What's SEVENS


SEVENS is a database including 7-TMR (7 transmembrane helix receptor) candidates predicted from whole human genome sequences, using sequence search, motif and domain assignment, transmembrane helix prediction and the gene quality refinement. This system is intended to detect sequences of multi-exon or remote homologues that can not be detected by using conventional sequence search tools alone. With the careful assessment of the analyzing components, we obtained candidate datasets, with several predicting accuracy, among which we found at least 1,033 and at most 1,916 candidate 7-TMR genes from human genome.

Contents

  • Content Search
  • Retrieve 7-TMR candidate sequences by the "AND" combination of (1) Keyword in nr.aa database search results, (2)Chromosome number, (3)Data Level, (4)Predicted exon number, (5) Gene Length, (6)Protein length, (7)E-value of sequence search against SWISSPROT or nr.aa, (8) Prosite motifs, and (10) Pfam domains.

    After selection with some of contents, 7-TMR candidates will be appear at the chromosomal viewer and the gene lists which navigate to the detailed analysis for each gene.


  • News
  • Release Information and news concerning updates of analysis.


  • What's SEVENS
  • Introduction for SEVENS database and it's usage.


  • Statistics
  • Release information of data statistics.


    How we found 7-TMR sequences.

    7-TMR are detected using automated system which is specified for finding this family. The automated discovery was performed in three stages:

    1)Gene prediction  (i.e., translation of genomic sequences into amino acid sequences).
    2)Screening  of 7-TMR candidates by assessing genes with sequence search, motif- and domain assignment, and transmembrane helix (TMH) prediction.
    3)Quality improvement  of the candidate 7-TMR dataset by eliminating duplication and redundancy.

    (1) Gene prediction stage:

    Genomic sequences were obtained from human sequences ( Human Genome Resources of the NCBI). To maximize the number of gene candidates, we detected three kinds of sequence sets,

    (a)"6f-sequences" which were all possible combination between initial and stop codons in 6 reading frames with the rule of using the most upstream ATG possible.
    (b)"ALN-sequences" which were aligned with known protein sequences by ALN.
    (c)"GD-sequences" which were generated by GeneDecoder.


  • ALN
  • Using a new convention for encoding a DNA sequence into a series of 23 possible letters, a dynamic programming algorithm ('aln' written in ANSI-C) was developed to align a DNA sequence and a protein sequence or profile so that the spliced and translated sequence optimally matches the reference the same as the standard protein sequence alignment allowing for long gaps. The objective function also takes account of frame shift errors, coding potentials, and translation initiation, termination and splicing signals. This method was tested on Caenorhabditis elegans genes of known structures. The accuracy of prediction measured in terms of a correlation coefficient was about 95% at the nucleotide level for the 288 genes tested, and 97.0% for the 170 genes whose product and closest homologue share more than 30% identical amino acids. (Gotoh, O., Bioinformatics.2000 Mar;16(3):190-202.).


  • GeneDecoder
  • Gene-finding system based on the Hidden Markov Model "HMM" (Asai, K., Itou, K., Ueno, Y. & Yada, T., Pacific Symposium on Biocomputing 98, pp. 228-239 (PSB98, 1998)). This system allows multiple inputs: not only sequence information, but homology scores and other data may be integrated for prediction. The prediction accuracy was evaluated with "genesets98"(http://bioinformatics.weizmann.ac.Il/databases/genesets/Human/). The sensitivity was 83% and the specificity was 74% for the detection of gene position without using homology scores.


    (2) Screening stage:

    Each analysis tool was first assessed to determine two threshold settings, best specificity and best sensitivity, with a reference dataset: 1,242 7-TMR sequences and 73,493 non-7-TMR sequences in the SWISS-PROTdatabase. The best specificity threshold is intended to achieve, when applied to the reference dataset, almost 100% specificity and with minimum false-negatives. On the other hand, the best sensitivity threshold is intended to achieve almost 100% sensitivity and with minimum false-positives.
    Using the thresholds shown in Table 1, those 7-TMR candidates were selected that showed significant sequence similarity or contained characteristic motifs and domains, and transmembrane helices. Four confidence levels of the datasets were determined by combining the best specificity and best sensitivity thresholds. Level A data, expected to show the best specificity, were obtained by adding the candidate sequences given by best specificity thresholds of the sequence similarity search, motif- and domain assignments. To discover remote 7-TMR homologues, we combined candidates from the three-level thresholds for TMH prediction (see Table 1) with the sequences that were obtained by the best sensitivity thresholds of sequence search and motif- and domain assignment, and level D data are expected to show the best sensitivity.


    Table 1. Thresholds used for 7-TMR discovery.
      Level A
    (Best specificity)
    Level B Level C Level D
    (Best sensitivity)
    Sequence search
    with BLASTP
    E < 10-80 E < 10-30 E < 10-30 E < 10-30
    Domain assignment
    with PFAM
    E < 10-10 E < 1.0 E < 1.0 E < 1.0
    Motif assignment
    with PROSITE
    Not used Match Match Match
    TMH Prediction Not used TMwindows(7)
    AND
    Hirokawa(7)
    TMwindows(7)
    AND
    Hirokawa(6-8)
    {TMwindows(7) AND Hirokawa(6-8)}
    OR
    TMwindows(7)
    OR
    Hirokawa(7)
    Sensitivity 99.4% 99.8% 99.9% 99.9%
    Specificity 96.6% 70.0% 48.4% 20.0%

    Thresholds of the programs are shown

    Using BLASTP (Altschul, S. F., et al Nucleic Acids Res.25,3389-3402 (1997)) known 7-TMR seguences were searched against the reference dataset, and the sensitivity and specificity of E values were computed for discriminating correct pairs.
    Using HMMER (Bateman, A., Birney, E., Durbin, R., Eddy, S. R., Howe, K, L. & Sonnhammer, E. L. Nucleic Acids Res.28,263-266 (2000).), 7-TMR specific Hidden Markov Models ( PFAM domain ) were assigned to reference sequences, and the sensitivity and specificity of E values were computed for correct assignment.
    Since PROSITE patterns are written by regular expression, we determined the P value, which is calculated as the multiplication of each residue frequency in the SWISS-PROT database; the sensitivity and specificity of P values were computed for correct assignment.
    For TMH prediction we used the TMwindows program, our original program along with the method of Hirokawa, et al . We treated the results as 7-TMR outputs when the predicted helix number was dispersed between n and m. Here we used n-m ranges 7-7, 6 -8, 5-9, and 4-10 and combined the sequences obtained from each range of the two programs. For example, the descriptor {TMwindows(7) OR Hirokawa(6-8) } unifies ("OR"), the sequences within range 7-7 that were obtained by TMwindows and the sequences within range 6-8 that were obtained by Hirokawa`s method.


  • Hirokawa Method

  • A useful tool for secondary structure prediction of membrane proteins from a protein sequence. The basic idea of prediction in this system is based on the physicochemical properties of amino acid sequences such as hydrophobicity and charges. The system deals with three types of prediction: discrimination of membrane proteins from soluble ones, prediction of the existence of transmembrane helices and determination of transmembrane helical regions.
    (Hirokawa, T., Boon-Chieng, S. & Mitaku, S. Bioinformatics.14,378-379 (1998).)


  • TMwindows

  • Predicts transmembrane helices by the following procedures.
    (1) It assigns the Engelman-Steitz-Goldman (Annual Review of Biophysics and Biophysical Chemistry.15,321-353 (1986).) hydropathy index to amino acid sequences and calculates average hydrophobicity within a pre-determined window. The index was selected, after comparing all indices in the AAindex database (Protein Eng. 9, 27-36 (1996). as the most powerful for discriminating membrane proteins from others using total average hydrophobicity.
    (2) The window size is changed from 19 to 27 and if the average hydrophobicity within each window exceeds 2.5, the region is regarded as a transmembrane helix. The total number of helices computed for each window size gives the range of predicted helix number.


    (3) Quality improvement stage:

    Candidate sequences selected by the above process still contain the following redundancies. (1) Perfect matches or overlaps at the same genomic position (chromosome number, contig order number and relative position within the contig). They originate in two independent sequence predictions: the 6-frame translation and the prediction by GeneDecoder. We regarded them as the same gene and adjusted the double count accordingly. (2) Multiple sequence copies in different genomic positions. We regarded them as different genes. (3) Separate sequence fragments linked by a known protein sequence. They originate in an erroneous prediction by the gene finding programs. We merged them using the linker sequence.
    These redundancies were detected by the following clustering method for each level. First, Smith-Waterman sequence alignment was applied to the candidate sequences in an all-against-all fashion. Then sequences were linked together only when they hit for more than 50 amino acids with more than 95% identity, and shared the same chromosome number, contig order number, and overlapping genomic position. If chromosome numbers were unknown for (either/both) sequences, they were linked with more than 99% identity. After computing transitive closures of the links, each of the known human 7-TMR sequences from the SWISS-PROT was aligned against all the candidate sequences. All clusters that hit for more than 50 amino acids with more than 99% identity were merged. Finally, in each cluster, the longest sequence was selected as the representative.


    (4) Databases used for analysis:

  • Human Genome Resources (Apr,2003)
  • SWISSPROT ver. 41
  • Prosite release 17.41
  • Pfam release 8.0
  • GPCRDB release 7.0
  • nr.aa (Jul,2003)
  • UniGene (Jul,2003)

  • Comments or questions to m-suwa@aist.go.jp
    Recent Revise on 2004/02/06.