O R I G I N A L A R T I C L E
Characterization and prediction of mRNA polyadenylation
sites in human genes
Tzu-Hao Chang• Li-Ching Wu•Yu-Ting Chen• Hsien-Da Huang•Baw-Jhiune Liu•
Kuang-Fu Cheng•Jorng-Tzong Horng
Received: 28 March 2009 / Accepted: 2 January 2011 / Published online: 1 February 2011 Ó International Federation for Medical and Biological Engineering 2011
Abstract The accurate identification of potential poly(A) sites has contributed to all many studies with regard to alternative polyadenylation. The aim of this study was the development of a machine-learning methodology that will help to discriminate real polyadenylation signals from randomly occurring signals in genomic sequence. Since previous studies have revealed that RNA secondary struc-ture in certain genes has significant impact, the authors
tried to computationally pinpoint common structural pat-terns around the poly(A) sites and to investigate how RNA secondary structure may influence polyadenylation. This involved an initial study on the impact of RNA structure and it was found using motif search tools that hairpin structures might be important. Thus, it was propose that, in addition to the sequence pattern around poly(A) sites, there exists a widespread structural pattern that is also employed during human mRNA polyadenylation. In this study, the authors present a computational model that uses support vector machines to predict human poly(A) sites. The results show that this predictive model has a comparable perfor-mance to the current prediction tool. In addition, it was identified common structural patterns associated with pol-yadenylation using several motif finding programs and this provides new insight into the role of RNA secondary structure plays in polyadenylation.
Keywords Bioinformatics Data mining Polyadenylation poly(A) Support vector machines (SVMs)
1 Introduction
The polyadenylation of mRNA is an essential cellular process by which most eukaryotic pre-mRNAs form their 30 ends (histone mRNAs are the major exceptions). Pre-vious studies have indicated that several cis elements of great importance participate in signaling most events of human polyadenylation. In general, there are two core elements essential for polyadenylation. One is the highly conserved AAUAAA hexamer (or a close variant), which is usually referred to as the polyadenylation signal (PAS) and is located 10–40 nucleotides (nt) upstream of the
T.-H. Chang H.-D. Huang
Institute of Bioinformatics and Systems Biology, National Chiao-Tung University, Hsin-Chu, Taiwan L.-C. Wu J.-T. Horng
Institute of Systems Biology and Bioinformatics, National Central University, Jhongli, Taiwan Y.-T. Chen J.-T. Horng (&)
Department of Computer Science and Information Engineering, National Central University, Jhongli, Taiwan
e-mail: [email protected] H.-D. Huang
Department of Biological Science and Technology, National Chiao-Tung University, Hsin-Chu, Taiwan B.-J. Liu
Department of Computer Science and Information Engineering, Yuan Ze University, Jhongli, Taiwan
K.-F. Cheng
Biostatistics Center, China Medical University, Taichung, Taiwan
K.-F. Cheng
Institute of Statistics, National Central University, Jhongli, Taiwan
J.-T. Horng
Department of Bioinformatics, Asia University, Taichung, Taiwan
poly(A) site [2, 12, 18, 26]. The other element is often referred to as a poorly conserved GU- or U-rich down-stream element and is located 20–40 nt downdown-stream of the poly(A) sites [16,26].
Traditional bioinformatics collects a large amount of cDNA sequences and Expressed sequenced tags (ESTs) with the aim of aligning the cDNA/ESTs and the genome sequence [4, 26, 29]. This has provided a systematic approach to the identification of poly(A) sites in genomes. A substantial amount of data generated computationally via cDNA/ESTs alignment is considered valid and, conse-quently, the dataset serves as an excellent resource for the studies related to polyadenylation machinery. The prediction of poly(A) sites takes advantage of cDNA/ESTs and because of their availability on a large scales, this approach has became practical. In early studies, the problem of poly(A) site prediction was transformed to the identification of a putative polyadenylation signal, which was thought to be primarily defined by the location of the poly(A) sites [16,25]. Since PASes are highly conserved elements in the region upstream of a poly(A) sites, a correctly identified PAS indicates that a real poly(A) sites is not far away. In view of this, recognition of PAS is considered to be an alternative solution to solve the problem of poly(A) site prediction. Reliable prediction of poly(A) sites plays a enhancing role in the exploration of the complex mechanism of alternative polyadenylation, since it involves the identification of cis elements and characterization of the flanking regions. The information revealed via prediction can be of great value when studying the mechanisms involved in polyadenylation as well as how gene regulation occurs due to alternative polyadenylation. The objective of this study was the devel-opment of a machine-learning methodology that will help to discriminate real polyadenylation signals from randomly occurring signals in genomic sequence.
2 Methods
2.1 Datasets
In this study, it was used a large number of positive sequences and negative sequences to train and test this model. A positive sequence is the human genomic sequence surrounding a poly(A) site. All the positive sequences were obtained from the PolyA_DB 2 database [15], which contains poly(A) sites identified for genes from several vertebrate species using alignments between cDNA/ESTs and the genome sequences. The authors retrieved 33,745 positive sequences from PolyA_DB 2 in total, which correspond to 14,078 human genes. Each positive sequence is 250 nt in length, spanning -125 to ?125 nt relative to the poly(A) site. A sequence was
defined as a single-site type if its associated poly(A) site was unique, otherwise it was be defined as a multiple-site type. Among all the positive sequences, 5,275 sequences are denoted as single-site type and 28,470 sequences were denoted as multiple-site type. In addition, it was obtained 2,327 sequences from the Erpin training data [16] to per-form an independent test. Each of these sequences is 200 nt in length with a candidate PAS in the middle. To test this model, it was prepared several types of negative sequences, that is sequences without real poly(A) sites. Each of neg-ative sequences is also 250 nt in length. The negneg-ative sets the authors used 6,000 sequence included randomized poly(A) regions (produced by randomizing the sequence surrounding a poly(A) site), 313,454 human mRNA coding sequences (CDS), 25,700 human 50-untranslated regions (50-UTRs) and randomized genome sequences. The human RefSeq mRNA coding sequences were obtained from NCBI Build 36 [23] (ftp://ftp.ncbi.nih.gov/genomes/
H_sapiens/). The 50-untranslated regions were
down-loaded from UTRdb release 22 [20] (http://www.ba.itb.
cnr.it/UTR/). The chromosome 1 sequence of the human
genome (hg17 version) was downloaded from the UCSC genome bioinformatics site (http://genome.ucsc.edu). It was also generated randomized sequences of same first order Markov model as human CDS, 50-UTRs and chro-mosome 1 sequence of human genome.
2.2 Test procedure
In this study, the authors tested the prediction model using all the positive sequences and all the categories of negative sequences, and compared its performance with polya_svm [8], the most current tool for poly(A) site prediction. It must be noted that the prediction model uses PAS location for prediction while polya_svm predicts the location of a potential poly(A) site directly. Therefore, the procedure used to evaluate this model and polya_svm should be clearly defined. Since this model could easily reject a sequence without a PAS, this would cause a large number of true negative or false negative predictions depending on the testing data. To avoid possible bias relative to the testing data, only sequences with PASes were taken into account. For positive and negative sequences, it was filtered out those without PASes through the simple approach of putative PAS detection, which will be illustrated later. The testing data the authors used is shown in Table3. A total of 27,573 sequences (4,908 single type, 22,665 multiple-type) were detected to have putative PASes. For each negative set, 500 sequences were randomly selected (from 14,958 Poly(A) region sequences randomized by 1st order Markov chain, 45,203 CDS, 16,368 CDS randomized by 1st order Markov chain, 3,156 50-UTR sequences, 4,645 50-UTR sequences randomized by 1st order Markov chain, 10,113 Genomic
sequences randomized by 1st order Markov chain and) pre-dicted by the model and polya_svm, which was repeated ten times to calculate mean values. Predictive accuracy was then measured as follows: Sensitivity: SN = (TP/(TP ? FN)), Specificity: SP = (TP/(TP ? FP)) Correlation coefficient: CC¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiTPTNFPFN
ðTPþFPÞðTPþFNÞðTNþFPÞðTNþFNÞ
p where TP is
true positive, TN is true negative, FN is false negative and FP is false positive.
Due to the differences between the predictive models, for polya_svm, a prediction was considered to be TP if the reported site is within 24 nt of a real poly(A) site, and is otherwise FN. For the model, a sequence was predicted as positive if a PAS is detected within 40 nt upstream of a real poly(A) site according to a previous study [26]. To yield Specificity, the number of TP derived from all of the positive testing set was scaled so that the size of the positive testing set was equal to the size of negative testing set.
2.3 Detection of candidate PASes
To prepare for training and testing, only the sequences with candidate PASes were retained. The authors referred to previous studies and selected frequently occurring PASes [2, 26]. The candidate PASes consist of the canonical AAUAAA hexamer, other 11 single-base variants of AAUAAA and one two-base variant of AAUAAA. A sequence having any of the 13 hexamers within 40 nt upstream of a poly(A) site was retained, otherwise the sequence was
discarded. To this end, it was implemented a filter to sequentially detect these 13 types of candidate PAS; they are shown in Table1in terms of their frequencies. For single-type poly(A) sites, *70% of them have AAUAAA, 14% have AUUAAA, *12% have one of the other 11 types of PAS hexamers and *7% of them do not have any known PAS. The pattern of PAS usage is consistent with observation reported in previous studies [16,26] and the ranking is approximately the same. Thus, in this study, the authors focused on the detection of these 13 hexamers, which are found in 81.7% of all poly(A) sites.
2.4 Extraction of k-mer features
For the upstream and downstream regions of a candidate PAS, the features the authors used are occurrences of k-length contiguous subsequences, that is k-mers [17]. In this study, it was used k-mer nucleotide patterns (k = 1, 2, 3) as the features, each of which has a frequency value. The same patterns, but appearing on different sides of a can-didate PAS, were treated as two distinct features. For example, the frequency of GC, a 2-mer pattern, should be counted separately when upstream and when downstream. Thus, a total number of 168 ð¼ðð4þ 42þ 43Þ 2ÞÞ
pos-sible words, that is features, were used in the first training stage. For a sequence with a candidate PAS, the authors retrieved 78 bases upstream as well as 78 bases down-stream of the PAS for generation of the k-mer features.
2.5 Characterization of the sub-regions around the PAS
The nucleotide composition of positive sequences as well as negative sequences was examined in order to charac-terize the sub-regions upstream and downstream of the polyadenylation signals. The nucleotide frequencies were visualized at each position in order to highlight regions that can significantly discriminate real PASes from look-alikes.
2.6 Detection of the core elements involved in the RNA secondary structure
To discover the structural patterns around poly(A) sites, the authors used several well-known motif finding programs, including Sfold [10], RNAfold [14], and RNAMotif [19], to identify possible RNA secondary structures that may be involved in polyadenylation. Based on previous observa-tions in the literature, it was assumed that there exists simple structures that flank the poly(A) sites and are to a certain extent currently unknown. Thus, the authors focused on a simple hairpin structure that contains a PAS in its loop or has a U-rich stem; this was because stem-loop structures commonly define protein-RNA binding sites.
Table 1 Top detected PAS hexamers
Frequency (%) Rank Single Multiple Hs Hs Hs.B Hs.B* AAUAAA 67.00 41.52 45.51 1 1 1 AUUAAA 14.14 14.57 14.51 2 2 2 UAUAAA 2.26 4.21 3.91 3 3 3 AGUAAA 2.60 3.49 3.35 4 4 4 AAGAAA 0.74 3.01 2.65 5 10 5 AAUAUA 1.00 2.13 1.96 7 5 6 AAUACA 1.19 2.02 1.89 8 8 7 CAUAAA 1.23 1.67 1.60 9 6 8 GAUAAA 0.89 1.50 1.40 10 7 9 AAUGAA 0.64 1.54 1.40 11 11 10 UUUAAA 0.74 2.39 2.13 6 9 11 ACUAAA 0.28 0.91 0.81 12 13 12 AAUAGA 0.32 0.64 0.60 13 12 13 Coverage 93.04 79.61 81.71
Human sequences located -40 to -1 nt upstream of poly(A) sites were used to detect hexamers that may function as polyadenylation signals. Single single-type poly(A) sites, Multiple multiple-type poly(A) site, Hs all human poly(A) sites in the material, Hs.B human result reported by Beaudoing et al. [2], Hs.B* human result reported by Tian et al. [26]
2.7 Machine learning
In this study, the authors used support vector machines (SVMs) as the machine learning method. As state-of-the-art classifiers, SVMs have been shown to have excellent empirical performance in prediction tasks. In addition, machine learning via SVMs is known to achieve good performance when identifying biological signals, such as translation initiation sites [34] and splice sites [30,32,33]. Thus, the authors used the SVM library LIBSVM for binary classification (http://www.csie.ntu.edu.tw/*cjlin/
libsvm) in which the C-support vector classification
(C-SVC) method and the radial basis kernel function (RBF) were applied at the default settings, i.e., cost = 1 and gamma = 1/15.
2.8 Integration of the different types of features
The authors designed a predictive model that was con-structed using two SVMs. In this model, the first SVM employs k-mer features (k = 1,2,3) and outputs a probability
value, which serves as an input value for the second SVM. The second SVM employs the contents of the characteristic sub-regions as features, which will be mentioned in Results. To train this model, it was randomly selected 4,000 positive sequences in addition to 6,000 negative sequences from the six types of negative set (1000*6).
3 Results
3.1 Characterization of the polyadenylation signals
First, the nucleotide composition of the genomic sequences of the single-type and multiple-type poly(A) sites were examined. For each poly(A) site, it was selected terminal sequences spanning -125 to ?125 nt surrounding the poly(A) site (Fig.1a, b). Both types of poly(A) sites have similar patterns in the -35 to ?35 region, in which the curve for each nucleotide acid reveals quick rises and falls. Upstream of the poly(A) site, an A-rich region is located from -25 to -15 and causes a drop in U-content (%U);
Fig. 1 Nucleotide composition across the -125/?125 region of asingle-type poly(A) sites bmultiple-type poly(A) sites. cThe difference between AU- and GC-ratio. The difference at each position is calculated from (AU- to GC-ratio) single_hs single-type poly(A) sites, multiple_hs multiple-type poly(A) sites
this is closely followed by a U-rich region (-15 to -5). Downstream of the poly(A) site, there is a visible rise in %U with a peak at around ?20; this spans a wider region and is closely mirrored by a decline in %A. This U-rich region is generally regarded as the area containing CstF-binding sites. Meanwhile, a sudden rise in %A at -1 indicates that cleavage preferentially occurs next to an Adenine. When multiple-type poly(A) sites are compared with single-type poly(A) sites, the difference between the AU- and GC-ratio is larger at almost each position in the vicinity of the former type of site with the exception of the cleavage site and the region containing PAS hexa-mers, which are both shown to be highly conserved (Fig.1c).
As the next step, in order to discriminate positive sequences from the various categories of negative sequences, it was analyzed nucleotide composition in each type. As Fig.2a shows, the Adenine peak at around posi-tion 1 corresponds to the upstream A-rich region in Fig.1a, b. Similar peaks are found in negative sequences, espe-cially those from CDS and hs_MC; however, these seem to reflect how often a randomly occurring PAS hexamer is found within an A-rich region. There is a decline in %A in the downstream region, which corresponds to the U-rich region (Fig.2b), and this can help significantly to dis-criminate real sequences from negative sequences. In addition, it was noticed a minor U-rich region was located between the PAS and the major U-rich region and that this results in a lower %C and %G relative to the whole sequence, as shown in Fig.2c, d.
To summarize, the authors identified the characteristics of polyadenylation signals as made up of the following sub-regions, which are shown in Fig.3:
(1) A non-G-rich region, spanning -20 to ?20 across the PAS.
(2) A major U-rich region, spanning ?20 to ?45. (3) A minor U-rich region, spanning ?3 to ?12. (4) A non-C-rich region, spanning ?6 to ?15. (5) A non-A-rich region, spanning ?17 to ?55.
Note that the positions described are relative to the PAS. The content of these five sub-regions, for example, the G-richness in the non-G-rich regions, was used by the SVM model for prediction.
3.2 Prediction of poly(A) sites by the SVM
The authors conducted an independent test using 2327 positive sequences and six types of negative sequences as previously described in Methods. For each negative set, 500 sequences were generated and predicted by the model and polya_svm version 1.1 using the default settings. The process was repeated ten times and means values are
presented. As shown in Table 2, this model is more sen-sitive than polya_svm, but only by a small amount. Com-parable false positive (FP) levels were predicted by the model and by polya_svm for the randomized sequences. Using most types of randomized sequences, our model showed a high specificity and Correlation coefficient, the exception being randomized poly(A) region sequences. Interestingly, the model outperforms polya_svm when randomized CDS and randomized 50-UTRs are used but shows an unexpected difference with real CDS and 50-UTRs, which requires further discussion.
3.3 RNA secondary structure
Here, the authors firstly tested the hypothesis that it is RNA secondary structures that make a real polyadenylation signal what it is, one key factor being recognition by the CPSF. To this end, several computer programs were used to pinpoint possible secondary structures. It was used RNAfold [14] with default parameters for the structure folding. Fig.4a shows the probability distribution at each site along the -40/? 40 region of PAS. In contrast with CDS, the result suggests that the PAS hexamers in the poly(A) sites and 50-UTRs have a high probability of being involved in a single-stranded structure, for instance, lying in a loop. This result was then verified using Sfold [10] with default setting to assess statistical folding. As shown in Fig. 4b, it was found that the AAUAAA hexamers (the middle of the sequence) tend to be unpaired when sequences were compared across all datasets, including single-type poly(A) sites, CDS and 50-UTRs. The same pattern was revealed when window sizes of 2 and 4 nt were used (data not shown).
Based on the above, it seems likely that polyadenyla-tion signals may stay unpaired during processing if no other factors interfere. However, the results for CDS and 50-UTRs showed that this property did not distinguish positive sequences from negative ones. Consequently, the authors turned the spotlight on the downstream U-rich region and used RNAMotif [19], which is a common RNA secondary structure search program. In this test, it was focused on simple hairpin structures in which the loop and the stem both have a flexible length of 6–10. G:U pairing was permitted in the stem in addition to the default Watson/Crick paring rule. Mispairs are allowed in the stem, with the base-paired limit set at 80% for the stem. Based on the positions in the literature, the authors counted the occurrence of PAS hexamers being entirely in a loop or just being part of it. As Table 3 shows, for example, the value 50.16 represents 50.16% of AAUAAA hexamers in poly(A) sites being present in the loop and therefore the results suggests that there is no obvious preference for AAUAAA and AUUAAA to lie in the
loop. Again, it was found a high percentage was found in negative sequences, especially randomized poly(A) region sequences.
For the downstream U-rich regions in the stem, the authors tried several Figures and eventually choose to set the threshold of U-richness at 60%. As shown in Table4,
Fig. 2 aAdenine. b Uracil. cCytosine. d Guanine frequencies at each position. In the vicinity of the poly(A) site (all_hs), CDS (hsCDS), randomized human CDS (hsCDS_MC), randomized poly(A) region (hs_MC), 50-UTRs (5UTR_nr), randomized 50-UTRs (5UTR_nr_MC) and
randomized genomic sequences (chr1_MC)
about 60% of real polyadenylation signals have down-stream U-rich regions that form the stems of hairpins and this is a relatively high correlation compared to other negative sequences. The value of Diff varies form *10 to *43%. The authors supposed such differences are mainly due to the U-content downstream of the different types of sequences.
Finally, it was explored the interaction between the two cases, that is, the authors identified those poly(A) sites with a PAS hexamer involved in the loop of one hairpin struc-ture and a downstream U-rich region forming the stem of another loop. Such an arrangement is exemplified by the SV40 L polyadenylation signal [31]. As a result, out of 27,573 sequences there were 8,977 matches, which is approximately one-third of the sequences that have a PAS. When examined in detail, most of the matches are found to
be multiple-type poly(A) sites (Table5). Given that the number of multiple-type sites is five times that of single-type sites, the association with this structure can not be inferred rigorously. For genes related to those matches, it was observed a 46% coverage of 13,756 human genes, which suggests that such a structural pattern might com-monly exist around poly(A) sites. In addition, preference in usage could be found in multiple-type genes and this could be associated with the role RNA structure plays in the selection of multiple poly(A) sites.
4 Discussion
In order to discriminate real polyadenylation signals from false ones, the characteristic of both the whole 30-UTR region and in its sub-regions motifs are curial. The tradi-tional conserved AAUAAA PAS hexamer located 10–40 nucleotides (nt) upstream of the poly(A) site [2,12,18,26] is not the only factor. Other features such as structure and small sequence variants are also important. In this predic-tive model, the authors took into account not only the general AU-rich environment around the poly(A) site but also the characteristic sub-regions, which reveal significant positional dependency. In a manner consistent with previ-ous findings, those sub-regions are supposed to harbor simple but important cis elements. A notable example is the specific AU-rich elements known as AREs, which represent the most common determinant of RNA stability in mammalian cells [7, 24]. This predictive model was found to be comparable to the most current prediction tool, polya_svm [8] and may have in many cases a higher sen-sitivity and specificity depending on the context of sequences in evaluation. It is noteworthy that when testing with CDS and 50-UTRs, both the model and polya_svm predicted a surprising number of false positives, but this was not the case with randomized CDS and 50-UTRs, where both showed excellent specificity. To explain this result, it was presume that a large number of ‘‘real sites’’ might in fact exist that are capable of satisfying the feature definitions of the predictive model and of polya_svm. For CDS, this would be consistent with previous findings that there are poly(A) sites in internal exons [26,29]. On the other hand, the false positives in the 50-UTRs may actually act as some form of regulatory element. However, this hypothesis will require considerable experimental evalua-tion to assess its validity.
The prediction result indicates that other unidentified features, such as RNA structures, may account for poly-adenylation activity among the false negative sequences. In the analysis of RNA secondary structure, it seems that PAS hexamers tend to be in single-stranded form and unlikely to be affected by surrounding sequences. Notably, the U-rich
Fig. 3 Characteristic sub-regions. PAS polyadenylation signal
Table 2 Comparison of the predictive model with the polya_svm approach
Our model Polya_svm
TP FN SN (%) TP FN SN(%) Poly(A) sites 1306 1021 56.12 1278 1049 54.92 Negative set Our model Polya_svm
TN FP SP (%) CC TN FP SP (%) CC Poly(A) region first-oder MC 424 76 78.65 0.312 446 54 83.54 0.332 CDS 417 83 77.13 0.302 432 68 80.12 0.330 CDS first-oder MC 483 17 94.28 0.403 469 31 89.84 0.363 50-UTR 408 92 75.27 0.288 441 59 82.28 0.345 50-UTR first-oder MC 482 18 93.96 0.402 482 18 93.84 0.393 Genome first-oder MC 473 27 91.21 0.388 481 19 93.52 0.397 MC Markov chain, Poly(A) region first-oder MC randomized sequences surrounding poly(A) sites, CDS coding region sequences, CDS first-oder MC randomized CDS, 50-UTR 50-UTR sequences, 50-UTR first-oder MC randomized 50-UTRs, TP true positives, FP false positives, TN true negatives, FP false positives, SN sensitivity, SP specificity, CC correlation coefficient
region downstream of polyadenylation signals probable form the stem of a simple hairpin structure, which may be regarded as a feature that will be able to improve the prediction of poly(A) sites in the future. In the last test, it was found that 32% of poly(A) sites with PAS hexamers have their PAS and downstream U-rich region involved in two separate hairpins. Overall, it was found that 46% of 13,756 human genes would seem to have this structural pattern around their poly(A) sites; clearly this might be related to functionality. Based on these observations, it was suggest that this simple hairpin structure is common across
human genes and such this structural pattern could be one of a number of functional RNA structures associated with polyadenylation. Since this structural pattern has not been pinpointed as important in the past, an extensive study is needed to delineate the significance of hairpin structures during mRNA polyadenylation. The authors hope this study has shed some light on the role that common RNA structures play in the complex mechanism of polyadenylation.
In most cases, the PAS serves as the binding site for the CPSF as soon as it is transcribed, while the GU- or U-rich element is bound by the CstF. There may be multiple
Fig. 4 Probability profiling of the loops by a RNAfold [14] and b Sfold [10] single_hs single-type poly(A) site (4,908 sequences), hsCDS human CDS (5,000 sequences), 5UTR_nr 50-UTRs (3,156 sequences). Note that in this test each sequence has the AAUAAA hexamer in the middle
Table 3 Statistics of PAS involvement in hairpin loops for the dif-ferent types of sequences
Percentage AAUAAA AUUAAA Other 11 types All types All_hs 50.16 53.99 50.08 50.82 hsCDS 37.07 44.01 40.23 40.23 hsCDS_MC 33.35 39.55 39.03 38.42 hs_MC 45.28 49.57 48.82 48.52 5UTR_nr 36.60 43.88 38.96 39.13 5UTR_nr_MC 27.09 37.11 36.56 35.89 chr1_MC 30.84 40.82 35.87 35.84 all_hs human poly(A) site, hsCDS human CDS, hsCDS_MC ran-domized CDS, hs_MC ranran-domized sequence of poly(A) region, 5UTR_nr 50-UTRs, 5UTR_nr_MC randomized 50-UTRs, chr1_MC, randomized genomic sequence of chromosome 1
Table 4 Statistics of downstream U-rich region involvement in hairpin stems for the different types of sequences
#Reported #All Percentage Diff (%) all_hs 16,532 27,573 59.96 hsCDS 16,027 45,203 35.46 24.50 hsCDS_MC 5,477 16,368 33.46 26.50 hs_MC 7,498 14,958 50.13 9.83 5UTR_nr 1,096 3,156 34.73 25.23 5UTR_nr_MC 792 4,645 17.05 42.91 chr1_MC 1,717 10,113 16.98 42.98 #Reported the number of sequences reported by RNAMotif [19], #All the size of dataset, Diff difference between all_hs and each negative sets
GU/U-rich downstream elements associated with a single poly(A) site, suggesting configuration may control the efficiency of polyadenylation [31]. When bound, coopera-tion between the CstF and the CPSF produces a greatly enhanced binding to the pre-mRNA substrate, because a weak interaction of the PAS with a CPSF can be com-pensated for by a strong interaction of the GU/U –rich element with a CstF, and vice versa [9,28]. Several human diseases have been reported to be caused by a malfunction of polyadenylation. The system involved include simian virus 40 (SV40), human immunodeficiency virus type 1 (HIV-1), human C2 complement, collagen and cyclooxy-genase-2 [1, 5, 6, 13, 21, 22, 27]. One example is the FOXP3 gene, where a point mutation with a polyadenyla-tion signal (AAUAAA–AAUGAA) can lead to IPEX syndrome [3]. Furthermore, some diseases may be ascribed to an abnormal level of mRNA 30end formation during the process of polyadenylation, such as hereditary thrombo-philia [11]. It was developed a comprehensive methodol-ogy for human poly(A) site prediction in this study and the authors hope this study assist the current understanding of features related to the polyadenylation.
References
1. Arhin GK et al (2002) Downstream sequence elements with different affinities for the hnRNP H/H’ protein influence the processing efficiency of mammalian polyadenylation signals. Nucleic Acids Res 30(8):1842–1850
2. Beaudoing E et al (2000) Patterns of variant polyadenylation signal usage in human genes. Genome Res 10(7):1001–1010 3. Bennett CL et al (2001) A rare polyadenylation signal mutation
of the FOXP3 gene (AAUAAA– [ AAUGAA) leads to the IPEX syndrome. Immunogenetics 53(6):435–439
4. Brockman JM et al (2005) PACdb: polya cleavage site and 30 -UTR database. Bioinformatics 21(18):3691–3693
5. Brown PH, Tiley LS, Cullen BR (1991) Efficient polyadenylation within the human immunodeficiency virus type 1 long terminal repeat requires flanking U3-specific sequences. J Virol 65(6): 3340–3343
6. Carswell S, Alwine JC (1989) Efficiency of utilization of the simian virus 40 late polyadenylation site: effects of upstream sequences. Mol Cell Biol 9(10):4248–4258
7. Chen CY, Shyu AB (1995) AU-rich elements: characterization and importance in mRNA degradation. Trends Biochem Sci 20(11): 465–470
8. Cheng Y, Miura RM, Tian B (2006) Prediction of mRNA poly-adenylation sites by support vector machine. Bioinformatics 22(19):2320–2325
9. Colgan DF, Manley JL (1997) Mechanism and regulation of mRNA polyadenylation. Genes Dev 11(21):2755–2766 10. Ding Y, Chan CY, Lawrence CE (2004) Sfold web server for
statistical folding and rational design of nucleic acids. Nucleic Acids Res 32(Web Server issue):W135–W141
11. Gehring NH et al (2001) Increased efficiency of mRNA 30 end formation: a new genetic mechanism contributing to hereditary thrombophilia. Nat Genet 28(4):389–392
12. Graber JH et al (1999) In silico detection of control signals: mRNA 30-end-processing sequences in diverse species. Proc Natl Acad Sci USA 96(24):14055–14060
13. Hall-Pogar T et al (2005) Alternative polyadenylation of cyclo-oxygenase-2. Nucleic Acids Res 33(8):2565–2579
14. Hofacker IL (2003) Vienna RNA secondary structure server. Nucleic Acids Res 31(13):3429–3431
15. Lee JY et al (2007) PolyA_DB 2: mRNA polyadenylation sites in vertebrate genes. Nucleic Acids Res 35(Database issue):D165– D168
16. Legendre M, Gautheret D (2003) Sequence determinants in human polyadenylation site selection. BMC Genomics 4(1):7 17. Liu H et al (2003) An in-silico method for prediction of
polyade-nylation signals in human sequences. Genome Inform 14: 84–93 18. MacDonald CC, Redondo JL (2002) Reexamining the
polyade-nylation signal: were we wrong about AAUAAA? Mol Cell Endocrinol 190(1–2):1–8
19. Macke TJ et al (2001) RNAMotif, an RNA secondary structure definition and search algorithm. Nucleic Acids Res 29(22): 4724–4735
20. Mignone F et al (2005) UTRdb and UTRsite: a collection of sequences and regulatory motifs of the untranslated regions of eukaryotic mRNAs. Nucleic Acids Res 33(Database issue): D141–D146
21. Moreira A et al (1995) Upstream sequence elements enhance poly(A) site efficiency of the C2 complement gene and are phylogenetically conserved. EMBO J 14(15):3809–3819 22. Natalizio BJ et al (2002) Upstream elements present in the 30
-untranslated region of collagen genes influence the processing efficiency of overlapping polyadenylation signals. J Biol Chem 277(45):42733–42740
23. Pruitt KD, Maglott DR (2001) RefSeq and locuslink: NCBI gene-centered resources. Nucleic Acids Res 29(1):137–140
24. Shaw G, Kamen R (1986) A conserved AU sequence from the 30 untranslated region of GM-CSF mRNA mediates selective mRNA degradation. Cell 46(5):659–667
25. Tabaska JE, Zhang MQ (1999) Detection of polyadenylation signals in human DNA sequences. Gene 231(1–2):77–86 Table 5 Statistics of poly(A) sites with PAS involvement in hairpin loops and the presence of downstream U-rich regions in hairpin stems
#Reported sites Percentage of all reported sites
#Related genes #Genes Percentage of all genes
All 8,977 6,390 13756 46.45
Single-type 1,378 15.35 1,378 5272 26.14
Multiple-type 7,599 84.65 5,012 8484 59.08 Note that the percentage of all reported sites (column 3) derives from values in column 2, e.g., 15.35 = 1378/8977 * 100. Percentage of all genes (column 6) is derived from the same row, e.g., 45.39 = 6390/14078 * 100
26. Tian B et al (2005) A large-scale analysis of mRNA polyade-nylation of human and mouse genes. Nucleic Acids Res 33(1): 201–212
27. Valsamakis A et al (1991) The human immunodeficiency virus type 1 polyadenylylation signal: a 3’ long terminal repeat element upstream of the AAUAAA necessary for efficient polyadenyly-lation. Proc Natl Acad Sci USA 88(6):2108–2112
28. Wahle E (1995) 30-end cleavage and polyadenylation of mRNA
precursors. Biochim Biophys Acta 1261(2):183–194
29. Yan J, Marr TG (2005) Computational analysis of 30-ends of
ESTs shows four classes of alternative polyadenylation in human mouse, and rat. Genome Res 15(3):369–375
30. Yeo G et al (2004) Variation in alternative splicing across human tissues. Genome Biol 5(10):R74
31. Zarudnaya MI et al (2003) Downstream elements of mammalian pre-mRNA polyadenylation signals: primary, secondary and higher-order structures. Nucleic Acids Res 31(5):1375–1386 32. Zhang MQ (2000) Discriminant analysis and its application in
DNA sequence motif recognition. Brief Bioinform 1(4):331–342 33. Zhang XH et al (2003) Sequence information for the splicing of human pre-mRNA identified by support vector machine classifi-cation. Genome Res 13(12):2637–2650
34. Zien A et al (2000) Engineering support vector machine kernels that recognize translation initiation sites. Bioinformatics 16(9):799–807
