- Methodology article
- Open Access
Predicting functional upstream open reading frames in Saccharomyces cerevisiae
© Selpi et al; licensee BioMed Central Ltd. 2009
Received: 4 November 2008
Accepted: 30 December 2009
Published: 30 December 2009
Some upstream open reading frames (uORFs) regulate gene expression (i.e., they are functional) and can play key roles in keeping organisms healthy. However, how uORFs are involved in gene regulation is not yet fully understood. In order to get a complete view of how uORFs are involved in gene regulation, it is expected that a large number of experimentally verified functional uORFs are needed. Unfortunately, wet-experiments to verify that uORFs are functional are expensive.
In this paper, a new computational approach to predicting functional uORFs in the yeast Saccharomyces cerevisiae is presented. Our approach is based on inductive logic programming and makes use of a novel combination of knowledge about biological conservation, Gene Ontology annotations and genes' responses to different conditions. Our method results in a set of simple and informative hypotheses with an estimated sensitivity of 76%. The hypotheses predict 301 further genes to have 398 novel functional uORFs. Three (RPC11, TPK1, and FOL1) of these 301 genes have been hypothesised, following wet-experiments, by a related study to have functional uORFs. A comparison with another related study suggests that eleven of the predicted functional uORFs from genes LDB17, HEM3, CIN8, BCK2, PMC1, FAS1, APP1, ACC1, CKA2, SUR1, and ATH1 are strong candidates for wet-lab experimental studies.
Learning based prediction of functional uORFs can be done with a high sensitivity. The predictions made in this study can serve as a list of candidates for subsequent wet-lab verification and might help to elucidate the regulatory roles of uORFs.
Different genes are expressed differently in different places, at different times and in different amounts. Misregulation of gene expression can cause an abnormality, leading to disease(s) or even cancer . Therefore, a complete understanding of gene regulation is important; one step towards this is to elucidate the roles of post-transcriptional regulatory elements.
Functional uORFs have been shown to play important roles in keeping organisms healthy, usually by controlling the synthesis of certain proteins which are harmful if over-synthesised [11, 12]. One example of this is the condition thrombocythaemia [, accessed on 12 September 2007] where blood contains too many platelets, a type of blood cell involved in blood clotting. People with this condition have a higher risk of developing a blood clot, a stroke or heart attack. The production of platelets is controlled by the hormone expressed from the gene thrombopoietin. According to a review , based on , under normal conditions, the uORFs of thrombopoietin mRNA act to limit the translation of the thrombopoietin gene and thus limit the production of the platelets in the blood cells. When uORFs are somehow eliminated from the thrombopoietin mRNA, the translation of thrombopoietin gene is increased and thus the amount of the platelets, causing thrombocythaemia.
To date, transcribed uORFs have only been verified in a small number of genes in several organisms. From these data, a partial understanding of how uORFs can regulate protein expression has been achieved. However, as more and more uORFs have been found in the mRNA of genes with critical roles, it has become important to get a complete understanding of how uORFs are involved in gene regulation. To be able to draw a complete understanding of the mechanism, we expect that a large number of experimentally verified functional uORFs will be needed.
Until recently, studies on uORFs have been largely limited to lab-based experiments. The most direct test to verify that uORFs are transcribed and whether they are functional is by comparing the amount of mRNA and the amount of protein produced from the main gene in its proper chromosomal context with and without site-specific mutation(s) on the uORF(s) of interest. The site-specific mutation is usually done on one of the bases of a uORF's start codon to remove the uORF. In general, these experiments to verify that uORFs are transcribed and whether they are functional, are costly and time-consuming (≈ 4 man-months per gene). As a result, the simplest approach to searching for functional uORFs, i.e., by sampling genes at random and testing their uORFs in the laboratory, is not effective, even for the simplest eukaryotic (the yeast Saccharomyces cerevisiae) genome. It has been suggested that no more than 10% of the 6000 yeast genes will have one or more functional uORFs and each of these genes will on average have two functional uORFs. Thus, if one searched for functional uORFs by selecting genes at random and testing them in the lab, then on average it would take ≈ 20 man-months to find a single functional uORF. Therefore, an in silico prediction method which can help in selecting sets of candidate functional uORFs for lab experimental studies is essential.
This study sets out to develop such a method using a machine learning technique and the yeast Saccharomyces cerevisiae as the model organism. Given relevant data, the method should automatically generate hypotheses which can then be used to predict novel functional uORFs.
Although a large number of genomic (DNA) sequences are now available, the task of computationally identifying functional uORFs is still very challenging. As explained earlier, a functional uORF is a transcribed uORF which can regulate the translation of the main gene (the associated CDS). Thus, in order to find functional uORF(s), one should ideally look for uORF(s) in the 5' UTR sequences. However, until recently, the start of 5' UTRs were only known for a small number of genes. Therefore, previous genome-wide computation studies trying to identify functional uORFs used intergenic sequences (sequences between two genes) instead [16, 17]. Determining the start of 5' UTRs computationally becomes even more difficult due to the fact that some genes have multiple transcription start sites . This situation makes the task of determining which genes contain uORF(s) in their 5'UTRs very challenging, not to mention identifying which of these uORFs are functional.
In this paper, a new approach to predicting functional uORFs is presented. There are three main differences between this work and the work described in . First, we employ a different machine learning system, Aleph . Second, instead of intergenic sequences, we use 5' UTR sequences. Third, in addition to the knowledge from S. cerevisiae's sequences, knowledge derived from sequences of other yeast species, an analysis of expression data sets, and Gene Ontology annotations are also used to form the background knowledge for Aleph. Why we think these heterogeneous data could be useful for this study and how we transform them into a suitable format for Aleph are discussed in the methods section.
Among many machine learning techniques, we chose inductive logic programming (ILP)  for the following reasons. First, ILP provides a richer representation than other machine learning techniques which are based on an attribute-value representation that cannot concisely represent the relationships between attributes in the uORF domain. For example, the attribute-value representation cannot concisely represent relationships between uORFs and UTRs because a UTR may have an arbitrary number of uORFs. Second, unlike other machine learning techniques, ILP has a beneficial feature in that it is able to bias inference to take into account background (domain) knowledge from domain experts and/or the literature [21, 22]. Third, ILP's input (examples and background knowledge) and output (hypotheses/rules) are all represented in predicate logic. This representation can be easily translated into English. Consequently domain experts can help with the selection and integration of potentially helpful knowledge and the final dissemination of discoveries to the wider scientific community. Lastly, ILP has been successfully applied to a diverse range of real-world problems, such as those reviewed in [23–25], and has been shown to have the potential to help in selecting sets of candidate functional uORFs for lab experimental studies .
Knowledge from 5'UTR sequences of S. cerevisiae
Detailed composition of uORFs obtained using getorf of the EMBOSS package.
Number of Genes
1,475 other genes
Detailed uORF composition from 18 studied genes within the collection obtained using getorf of the EMBOSS package.
Features that can determine the impact of a uORF on gene regulation, as suggested in  and , were extracted from the 5' UTR sequences. These include distance from the uORF to the start of the CDS in bases, frequency of AU and GC base pairs immediately upstream and downstream of each uORF, the length of the uORF in codons, number of uORFs found in the UTR, and the length of the UTR. The bases in positions -3 and +4 relative to each uORF were also extracted. Through experiments with mammalian sequences, these positions were found to give an optimum context for an AUG to be recognised by ribosome [30, 31]. All of this information was represented as ILP background knowledge [see Additional file 1: Table S-1].
Beside listing the instances of the background predicates, background knowledge in ILP can also be represented as rules. Here, some declarative rules, that are potentially useful for helping to identify functional uORFs, were created to examine uORF's and UTR's features and their relationships. These include rules that examine the abundance of AU and GC base pairs immediately upstream and downstream of each uORF, whether the base in position +4 relative to the uORF's start codon is G and whether the base A or G is present at position -3 relative to the uORF's start codon. These rules are given in [Additional file 1: Table S-2].
Knowledge from sequences of other yeast species
Most functional genomic elements are needed to preserve fitness and thus are conserved between closely related species. This insight may also apply to functional uORFs, that is to say, uORFs which are functional are likely to be conserved in closely related species; this has been demonstrated by the uORFs of GCN4 and CPA1 which are conserved in multiple fungal species [, Supplementary Material 1], [, Figure 1]. Therefore, information about uORFs in closely related species to S. cerevisiae could be beneficial for this study. Saccharomyces phylogeny is given in [Additional file 1: Figure S-1].
5' UTR sequences of several other species in the genus Saccharomyces would be ideal sources, but these are not available. However, upstream sequences of six other Saccharomyces species i.e., S. bayanus, S. castellii, S. kluyveri, S. kudriavzevii, S. mikatae and S. paradoxus are available. S. castellii and S. kluyveri are considered quite far from S. cerevisiae, and thus would be expected to have lesser degree of conservation to S. cerevisiae. Of the four closer to S. cerevisiae, S. paradoxus, S. mikatae and S. bayanus were studied recently in . It was found that the order of genes in these three genomes and S. cerevisiae is well conserved. It was also suggested that the three genomes and S. cerevisiae have diverged enough to allow functional elements to be recognised. Therefore, we chose to use these three species.
For each S. cerevisiae gene, the upstream sequences of 500 bases of the orthologs from S. paradoxus [, accessed on 3 May 2007], S. mikatae [, accessed on 3 May 2007] and S. bayanus [, accessed on 2 May 2007] were downloaded from Saccharomyces Genome Database (SGD) . 500 bases was chosen because 5' UTR lengths in S. cerevisiae are mainly distributed below 500 bases [18, 26]. Furthermore, according to , the mean length of intergenic regions in the Saccharomyces family is around 500 bases. From these upstream sequences, uORFs with minimum length 3 codons as well as their features were extracted. This information, which is similar to that of S. cerevisiae, was represented as background knowledge [see Additional file 1: Table S-3].
In determining whether a uORF is conserved in ortholog genes from different species, we looked for the presence of uORFs of approximately the same length (i.e., the difference in the lengths is not more than three codons; this criterion was used in  and ) at the same position in the sequence of uORFs relative to the CDS. This was done in preference to using a conventional sequence comparison approach that considers nucleotide sequences in a base-by-base manner. Our new method of testing for conservation is particularly useful for finding regulatory motifs in distantly related species where sequence similarity between the species can be low but the presence of regulatory motifs remains. We implemented this by defining several rules [see Additional file 1: Table S-4].
Knowledge from Gene Ontology annotations
Number of nodes in the first five levels in each of the GO categories.
In this work, we used the GO annotations for yeast genes [, version 24 March 2007] provided by SGD . Although not all of the GO terms are used for annotating yeast genes, the GO annotations for yeast can be very specific. There are some annotations that only cover one gene. For the purpose of our study, terms as specific as this are not useful. We want more general annotations for yeast genes, so that each used annotation covers more genes. GO slim [, accessed on 11 April 2007] provides such mapping. However, this mapping is too general for our study. Therefore, for each of the GO categories, we mapped the yeast genes to the third level terms of their GO annotations (if the original level of GO annotation is fourth, fifth and so on). The new mapping was given to ILP as background knowledge [see Additional file 1: Table S-5]. Several rules [see Additional file 1: Table S-6] were defined to allow ILP to relate a uORF with one or more function/process/component-annotation(s) of the main gene associated to that uORF.
Knowledge from expression data
Microarray data can be viewed as a gene expression matrix where each row represents a gene and each column represents a condition, and the value of each position in the matrix represents the expression of a certain gene in a certain condition. Such data allow us not only to study the expression of individual genes under different conditions in a genome-scale, but also allow us to group genes which respond similarly to a set of conditions. With regard to predicting functional uORFs, it has been suggested that the polysomal association study integrating microarray data sets from several different stress conditions can provide an efficient way to experimentally verify the predictions of functional uORFs .
Here, derived knowledge from an analysis of four microarray data sets measuring translational activity under rapamycin stress , oxidative (H2O2) stress , butanol stress and amino acid starvation  were included as background knowledge for ILP. This was done to investigate whether functional uORFs could be explained in terms of how genes respond to different stress conditions. The polysome-to-monosome log-fold change between stressed and normal condition was used to determine whether the expression of a gene is up/down/not-regulated under each stress. The latter information was given to ILP as background knowledge [see Additional file 1: Table S-7]. We also defined two rules [see Additional file 1: Table S-8] to relate a uORF with information on whether the main gene associated to that uORF is regulated (either up or down) or not regulated under certain stress. A more detailed description of how the microarray data were analysed is given in [Additional file 1: Part III].
Since our goal is to learn how to recognise which uORFs regulate gene expression, we can consider this learning task to be a classification problem. Ideally, a typical classification system in ILP (or machine learning in general) learns from a mixture of positive and negative examples. In this domain, positive examples would be uORFs that are transcribed and regulate gene expression (i.e., functional) and negative examples would be uORFs that are transcribed but do not regulate gene expression (i.e., non-functional). The uORF data from 1,475 genes (Table 1) are all unlabelled. Hence, for the training stage in this study, only the uORF data of the 18 studied genes were used.
As summarised in Table 2, among the uORF data of the 18 studied genes, 21 uORFs have been verified experimentally as functional . These were used as positive examples. [, p. 32] pointed out that there are only two uORFs from two genes which have been verified to be non-functional and the 5' UTR of each of these genes does not contain any other uORFs known to be functional. Therefore, there were only two negative examples in our data set. Given the characteristics of the data, we explored the positive-only setting  of an ILP system Aleph . The positive-only setting of Aleph allows induction of hypotheses in the absence of negative examples.
We investigated whether our new approach could generate hypotheses with good performance. Only the positive examples of the 18 studied genes were used. Aleph's parameter settings and the definition of hypotheses space are given in [Additional file 1: Tables S-9 and S-10]. Since there were only 21 positive examples, evaluation was done using leave-one-out cross-validation. This means that each example in turn is used as a test set, while the other 20 examples are used as a training set. Thus in total, we did 21 executions.
Since we only used positive examples, the performance of the hypotheses was measured using sensitivity (or recall), which measures the fraction of positives which are recognised by the hypotheses as positives. In 16 executions, from a total of 21, the hypotheses can correctly recognise the test set. Thus, the estimate of how well our hypotheses can correctly identify functional uORFs is 76%.
Predicting novel functional uORFs
The English translations of the hypotheses generated from the set of 21 positive examples.
A uORF has functional role if it satisfies at least one of the following rules.
the main gene is regulated under butanol stress and the product of the main gene is involved in nucleic acid binding;
the uORF is conserved in two other species, the main gene is localised in intracellular (or protoplasm), and the UTR length > = 463;
the uORF is conserved in three other species and the main gene is localised in intracellular (or protoplasm);
its length < = 7 and the product of the main gene is involved in nucleic acid binding;
the base in position +4 relative to the uORF's start codon is 'G' and the main gene is involved in regulation of biological process;
its length < = 6, the main gene is localised in intracellular (or protoplasm), and the main gene is involved in regulation of biological process;
the uORF is conserved in two other species, the main gene is not regulated under low concentration of H2O2, the main gene is localised in intracellular (or protoplasm), and the UTR length > = 244;
the product of the main gene is involved in translation regulator activity.
The fact that negative examples were not used during training raises a suspicion that the hypotheses that have been generated could have been overly general. Furthermore, the performance measure used (sensitivity) does not penalise over-generalisation and so will not indicate if over-generalisation has arisen. Note that a hypothesis which simply states that any uORF is a functional uORF would have a sensitivity of 100%. Hence there was a need for an additional test to determine whether this set of hypotheses were overly general i.e., tend to clasify any example as functional uORF.
The uORF data on these 18 studied genes is precious due to its scarcity and therefore it made sense to utilise every part of the limited data available. The set of hypotheses shown in Table 4 was used to classify the negative (2) and unlabelled (8) examples within the 18 studied genes. Only 2 of 10 examples were classified as positives; one from the negative set and one from the unlabelled set. Thus, we believe that the high sensitivity is not because the hypotheses tend to classify any example as positive.
When the same set of hypotheses was used to clasify 3,616 unlabelled examples from 1,475 genes, they predict 398 uORFs from 301 genes as functional. The 398 predicted functional uORFs are listed in additional file 1: Table S-11. Generally, more precise mapping of transcription start sites in yeast will in some cases help confirm whether these uORFs are real or not, and some of these predicted functional uORFs may turn out to be artifact due to errors in the transcription start site prediction. Thus, extensive lab work would be required to verify whether these 398 uORFs from 301 genes are indeed functional. However, there are other observations that provide support for our predictions (see discussion section).
The novel approach to predicting functional uORFs in the yeast S. cerevisiae presented here makes use of knowledge about biological conservation, GO annotations, and genes' response to different stress conditions; while there have been several studies involving machine learning which make use of expression data and/or GO annotations (e.g. [48–51]), such a combination of knowledge has not been explored previously for learning yeast functional uORFs.
To date, there are very few computational studies on uORFs. The most closely related work are the studies by  and , and these are discussed in detail below.  Studied uORFs in the genome of fungal pathogen Cryptococcus neoformans with the aim of finding the proportion of uORFs conserved in four strains of C. neoformans. Similar to  looked for conserved uORFs between human and mouse genomes.
Comparison between predictions made by our hypotheses (Table 4) and by  for the seven genes that they wet-experimentally tested.
Predicted as functional in
The work described here provides some improvements compared to , and in the remainder of this section we discuss the differences in the methodology, hypotheses and predictions. Concerning methodology, the computational system used in  was an expert system shell with a certainty factor model for representing uncertainty in both the data and the rules/hypotheses. To generate the initial hypotheses, a rule base is constructed manually. The final set of hypotheses was generated after several cycles of running the expert system with the input data, analysing the results, and manually modifying the rule base for better classification. In contrast, in the work described here, the hypotheses are generated automatically by the ILP system. Thus, when applying the method for learning functional uORFs in other organisms, applying the method described here will be more practical than the one described in .
In , each rule for inferring whether a uORF was likely to affect gene expression was assigned a certainty factor representing the confidence in a consequent of the rule being true if all of the antecedents are true. If a uORF was predicted to be functional using two or more different lines of inference, then the certainty factors associated with these were combined, and the resulting combined certainty factor was used to score the uORF. The highest certainty factor value for any one of a gene's uORFs was used as the score for the gene itself, and those genes with a score above a selected threshold were classified as having a functional uORF. In that work, genes could be ranked according to their scores. In contrast, the ILP approach used in this paper gives a boolean "yes or no" prediction for whether each uORF has a functional role, and there are no scores that can be used for ranking. Thus, unlike in , there is no notion of having "strongly predicted" genes in the present work. The fact that one approach produced a ranked list of predictions while the other resulted in an unordered set of predictions makes a direct comparison of their results difficult.
While it is plausible that collectively, the rules listed in Table 4 will have a relation to the probability of uORFs having a functional role, it is uncertain at this point how the specific combinations of criteria that make up the rules relate to a biological context. This is primarily due to the small number of examples underlying each rule. Several of the derived criteria are generally in agreement with previously identified properties of genes with functional uORFs. Thus, two of the rules require the 5' UTR length to be over a certain value (Rules 2 and 7), which is consistent with uORFs generally being placed in genes with long untranslated regions. Rule 5 suggested that the base G in position +4 relative to the uORF's start codon is a favourable context for yeast. Previous research on mammals [30, 31] have shown that AUG context has influence on the recognition of AUG by the ribosome. The optimum AUG context for mammalian genes was found to be A or G in position -3 and G in position +4, where the A of uORF's AUG is position +1. A or G in position -3 have also been shown as favourable context in yeast too . A requirement for short uORFs has been marked in two rules (Rules 4 and 6). This is in agreement with the finding in  that uORFs conserved in evolution between Saccharomyces species are shorter than non-conserved (and presumably non-functional) uORFs. It is also interesting to note that several of the rules (Rules 1, 4, 5, 6, and 8) imply a role in regulation of transcription for the gene product. In , among genes with predicted functional uORFs, an overrepresentation with products implicated in transcription was found by analysis of GO terms.
By comparing the hypotheses in Table 4 with those described in  further, the following observation was made. Although 50 to 250 nucleotides was considered as the optimal distance between a functional uORF and the CDS in , this feature does not appear in the hypotheses in Table 4, indicating that the learning system used here considers this feature to be less important. This is due to the use of 5' UTR sequences; around 88% of the 5' UTR sequences of S. cerevisiae used here are not more than 250 nucleotides, whereas the intergenic regions used in  are generally much longer than 5' UTRs.
There are differences in the predictions described in  and those described here. In , 245 additional genes were predicted to have 367 new functional uORFs. Among these, 34 uORFs from 32 genes were strongly predicted to be functional. Among the strongly predicted ones, 24 uORFs from 23 genes lie within the 5'UTRs based on calculation from . When we checked how many of these 24 uORFs from 23 genes were also predicted as functional by the hypotheses in Table 4, we found eleven uORFs from genes LDB17, HEM3, CIN8, BCK2, PMC1, FAS1, APP1, ACC1, CKA2, SUR1, and ATH1 [see Additional file 1: Table S-12]. This suggests that these eleven genes are strong candidates for lab experimental studies. Moreover, HEM3 has been investigated by  and has been confirmed to have one real uORF.
We have taken a new approach to learning functional uORFs in the yeast S. cerevisiae. The method, which can help to select sets of candidate functional uORFs for lab experimental studies, uses the positive-only setting of an ILP system called Aleph and makes use of knowledge derived from biological sequences of several different yeast species, an analysis of several publicly available expression data sets, and Gene Ontology annotations; this is the first time such a combination of knowledge has been explored for learning yeast functional uORFs. With only a little adjustment and provided the relevant data are available, our method can be applied to the task of learning functional uORFs in other organisms. The heterogeneous knowledge used here allows Aleph to generate a set of hypotheses with reasonably high sensitivity (76%). While the idea of using conservation for learning functional uORFs is not new, the way in which conservation is tested here is new.
Our hypotheses are simple and informative. They are quite specific yet general enough to cover different types of functional uORFs. The hypotheses provide provisional insights into biological characteristics of functional uORFs. These may include being conserved in at least two other yeast species, the main gene's product being involved in regulation of biological process, translation regulator activity, and in nucleic acid binding, as well as the main gene being regulated or not regulated under certain stress.
When the hypotheses were used to predict novel functional uORFs from a set of unlabelled uORFs within the genome of S. cerevisiae, they predict 301 further genes to have 398 novel functional uORFs. Three (RPC11, TPK1, and FOL1) of these 301 genes have been hypothesised, following wet-experiments, to have functional uORFs . Finally, a comparison of our predictions here and those in  suggests that a set of eleven predicted functional uORFs from genes LDB17, HEM3, CIN8, BCK2, PMC1, FAS1, APP1, ACC1, CKA2, SUR1, and ATH1 are strong candidates for lab experimental studies. The predicted functional uORFs have yet to be tested biologically. Positive results are certainly hoped for. However, whatever the biological test results would be, we believe these could be used to improve the computational research on uORFs as well as to advance the current knowledge in biology.
Part of this work was done when the first author (at that time at the School of Computing, the Robert Gordon University, UK) was on a three-month research visit at Chalmers University of Technology, funded by a Marie Curie Intra-European Fellowship (MEST-2-CT-2004-514169). We thank Marija Cvijovic, Alexandra Jauhiainen and Professor Olle Nerman for useful discussions.
- Mata J, Marguerat S, Bähler J: Post-transcriptional control of gene expression: a genome-wide perspective. Trends Biochem Sci 2005, 30(9):506–514. 10.1016/j.tibs.2005.07.005View ArticlePubMedGoogle Scholar
- Pesole G, Mignone F, Gissi C, Grillo G, Licciulli F, Liuni S: Structural and functional features of eukaryotic mRNA untranslated regions. Gene 2001, 276: 78–81. 10.1016/S0378-1119(01)00674-6View ArticleGoogle Scholar
- Kwon HS, Lee DK, Lee JJ, Edenberg HJ, ho Ahn Y, Hur MW: Posttranscriptional Regulation of Human ADH5 /FDH and Myf6 Gene Expression by Upstream AUG Codons. Arch Biochem Biophys 2001, 386(2):163–171. 10.1006/abbi.2000.2205View ArticlePubMedGoogle Scholar
- Vilela C, McCarthy JEG: Regulation of fungal gene expression via short open reading frames in the mRNA 5' untranslated region. Mol Microbiol 2003, 49(4):859–867. 10.1046/j.1365-2958.2003.03622.xView ArticlePubMedGoogle Scholar
- Vilela C, Ramirez CV, Linz B, Rodrigues-Pousada C, McCarthy JEG: Post-termination ribosome interactions with the 5' UTR modulate yeast mRNA stability. EMBO J 1999, 18(11):3139–3152. 10.1093/emboj/18.11.3139PubMed CentralView ArticlePubMedGoogle Scholar
- Hinnebusch AG: Translational Regulation of Yeast GCN4. A Window on Factors that Control Initiator-tRNA Binding to the Ribosome. J Biol Chem 1997, 272(35):21661–21664. 10.1074/jbc.272.35.21661View ArticlePubMedGoogle Scholar
- Fiaschi T, Marzocchini R, Raugei G, Veggi D, Chiarugi P, Ramponi G: The 5'-untranslated region of the human muscle acylphosphatase mRNA has an inhibitory effect on protein expression. FEBS Lett 1997, 417: 130–134. 10.1016/S0014-5793(97)01270-2View ArticlePubMedGoogle Scholar
- Iacono M, Mignone F, Pesole G: uAUG and uORFs in human and rodent 5' untranslated mRNAs. Gene 2005, 349: 97–105. 10.1016/j.gene.2004.11.041View ArticlePubMedGoogle Scholar
- Morris DR, Geballe AP: Upstream Open Reading Frames as Regulators of mRNA Translation. Mol Cell Biol 2000, 20(23):8635–8642. 10.1128/MCB.20.23.8635-8642.2000PubMed CentralView ArticlePubMedGoogle Scholar
- Krummeck G, Gottenöf T, Rödel G: AUG codons in the RNA leader sequences of the yeast PET genes CBS1 and SCO1 have no influence on translation efficiency. Curr Genet 1991, 20: 465–469. 10.1007/BF00334773View ArticlePubMedGoogle Scholar
- Kozak M: An Analysis of Vertebrate mRNA Sequences: Intimations of Translational Control. J Cell Biol 1991, 115(4):887–903. 10.1083/jcb.115.4.887View ArticlePubMedGoogle Scholar
- Willis AE: Translational control of growth factor and proto-oncogene expression. Int J Biochem Cell Biol 1999, 31: 73–86. 10.1016/S1357-2725(98)00133-2View ArticlePubMedGoogle Scholar
- Cancer Research UK[http://www.cancerhelp.org.uk/help/default.asp?page=6412]
- Kozak M: Initiation of translation in prokaryotes and eukaryotes. Gene 1999, 234: 187–208. 10.1016/S0378-1119(99)00210-3View ArticlePubMedGoogle Scholar
- Wiestner A, Schlemper RJ, Maas APC, Skoda RC: An activating splice donor mutation in the thrombopoietin gene causes hereditary thrombocythaemia. Nat Genet 1998, 18: 49–52. 10.1038/ng0198-49View ArticlePubMedGoogle Scholar
- Cvijovic M, Dalevi D, Bilsland E, Kemp GJL, Sunnerhagen P: Identification of putative regulatory upstream ORFs in the yeast genome using heuristics and evolutionary conservation. BMC Bioinformatics 2007, 8: 295. 10.1186/1471-2105-8-295PubMed CentralView ArticlePubMedGoogle Scholar
- Selpi , Bryant CH, Kemp GJL, Cvijovic M: A First Step towards Learning which uORFs Regulate Gene Expression. Journal of Integrative Bioinformatics 2006, 3(2):31. [http://journal.imbio.de/index.php?paper_id=31]Google Scholar
- Miura F, Kawaguchi N, Sese J, Toyoda A, Hattori M, Morishita S: A large-scale full-length cDNA analysis to explore the budding yeast transcriptome. Proc Natl Acad Sci USA 2006, 103(47):17846–17851. 10.1073/pnas.0605645103PubMed CentralView ArticlePubMedGoogle Scholar
- Srinivasan A: The Aleph Manual.1999. [http://web.comlab.ox.ac.uk/oucl/research/areas/machlearn/Aleph/]Google Scholar
- Muggleton S, Raedt LD: Inductive Logic Programming: Theory and Methods. J Log Program 1994, 19/20: 629–679. 10.1016/0743-1066(94)90035-3View ArticleGoogle Scholar
- Lavrač N, Džeroski S: Inductive Logic Programming: Techniques and Applications. Ellis Horwood 1994.Google Scholar
- Srinivasan A: Four Suggestions and a Rule Concerning the Application of ILP. In Džeroski S and Lavrač N, Eds: Relational Data Mining, p365–374. Springer-Verlag; 2001.Google Scholar
- Bratko I, Muggleton SH: Applications of Inductive Logic Programming. Commun ACM 1995, 38(11):65–70. 10.1145/219717.219771View ArticleGoogle Scholar
- Muggleton S: Scientific Knowledge Discovery Using Inductive Logic Programming. Commun ACM 1999, 42(11):42–46. 10.1145/319382.319390View ArticleGoogle Scholar
- Džeroski S: Relational Data Mining Applications: An Overview. In Džeroski S and Lavrač N, Eds: Relational Data Mining, p339–364. Springer-Verlag; 2001.View ArticleGoogle Scholar
- David L, Huber W, Granovskala M, Toedling J, Palm CJ, Bofkin L, Jones T, Davis RW, Steinmetz LM: A high-resolution map of transcription in the yeast genome. Proc Natl Acad Sci USA 2006, 103(14):5320–5325. 10.1073/pnas.0601091103PubMed CentralView ArticlePubMedGoogle Scholar
- ENSEMBL Database at BioMart[http://www.biomart.org/biomart/martview/]
- Rice P, Longden I, Bleasby A: EMBOSS: The European Molecular Biology Open Software Suite. Trends Genet 2000, 16(6):276–277. 10.1016/S0168-9525(00)02024-2View ArticlePubMedGoogle Scholar
- Cvijovic M: Comparative Genomic Study of upstream Open Reading Frames. Online Masters Thesis Chalmers University of Technology; 2005. Accessed 3 May 2006 [http://www.math.chalmers.se/Stat/Bioinfo/Master/Theses/2005/2.pdf] Accessed 3 May 2006Google Scholar
- Kozak M: Possible role of flanking nucleotides in recognition of the AUG initiator codon by eukaryotic ribosomes. Nucleic Acids Res 1981, 9(20):5233–5252. 10.1093/nar/9.20.5233PubMed CentralView ArticlePubMedGoogle Scholar
- Kozak M: Regulation of translation via mRNA structure in prokaryotes and eukaryotes. Gene 2005, 361: 13–37. 10.1016/j.gene.2005.06.037View ArticlePubMedGoogle Scholar
- Zhang Z, Dietrich FS: Identification and characterization of upstream open reading frames (uORF) in the 5' untranslated regions (UTR) of genes in Saccharomyces cerevisiae. Curr Genet 2005, 48: 77–87. 10.1007/s00294-005-0001-xView ArticlePubMedGoogle Scholar
- Cliften P, Sudarsanam P, Desikan A, Fulton L, Fulton B, Majors J, Waterson R, Cohen BA, Johnston M: Finding Functional Features in Saccharomyces Genomes by Phylogenetic footprinting. Science 2003, 301: 71–76. 10.1126/science.1084337View ArticlePubMedGoogle Scholar
- Kellis M, Patterson N, Endrizzi M, Birren B, Lander ES: Sequencing and comparison of yeast species to identify genes and regulatory elements. Nature 2003, 423: 241–254. 10.1038/nature01644View ArticlePubMedGoogle Scholar
- Upstream sequences of Sacchamomyces paradoxus[ftp://genome-ftp.stanford.edu/pub/yeast/sequence/fungal_genomes/S_paradoxus/MIT/orf_dna/utr5_500.fasta.gz]
- Upstream sequences of Saccharomyces mikatae[ftp://genome-ftp.stanford.edu/pub/yeast/sequence/fungal_genomes/S_mikatae/MIT/orf_dna/utr5_500.fasta.gz]
- Upstream sequences of Saccharomyces bayanus[ftp://genome-ftp.stanford.edu/pub/yeast/sequence/fungal_genomes/S_bayanus/MIT/orf_dna/utr5_500.fasta.gz]
- Hong EL, Balakrishnan R, Christie KR, Costanzo MC, Dwight SS, Engel SR, Fisk DG, Hirschman JE, Livstone MS, Nash R, Oughtred R, Park J, Skrzypek M, Starr B, Andrada R, Binkley G, Dong Q, Hitz BC, Miyasato S, Schroeder M, Weng S, Wong ED, Zhu KK, Dolinski K, Botstein D, Cherry JM: Saccharomyces Genome Database.2007. [ftp://ftp.yeastgenome.org/yeast/] Accessed 2-3 May 2007Google Scholar
- Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, Davis AP, Dolinski K, Dwight SS, Eppig JT, Harris MA, Hill DP, Issel-Tarver L, Kasarskis A, Lewis S, Matese JC, Richardson JE, Ringwald M, Rubin GM, Sherlock G: Gene Ontology: tool for the unification of biology. Nat Genet 2000, 25: 25–29. 10.1038/75556PubMed CentralView ArticlePubMedGoogle Scholar
- GO website[http://www.geneontology.org/GO.downloads.ontology.shtml]
- GO annotations for yeast genes[ftp://genome-ftp.stanford.edu/pub/yeast/data_download/literature_curation/gene_association.sgd.gz]
- Dwight SS, Harris MA, Dolinski K, Ball CA, Binkley G, Christie KR, Fisk DG, Issel-Tarver L, Schroeder M, Sherlock G, Sethuraman A, Weng S, Botstein D, Cherry JM: Saccharomyces Genome Database (SGD) provides secondary gene annotation using the Gene Ontology (GO). Nucleic Acids Res 2002, 30: 69–72. 10.1093/nar/30.1.69PubMed CentralView ArticlePubMedGoogle Scholar
- GO Slim Mapper[http://www.yeastgenome.org/help/goslimhelp.html]
- Preiss T, Baron-Benhamou J, Ansorge W, Hentze MW: Homodirectional changes in transcriptome composition and mRNA translation induced by rapamycin and heat shock. Nat Struct Biol 2003, 10(12):1039–1047. 10.1038/nsb1015View ArticlePubMedGoogle Scholar
- Shenton D, Smirnova JB, Selley JN, Carroll K, Hubbard SJ, Pavitt GD, Ashe MP, Grant CM: Global Translational Responses to Oxidative Stress Impact upon Multiple Levels of Protein Synthesis. J Biol Chem 2006, 281(39):29011–29021. 10.1074/jbc.M601545200View ArticlePubMedGoogle Scholar
- Smirnova JB, Selley JN, Sanchez-Cabo F, Carroll K, Eddy AA, McCarthy JEG, Hubbard SJ, Pavitt GD, Grant CM, Ashe MP: Global Gene Expression Profiling Reveals Widespread yet Distinctive Translational Responses to Different Eukaryotic Translation Initiation Factor 2B-Targeting Stress Pathways. Mol Cell Biol 2005, 25(21):9340–9349. 10.1128/MCB.25.21.9340-9349.2005PubMed CentralView ArticlePubMedGoogle Scholar
- Muggleton S: Learning from Positive Data. In Inductive Logic Programming Workshop, of Lect Notes Comput Sci. Volume 1314. Edited by: Muggleton S. Springer; 1996:358–376.Google Scholar
- Tran TN, Satou K, Ho TB: Using Inductive Logic Programming for Predicting Protein-Protein Interactions from Multiple Genomic Data. In PKDD, of Lect Notes Comput Sci. Volume 3721. Edited by: Jorge A, Torgo L, Brazdil P, Camacho R, Gama J. Springer; 2005:321–330.Google Scholar
- Badea L: Functional Discrimination of Gene Expression Patterns in Terms of the Gene Ontology. Pacific Symposium on Biocomputing 2003, 565–576.Google Scholar
- Clare A, King RD: Predicting gene function in Saccharomyces cerevisiae. Bioinformatics 2003, 19(Suppl 2):ii42-ii49.View ArticlePubMedGoogle Scholar
- Trajkovski I, Železný F, Lavrač N, Tolar J: Learning Relational Descriptions of Differentially Expressed Gene Groups. IEEE Trans Syst Man Cybern C Appl Rev 2008, 38: 16–25. 10.1109/TSMCC.2007.906059View ArticleGoogle Scholar
- Neafsey DE, Galagan JE: Dual Modes of Natural Selection on Upstream Open Reading Frames. Mol Biol Evol 2007, 24(8):1744–1751. 10.1093/molbev/msm093View ArticlePubMedGoogle Scholar
- Crowe ML, Wang XQ, Rothnagel JA: Evidence for conservation and selection of upstream open reading frames suggests probable encoding of bioactive peptides. BMC Genomics 2006., 7(16):Google Scholar
- Baim SB, Sherman F: mRNA Structures Influencing Translation in the Yeast Saccharomyces cerevisiae. Molecular and Cellular Biology 1988, 8(4):1591–1601.PubMed CentralView ArticlePubMedGoogle Scholar
- Džeroski S, Lavrač N, Eds: Relational Data Mining. Springer-Verlag; 2001.Google Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.