Skip to main content
Wiley Open Access Collection logoLink to Wiley Open Access Collection
. 2025 Jun 4;16(3):e70017. doi: 10.1002/wrna.70017

The Advances in Deep Learning Modeling of Polyadenylation Codes

Emily Kunce Stroup 1, Tianjiao Sun 1, Qianru Li 1, John Carinato 2, Zhe Ji 1,2,✉
PMCID: PMC12138237  PMID: 40468587

ABSTRACT

3′‐end cleavage and polyadenylation is an essential step of eukaryotic mRNA and lncRNA expression. The formation of a polyadenylation (polyA) site is determined by combinatory effects of multiple tandem motifs (~6 motifs in humans), each of which is bound by a protein subcomplex. However, motif occurrences and compositions are quite variable across individual polyA sites, leading to the technical challenge of quantifying polyadenylation activities and defining cleavage sites. Although conventional motif enrichment analyses and machine learning models identified contributing polyadenylation motifs, these cannot unbiasedly quantify motif crosstalk. Recently, several groups developed deep learning models to resolve sequence complexity, capture complex positional interactions among cis‐regulatory motifs, examine polyA site formation, predict cleavage probability, and calculate site strength. These deep learning models have brought novel insights into polyadenylation biology, such as site configuration differences across species, cleavage heterogeneity, genomic parameters regulating site expression, and human genetic variants altering polyadenylation activities. In this review, we summarize the advances of deep learning models developed to address facets of polyadenylation regulation and discuss applications of the models.

Keywords: deep learning, genetics, polyadenylation


Narrative sentence: Deep learning models have recently been developed to examine polyA site formation and calculate site strength. These models have enabled the analyses of different aspects of polyadenylation biology: the identification of polyA sites from the genomic sequences, the cleavage heterogeneity of individual sites, the alternative polyadenylation (APA) regulation in the same gene, and genetic mutations impacting polyadenylation activity. These deep learning models opened up novel avenues to study fundamental principles underlying polyadenylation regulation at the genome scale. Caption: The deep learning models have provided unprecedented opportunities to study different aspects of polyadenylation regulation.

graphic file with name WRNA-16-e70017-g004.jpg

1. Introduction

Cleavage and polyadenylation is a co‐transcriptional process, where the 3′‐end of a pre‐mRNA is cleaved, and the polyA tail is synthesized. It is an essential process for mRNA and long noncoding RNA (lncRNA) maturation, as well as proper transcription termination of RNA polymerase II (Richard and Manley 2009). Near 80% of mammalian genes contain multiple polyA sites in 3′ untranslated regions (3′ UTRs) or introns (Hoque et al. 2013). Usage of the sites via alternative polyadenylation (APA) can produce RNA isoforms with different 3′ UTRs or coding sequences. APA facilitates extensive transcriptome diversity and isoform‐specific regulation of RNA stability, localization, and translation efficiency, and plays important regulatory roles in development, tissue identity, and diseases (Gruber and Zavolan 2019; Ji et al. 2009; Mayr and Bartel 2009; Mitschka and Mayr 2022; Sandberg et al. 2008; Tian and Manley 2017).

During gene transcription, some cleavage and polyadenylation factors scaffold along the C‐terminal domain of RNA polymerase II (Pol II) and scan the nascent RNA transcripts to identify potential polyA sites (McCracken et al. 1997). A polyA site is defined by multiple tandem motifs, each of which is recognized by a protein subcomplex (Figure 1A). In mammals, the core motifs include: (1) the polyadenylation signal (PAS) AAUAAA or close variants, typically found 10–30 nt upstream of the cleavage site and bound by the polyadenylation specificity factor (CPSF) complex; (2) the U‐rich and GU‐rich motifs located ~20 nt downstream, bound by the cleavage stimulation factor (CstF) complex dimers (Colgan and Manley 1997; Mandel et al. 2008; Tian et al. 2005; Zhang et al. 2020). PolyA sites are minimally functional with only CPSF and CstF binding motifs. A few other auxiliary motifs can augment the cleavage and polyadenylation activity, such as the UGUA motif bound by cleavage factor I (CFI) and the U‐rich motif recognized by FIP1 around the PAS, as well as the G‐rich motifs bound by heterogeneous nuclear ribonucleoproteins H/F (hnRNP H/F) and/or PCF11 downstream of CstF binding (Arhin et al. 2002; Brown and Gilmartin 2003; Kaufmann et al. 2004; Schafer et al. 2018; Zhu et al. 2018). The strength and cleavage of a polyA site are determined by the combinatory effects of these motifs, such as occurrences, binding affinities, and their relative positions.

FIGURE 1.

FIGURE 1

Composition of the human and yeast polyadenylation complexes. (A) Composition of the human polyadenylation complex and the cis‐regulatory elements recognized by the subprotein complexes. Individual proteins are colored by the associated complex. (B) Composition of the Saccharomyces cerevisiae polyadenylation complex, as in (A). Individual proteins are colored to match the homologous human protein.

The budding yeast Saccharomyces cerevisiae is another model species frequently used to study polyadenylation regulation. Although some protein factors are conserved between humans and S. cerevisiae based on sequence homology, the assembled subcomplexes and their recognized motifs are different (Figure 1B) (Mandel et al. 2008; Tian and Graber 2012). The yeast cleavage factor IA (CFIA) subcomplex contains proteins homologous to mammalian CstF and CFII components, which recognize A‐rich “positioning elements” 10–30 nt upstream of the cleavage site. Further upstream, UA‐rich “efficiency elements” engage with yeast‐specific cleavage factor IB (CFIB) to regulate the efficiency of polyadenylation (Guo and Sherman 1996). Subunits homologous to the human CPSF complex are reorganized into the yeast cleavage and polyadenylation factor (CPF), which further consists of proteins with yeast‐specific functions. The CPF complex binds to the U‐rich elements around some cleavage sites (Dichtl and Keller 2001). Although U‐rich motifs can increase polyA site strengths, they do not uniformly exist for all polyA sites. These contributing motifs can be quite degenerate and variable across genes. Moreover, some yeast genes show extensive cleavage heterogeneity spanning dozens of nucleotides at the 3′‐ends, which rarely happens in humans (Moqtaderi et al. 2013, 2018; Pelechano et al. 2013).

While the polyadenylation complexes in humans and S. cerevisiae were relatively well characterized, the site motif configurations and associated proteins remain understudied in most other model species. Existing studies were mostly based on computational analyses to identify cis‐elements overrepresented in different regions around polyA sites compared to random sequences, based on the assumption that enriched motifs would not emerge by chance and are likely to be functionally bound by protein factors. Studies have shown that polyA site motif compositions are divergent across distantly related species in animals, fungi, plants, and protists, and the differences were previously described in reviews (Boreikaitė and Passmore 2023; Mandel et al. 2008; Tian and Graber 2012).

While the motif enrichment analyses can identify cis‐elements contributing to polyA site formation, they are hardly applicable to studying the motif crosstalk in a position‐specific and quantitative manner. However, building quantitative models can be valuable for studying polyadenylation regulation. First, the motif compositions and strengths of APA sites in the same gene dictate their relative usage and differential expression during biological processes (Hoque et al. 2013; Jenal et al. 2012; Masamha et al. 2014). Second, inappropriate cleavage and polyadenylation can impair transcription termination and lead to the production of aberrant mRNA isoforms and proteins (Kamieniarz‐Gdula and Proudfoot 2019). In particular, many randomly occurring polyA signals are located in introns, and efficient splicing plays a major role in suppressing unwanted intronic polyadenylation (Berg et al. 2012; Kaida et al. 2010; Tian et al. 2007). The locations and efficiencies of polyA sites must be evolutionarily co‐optimized within the context of the surrounding genomic landscape. Third, mutations in key PASs can cause human diseases. For example, mutations of polyA sites in TP53 (linked with cancer susceptibility), HBA2 (α‐thalassemia), and INS (diabetes) can cause mRNA and protein downregulation and lead to disease initiation (Garin et al. 2010; Higgs et al. 1983; Stacey et al. 2011). It is imperative to build quantitative models to measure the polyadenylation activity and understand the complex motif “grammar” that regulates polyA sites genome‐wide.

Defining polyA sites represents a computational challenge. Multiple motifs must be present in specific combinations and positioned appropriately to facilitate polyadenylation complex recruitment. Additionally, some polyA motifs are highly degenerate, and multiple auxiliary signals in combination can overcome missing or weak other cis‐regulatory elements (Sheets et al. 1990; Tian et al. 2005). The historical bioinformatics approaches using machine learning cannot quantitatively model site formation and strengths. To address this technical challenge, during the past few years, several groups developed deep learning models to address facets of polyadenylation regulation leveraging that the coevolutionary neural network can unbiasedly and quantitatively capture the dynamic interactions among cis‐regulatory motifs in a position‐specific manner. In this review, we summarize the historical bioinformatics analyses of polyA sites and the recent development of deep learning models. We discuss the advances using these deep learning models to decipher the motif grammars mediating polyadenylation, study polyA site expression, examine PASs on an evolutionary timescale, and quantify the impacts of genetic variants on polyadenylation activity.

2. Historical Bioinformatic Analyses of PASs

The earliest analyses of PASs relied on manual examination and identified AAUAAA as the canonical signal upstream of several eukaryotic polyA sites (Fitzgerald and Shenk 1981; Proudfoot and Brownlee 1976). Later on, the genome‐wide polyA sites were identified by polyA‐tailed RNA fragments using the expressed sequence tags (ESTs) (Graber et al. 1999a; Hajarnavis et al. 2004; Tian et al. 2005; van Helden et al. 2000). With the advances in deep sequencing, various approaches were developed to sequence 3′‐end fragments of polyA‐tailed RNAs and identify genome‐wide polyA sites, such as 3′READS (Hoque et al. 2013), 3P‐seq (Jan et al. 2011), PAS‐seq (Shepard et al. 2011), A‐seq (Martin et al. 2012), 3′‐seq (Lianoglou et al. 2013), 3′‐end‐seq (Haenni et al. 2012), and QuantSeq (Moll et al. 2014). Several databases were established to present the annotations of genome‐wide polyA sites identified from deep sequencing, such as PolyaDB (Wang, Nambiar, et al. 2018), PolyASite (Herrmann et al. 2020), and APADB (Muller et al. 2014).

With the identification of genome‐wide polyA sites, the commonly used bioinformatics approach was to identify motifs enriched around the sites. Because the protein subcomplexes bind to specific regions, polyA site sequences were split into subregions located upstream and downstream. For example, in humans, the sequences were divided into four subregions: −100/−41, −40/−1, +1/+40, and +41/+100 (Hu et al. 2005). Then, the frequency of a motif in each region is calculated. The significance of enrichment can be calculated using different statistical tests (e.g., Fisher's exact test), comparing motif occurrences in polyA sequences vs. other genomic sequences or randomized nucleotides. The analyses identified motifs contributing to the polyA site formation (Beaudoing et al. 2000; Hu et al. 2005; Tabaska and Zhang 1999). Similar analyses have been carried out to identify polyA site configurations in different model species such as humans, Drosophila, S. cerevisiae , Schizosaccharomyces pombe, and Arabidopsis (Beaudoing et al. 2000; Graber et al. 1999b; Hu et al. 2005; Hutchins et al. 2008; Liu et al. 2017; Loke et al. 2005; Retelska et al. 2006). Motif enrichment analyses are still frequently utilized for unbiased motif discovery or to validate significant motifs found via other approaches. However, the analyses have limited resolution since the sequence must be split into subregions; they do not elucidate the nucleotide‐level positional importance and are hardly applicable to examining motif interactions on polyadenylation regulation.

Using the enriched motifs in different polyA regions, groups developed machine learning models based on the motif occurrences to classify polyA sequences vs. random nucleotides. The earliest models used linear and quadratic discriminant functions to separate functional from non‐functional polyA signals based on the observed PAS AAUAAA and variants, downstream GU/U‐rich elements, and the surrounding nucleotide composition (Salamov and Solovyev 1997; Tabaska and Zhang 1999). The PolyaSVM model used the support vector machine to examine the combinatory effect of 15 motifs identified from the motif enrichment analyses and the strengths of motifs were measured using position‐weight matrices (Cheng et al. 2006). Later on, to further boost model performance, more complex modeling approaches combined multiple classifiers (Wu et al. 2012; Xie et al. 2013) and incorporated additional features of the polyA region, such as RNA secondary structures (Chang et al. 2011) and the thermodynamic properties of DNA and DNA‐protein interactions (Kalkatawi et al. 2012). A study also examined tissue‐specific APA using machine learning models with input features such as cis‐elements, distances between tandem polyA sites, and distances between polyA sites and splice sites (Weng et al. 2016). More detailed reviews of the statistical and machine learning polyA models were described by Ji et al. (2015) and Ye et al. (2023). However, the machine learning approaches have limitations. First, they require pre‐identification of predictive sequence features, which were curated manually or discovered through motif enrichment analyses. Second, although machine learning models can classify polyA sequences vs. random nucleotides, they cannot perform nucleotide‐resolution identification of cleavage sites and cannot quantitatively measure site strengths.

3. Deep Learning Models of Polyadenylation

The improved accessibility of deep learning approaches (e.g., Python frameworks Pytorch and TensorFlow) and the proliferation of large‐scale deep sequencing data sets in the past several years have ushered in a new era of studying polyadenylation regulation. Importantly, deep learning models are not limited to traditional tabular data forms and can directly process genomic sequences as input to model biological phenomena. Sequences are one‐hot encoded into matrices to create a mathematical representation of the input, and the convolutional neural network (CNN) can quantitatively capture the dynamic interactions among cis‐regulatory motifs in a nucleotide‐level and position‐specific manner. As a result, sequence‐based deep learning models do not require predefined features and outperform traditional machine learning approaches based on manually curated motif inputs. Pytorch and TensorFlow have commonly used Python deep learning frameworks.

For the model training, large‐scale polyA site expression data can be obtained from deep sequencing experiments, measuring endogenous polyA site expression or synthetic sites from massively parallel reporter assays (MPRAs). As described earlier, several 3′‐end sequencing techniques were developed to quantify polyA site expression. The sequencing reads can discern individual cleavage sites at single‐nucleotide resolution and quantitatively measure polyA site isoform abundance in the native context. Additionally, experimental results obtained from different cell types can be used to study APA regulation during different biological conditions. Alternatively, the MPRAs introduce hundreds of thousands of synthetic sequences and use deep sequencing to quantify their polyadenylation activities (Bogard et al. 2019; La Fleur et al. 2024; Vainberg Slutskin et al. 2019). The pool of sequences used in MPRAs can contain extensive sequence diversity to maximize the regulatory information captured, although this could also lead to the discovery of spurious regulatory elements that are not present in the genome. Because of the number of unique sequences possible and the control over sequence design, MPRAs are a popular source of training data for complex deep learning models.

Current deep learning models can be grouped into four categories by their functions: (1) classify polyA site sequences vs. random nucleotides; (2) quantify polyA site strength from sequence content; (3) predict the cleavage sites; and (4) examine APA regulation, typically by predicting APA site usage levels of the same genes. Here, we review the deep learning models addressing different aspects of polyadenylation regulation.

3.1. Models That Classify polyA Site Sequences vs. Random Nucleotides

These deep learning models take as input the sequences surrounding polyA sites and negative control sequences derived from other genomic regions or shuffled sequences. Deep learning couples feature extraction with prediction, and models built with top‐level convolutional layers can distill meaningful, interpretable motifs from input sequences in the form of position‐weight matrices. As a result, CNNs are commonly used to study polyA site formation, such as DeepPolyA (Gao et al. 2018), DeepGSR (Kalkatawi et al. 2019), and DeeReCT‐PolyA (Xia et al. 2019) (Figure 2A). Deep learning models with diverse architectures can also achieve highly accurate predictions, such as the combined CNN‐recurrent neural network (RNNs) used by DeepPASTA (Arefeen et al. 2019), PolyaID (Stroup and Ji 2023), and PolyaClassifier (Stroup and Ji 2024) (Figure 2B). Other model variants include the self‐attentive deep neural network SANPolyA (Yu and Dai 2020) and the self‐attention gated convolutional highway network PASNet (Guo et al. 2021). To optimize the classification accuracy, models that calculate the probability that a sequence represents a polyA site are typically trained using the binary cross‐entropy (BCE) loss function, which can also be modified to suit the requirements of individual models (Figure 2C). The models are adept at classifying polyA site sequences vs. negative controls, and the performance can be measured by area under the receiver‐operating characteristic curve (AUROC) and area under the precision–recall curve (AUPRC) values.

FIGURE 2.

FIGURE 2

Deep learning models of polyadenylation. (A and B) Common deep learning architectures used to model polyadenylation regulation include convolutional‐convolutional (A) and convolutional‐recurrent (B) neural networks. (C) Schematic diagram of models that classify polyA site sequences vs. random nucleotides. Relevant models are listed beside the illustration. (D) Similar to (C), except describing models that quantify polyA site strength. (E) Similar to (C), except describing models that examine APA regulation. (F) Similar to (C), except describing models that predict cleavage sites. (G) Similar to (C), except describing a model that combines polyA site classification and cleavage prediction.

3.2. Models That Quantify polyA Site Strengths

Endogenous polyA sites exhibit a wide range of activities/strengths in their sequence context despite being functional polyA sites. Deep learning models trained to classify polyA sites from random sequences often cannot distinguish strong vs. weak ones. Multiple methods were used to measure relative polyA site strengths. For some MPRA designs, the proximal polyA sites were randomly synthesized, while the distal site sequences were fixed (Bogard et al. 2019). The relative usage levels of proximal sites were used to measure site strengths. For endogenous polyA sites, the relative isoform abundance of APA sites in 3′ terminal exons is mostly determined by their relative strengths with stronger sites having higher usage levels, although other parameters play fine‐tuned roles in isoform expression, such as differential isoform stability and tandem site distances (Stroup and Ji 2023). The relative usage levels of APA sites can be calculated using the ratios between read counts supporting one site vs. the sum read counts of top expressed APA sites. The isoform ratio provides a rough estimate of the site strength.

The deep learning architectures take one‐hot encoded polyA site sequences as input, followed by CNN and/or RNN, and the output is the relative polyA site usage levels (Figure 2D). Several models were developed, such as Polyadenylation Code by (Leung et al. 2018), PolyaPredictors (Vainberg Slutskin et al. 2019), APARENT (Bogard et al. 2019), APARENT2 (Linder et al. 2022), and PolyaStrength (Stroup and Ji 2023, 2024). Such models generally use the mean squared error or modified cross‐entropy loss functions, depending on the scale and framing of the prediction task. The correlation between predicted and observed values and/or the binary classification of highly vs. lowly expressed APA sites in the same genes can be used to examine the model performance. By learning from the sequences around polyA sites with various usage levels, these models quantify the relative strength of polyA signals in the sequences and learn contributing cis‐regulatory elements.

3.3. Models That Examine APA Regulation

Many mammalian genes contain multiple polyA sites in 3′ terminal exons. As the selection of competing APA sites dictates the composition of 3′ UTR, models of APA regulation are useful for predicting the most likely transcript isoforms. The DeeReCT‐APA model (Li et al. 2022) quantifies the relative usage of multiple terminal exonic polyA sites within the same gene (Figure 2E). For a gene with N polyA sites, DeeReCT‐APA is composed of N small convolutional networks using 455 nt sequence centered around each polyA site as the input, which are joined by a bidirectional LSTM to learn interactions between polyA sites and then predict the relative usage level of each site. During training, DeeReCT‐APA minimizes the cross‐entropy error between the observed and predicted site usage. The model mostly considers local polyA site motif strengths in determining their relative site usage levels. However, multiple other parameters can regulate APA, such as the tandem polyA site distances and downstream gene distances (Stroup and Ji 2023; Weng et al. 2016), and those were not incorporated in the model training. It will be interesting to build new APA models incorporating these features. Moreover, recent studies examined the anticancer effects of the CPSF73 inhibitor JTE‐607 (Cui, Wang, et al. 2023; Liu et al. 2023). To understand the molecular basis of the specificity of JTE‐607 on polyA site targeting, a deep learning model named C3PO was built to identify sequence features that determine drug sensitivity using the MPRA data (Liu et al. 2023). It will be valuable to develop deep learning models to examine the molecular mechanisms mediating APA regulation in different biological conditions.

3.4. Models That Predict Site Cleavage

The relative polyA motif locations in the primary sequences determine the choices of cleavage sites. In humans, the cleavage happens between the CPSF and CstF binding sites with nucleotide preferences in A‐rich regions and after the CA dinucleotides (Wang, Zheng, et al. 2018). For some polyA sites, the 3′‐end cleavage is known to be imprecise, where cleavage heterogeneity can produce RNA isoforms with different 3′‐ends varying by only a few nucleotides. In S. cerevisiae , the 3′‐end isoforms are particularly heterogeneous, and the cleavage can happen in a region spanning dozens of nucleotides for some genes (Moqtaderi et al. 2013). The transcripts with a few nucleotide differences in 3′‐ends can have profound changes in stability (Geisberg et al. 2014) and structure (Moqtaderi et al. 2018). As a result, it is valuable to understand how cleavage events are determined by primary DNA sequences.

A few deep learning models have been developed to predict cleavage events using the DNA sequence as the input, including PolyApredictors (Vainberg Slutskin et al. 2019), APARENT (Bogard et al. 2019), and PolyaCleavage (Stroup and Ji 2024) (Figure 2F). During training, these models generally use the Kullback–Leibler divergence loss function. APARENT makes cleavage distribution predictions across the full length of the input sequence using CNN–CNN. In contrast, PolyaCleavage predicts the cleavage site distribution in a smaller 50 nt window around the center point of the input sequence using CNN–LSTM. The 50 nt window size was chosen based on the known spacing of PASs, as cleavage sites outside of this span are unlikely to be related. Processing of the input data can influence where predicted cut sites are located. APARENT tends to predict cleavage at non‐A nucleotides, while PolyaCleavage identifies cleavage sites at genomic As in agreement with earlier analysis of cleavage events (Wang, Zheng, et al. 2018).

To measure the model performance, the observed and predicted cleavage heterogeneity, as well as mean cleavage position (MCP), were calculated based on the sequencing reads' measured cleavage probability vector and the predicted cleavage vector, respectively (Bogard et al. 2019; Stroup and Ji 2023, 2024). The resulting MCP measures the weighted mean and indicates the most likely cleavage position. The cleavage heterogeneity was quantified by the entropy score (Stroup and Ji 2023, 2024). The Pearson correlation coefficients between the observed and predicted entropy scores and MCP values were used to examine algorithm performance.

3.5. The Model That Combines polyA Site Classification and Cleavage Prediction

The PolyaID model takes the 240 nt sequences as the input followed by CNN and LSTM layers, and it has two independent output branches: (1) calculate the probability of classifying polyA sequences vs. random nucleotides and (2) the cleavage probability vector in a 50 nt window in the middle of input sequences (Figure 2G) (Stroup and Ji 2023). The model used the BCE loss function for the classification branch and the Kullback–Leibler divergence loss function for the cleavage branch. By combining the sequence classification and cleavage site predictions, the model can scan the human genome and identify the putative polyA sites at the single nucleotide resolution. This advance opened up the possibility of identifying polyA sites from any input sequences and de novo sites derived from genetic mutations.

4. Extracting Predictive Information From the Deep Learning Models

When building a deep learning model, the input data set is commonly split into training, validation, and testing sets. The validation set was intricately considered during the model training. To avoid the overfitting issue, the model performance should be comparable among these three split data sets, especially the independent testing set. The well‐performed models should have extracted predictive information from the input nucleotide sequences. These models should be interpretable rather than be considered “black boxes.” Groups developed post hoc techniques to extract biological insight from trained models.

Some methods, such as visualizing convolutional filters and gradient saliency maps, are intrinsic to the model design. For models that begin to process the encoded genomic sequences with a convolutional layer, the model will “learn” a predetermined number of convolutional filters. These filters resemble common position‐weight matrices and can be visualized as sequence logos to expose the features within the input sequences that are used to make model predictions. However, the convolutional filter logos contain no information about the position or direction of their influence on model predictions.

Saliency maps are an alternative approach to revealing the internal reasoning of deep learning models and are commonly used in biomedical imaging contexts. These heatmaps visualize the influence of the input sequence by calculating the gradient of the model prediction to each nucleotide. Although saliency maps highlight regions that receive the most “attention” from the model, gradient‐based importance calculations are often difficult to interpret and may be misleading when the input features are complex (Majdandzic et al. 2023). As such, they need to be used with caution.

Alternatively, the input sequences can be engineered/mutated in specific ways, and the changes in the model predictions can be used to infer the importance of regulatory elements in the sequences. Studies performed systematic single nucleotide substitutions and calculated the changes of model predicted probabilities (Bogard et al. 2019; Leung et al. 2018; Li et al. 2022). If a mutation disrupts an important motif, the predicted probability will decrease. On the other hand, the predicted value will increase if a mutation creates a de novo enhancer motif.

To identify cis‐regulatory elements mediating genome‐wide polyA site formation, a motif disruption algorithm was used to systematically disrupt motifs around well‐expressed polyA sites (Stroup and Ji 2023, 2024). Individual hexamers are repeatedly replaced with randomized nucleotides. The resulting change in model predictions should correlate with the importance of the motif at a specific location. For example, a motif that is important to defining the polyA site formation should induce a large decrease in model predictions when it is replaced with random nucleotides. As a result, this motif disruption approach is model agnostic and can be universally applied to any quantitative predictor whose input is a genomic sequence. Its performance can be better than that of single nucleotide substitution because the motifs (e.g., hexamers) can better represent functional element units. Additionally, we can study motif crosstalk by examining the motif's importance as a function of the distance between another fixed motif. The analyses can reveal the complex crosstalk between both neighboring and distant PASs. Alternatively, the motifs can also be introduced to polyA sites, and the changes of predicted values will be examined. In combination, the approaches capture the position‐specific importance of PASs and the interaction/interdependence between them. When applied to a large pool of well‐expressed polyA sites, these interpretive techniques reveal the biological principles that shape the polyadenylation landscape genome‐wide.

The authors of the APARENT models developed a sequence engineering method called “SeqProp” to create polyA sequences that have specific isoform ratios or cleavage sites by maximally activating a segment of the model through gradient ascent (Bogard et al. 2019; Linder et al. 2022). A sequence position‐weight matrix is randomly initialized and then iteratively optimized through small changes to reach the target isoform ratios or cleavage profile. The resulting position‐weight matrics can then be sampled to produce many potential sequences with the desired APA characteristics.

5. Polyadenylation Biology Revealed by Deep Learning Models

5.1. Optimal polyA Site Configurations in Humans

PolyA sites are defined by surrounding cis‐regulatory elements, which are recognized by the protein subcomplexes forming the polyadenylation machinery to trigger endonucleolytic cleavage and polyA tail synthesis. As the conventional biochemical experiments are labor‐intensive and restricted to the predesigned small number of polyA site sequences, the systematic motif disruption/mutation analyses using the deep learning models can provide quantitative and unbiased measurements of motif importance.

AAUAAA is the canonical PAS, while one or two nucleotide variants can function as weaker PAS variants. Using deep learning models, the relative strengths of alternative PAS were examined by in silico experiments where targeted conversions of AAUAAA to other PAS variants were performed (Stroup and Ji 2023). When AAUAAA is changed to AUUAAA in the same sequence context, the PolyaStrength score decreased threefold, medium across well‐expressed sites (Figure 3A). Conversion of AAUAAA to other PAS variants decreased the predicted sequence strength 4.6‐ to 21‐fold. These differences are attributed to the specific binding affinities of CPSF to the motifs.

FIGURE 3.

FIGURE 3

The configurations of human polyA sites. (A) Bar plot showing the fold‐change in PolyaStrength scores after AAUAAA PAS signals are replaced with alternative variants. For each motif, we plot the median fold‐change with the 90% confidence interval (error bars). (B) The optimal arrangement of motifs around a human polyA site, as determined by motif disruption and crosstalk analyses. (C) Histogram showing the predicted PolyaStrength scores for 3′ READS‐supported human polyA sites from (N = 109,814).

Optimal motif locations related to polyA sites and motif crosstalk can be inferred based on deep learning models. For example, using the APARENT and PolyApredictors models, the first convolutional layer can be used to extract filter position‐weight matrices representing protein complex binding sites in different regions around polyA sites (Bogard et al. 2019; Vainberg Slutskin et al. 2019). The second convolutional layer can infer interactions of motifs located nearby.

Using the PolyaID and PolyaStrength models, the hexamer disruption algorithm captured nucleotide‐resolution motif importance around polyA sites (Figure 3B) (Stroup and Ji 2023). The PAS AAUAAA signal bound by the CPSF complex is optimal when located ∼20 nt upstream of the cleavage site. The CstF dimer binds both GU‐rich and U‐rich motifs downstream of cleavage sites. The GU‐rich motif should precede the U‐rich motif, with no more than 10 nt separating the two elements. The GU‐rich motif showed peak importance 10 nt downstream of the cleavage site vs. 18 nt for U‐rich motifs. Upstream of the cleavage site, the auxiliary UGUA motifs bound by CFI are strongest when 10–20 nt upstream of the AAUAAA, and the U‐rich FIP1 binding motifs should immediately follow the PAS 1–6 nt downstream. G‐rich motifs bound by hnRNP h/f or PCF11 contribute most to polyA site formation when found ∼40 nt downstream of the cleavage site and 10–25 nt downstream of the GU‐rich motif.

However, most endogenous human polyA sites do not contain all six key motifs in their optimal positions described above, even for the well‐expressed ones. Considering the distribution of PolyaStrength scores, the polyA site strengths vary > 1000‐fold (Figure 3C) (Stroup and Ji 2023). Diverse polyA site configurations are observed across the sites. 35% of polyA sites contain only the CPSF binding site and one CstF binding motif. 45% contain three motifs, 17% contain four motifs, and only < 3% of sites contain ≥ 5 motifs. The results indicate that most endogenous polyA sites are weak, and the diverse configurations can allow potential differential regulation of the site expression during biological processes.

5.2. Evolution of PASs Across Eukaryotic Kingdoms

Previous computational results using the motif enrichment analyses have shown that polyA site compositions are different across distantly related species, which were described in the review (Tian and Graber 2012). However, the detailed site configurations in most nonmammalian species remain largely unknown. Recently, the PolyaClassifer deep learning models were developed for two yeast species ( S. cerevisiae and S. pombe), as well as for the plant A. thaliana (Stroup and Ji 2024). The in silico motif disruption analyses identified the positional importance of the motifs contributing to the polyA site formation across the species.

Following this, we summarize polyA site composition differences across the four species: humans, S. cerevisiae, S. pombe, and A. thaliana (Figure 4A,B). First, although all species contain a strong A‐rich signal located 20 nt upstream of polyA sites like the mammalian PAS, the A‐rich motifs are quite degenerate and show relatively lower overall importance for polyA site formation in yeast and plant species. In humans, single nucleotide variants in the AAUAAA signal drastically reduce the motif importance (Figure 4B). In contrast, the PAS in S. cerevisiae shows only a slight preference for the UAAUAA hexamer, while AAUAAU and AUAAUA are the next strongest motifs in the PAS region (91%–96% relative importance vs. UAAUAA). Unlike for S. cerevisiae , AAUAAA is the strongest PAS for the fission yeast S. pombe; however, alternative PAS demonstrate equivalent importance, and there is very little sequence specificity beyond A/U‐richness (∆relative importance < 4% across the top five motifs). For A. thaliana , the top four PAS motifs exhibit nearly equivalent importance (UAAUAA, AAUAAA, AAUGAA, and AUAAUA).

FIGURE 4.

FIGURE 4

Motif importance profiles around human, yeast, and Arabidopsis polyA sites. (A) Left: Phylogenetic tree comparing the evolutionary relationship between humans, two yeast species Saccharomyces cerevisiae and Schizosaccharomyces pombe, and the plant Arabidopsis thaliana. Right: Motif importance profiles for significant motifs surrounding well‐expressed polyA sites in humans, yeasts, and plants. Motifs are grouped into families based on nucleotide composition. (B) Relative importance of the top five hexamers in the PAS region. For humans, the relative motif importance was calculated using the PolyaID model. The PolyaClassifier models were used for yeasts and plants.

Second, for human sites, the U‐rich elements show a maximum importance score located 20 nt downstream polyA sites and are bound by the CstF complex (Figure 4A). However, for all other species, the U‐rich elements show the highest importance when located surrounding the polyA sites. The data suggested that, unlike the human sites where the CstF complex binds downstream, the polyadenylation subcomplex bound to U‐rich elements is flanking the polyA sites.

Third, nonmammalian polyA sites show unique species‐specific contributing motifs (Figure 4A). In S. cerevisiae , the UA‐rich motifs (UAUAUA/AUAUAU or nucleotide variants) located upstream of PAS show a broad range of importance (to 150 nt upstream polyA sites). These UA‐rich motifs show higher overall importance scores in determining polyA site formation than the downstream PAS and U‐rich elements around polyA sites. In A. thaliana , the UGUA elements located 25–200 nt upstream of polyA sites show higher importance than PAS, and U‐rich elements also show importance in this far upstream region. In S. pombe, the polyA sites contain the unique UAG motifs right upstream of PAS, and GUA signals from 10 to 120 nt downstream of the polyA sites.

Currently, deep learning models have been developed for polyA sites of some species from the three kingdoms: animals, fungi, and plants. For the other eukaryotic kingdom, protist, polyA site configurations are unstudied. A recent study used long‐read direction RNA sequencing to study polyA sites of six protist species and found that the PAS and other auxiliary motifs were quite variable across species (Sajek et al. 2024). Overall, the PAS is degenerate and tends to be A‐rich, containing Us, which is similar to fungi and plants. However, other auxiliary motifs remain unknown. It will be interesting to use deep learning models to deconvolute full polyA site compositions of the species.

5.3. Cleavage Heterogeneity

The endonuclease cleavage of mRNA 3′‐ends can be imprecise, where cleavage heterogeneity can produce RNA isoforms that differ by only a few to dozens of nucleotides. Global differences in cleavage heterogeneity were observed across species. Comparing polyA sites in homologous genes, the cleavage of human sites is generally much more precise than that of yeast and A. thaliana , while S. cerevisiae polyA sites show the greatest cleavage heterogeneity (Figure 5A) (Stroup and Ji 2024). The earlier deep learning models, including PolyApredictors and APARENT models, showed that the cleavage positions can be predicted based on primary DNA sequences (Bogard et al. 2019; Vainberg Slutskin et al. 2019).

FIGURE 5.

FIGURE 5

Mechanisms driving cleavage heterogeneity around polyA sites. (A) The cleavage entropy around polyA sites in each species. Data are shown as the mean ± 95% confidence interval. (B) Schematic summary of the two mechanisms underlying cleavage heterogeneity in humans and Saccharomyces cerevisiae .

Recent work of PolyaID and PolyaCleavage models investigated detailed molecular mechanisms mediating the cleavage heterogeneity in humans and S. cerevisiae, respectively (Stroup and Ji 2023, 2024). The motif analyses and deep learning predictions were combined to compare the differences of polyA sites showing high vs. low cleavage heterogeneity. In humans, the cleavage heterogeneity is promoted by two different mechanisms (Figure 5B): (1) the multiple occurrences of upstream AAUAAA, which can lead to different choices of CPSF binding sites; (2) the longer distance between upstream AAUAAA and downstream U/GU‐rich motifs, which can result in far apart binding of CPSF and CstF complex binding leading to the heterogeneous cleavage in the middle. The cleavage branch of the PolyaID model can recapture the regulation by mutating input sequences and examining prediction changes (Stroup and Ji 2023).

In S. cerevisiae, two distinct types of polyA sites were observed (Figure 5C). One type shows the U‐rich motifs flanking the polyA sites. For these sites, the cleavage is pretty precise and only occurs in a few positions. Another type of polyA site does not have the U‐rich motif around cleavage sites. Instead, these cleavage sites are enriched with AU‐rich motifs. Extensive cleavage heterogeneity spanning dozens of nucleotides was observed for these polyA sites. The results indicated that the U‐rich motifs flanking the cleavage sites, which can be recognized by the CPF complex, are not essential for the recruitment of the cleavage and polyadenylation machinery, but function as the regulatory factor to strengthen polyA sites and induce more precise cleavage. Another parameter promoting cleavage heterogeneity is the multiple occurrence of upstream UA‐rich motifs which are bound by the CFIB complex. The UA‐rich motifs are essential for yeast polyadenylation complex recruitment and activity. The sequence disruption/mutation analyses showed that the PolyaCleavage model can capture these regulatory parameters in regulating cleavage heterogeneity (Stroup and Ji 2024).

5.4. The APA Regulation in 3′‐Most Exons

Multiple deep learning models were developed to examine differential polyA site expression of APA sites in 3′‐terminal exons (Figure 6A) (Bogard et al. 2019; Li et al. 2022; Stroup and Ji 2023, 2024). The occurrences of core polyA site motifs or site strengths are major determinants of relative site usage. Site strengths measured by the PolyaStrength and APARENT models are correlated, although they were trained based on endogenous and synthetic polyA sites, respectively.

FIGURE 6.

FIGURE 6

APA regulation in humans. (A) The types of APA are based on the polyA site location within the gene. Alternative sites may be used in the terminal exon or downstream region, affecting the 3′ UTR composition (terminal exonic APA). Upstream intronic sites may also be used, which alter the coding sequence in addition to the 3′ UTR (composite or skipped intronic APA). (B) Factors regulating the selection of proximal or distal polyA sites during terminal exonic APA. Terminal exonic polyA sites were annotated with relevant genomic information describing the gene structure, their relative location, and the surrounding transcriptional landscape. Proximal polyA sites demonstrating high vs. low usage were compared to their paired distal sites to determine the factors promoting proximal or distal APA site selection. (C) Factors promoting composite intronic APA. Putative composite intronic polyA sites were annotated with relevant genomic information describing their position within the gene structure. Used sites were compared to unused sites to determine the factors promoting composite intronic APA. (D) Similar to (C), except describing the factors promoting skipped intronic APA.

Using the resampling approach, the absolute and relative strengths of APA sites can be controlled, and other genomic parameters regulating relative polyA site usage were identified (Figure 6B) (Stroup and Ji 2023). First, the distance between tandem polyA sites matters. The proximal polyA sites have higher usage levels when located distant from the distal sites, presumably because the longer distance allows a longer time window for the polyadenylation complex to be recruited and assembled in the proximal. Second, tandem gene distances can also regulate APA site expression. The distal polyA sites are suppressed when the downstream gene is located on the opposite strand, and the distance is very short, which can cause the RNA polymerase II from two strands to collide and slow down in 3′ UTRs. Finally, the splicing activity of the last intron can impact the polyadenylation activity. For genes with smaller last introns, the proximal polyA sites in terminal exons show higher usage levels. The finding is in line with previous biochemical assays showing that efficient splicing of the last intron facilitates proper 3′‐end formation (Nesic et al. 1993; Nesic and Maquat 1994).

Deep learning models can be useful to examine the molecular mechanisms mediating APA regulation during biological processes. For example, in S. cerevisiae , under diauxic stress, the promoter‐proximal polyA sites show overall higher usage levels due to slower Pol II elongation, but the regulatory levels are not uniform across genes (Geisberg et al. 2022). Using the PolyaStrength model, the analysis results showed that weaker proximal sites exhibit more drastic upregulation of usage levels compared to the stronger ones (Stroup and Ji 2024). Similarly, the deep learning models can be applied to study the APA regulation during other processes.

5.5. Intronic Polyadenylation

A minimum functional polyA site can only contain the CPSF binding site (AAUAAA or variants) and downstream CSTF binding (U‐rich/GU‐rich elements) within ~40 nt. By random chance, the two motifs can co‐occur in the genome sequences mimicking a possible polyA site. The PolyaID model represents the first to unbiasedly scan a human genome and obtain the possible polyA sites at the single nucleotide resolution (Stroup and Ji 2023). Considering the gene body of protein‐coding regions, the median gene length is ~26 kb, and there are two PolyaID‐predicted sites per kb. Eighty‐nine percent of these sites are located in the introns, 3.2% are located in RefSeq‐annotated terminal exons, and 6.8% are in the 5 kb extended downstream regions. Thirty‐nine percent of terminal exonic sites and 6.0% of sites located in the downstream extended regions are expressed. For intronic sites, only 1.5% are expressed, supported by 3′‐end sequencing reads, indicating the suppression of intronic polyadenylation by splicing.

The deep learning models provided a novel avenue to address the question of what the differences are between expressed vs. nonexpressed intronic polyA sites (Stroup and Ji 2023). The intronic sites were divided into two groups: composite intronic sites and skipped intronic sites (Figure 6A). The skipped sites show splicing between the first upstream 5′ splice site (SS) and polyA site, while the composite type does not. The expressed intronic sites are stronger than nonexpressed ones, as measured by the PolyaStrength scores. To examine other genomic parameters regulating intronic polyadenylation, the site strengths can be controlled between the expressed and nonexpressed groups using the resampling approach. Then, the splicing parameters can be compared between the two sampled groups. For the composite type, the sites are more likely to be expressed when the surrounding splicing activity is lower, with weaker upstream 5′ SS and downstream 3′ SS and larger surrounding exons (Figure 6C). For the skipped type, a site tends to be expressed when the upstream splicing efficiency is higher (stronger SS and smaller exons), and the downstream splicing activity is lower (Figure 6D). The polyA sites located in the last introns have a higher probability of being expressed than those in the upstream introns. The reason is that the splicing of the last introns relies on the intronic definition, and the efficiency is lower than the upstream introns for which the splicing is dependent on the exon definition (Berget 1995).

5.6. The Genetic Variants Regulating Polyadenylation Activity

As whole genome sequencing and paired RNA sequencing data have been generated by different consortiums, studies have intersected the two types of data sets to identify genetic variants impacting polyadenylation activity. A study analyzed the genetic variants from the Genotype‐Tissue Expression (GTEx) project (Consortium 2020), identified those located near database‐annotated polyA sites, and selected the variants altering the PAS AAUAAA or AUUAAA. Further, using the RNA‐seq data, they used the expression quantitative trait loci (eQTL) analysis and identified 330 SNPs linked to significant changes in APA regulation in at least one tissue type (Shulman and Elkon 2020). However, the analyses were restricted to database‐annotated polyA sites and the variants mutating AAUAAA/AUUAAA.

Alternative approaches were developed to identify genetic variants linked to APA regulation. The studies identified the APA events using the RNA‐seq data, calculated the relative usage levels of polyA sites, and then used eQTL analyses to identify genetic variants significantly correlated with the APA regulation (Cui, Arnold, et al. 2023; Li et al. 2021). The analyses did not require that the variants alter polyadenylation motifs, and the underlying reasons linking the variants and APA regulation may not be clear.

The deep learning models have provided the opportunity to quantitatively examine the impacts of genetic variants on polyadenylation in silico, which do not require RNA‐seq data. The APARENT model analyzed the sequences around database‐annotated polyA sites and showed that hundreds of ClinVar variants (Landrum et al. 2025) can alter polyadenylation activity by > 2‐fold (Bogard et al. 2019; Linder et al. 2022). In the recent work, using the PolyaID and PolyaStrength models, the analyses expanded to polyA sites with high PolyaID‐predicted probabilities, including those created de novo by genetic mutations (Stroup and Ji 2023). Thousands of ClinVar‐annotated genetic variants were predicted to alter polyadenylation activity in terminal exons and introns. Many more variants impacting polyadenylation can be identified if considering genetic variants annotated by other databases. These model‐predicted regulatory variants represent a valuable catalog to be followed experimentally to examine their functional roles in diseases and trait formation.

6. Discussion

Deep learning represents an exciting advance in our ability to comprehensively investigate polyadenylation regulation at the genome scale. By combining novel modeling approaches with large endogenous sequencing data sets or synthetic MPRAs, researchers have uncovered critical regulatory principles that determine polyadenylation activity across the genome. In general, existing deep learning models encode genomic sequences and extract features using convolutions at the first layer, which identifies predictive position‐weight matrices that resemble motifs. These motif features are then fed into either a second convolutional layer or recurrent bidirectional LSTM units. These second layers learn complex interactions between motifs. Finally, the abstract representations are sent through dense, fully connected layers to make final predictions for polyA site classification, strength regression, cleavage distribution vectors, or some combination of these outputs. The exception is the APARENT2 model, which is composed of nested ResNet blocks to create a deep, fully convolutional model. In all cases, these sequence‐based deep learning approaches outperform machine learning models and do not rely on predefined input features. This highlights the advantage of flexible, dynamic deep learning models that can ingest large data sets to reveal patterns of polyadenylation regulation.

Multiple parameters are important to consider during data analyses in order to obtain a robust deep learning model. The high‐quality training data, including the positive and negative sets, need to be curated. Especially, there can be many possible negative examples across the genome, and the design of negative sets should cover many possible scenarios to reduce false positives. The bagging approach, taking the averaged predictions across multiple models using different negative sets, can mitigate the issue. Additionally, to avoid the overfitting issue, the model performance should be comparable across the independent training, validation, and testing data sets. Both AUROC and AUPRC values need to be considered for the classification task. Other types of loss functions/matrices, such as regression correlation values, can be used for modeling the continuous data types. To compare the performances of multiple models, the same benchmark data set should be used. The model predictive features can be extracted from the position‐weight matrices in the convolutional layers and/or using the motif disruption approaches. However, some hidden complex features can be hard to interpret and miss.

By establishing a powerful and extensible deep learning modeling platform such as those reviewed here, there are many avenues for future studies that can address more complex open questions in the polyadenylation field. Although polyA tails are universally present on mature mRNA, the cleavage and polyadenylation reactions are subject to species‐specific regulation that has not been fully appreciated. Additionally, polyadenylation is implicated in developmental processes and tissue identity, as well as disease initiation and progression. Sequence‐based deep learning models can accelerate scientific investigation into the regulatory factors mediating polyadenylation regulation in these settings and augment our experimental capabilities.

PolyA tails produced by the biochemical cleavage and polyadenylation reactions are universally present at the 3′‐ends of mature eukaryotic mRNAs. However, the evolutionary principles of decoding mRNA 3′‐end formation and its regulation across phylogenetic kingdoms are insufficiently explored. The few existing studies consider only small sets of closely related species, typically Metazoan species, or perform limited downstream analysis that cannot resolve the complex motif interactions present around polyA sites. These studies could be expanded to comprehensively examine polyadenylation regulation across evolutionary kingdoms. One could leverage multiple sources of 3′‐end data in conjunction with the deep learning approaches reviewed here to model species‐specific polyA site compositions. Then, by applying interpretive methods to these models, the cis‐regulatory elements that define polyA sites would be uncovered, even for species where direct quantification of 3′‐end formation is sparse.

Although 3′‐end formation is regulated by optimized cis‐regulatory elements that can tolerate only a certain degree of sequence degeneracy, mutations within PASs are associated with human disease. PolyA sites often exhibit tissue‐specific usage and regulation due to the variable expression of PAC subunits and/or other RNA‐binding proteins. This should be considered during future analyses as some variants may disrupt polyadenylation only in certain tissues. Currently, the molecular mechanisms underlying tissue‐specific APA regulation remain elusive. For example, the brain tends to use promoter distal polyA sites generating long 3′ UTRs, while the testis expresses promoter proximal sites (Ji et al. 2009; Zhang et al. 2005). It will be interesting to develop new deep learning models to examine polyA sites showing tissue‐specific expression. Additionally, future work in this area could investigate disease‐associated variants from large consortia, such as cancers and neuronal diseases. Clinically relevant mutations that reshape the polyadenylation landscape during disease initiation or progression could present therapeutic targets for experimental validation.

Altogether, the deep learning models have opened up novel avenues to study different aspects of polyadenylation regulation and quantitatively explore the determinants of polyA site formation, strength, cleavage heterogeneity, and site competition during APA. These models can be applied to study real‐world questions, such as APA regulation in clinical settings, genetic mutations altering polyadenylation activity in diseases, and RNA therapeutics design. These advances set the stage for future work to address polyadenylation regulation underlying different biological processes and human disease progression.

Author Contributions

Emily Kunce Stroup: conceptualization (equal), data curation (equal), formal analysis (equal), funding acquisition (supporting), investigation (equal), methodology (equal), software (lead), visualization (equal), writing – original draft (equal), writing – review and editing (equal). Tianjiao Sun: investigation (supporting). Qianru Li: visualization (supporting). John Carinato: investigation (supporting). Zhe Ji: conceptualization (equal), funding acquisition (lead), investigation (equal), project administration (lead), supervision (lead), visualization (equal), writing – original draft (equal), writing – review and editing (equal).

Conflicts of Interest

The authors declare no conflicts of interest.

Related WIREs Articles

Signals for pre‐mRNA cleavage and polyadenylation

Acknowledgments

This work was supported by grants to Z.J. from the National Institutes of Health (National Institute of General Medical Sciences: R35GM138192). E.K.S. was supported by the predoctoral training program in Biomedical Data Driven Discovery (U.S. National Library of Medicine: T32LM012203). We thank Bin Tian and the members of the Ji laboratory for their helpful discussions and comments.

Associate Editor: Carol Lutz

Editor‐in‐Chief: Jeff Wilusz

Funding: This work was supported by grants to Z.J. from the National Institutes of Health (National Institute of General Medical Sciences: R35GM138192). E.K.S. was supported by the predoctoral training program in Biomedical Data Driven Discovery (U.S. National Library of Medicine: T32LM012203).

Data Availability Statement

Data sharing is not applicable as no new data were created in this study.

References

  1. Arefeen, A. , Xiao X., and Jiang T.. 2019. “DeepPASTA: Deep Neural Network Based Polyadenylation Site Analysis.” Bioinformatics 35: 4577–4585. 10.1093/bioinformatics/btz283. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Arhin, G. K. , Boots M., Bagga P. S., Milcarek C., and Wilusz J.. 2002. “Downstream Sequence Elements With Different Affinities for the hnRNP H/H' Protein Influence the Processing Efficiency of Mammalian Polyadenylation Signals.” Nucleic Acids Research 30, no. 8: 1842–1850. 10.1093/nar/30.8.1842. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Beaudoing, E. , Freier S., Wyatt J. R., Claverie J.‐M., and Gautheret D.. 2000. “Patterns of Variant Polyadenylation Signal Usage in Human Genes.” Genome Research 10, no. 7: 1001–1010. 10.1101/gr.10.7.1001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Berg, M. G. , Singh L. N., Younis I., et al. 2012. “U1 snRNP Determines mRNA Length and Regulates Isoform Expression.” Cell 150, no. 1: 53–64. 10.1016/j.cell.2012.05.029. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Berget, S. M. 1995. “Exon Recognition in Vertebrate Splicing.” Journal of Biological Chemistry 270, no. 6: 2411–2414. 10.1074/jbc.270.6.2411. [DOI] [PubMed] [Google Scholar]
  6. Bogard, N. , Linder J., Rosenberg A. B., and Seelig G.. 2019. “A Deep Neural Network for Predicting and Engineering Alternative Polyadenylation.” Cell 178, no. 1: 91–106.e23. 10.1016/j.cell.2019.04.046. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Boreikaitė, V. , and Passmore L. A.. 2023. “3′‐End Processing of Eukaryotic mRNA: Machinery, Regulation, and Impact on Gene Expression.” Annual Review of Biochemistry 92, no. 1: 199–225. 10.1146/annurev-biochem-052521-012445. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Brown, K. M. , and Gilmartin G. M.. 2003. “A Mechanism for the Regulation of Pre‐mRNA 3′ Processing by Human Cleavage Factor Im.” Molecular Cell 12, no. 6: 1467–1476. 10.1016/S1097-2765(03)00453-2. [DOI] [PubMed] [Google Scholar]
  9. Chang, T.‐H. , Wu L.‐C., Chen Y.‐T., et al. 2011. “Characterization and Prediction of mRNA Polyadenylation Sites in Human Genes.” Medical & Biological Engineering & Computing 49, no. 4: 463–472. 10.1007/s11517-011-0732-4. [DOI] [PubMed] [Google Scholar]
  10. Cheng, Y. , Miura R. M., and Tian B.. 2006. “Prediction of mRNA Polyadenylation Sites by Support Vector Machine.” Bioinformatics 22, no. 19: 2320–2325. 10.1093/bioinformatics/btl394. [DOI] [PubMed] [Google Scholar]
  11. Colgan, D. F. , and Manley J. L.. 1997. “Mechanism and Regulation of mRNA Polyadenylation.” Genes & Development 11, no. 21: 2755–2766. 10.1101/gad.11.21.2755. [DOI] [PubMed] [Google Scholar]
  12. Consortium . 2020. “The GTEx Consortium Atlas of Genetic Regulatory Effects Across Human Tissues.” Science 369, no. 6509: 1318–1330. 10.1126/science.aaz1776. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Cui, Y. , Arnold F. J., Peng F., et al. 2023. “Alternative Polyadenylation Transcriptome‐Wide Association Study Identifies APA‐Linked Susceptibility Genes in Brain Disorders.” Nature Communications 14, no. 1: 583. 10.1038/s41467-023-36311-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Cui, Y. , Wang L., Ding Q., et al. 2023. “Elevated Pre‐mRNA 3′ End Processing Activity in Cancer Cells Renders Vulnerability to Inhibition of Cleavage and Polyadenylation.” Nature Communications 14, no. 1: 4480. 10.1038/s41467-023-39793-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Dichtl, B. , and Keller W.. 2001. “Recognition of Polyadenylation Sites in Yeast Pre‐mRNAs by Cleavage and Polyadenylation Factor.” EMBO Journal 20, no. 12: 3197–3209. 10.1093/emboj/20.12.3197. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Fitzgerald, M. , and Shenk T.. 1981. “The Sequence 5'‐AAUAAA‐3' Forms Parts of the Recognition Site for Polyadenylation of Late SV40 mRNAs.” Cell 24, no. 1: 251–260. 10.1016/0092-8674(81)90521-3. [DOI] [PubMed] [Google Scholar]
  17. Gao, X. , Zhang J., Wei Z., and Hakonarson H.. 2018. “DeepPolyA: A Convolutional Neural Network Approach for Polyadenylation Site Prediction.” IEEE Access 6: 24340–24349. 10.1109/ACCESS.2018.2825996. [DOI] [Google Scholar]
  18. Garin, I. , Edghill E. L., Akerman I., et al. 2010. “Recessive Mutations in the INS Gene Result in Neonatal Diabetes Through Reduced Insulin Biosynthesis.” Proceedings of the National Academy of Sciences of the United States of America 107, no. 7: 3105–3110. 10.1073/pnas.0910533107. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Geisberg, J. V. , Moqtaderi Z., Fong N., Erickson B., Bentley D. L., and Struhl K.. 2022. “Nucleotide‐Level Linkage of Transcriptional Elongation and Polyadenylation.” eLife 11: e83153. 10.7554/eLife.83153. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Geisberg, J. V. , Moqtaderi Z., Fan X., Ozsolak F., and Struhl K.. 2014. “Global Analysis of mRNA Isoform Half‐Lives Reveals Stabilizing and Destabilizing Elements in Yeast.” Cell 156, no. 4: 812–824. 10.1016/j.cell.2013.12.026. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Graber, J. H. , Cantor C. R., Mohr S. C., and Smith T. F.. 1999a. “Genomic Detection of New Yeast Pre‐mRNA 3′‐End‐Processing Signals.” Nucleic Acids Research 27, no. 3: 888–894. 10.1093/nar/27.3.888. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Graber, J. H. , Cantor C. R., Mohr S. C., and Smith T. F.. 1999b. “In Silico Detection of Control Signals: mRNA 3′‐End‐Processing Sequences in Diverse Species.” Proceedings of the National Academy of Sciences of the United States of America 96, no. 24: 14055–14060. 10.1073/pnas.96.24.14055. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Gruber, A. J. , and Zavolan M.. 2019. “Alternative Cleavage and Polyadenylation in Health and Disease.” Nature Reviews. Genetics 20, no. 10: 599–614. 10.1038/s41576-019-0145-z. [DOI] [PubMed] [Google Scholar]
  24. Guo, Y. , Zhou D., Li W., et al. 2021. “Identifying Polyadenylation Signals With Biological Embedding via Self‐Attentive Gated Convolutional Highway Networks.” Applied Soft Computing 103: 107133. 10.1016/j.asoc.2021.107133. [DOI] [Google Scholar]
  25. Guo, Z. , and Sherman F.. 1996. “Signals Sufficient for 3′‐End Formation of Yeast mRNA.” Molecular and Cellular Biology 16, no. 6: 2772–2776. 10.1128/MCB.16.6.2772. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Haenni, S. , Ji Z., Hoque M., et al. 2012. “Analysis of C. elegans Intestinal Gene Expression and Polyadenylation by Fluorescence‐Activated Nuclei Sorting and 3′‐End‐Seq.” Nucleic Acids Research 40, no. 13: 6304–6318. 10.1093/nar/gks282. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Hajarnavis, A. , Korf I., and Durbin R.. 2004. “A Probabilistic Model of 3′ End Formation in Caenorhabditis elegans .” Nucleic Acids Research 32, no. 11: 3392–3399. 10.1093/nar/gkh656. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Herrmann, C. J. , Schmidt R., Kanitz A., Artimo P., Gruber A. J., and Zavolan M.. 2020. “PolyASite 2.0: A Consolidated Atlas of Polyadenylation Sites From 3′ End Sequencing.” Nucleic Acids Research 48, no. D1: D174–D179. 10.1093/nar/gkz918. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Higgs, D. R. , Goodbourn S. E., Lamb J., Clegg J. B., Weatherall D. J., and Proudfoot N. J.. 1983. “Alpha‐Thalassaemia Caused by a Polyadenylation Signal Mutation.” Nature 306, no. 5941: 398–400. 10.1038/306398a0. [DOI] [PubMed] [Google Scholar]
  30. Hoque, M. , Ji Z., Zheng D., et al. 2013. “Analysis of Alternative Cleavage and Polyadenylation by 3′ Region Extraction and Deep Sequencing.” Nature Methods 10, no. 2: 133–139. 10.1038/nmeth.2288. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Hu, J. , Lutz C. S., Wilusz J., and Tian B.. 2005. “Bioinformatic Identification of Candidate Cis‐Regulatory Elements Involved in Human mRNA Polyadenylation.” RNA 11, no. 10: 1485–1493. 10.1261/rna.2107305. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Hutchins, L. N. , Murphy S. M., Singh P., and Graber J. H.. 2008. “Position‐Dependent Motif Characterization Using Non‐Negative Matrix Factorization.” Bioinformatics 24, no. 23: 2684–2690. 10.1093/bioinformatics/btn526. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Jan, C. H. , Friedman R. C., Ruby J. G., and Bartel D. P.. 2011. “Formation, Regulation and Evolution of Caenorhabditis elegans 3'UTRs.” Nature 469, no. 7328: 97–101. 10.1038/nature09616. [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Jenal, M. , Elkon R., Loayza‐Puch F., et al. 2012. “The Poly(A)‐Binding Protein Nuclear 1 Suppresses Alternative Cleavage and Polyadenylation Sites.” Cell 149, no. 3: 538–553. 10.1016/j.cell.2012.03.022. [DOI] [PubMed] [Google Scholar]
  35. Ji, G. , Guan J., Zeng Y., Li Q. Q., and Wu X.. 2015. “Genome‐Wide Identification and Predictive Modeling of Polyadenylation Sites in Eukaryotes.” Briefings in Bioinformatics 16, no. 2: 304–313. 10.1093/bib/bbu011. [DOI] [PubMed] [Google Scholar]
  36. Ji, Z. , Lee J. Y., Pan Z., Jiang B., and Tian B.. 2009. “Progressive Lengthening of 3′ Untranslated Regions of mRNAs by Alternative Polyadenylation During Mouse Embryonic Development.” Proceedings of the National Academy of Sciences of the United States of America 106, no. 17: 7028–7033. 10.1073/pnas.0900028106. [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Kaida, D. , Berg M. G., Younis I., et al. 2010. “U1 snRNP Protects Pre‐mRNAs From Premature Cleavage and Polyadenylation.” Nature 468, no. 7324: 664–668. 10.1038/nature09479. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Kalkatawi, M. , Magana‐Mora A., Jankovic B., and Bajic V. B.. 2019. “DeepGSR: An Optimized Deep‐Learning Structure for the Recognition of Genomic Signals and Regions.” Bioinformatics 35, no. 7: 1125–1132. 10.1093/bioinformatics/bty752. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Kalkatawi, M. , Rangkuti F., Schramm M., et al. 2012. “Dragon PolyA Spotter: Predictor of Poly(A) Motifs Within Human Genomic DNA Sequences.” Bioinformatics 28, no. 1: 127–129. 10.1093/bioinformatics/btr602. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Kamieniarz‐Gdula, K. , and Proudfoot N. J.. 2019. “Transcriptional Control by Premature Termination: A Forgotten Mechanism.” Trends in Genetics 35, no. 8: 553–564. 10.1016/j.tig.2019.05.005. [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Kaufmann, I. , Martin G., Friedlein A., Langen H., and Keller W.. 2004. “Human Fip1 Is a Subunit of CPSF That Binds to U‐Rich RNA Elements and Stimulates Poly(A) Polymerase.” EMBO Journal 23, no. 3: 616–626. 10.1038/sj.emboj.7600070. [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. La Fleur, A. , Shi Y., and Seelig G.. 2024. “Decoding Biology With Massively Parallel Reporter Assays and Machine Learning.” Genes & Development 38, no. 17–20: 843–865. 10.1101/gad.351800.124. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Landrum, M. J. , Chitipiralla S., Kaur K., et al. 2025. “ClinVar: Updates to Support Classifications of Both Germline and Somatic Variants.” Nucleic Acids Research 53, no. D1: D1313–D1321. 10.1093/nar/gkae1090. [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Leung, M. K. K. , Delong A., and Frey B. J.. 2018. “Inference of the Human Polyadenylation Code.” Bioinformatics 34, no. 17: 2889–2898. 10.1093/bioinformatics/bty211. [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Li, L. , Huang K. L., Gao Y., et al. 2021. “An Atlas of Alternative Polyadenylation Quantitative Trait Loci Contributing to Complex Trait and Disease Heritability.” Nature Genetics 53, no. 7: 994–1005. 10.1038/s41588-021-00864-5. [DOI] [PubMed] [Google Scholar]
  46. Li, Z. , Li Y., Zhang B., et al. 2022. “DeeReCT‐APA: Prediction of Alternative Polyadenylation Site Usage Through Deep Learning.” Genomics, Proteomics & Bioinformatics 20, no. 3: 483–495. 10.1016/j.gpb.2020.05.004. [DOI] [PMC free article] [PubMed] [Google Scholar]
  47. Lianoglou, S. , Garg V., Yang J. L., Leslie C. S., and Mayr C.. 2013. “Ubiquitously Transcribed Genes Use Alternative Polyadenylation to Achieve Tissue‐Specific Expression.” Genes & Development 27, no. 21: 2380–2396. 10.1101/gad.229328.113. [DOI] [PMC free article] [PubMed] [Google Scholar]
  48. Linder, J. , Koplik S. E., Kundaje A., and Seelig G.. 2022. “Deciphering the Impact of Genetic Variation on Human Polyadenylation Using APARENT2.” Genome Biology 23, no. 1: 232. 10.1186/s13059-022-02799-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
  49. Liu, L. , Yu A. M., Wang X., et al. 2023. “The Anticancer Compound JTE‐607 Reveals Hidden Sequence Specificity of the mRNA 3′ Processing Machinery.” Nature Structural & Molecular Biology 30, no. 12: 1947–1957. 10.1038/s41594-023-01161-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  50. Liu, X. , Hoque M., Larochelle M., et al. 2017. “Comparative Analysis of Alternative Polyadenylation in S. cerevisiae and S. pombe .” Genome Research 27, no. 10: 1685–1695. 10.1101/gr.222331.117. [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Loke, J. C. , Stahlberg E. A., Strenski D. G., Haas B. J., Wood P. C., and Li Q. Q.. 2005. “Compilation of mRNA Polyadenylation Signals in Arabidopsis Revealed a New Signal Element and Potential Secondary Structures.” Plant Physiology 138, no. 3: 1457–1468. 10.1104/pp.105.060541. [DOI] [PMC free article] [PubMed] [Google Scholar]
  52. Majdandzic, A. , Rajesh C., and Koo P. K.. 2023. “Correcting Gradient‐Based Interpretations of Deep Neural Networks for Genomics.” Genome Biology 24, no. 1: 109. 10.1186/s13059-023-02956-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  53. Mandel, C. R. , Bai Y., and Tong L.. 2008. “Protein Factors in Pre‐mRNA 3′‐End Processing.” Cellular and Molecular Life Sciences 65, no. 7: 1099–1122. 10.1007/s00018-007-7474-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  54. Martin, G. , Gruber A. R., Keller W., and Zavolan M.. 2012. “Genome‐Wide Analysis of Pre‐mRNA 3′ End Processing Reveals a Decisive Role of Human Cleavage Factor I in the Regulation of 3' UTR Length.” Cell Reports 1, no. 6: 753–763. 10.1016/j.celrep.2012.05.003. [DOI] [PubMed] [Google Scholar]
  55. Masamha, C. P. , Xia Z., Yang J., et al. 2014. “CFIm25 Links Alternative Polyadenylation to Glioblastoma Tumour Suppression.” Nature 510, no. 7505: 412–416. 10.1038/nature13261. [DOI] [PMC free article] [PubMed] [Google Scholar]
  56. Mayr, C. , and Bartel D. P.. 2009. “Widespread Shortening of 3'UTRs by Alternative Cleavage and Polyadenylation Activates Oncogenes in Cancer Cells.” Cell 138, no. 4: 673–684. 10.1016/j.cell.2009.06.016. [DOI] [PMC free article] [PubMed] [Google Scholar]
  57. McCracken, S. , Fong N., Yankulov K., et al. 1997. “The C‐Terminal Domain of RNA Polymerase II Couples mRNA Processing to Transcription.” Nature 385, no. 6614: 357–361. 10.1038/385357a0. [DOI] [PubMed] [Google Scholar]
  58. Mitschka, S. , and Mayr C.. 2022. “Context‐Specific Regulation and Function of mRNA Alternative Polyadenylation.” Nature Reviews. Molecular Cell Biology 23: 779–796. 10.1038/s41580-022-00507-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  59. Moll, P. , Ante M., Seitz A., and Reda T.. 2014. “QuantSeq 3′ mRNA Sequencing for RNA Quantification.” Nature Methods 11, no. 12: i–iii. 10.1038/nmeth.f.376. [DOI] [Google Scholar]
  60. Moqtaderi, Z. , Geisberg J. V., and Struhl K.. 2018. “Extensive Structural Differences of Closely Related 3' mRNA Isoforms: Links to Pab1 Binding and mRNA Stability.” Molecular Cell 72, no. 5: 849–861.e6. 10.1016/j.molcel.2018.08.044. [DOI] [PMC free article] [PubMed] [Google Scholar]
  61. Moqtaderi, Z. , Geisberg J. V., Jin Y., Fan X., and Struhl K.. 2013. “Species‐Specific Factors Mediate Extensive Heterogeneity of mRNA 3′ Ends in Yeasts.” Proceedings of the National Academy of Sciences of the United States of America 110, no. 27: 11073–11078. 10.1073/pnas.1309384110. [DOI] [PMC free article] [PubMed] [Google Scholar]
  62. Muller, S. , Rycak L., Afonso‐Grunz F., et al. 2014. “APADB: A Database for Alternative Polyadenylation and microRNA Regulation Events.” Database: The Journal of Biological Databases and Curation 2014: bau076. 10.1093/database/bau076. [DOI] [PMC free article] [PubMed] [Google Scholar]
  63. Nesic, D. , and Maquat L. E.. 1994. “Upstream Introns Influence the Efficiency of Final Intron Removal and RNA 3′‐End Formation.” Genes & Development 8, no. 3: 363–375. 10.1101/gad.8.3.363. [DOI] [PubMed] [Google Scholar]
  64. Nesic, D. , Cheng J., and Maquat L. E.. 1993. “Sequences Within the Last Intron Function in RNA 3′‐End Formation in Cultured Cells.” Molecular and Cellular Biology 13, no. 6: 3359–3369. 10.1128/mcb.13.6.3359-3369.1993. [DOI] [PMC free article] [PubMed] [Google Scholar]
  65. Pelechano, V. , Wei W., and Steinmetz L. M.. 2013. “Extensive Transcriptional Heterogeneity Revealed by Isoform Profiling.” Nature 497, no. 7447: 127–131. 10.1038/nature12121. [DOI] [PMC free article] [PubMed] [Google Scholar]
  66. Proudfoot, N. J. , and Brownlee G. G.. 1976. “3′ Non‐Coding Region Sequences in Eukaryotic Messenger RNA.” Nature 263, no. 5574: 211–214. 10.1038/263211a0. [DOI] [PubMed] [Google Scholar]
  67. Retelska, D. , Iseli C., Bucher P., Jongeneel C. V., and Naef F.. 2006. “Similarities and Differences of Polyadenylation Signals in Human and Fly.” BMC Genomics 7, no. 1: 176. 10.1186/1471-2164-7-176. [DOI] [PMC free article] [PubMed] [Google Scholar]
  68. Richard, P. , and Manley J. L.. 2009. “Transcription Termination by Nuclear RNA Polymerases.” Genes & Development 23, no. 11: 1247–1269. 10.1101/gad.1792809. [DOI] [PMC free article] [PubMed] [Google Scholar]
  69. Sajek, M. P. , Bilodeau D. Y., Beer M. A., et al. 2024. “Evolutionary Dynamics of Polyadenylation Signals and Their Recognition Strategies in Protists.” Genome Research 34, no. 10: 1570–1581. 10.1101/gr.279526.124. [DOI] [PMC free article] [PubMed] [Google Scholar]
  70. Salamov, A. A. , and Solovyev V. V.. 1997. “Recognition of 3′‐Processing Sites of Human mRNA Precursors.” Bioinformatics 13, no. 1: 23–28. 10.1093/bioinformatics/13.1.23. [DOI] [PubMed] [Google Scholar]
  71. Sandberg, R. , Neilson J. R., Sarma A., Sharp P. A., and Burge C. B.. 2008. “Proliferating Cells Express mRNAs With Shortened 3′ Untranslated Regions and Fewer microRNA Target Sites.” Science 320, no. 5883: 1643–1647. 10.1126/science.1155390. [DOI] [PMC free article] [PubMed] [Google Scholar]
  72. Schafer, P. , Tuting C., Schonemann L., et al. 2018. “Reconstitution of Mammalian Cleavage Factor II Involved in 3′ Processing of mRNA Precursors.” RNA 24, no. 12: 1721–1737. 10.1261/rna.068056.118. [DOI] [PMC free article] [PubMed] [Google Scholar]
  73. Sheets, M. D. , Ogg S. C., and Wickens M. P.. 1990. “Point Mutations in AAUAAA and the Poly (A) Addition Site: Effects on the Accuracy and Efficiency of Cleavage and Polyadenylation In Vitro.” Nucleic Acids Research 18, no. 19: 5799–5805. 10.1093/nar/18.19.5799. [DOI] [PMC free article] [PubMed] [Google Scholar]
  74. Shepard, P. J. , Choi E. A., Lu J., Flanagan L. A., Hertel K. J., and Shi Y.. 2011. “Complex and Dynamic Landscape of RNA Polyadenylation Revealed by PAS‐Seq.” RNA 17, no. 4: 761–772. 10.1261/rna.2581711. [DOI] [PMC free article] [PubMed] [Google Scholar]
  75. Shulman, E. D. , and Elkon R.. 2020. “Systematic Identification of Functional SNPs Interrupting 3'UTR Polyadenylation Signals.” PLOS Genetics 16, no. 8: e1008977. 10.1371/journal.pgen.1008977. [DOI] [PMC free article] [PubMed] [Google Scholar]
  76. Stacey, S. N. , Sulem P., Jonasdottir A., et al. 2011. “A Germline Variant in the TP53 Polyadenylation Signal Confers Cancer Susceptibility.” Nature Genetics 43, no. 11: 1098–1103. 10.1038/ng.926. [DOI] [PMC free article] [PubMed] [Google Scholar]
  77. Stroup, E. K. , and Ji Z.. 2023. “Deep Learning of Human Polyadenylation Sites at Nucleotide Resolution Reveals Molecular Determinants of Site Usage and Relevance in Disease.” Nature Communications 14, no. 1: 7378. 10.1038/s41467-023-43266-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  78. Stroup, E. K. , and Ji Z.. 2024. “Delineating Yeast Cleavage and Polyadenylation Signals Using Deep Learning.” Genome Research 34: 1066–1080. 10.1101/gr.278606.123. [DOI] [PMC free article] [PubMed] [Google Scholar]
  79. Tabaska, J. E. , and Zhang M. Q.. 1999. “Detection of Polyadenylation Signals in Human DNA Sequences.” Gene 231, no. 1–2: 77–86. 10.1016/S0378-1119(99)00104-3. [DOI] [PubMed] [Google Scholar]
  80. Tian, B. , and Graber J. H.. 2012. “Signals for Pre‐mRNA Cleavage and Polyadenylation.” Wiley Interdisciplinary Reviews: RNA 3, no. 3: 385–396. 10.1002/wrna.116. [DOI] [PMC free article] [PubMed] [Google Scholar]
  81. Tian, B. , and Manley J. L.. 2017. “Alternative Polyadenylation of mRNA Precursors.” Nature Reviews. Molecular Cell Biology 18, no. 1: 18–30. 10.1038/nrm.2016.116. [DOI] [PMC free article] [PubMed] [Google Scholar]
  82. Tian, B. , Hu J., Zhang H., and Lutz C. S.. 2005. “A Large‐Scale Analysis of mRNA Polyadenylation of Human and Mouse Genes.” Nucleic Acids Research 33, no. 1: 201–212. 10.1093/nar/gki158. [DOI] [PMC free article] [PubMed] [Google Scholar]
  83. Tian, B. , Pan Z., and Lee J. Y.. 2007. “Widespread mRNA Polyadenylation Events in Introns Indicate Dynamic Interplay Between Polyadenylation and Splicing.” Genome Research 17, no. 2: 156–165. 10.1101/gr.5532707. [DOI] [PMC free article] [PubMed] [Google Scholar]
  84. Vainberg Slutskin, I. , Weinberger A., and Segal E.. 2019. “Sequence Determinants of Polyadenylation‐Mediated Regulation.” Genome Research 29, no. 10: 1635–1647. 10.1101/gr.247312.118. [DOI] [PMC free article] [PubMed] [Google Scholar]
  85. van Helden, J. , del Olmo M., and Perez‐Ortin J. E.. 2000. “Statistical Analysis of Yeast Genomic Downstream Sequences Reveals Putative Polyadenylation Signals.” Nucleic Acids Research 28, no. 4: 1000–1010. 10.1093/nar/28.4.1000. [DOI] [PMC free article] [PubMed] [Google Scholar]
  86. Wang, R. , Zheng D., Yehia G., and Tian B.. 2018. “A Compendium of Conserved Cleavage and Polyadenylation Events in Mammalian Genes.” Genome Research 28, no. 10: 1427–1441. 10.1101/gr.237826.118. [DOI] [PMC free article] [PubMed] [Google Scholar]
  87. Wang, R. , Nambiar R., Zheng D., and Tian B.. 2018. “PolyA_DB 3 Catalogs Cleavage and Polyadenylation Sites Identified by Deep Sequencing in Multiple Genomes.” Nucleic Acids Research 46, no. D1: D315–D319. 10.1093/nar/gkx1000. [DOI] [PMC free article] [PubMed] [Google Scholar]
  88. Weng, L. , Li Y., Xie X., and Shi Y.. 2016. “Poly(A) Code Analyses Reveal Key Determinants for Tissue‐Specific mRNA Alternative Polyadenylation.” RNA 22, no. 6: 813–821. 10.1261/rna.055681.115. [DOI] [PMC free article] [PubMed] [Google Scholar]
  89. Wu, X. , Ji G., and Zeng Y.. 2012. “In Silico Prediction of mRNA Poly(A) Sites in Chlamydomonas reinhardtii .” Molecular Genetics and Genomics 287, no. 11–12: 895–907. 10.1007/s00438-012-0725-5. [DOI] [PubMed] [Google Scholar]
  90. Xia, Z. , Li Y., Zhang B., et al. 2019. “DeeReCT‐PolyA: A Robust and Generic Deep Learning Method for PAS Identification.” Bioinformatics 35, no. 14: 2371–2379. 10.1093/bioinformatics/bty991. [DOI] [PMC free article] [PubMed] [Google Scholar]
  91. Xie, B. , Jankovic B. R., Bajic V. B., Song L., and Gao X.. 2013. “Poly(A) Motif Prediction Using Spectral Latent Features From Human DNA Sequences.” Bioinformatics 29, no. 13: i316–i325. 10.1093/bioinformatics/btt218. [DOI] [PMC free article] [PubMed] [Google Scholar]
  92. Ye, W. , Lian Q., Ye C., and Wu X.. 2023. “A Survey on Methods for Predicting Polyadenylation Sites From DNA Sequences, Bulk RNA‐Seq, and Single‐Cell RNA‐Seq.” Genomics, Proteomics & Bioinformatics 21, no. 1: 67–83. 10.1016/j.gpb.2022.09.005. [DOI] [PMC free article] [PubMed] [Google Scholar]
  93. Yu, H. , and Dai Z.. 2020. “SANPolyA: A Deep Learning Method for Identifying Poly(A) Signals.” Bioinformatics 36, no. 8: 2393–2400. 10.1093/bioinformatics/btz970. [DOI] [PubMed] [Google Scholar]
  94. Zhang, H. , Lee J. Y., and Tian B.. 2005. “Biased Alternative Polyadenylation in Human Tissues.” Genome Biology 6, no. 12: R100. 10.1186/gb-2005-6-12-r100. [DOI] [PMC free article] [PubMed] [Google Scholar]
  95. Zhang, Y. , Sun Y., Shi Y., Walz T., and Tong L.. 2020. “Structural Insights Into the Human Pre‐mRNA 3′‐End Processing Machinery.” Molecular Cell 77, no. 4: 800–809.e6. 10.1016/j.molcel.2019.11.005. [DOI] [PMC free article] [PubMed] [Google Scholar]
  96. Zhu, Y. , Wang X., Forouzmand E., et al. 2018. “Molecular Mechanisms for CFIm‐Mediated Regulation of mRNA Alternative Polyadenylation.” Molecular Cell 69, no. 1: 62–74.e64. 10.1016/j.molcel.2017.11.031. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

Data sharing is not applicable as no new data were created in this study.


Articles from Wiley Interdisciplinary Reviews. RNA are provided here courtesy of Wiley

RESOURCES