Computational Gene Prediction

Practical exercise
Abstract: In this exercise, a previously annotated human (HS307871) will be used to assess the accuracy of different gene finding approaches will be used to annotate the sequence: ab-initio (GRAIL, GENSCAN, geneid, FGENESH) and homology-based (GenomeScan, GrailEXP and GENEWISE) Weak conservation of Start codons will lead to wrong prediction of initial exons in most cases.
There are two main methods for automatic gene prediction: Ab-initio methods and comparative methods:
  1. Ab-initio methods: they use the DNA sequence as only input. As they use no other information they are usually called "intrinsic". There are several elements in a gene that can be localized in the genomic sequence, which are used to identify genes computationally. These elements are related with the signals that regulate the biological mechanisms of gene expression, and with thesequence bias due to the encoding of a DNA that can be translated into protein. They are called signal sensors and content sensors, respectively.

    The signals are tipically the splice-sites (donor: GTRAGT, acceptor: NCAG, branch-site: CURAY), the start of translation (ATG), and the end of translation (TGA,TAA,TAG). The content sensor most commonly used is the bias in codon usage: coding regions use some codons more frequently than others. Both, signal sensors and content sensors must be trained, i.e., we must start from a set of observations (know genes) from which we build our model. Predicting a gene is therefore looking for new elements in the genomic sequence that look like our model. This can be established in terms of probabilities.

  2. Comparative methods: They are also called extrinsic. They include two strategies: those that use homologies with sequences from other genes, also called homology-based, and those that make comparisons with genomic sequence from other genomes, also called comparative.
    • Homology-based: these methods predict a gene using the alignment of a protein (or RNA sequence in the form of full-length mRNA, cDNA or EST) with the genome sequence that we want to annotate. The known sequence (also called evidence) guides the prediction. There are several ways to achieve this: the simplest way consist in accepting the alignment of the known sequence to the genome as the gene prediction. More sophisticated methods use the known sequence as a guide and try to complete the evidence into a complete gene structure. The efficacy of this method depends on the number of known gene sequences, hence it is limited by how complete the biolodical databases are.
    • Comparative-genomics-based: These methods are based in the hypothesis that the sequences conserved between to genomes relatively close to each other are functional, and therefore, possibly coding for a gene.

In fact, the annotation of a genome involves a combination of several methods of gene prediction, and perhaps the prediction of other biological signals (like transcription start sites, promoter regions, etc). In this practical we will try different methods of the type ab-initio and homology-based.

A. Gene annotation: using a genome annotation browser
Step 1. Identification of a known gene
  • Open the sequence for this genomic sequence in a different window: FASTA SEQUENCE.

  • Open the UCSC Genome Browser

  • Select the BLAT (human) link to locate the genomic coordinates of our anonymous sequence

  • Paste the DNA sequence

  • Submit the file

  • Click over the first hit: (browser link)

  • In this picture, can you find your sequence? Which gene is this one?

  • Try to understand the following options (use Refresh):

    1. RefSeq Genes
    2. Vega Genes
    3. SNPs
    4. Conservation

  • Use the Zoom in and out (specially to find the neighbour genes)

Figure 7. (a) UCSC genome browser representation of the region containing the gene uroporphyrinogen decarboxylase (URO-D) (b) UCSC genome browser representation of the contex (100Kbps) region around the gene uroporphyrinogen decarboxylase (URO-D).

Step 2. The coding annotation of the URO-D gene:

The FASTA SEQUENCE used corresponds to (a version of) the genomic region containing the URO-D gene. Defects in UROD are the cause of porphyrya, a group of rare, inherited blood disorders in which cells fail to change chemicals (porphyrins) to the substance (heme) that gives blood its color.


B. Exploring ab initio gene prediction
Step 3. Running geneid
  • Connect to the geneid server

  • Paste the FASTA sequence of the genomic region of the URO-D gene

  • Choose geneid output format

  • Run geneid with different parameters:

    1. Searching signals: Select acceptors, donors, start and stop codons. Look for them in the real annotation of the sequence

    2. Searching exons: Select All exons and try to find the real ones

    3. Finding genes: You do not need to select any option (default behaviour).

      Compare the predicted gene with the real gene, with CDS coordinates: 

       1107..1126
 1748..1860
 1976..2055
 2132..2194
 2434..2631
 2749..2910
 3279..3416
 3576..3676
 3780..3846
 4179..4340

    Figure 1. Signal, exons and genes predicted by geneid in the sequence HS307871

Step 4. Running other genefinders

Provided that there are several alternative programs to analyze a DNA sequence, we can run every application and observe the common parts of the predictions.

  1. GENSCAN:
    • Connect to the GENSCAN server

    • Paste DNA sequence

    • Press Run Genscan button

    • Compare annotations and predictions

  2. FGENESH:
    • Connect to Softberry homepage

    • On the left frame, select GENE FINDING in Eukaryota

    • Select the program FGENESH

    • Paste DNA sequence

    • Press Search button

    • Compare annotations and predictions

  3. GRAIL:
    • Connect to GrailEXP homepage

    • Activate Perceval Exon Candidates box

    • Paste DNA sequence

    • Press Go! button

    • Check the results

    • Compare annotations and predicted exons

  4. NOTE: First exon is always missed in the predictions and there are some problems to detect the donor site from exon 5. Detection of Start codons is a serious drawback in current gene finding programs (see Figure 2). However, this problem can be overcome by using homology information to complete the gene prediction.
Figure 2. EMBL annotation and genes predicted by Grail, GENSCAN, geneid and FGENESH in the sequence HS307871


C. Using EST/cDNA homology information
Step 5. Using GrailEXP
  • Connect to GrailExp homepage

  • Activate Galahad EST/mRNA/cDNA Alignments box

  • Select GrailEXP database (RefSeq/HTDB/dbEST/EGAD/Riken)

  • Activate exon assembly: Gawain Gene Models

  • Paste DNA sequence

  • Press Go! button

  • Check the results: predictions and supporting information

  • Compare annotations, ab initio GRAIL prediction and five predicted alternative spliced variants
Figure 3. Comparison between EMBL annotation and genes predicted ab inition by Grail Vs five alternative predictions supported by ESTs information in the sequence HS307871


Step 6. Using other gene finding programs + alignment of transcripts

Using blastn, we can search the database est_human for ESTs supporting future predictions. Filter this output in order to select those non-overlapping ESTs that could form a complete cDNA sequence (see Figure 4). Moreover, ESTs not divided into two or more pieces in the genomic sequence (containing a couple of splice sites) should be rejected.
  • Connect to the FGENESH-C server (on Gene finding with similarity menu)

  • Paste the sequence HS307871

  • Paste the cDNA sequence or EST you have selected

  • Press the search button

  • Notice that the predicted gene will be necessarily supported by homology information, so the prediction will most likely only cover in the genomic region overlapping your EST query.

Figure 4. Best human ESTs in the alignment mapped on the genomic sequence HS307871

D. Using protein homology information
Step 7. Spliced alignment

A tool capable to produce Spliced alignments is very useful in the homology-based prediction. For instance, if we have a homologous protein sequence. In this case, gene prediction is carried by producing an alignment of the protein sequence to the genome, and at the same time, fitting the alignment to the best splice sites predicted in the genomic sequence.
  • Open the NCBI blast server

  • Choose blastx program (genomic query versus protein database)

  • Paste the genomic sequence and press the Blast! and Format!

  • Select the first protein. Display the FASTA sequence or click here. Obviously, it is the known protein URO-D, which is already annotated in the genomic sequence (see UCSC browser).

  • Open the GeneWise web server to use this protein to predict the best gene structure

  • Paste both protein and genomic sequences and run the program

  • Compare the predicted gene and the EMBL annotations (above). Look at the splice-sites given by GeneWise to check whether the exon boundaries are correct.

    Figure 5. Best HSPs representing proteins homologues similar to the genomic sequence HS307871 obtained using blastx


Step 8. Spliced alignment using homologous proteins

From the blastx output, choose several homologous genes from other species (different from human) and run genewise for each one separately, again. Observe the loss of accuracy as the homologue is further from the original human protein:
Step 9. Using protein homology information: GenomeScan

Protein homology information can also be used to enhance ab initio predicted exons supported by blastx HSPs as in the case of GenomeScan and geneid improving therefore the final prediction GenomeScan:
  • Connect to the GenomeScan web server

  • Retrieve the protein from the previous blast search

  • Paste both genomic and protein sequences

  • Press the button GenomeScan

  • Check the results. It seems that the first exon has not been detected even using homology information. This is due to the fact that blast programs have a minimal word lenght.

Figure 6. GenomeScan output: first exon is not correctly predicted probably due to blast length restrictions


E. Results

Here you can find the solutions to every exercise:

EMBL annotation
EMBL annotation (plain text)
FASTA sequence
geneid results: signals
geneid results: exons
geneid results: genes
GENSCAN results
FGENESH results
GRAIL results
GrailEXP results
Blastn + human ESTs results
Blastx + protein results
Genewise (human protein)
Genewise (ovis protein)
Genewise (mouse protein)
Genewise (rat protein)
Genewise (Danio rerio protein)
Genewise (Drosophila melanogaster protein)
Genewise (Drosophila virilis protein)
Genewise (yeast protein)
Genewise (fission yeast protein)
GenomeScan results


F. Bibliography
  1. J.F. Abril and R. Guigó. gff2ps: visualizing genomic annotations. Bioinformatics 16:743-744 (2000).

  2. Altschul, S.F., Gish, W., Miller, W., Myers, E.W. & Lipman, D.J. Basic local alignment search tool. J. Mol. Biol. 215:403-410 (1990).

  3. Burge, C. and Karlin, S. Prediction of complete gene structures in human genomic DNA. J. Mol. Biol. 268, 78-94 (1997).

  4. E. Blanco, G. Parra and R. Guigó. Using geneid to Identify Genes. In A. D. Baxevanis and D. B. Davison, chief editors: Current Protocols in Bioinformatics. Volume 1, Unit 4.3. John Wiley & Sons Inc., New York. ISBN: 0-471-25093-7 (2002).

  5. G. Parra, E. Blanco, and R. Guigó. Geneid in Drosophila. Genome Research 10:511-515 (2000).

  6. Asaf A. Salamov and Victor V. Solovyev. Ab initio Gene Finding in Drosophila Genomic DNA Genome Res. 10: 516-522 (2000).

  7. Yeh, R.-F., Lim, L. P. and Burge, C. B. Computational inference of homologous gene structures in the human genome. Genome Res. 11: 803-816 (2001).

  8. D. Hyatt, J. Snoddy, D. Schmoyer, G. Chen, K. Fischer, M. Parang, I. Vokler, S. Petrov, P. Locascio, V. Olman, Miriam Land, M. Shah, and E. Uberbacher. Improved Analysis and Annotation Tools for Whole-Genome Computational Annotation and Analysis: GRAIL-EXP Genome Analysis Toolkit and Related Analysis Tools. Genome Sequencing & Biology Meeting (2000).

  9. Ewan Birney and Richard Durbin. Using GeneWise in the Drosophila Annotation Experiment. Genome Res. 10: 547-548 (2000).


created by Enrique Blanco - eblanco@imim.es
modified by E. Eyras