;13(3):526. pmid:36983708
- 44. O’Leary NA, Wright MW, Brister JR, Ciufo S, Haddad D, McVeigh R, et al. Reference sequence (RefSeq) database at NCBI: current status, taxonomic expansion, and functional annotation. Nucleic Acids Res. 2016;44(D1):D733-45. pmid:26553804
- 45. Kim D, Paggi JM, Park C, Bennett C, Salzberg SL. Graph-based genome alignment and genotyping with HISAT2 and HISAT-genotype. Nat Biotechnol. 2019;37(8):907–15. pmid:31375807
- 46. Li H, Handsaker B, Wysoker A, Fennell T, Ruan J, Homer N, et al. The Sequence Alignment/Map format and SAMtools. Bioinformatics. 2009;25(16):2078–9. pmid:19505943
- 47. Cheng S, Fahmi NA, Park M, Sun J, Thao K, Yeh H-S, et al. mTOR contributes to the proteome diversity through transcriptome-wide alternative splicing. Int J Mol Sci. 2022;23(20):12416. pmid:36293270
- 48. Chang J-W, Zhang W, Yeh H-S, de Jong EP, Jun S, Kim K-H, et al. mRNA 3’-UTR shortening is a molecular signature of mTORC1 activation. Nat Commun. 2015;6:7218. pmid:26074333
- 49. Langmead B, Trapnell C, Pop M, Salzberg SL. Ultrafast and memory-efficient alignment of short DNA sequences to the human genome. Genome Biol. 2009;10(3):R25. pmid:19261174
- 50. Li H. Minimap2: pairwise alignment for nucleotide sequences. Bioinformatics. 2018;34(18):3094–100. pmid:29750242
- 51. Pardo-Palacios FJ, Arzalluz-Luque A, Kondratova L, Salguero P, Mestre-Tomás J, Amorín R, et al. SQANTI3: curation of long-read transcriptomes for accurate identification of known and novel isoforms. Nat Methods. 2024;21(5):793–7. pmid:38509328
- 52. Nowicka M, Robinson MD. DRIMSeq: a Dirichlet-multinomial framework for multivariate count outcomes in genomics. F1000Res. 2016;5:1356. pmid:28105305