Gene Structure Prediction Research Articles

Analysis of the cDNAs of hypothetical genes on Arabidopsis chromosome 2 reveals numerous transcript variants.

In the fully sequenced Arabidopsis (Arabidopsis thaliana) genome, many gene models are annotated as "hypothetical protein," whose gene structures are predicted solely by computer algorithms with no support from either expressed sequence matches from Arabidopsis, or nucleic acid or protein homologs from other species. In order to confirm their existence and predicted gene structures, a high-throughput method of rapid amplification of cDNA ends (RACE) was used to obtain their cDNA sequences from 11 cDNA populations. Primers from all of the 797 hypothetical genes on chromosome 2 were designed, and, through 5' and 3' RACE, clones from 506 genes were sequenced and cDNA sequences from 399 target genes were recovered. The cDNA sequences were obtained by assembling their 5' and 3' RACE polymerase chain reaction products. These sequences revealed that (1) the structures of 151 hypothetical genes were different from their predictions; (2) 116 hypothetical genes had alternatively spliced transcripts and 187 genes displayed polyadenylation sites; and (3) there were transcripts arising from both strands, from the strand opposite to that of the prediction and possible dicistronic transcripts. Promoters from five randomly chosen hypothetical genes (At2g02540, At2g31270, At2g33640, At2g35550, and At2g36340) were cloned into report constructs, and their expressions are tissue or development stage specific. Our results indicate at least 50% of hypothetical genes on chromosome 2 are expressed in the cDNA populations with about 38% of the gene structures differing from their predictions. Thus, by using this targeted approach, high-throughput RACE, we revealed numerous transcripts including many uncharacterized variants from these hypothetical genes.

JIGSAW: integration of multiple sources of evidence for gene prediction

Computational gene finding systems play an important role in finding new human genes, although no systems are yet accurate enough to predict all or even most protein-coding regions perfectly. Ab initio programs can be augmented by evidence such as expression data or protein sequence homology, which improves their performance. The amount of such evidence continues to grow, but computational methods continue to have difficulty predicting genes when the evidence is conflicting or incomplete. Genome annotation pipelines collect a variety of types of evidence about gene structure and synthesize the results, which can then be refined further through manual, expert curation of gene models. JIGSAW is a new gene finding system designed to automate the process of predicting gene structure from multiple sources of evidence, with results that often match the performance of human curators. JIGSAW computes the relative weight of different lines of evidence using statistics generated from a training set, and then combines the evidence using dynamic programming. Our results show that JIGSAW's performance is superior to ab initio gene finding methods and to other pipelines such as Ensembl. Even without evidence from alignment to known genes, JIGSAW can substantially improve gene prediction accuracy as compared with existing methods. JIGSAW is available as an open source software package at http://cbcb.umd.edu/software/jigsaw.

An approach towards experimental cDNA sequence determination of predicted genes: An example from Arabidopsis U3-55k homologues

Experimental cDNA sequence determinations lag behind in silico gene structure predictions in some recently sequenced genomes. This may be due in part to low transcript abundance and/or the severely spatio-temporarily restricted expression pattern of some genes. Here we characterize the predicted repressed gene of Arabidopsis thaliana (At4g21130) that encodes a homologue of the Arabidopsis U3-55K-like protein (At4g05410) and of the U3-55K (RNU3IP2, Rrp9p) proteins from other eukaryotes. In man and yeast, U3-55K is involved in the processing of the pre-ribosomal RNA. Here we show that treatment with inhibitors of histone deacetylases (trichostatin A, sodium butyrate) or DNA methyltransferases (5-aza-2′-deoxycytidine) induces a low but distinct level of mRNA from the repressed Arabidopsis At4g21130 locus, which can be detected by RT-PCR amplification. Direct sequencing of PCR products reveals the open reading frame that differs, in part, from the hypothetical one and encodes a seven-WD-repeat protein highly conserved when compared to U3-55K proteins from various eukaryotic species. This suggests the conservation of its function. The described approach may help to determine the nucleotide sequences of transcripts from predicted genes with a low level of expression.

AUGUSTUS: a web server for gene prediction in eukaryotes that allows user-defined constraints

We present a WWW server for AUGUSTUS, a software for gene prediction in eukaryotic genomic sequences that is based on a generalized hidden Markov model, a probabilistic model of a sequence and its gene structure. The web server allows the user to impose constraints on the predicted gene structure. A constraint can specify the position of a splice site, a translation initiation site or a stop codon. Furthermore, it is possible to specify the position of known exons and intervals that are known to be exonic or intronic sequence. The number of constraints is arbitrary and constraints can be combined in order to pin down larger parts of the predicted gene structure. The result then is the most likely gene structure that complies with all given user constraints, if such a gene structure exists. The specification of constraints is useful when part of the gene structure is known, e.g. by expressed sequence tag or protein sequence alignments, or if the user wants to change the default prediction. The web interface and the downloadable stand-alone program are available free of charge at .

Annotating the Human Proteome

The completion of the human genome has shifted the attention from deciphering the sequence to the identification and characterization of the encoded components. The identification and functional annotation of the proteome is here of special interest and starts with the identification of genes and transcripts as a prerequisite of proteome annotation. Gene predictions are very powerful in predicting most of the exons in a genome, but reliable gene structure predictions of both known and novel genes are dependent on existing transcript and protein information. An enormous amount of data already exists on the function of many human proteins, but this is scattered over many resources. Public domain databases are required to manage and collate this information and present it to the user community in both a human and machine readable manner.

Experiments in Searching Small Proteins in Unannotated Large Eukaryotic Genomes

There is growing interest to use mass spectrometry data to search genome sequences directly. Previous work by other authors demonstrated that this approach is able to correct and complement available genome annotations. We discuss the practical difficulty of searching large eukaryotic genomes with peptide ion trap tandem mass spectra of small proteins (<40 kDa). The challenging problem of automatically identifying peptides that span across exon/intron boundaries is explored for the first time by using experimental data. In a human genome search, we find that roughly 30% of the peptides are missed, due to various reasons, compared to a Swiss-Prot search. We show that this percentage is significantly reduced with improved parent mass accuracy. We finally provide several examples of predicted gene structures that could be improved by proteomics data, in particular by peptides spanning across exon/intron boundaries.

Tiling microarray analysis of rice chromosome 10 to identify the transcriptome and relate its expression to chromosomal architecture

A transcriptome analysis of chromosome 10 of 2 rice subspecies identifies 549 new gene models and gives experimental evidence for around 75% of the previously unsupported predicted genes.

SpliceMachine: predicting splice sites from high-dimensional local context representations

In this age of complete genome sequencing, finding the location and structure of genes is crucial for further molecular research. The accurate prediction of intron boundaries largely facilitates the correct prediction of gene structure in nuclear genomes. Many tools for localizing these boundaries on DNA sequences have been developed and are available to researchers through the internet. Nevertheless, these tools still make many false positive predictions. This manuscript presents a novel publicly available splice site prediction tool named SpliceMachine that (i) shows state-of-the-art prediction performance on Arabidopsis thaliana and human sequences, (ii) performs a computationally fast annotation and (iii) can be trained by the user on its own data. Results, figures and software are available at http://www.bioinformatics.psb.ugent.be/supplementary_data/ sven.degroeve@psb.ugent.be; yves.vandepeer@psb.ugent.be.

Gene prediction and verification in a compact genome with numerous small introns.

The genomes of clusters of related eukaryotes are now being sequenced at an increasing rate, creating a need for accurate, low-cost annotation of exon-intron structures. In this paper, we demonstrate that reverse transcription-polymerase chain reaction (RT-PCR) and direct sequencing based on predicted gene structures satisfy this need, at least for single-celled eukaryotes. The TWINSCAN gene prediction algorithm was adapted for the fungal pathogen Cryptococcus neoformans by using a precise model of intron lengths in combination with ungapped alignments between the genome sequences of the two closely related Cryptococcus varieties. This approach resulted in approximately 60% of known genes being predicted exactly right at every coding base and splice site. When previously unannotated TWINSCAN predictions were tested by RT-PCR and direct sequencing, 75% of targets spanning two predicted introns were amplified and produced high-quality sequence. When targets spanning the complete predicted open reading frame were tested, 72% of them amplified and produced high-quality sequence. We conclude that sequencing a small number of expressed sequence tags (ESTs) to provide training data, running TWINSCAN on an entire genome, and then performing RT-PCR and direct sequencing on all of its predictions would be a cost-effective method for obtaining an experimentally verified genome annotation.

GeneScout: a data mining system for predicting vertebrate genes in genomic DNA sequences

Automated detection or prediction of coding sequences from within genomic DNA has been a major rate-limiting step in the pursuit of vertebrate genes. Programs currently available are far from being powerful enough to elucidate a gene structure completely. In this paper, we present a new system, called GeneScout, for predicting gene structures in vertebrate genomic DNA. The system contains specially designed hidden Markov models (HMMs) for detecting functional sites including protein-translation start sites, mRNA splicing junction donor and acceptor sites, etc. An HMM model is also proposed for exon coding potential computation. Our main hypothesis is that, given a vertebrate genomic DNA sequence S, it is always possible to construct a directed acyclic graph G such that the path for the actual coding region of S is in the set of all paths on G. Thus, the gene detection problem is reduced to that of analyzing the paths in the graph G. A dynamic programming algorithm is used to find the optimal path in G. The proposed system is trained using an expectation-maximization algorithm and its performance on vertebrate gene prediction is evaluated using the 10-way cross-validation method. Experimental results show that the proposed system performs well and is comparable to existing gene discovery tools.

Recent advances in gene structure prediction

Recent advances in gene structure prediction

Prediction of eukaryotic gene structures based on multilevel optimization

Computational gene structure prediction, which is valuable for finding new genes and understanding the composition of genomes, plays a very important role in various kinds of genome projects. For eukaryotic gene structures, however, the prediction accuracy of existing methods is still limited. This paper presents a method of predicting eukaryotic gene structures based on multilevel optimization. The complicated problem of predicting gene structure in eukaryotic DNA sequence containing multiple genes can be decomposed into a series of sub-problems at several levels with decreasing complexity, including the gene level (single-exon gene, multi-exon gene), the element level (exon, intron, etc.), and the feature level (functional site signals, codon usage preference, etc.). On the basis of this decomposition, a multilevel model for the prediction of complex gene structures is created by a multilevel optimization process, in which the models dealing with sub-problems at low complexity level are first optimized respectively, and then optimally combined together to form models for those sub-problems at higher complexity level. Based on the multi-level model, a dynamic programming algorithm is designed to search for optimal gene structures from DNA sequences, and a new program GeneKey (1.0) for the prediction of eukaryotic gene structures is developed. Testing results with widely used datasets demonstrate that the prediction accuracies of GeneKey (1.0) at the nucleotide level, exon level and gene level are all higher than that of the well known program GENSCAN. A web server of GeneKey(1.0) is available at http://infosci.hust.edu.cn

Gene Structure Prediction Using an Orthologous Gene of Known Exon-Intron Structure

Given the availability of complete genome sequences from related organisms, sequence conservation can provide important clues for predicting gene structure. In particular, one should be able to leverage information about known genes in one species to help determine the structures of related genes in another. Such an approach is appealing in that high-quality gene prediction can be achieved for newly sequenced species, such as mouse and puffer fish, using the extensive knowledge that has been accumulated about human genes. This article reports a novel approach to predicting the exon-intron structures of mouse genes by incorporating constraints from orthologous human genes using techniques that have previously been exploited in speech and natural language processing applications. The approach uses a context-free grammar to parse a training corpus of annotated human genes. A statistical training procedure produces a weighted recursive transition network (RTN) intended to capture the general features of a mammalian gene. This RTN is expanded into a finite state transducer (FST) and composed with an FST capturing the specific features of the human orthologue. This model includes a trigram language model on the amino acid sequence as well as exon length constraints. A final stage uses the free software package ClustalW to align the top n candidates in the search space. For a set of 98 orthologous human-mouse pairs, we achieved 96% sensitivity and 97% specificity at the exon level on the mouse genes, given only knowledge gleaned from the annotated human genome.

GeneScout: a data mining system for predicting vertebrate genes in genomic DNA sequences

Automated detection or prediction of coding sequences from within genomic DNA has been a major rate-limiting step in the pursuit of vertebrate genes. Programs currently available are far from being powerful enough to elucidate a gene structure completely. In this paper, we present a new system, called GeneScout, for predicting gene structures in vertebrate genomic DNA. The system contains specially designed hidden Markov models (HMMs) for detecting functional sites including protein-translation start sites, mRNA splicing junction donor and acceptor sites, etc. An HMM model is also proposed for exon coding potential computation. Our main hypothesis is that, given a vertebrate genomic DNA sequence S, it is always possible to construct a directed acyclic graph G such that the path for the actual coding region of S is in the set of all paths on G. Thus, the gene detection problem is reduced to that of analyzing the paths in the graph G. A dynamic programming algorithm is used to find the optimal path in G. The proposed system is trained using an expectation-maximization algorithm and its performance on vertebrate gene prediction is evaluated using the 10-way cross-validation method. Experimental results show that the proposed system performs well and is comparable to existing gene discovery tools.

Using TWINSCAN to Predict Gene Structures in Genomic DNA Sequences

AbstractTWINSCAN is a gene‐prediction system that combines the methods of ab initio predictors like GENSCAN with information derived from genome comparison. This unit describes the use of TWINSCAN to identify gene structures in eukaryotic genomic sequences. Protocols for using TWINSCAN through its Web interface and from the command line in a Linux environment are provided. Detailed discussion about the appropriate parameter settings, input sequence processing, and choice of genome for comparison are included.

GeneSeqer@PlantGDB: Gene structure prediction in plant genomes.

The GeneSeqer@PlantGDB Web server (http://www.plantgdb.org/cgi-bin/GeneSeqer.cgi) provides a gene structure prediction tool tailored for applications to plant genomic sequences. Predictions are based on spliced alignment with source-native ESTs and full-length cDNAs or non-native probes derived from putative homologous genes. The tool is illustrated with applications to refinement of current gene structure annotation and de novo annotation of draft genomic sequences. The service should facilitate expert annotation as a community effort by providing convenient access to all public plant sequences via the PlantGDB database, a simple four-step protocol for spliced alignment and visually appealing displays of the predicted gene structures in addition to detailed sequence alignments.

Cloning and sequencing of cDNAs for hypothetical genes from chromosome 2 of Arabidopsis.

About 25% of the genes in the fully sequenced and annotated Arabidopsis genome have structures that are predicted solely by computer algorithms with no support from either nucleic acid or protein homologs from other species or expressed sequence matches from Arabidopsis. These are referred to as "hypothetical genes." On chromosome 2, sequenced by The Institute for Genomic Research, there are approximately 800 hypothetical genes among a total of approximately 4,100 genes. To test their expression under various growth conditions and in specific tissues, we used six cDNA populations prepared from cold-treated, heat-treated, and pathogen (Xanthomonas campestris pv campestris)-infected plants, callus, roots, and young seedlings. To date, 169 hypothetical genes were tested, and 138 of them are found to be expressed in one or more of the six cDNA populations. By sequencing multiple clones from each 5'- and 3'-rapid amplification of cDNA ends (RACE) product and assembling the sequences, we generated full-length sequences for 16 of these genes. For 14 genes, there was one full-length assembly that precisely supported the intron-exon boundaries of their gene predictions, adding only 5'- and 3'-untranslated region sequences. However, for three of these genes, the other assemblies represent additional exons and alternatively spliced or unspliced introns. For the remaining two genes, the cDNA sequences reveal major differences with predicted gene structures. In addition, a total of six genes displayed more than one polyadenylation site. These data will be used to update gene models in The Institute for Genomic Research annotation database ATH1.

Comparative Methods for Gene Structure Prediction in Homologous Sequences

The increasing number of sequenced genomes motivates the use of evolutionary patterns to detect genes. We present a series of comparative methods for gene finding in homologous prokaryotic or eukaryotic sequences. Based on a model of legal genes and a similarity measure between genes, we find the pair of legal genes of maximum similarity. We develop methods based on genes models and alignment based similarity measures of increasing complexity, which take into account many details of real gene structures, e.g. the similarity of the proteins encoded by the exons. When using a similarity measure based on an exiting alignment, the methods run in linear time. When integrating the alignment and prediction process which allows for more fine grained similarity measures, the methods run in quadratic time. We evaluate the methods in a series of experiments on synthetic and real sequence data, which show that all methods are competitive but that taking the similarity of the encoded proteins into account really boost the performance.

Gene structure identification with MyGV using cDNA evidence and protein homologs to improve ab initio predictions.

MyGV is an application to visualize (potentially genome-scale) gene structure annotation and prediction. The output of any external gene prediction program can be easily converted to a generalized format for input into MyGV. The application displays all input simultaneously in graphical representation, with a toggle option for a text-based view. Zooming capabilities allow detailed comparisons for specific genome locations. The tool is particularly helpful for refinement of ab initio predicted gene structures by spliced alignment with cDNA or protein homologs. The program was written in Java and is freely available to non-commercial users by electronic download from http://bioinformatics.iastate.edu/bioinformatics2go/MyGV.

Isolation of Drosophila Activin and Follistatin cDNAs using novel MACH amplification protocols

With the genomic sequence of multicellular organisms such as Caenorhabditiselegans, Drosophilamelanogaster, and Homosapiens completed and others to be finished in the near future, the focus has shifted from accumulating sequence information to the prediction and analysis of genes within the completed genomes. Unfortunately, presently available computer programs do not always accurately predict gene structure such as mRNA and translation start sites or intron/exon boundaries. The only way to be certain about a gene's structure is to isolate and characterize its cDNA. Since the screening of libraries is a time-consuming, labor-intensive process that sometimes fails to yield the desired clone, we searched for faster, more efficient ways to isolate cDNAs. In this study, we describe two methods for amplification and isolation of cDNAs from plasmid libraries that requires no hybridization (MACH). With the polymerase chain reaction-based MACH-2 protocol, we present a strategy that requires little DNA sequence information to selectively isolate the longest cDNA variant from plasmid libraries in about 3 days. Our protocols were used to isolate cDNAs for the Drosophila activin and follistatin genes.