Difference between revisions of "Glucocerebrosidase sequence based prediction"

From Bioinformatikpedia
(Jpred3)
(General)
 
(46 intermediate revisions by 2 users not shown)
Line 5: Line 5:
   
 
=== General ===
 
=== General ===
[[File:glucocerebrosidase_secondary_structure.gif||thumb||right||beta sheets and a α-helix as examples for secondary structure <br/>Source: http://www.nature.com/horizon/proteinfolding/background/images/importance_f3.gif]]
+
[[File:glucocerebrosidase_secondary_structure.gif||thumb||right||'''Figure 1:''' Beta sheets, α-helices and stabilizing hydrogen bonds. <ref>http://www.nature.com/horizon/proteinfolding/background/images/importance_f3.gif</ref> ]]
The secondary structure of a protein describes the local conformation of its polypeptide chain, which is limited by the peptide bond and hydrogen bonding considerations. Two types of secondary structures are dominating the local conformations of a polypeptide chain: alpha (α) helices and beta (β) sheets. Some helix structures occuring less frequent are e.g. π-helices and 3,10-helices. Helices and sheets are stabilized by hydrogen bond interactions between the backbone atoms of the corresponding residues and are the only regular secondary structural elements present in proteins, which can also be seen in the picture. Loops or coils are examples for irregular structural elements. <ref>http://en.wikipedia.org/wiki/Biomolecular_structure#Secondary_structure</ref>
+
The secondary structure of a protein describes the local conformation of its polypeptide chain, which is limited by the peptide bond and hydrogen bonding considerations. Two types of secondary structures are dominating the local conformations of a polypeptide chain: alpha (α) helices and beta (β) sheets (cf. Figure 1). Some helix structures occuring less frequent are e.g. π-helices and 3,10-helices. Helices and sheets are stabilized by hydrogen bond interactions between the backbone atoms of the corresponding residues and are the only regular secondary structural elements present in proteins (cf. Figure 1). Loops or coils are examples for irregular structural elements. <ref>http://en.wikipedia.org/wiki/Biomolecular_structure#Secondary_structure</ref>
   
   
Line 163: Line 163:
   
 
Jpred3 predicts the majority of the secondary structure elements correctly: Sheets and helices get mixed up rarely. The same applies to both secondary structures being overlooked and certain regions being assigned to a certain structure by mistake.
 
Jpred3 predicts the majority of the secondary structure elements correctly: Sheets and helices get mixed up rarely. The same applies to both secondary structures being overlooked and certain regions being assigned to a certain structure by mistake.
For the prediction a lot of BLAST-Hits with an E-value of 0 and one Hit with an E-Value of 2e-52 were used: 2wkl, 3keh, 3ke0, 3gxm, 3gxi, 3gxf, 3gxd, 2wcg, 2vt0, 2v3f, 2v3e, 2v3d, 2nt1, 2nt0, 2nsx, 2j25, 2f61, 1y7v, 1ogs (self-hit), 2wnw. Of these proteins, all but 2wnw are glucocerebrosidase proteins of Homo Sapiens. The latter is a hydrolyse activated by a transcription factor from salmonella typhimurium, which seems to be a glucocerebrosidase as well. Some examples of the used proteins are shown in the pictures to the right.
+
For the prediction a lot of BLAST-Hits with an E-value of 0 and one Hit with an E-Value of 2e-52 were used: 2wkl, 3keh, 3ke0, 3gxm, 3gxi, 3gxf, 3gxd, 2wcg, 2vt0, 2v3f, 2v3e, 2v3d, 2nt1, 2nt0, 2nsx, 2j25, 2f61, 1y7v, 1ogs (self-hit), 2wnw. Of these proteins, all but 2wnw are glucocerebrosidase proteins of Homo Sapiens. The latter is a hydrolase activated by a transcription factor from salmonella typhimurium, which seems to be a glucocerebrosidase as well. Due to the high identity the prediction is very good. Some examples of the used proteins are shown in Figures 2 to 5.
   
   
  +
{| border="0" style="text-align:center; border-spacing:0;text-align:left;clear:all;"
[[Image:2wkl.jpg|thumb|right|2wkl [http://www.ebi.ac.uk/pdbsum/2wkl]]]
 
[[Image:3keh.jpg|thumb|right|3keh [http://www.ebi.ac.uk/pdbsum/3keh]]]
 
[[Image:1ogs.jpg|thumb|right|1ogs [http://www.ebi.ac.uk/pdbsum/1ogs]]]
 
[[Image:2wnw.jpg|thumb|right|2wnw [http://www.ebi.ac.uk/pdbsum/2wnw]]]
 
 
{| border="0" style="text-align:center; border-spacing:0;text-align:left;"
 
 
|OrigSeq||<code>MEFSSPSREECPKPLSRVSIMAGSLTGLLLLQAVSWASGARPCIPKSFGYSSVVCVCNATYCDSFDPPTFPALGTFSRYESTRSGRRMELSMGPIQANH</code>
 
|OrigSeq||<code>MEFSSPSREECPKPLSRVSIMAGSLTGLLLLQAVSWASGARPCIPKSFGYSSVVCVCNATYCDSFDPPTFPALGTFSRYESTRSGRRMELSMGPIQANH</code>
 
|-
 
|-
Line 241: Line 236:
 
|-
 
|-
 
|uniprot||<code>EEEEE-EEE-EEEEEEECCCEEEEEEE---EEEEEEE----</code>
 
|uniprot||<code>EEEEE-EEE-EEEEEEECCCEEEEEEE---EEEEEEE----</code>
  +
|}
  +
  +
  +
{| class="centered"
  +
| [[Image:2wkl.jpg|thumb|right|'''Figure 2:''' [http://www.ebi.ac.uk/pdbsum/2wkl 2WKL]]]
  +
| [[Image:3keh.jpg|thumb|right|'''Figure 3:''' [http://www.ebi.ac.uk/pdbsum/3keh 3KEH]]]
  +
| [[Image:1ogs.jpg|thumb|right|'''Figure 4:''' [http://www.ebi.ac.uk/pdbsum/1ogs 1OGS]]]
  +
| [[Image:2wnw.jpg|thumb|right|'''Figure 5:''' [http://www.ebi.ac.uk/pdbsum/2wnw 2WNW]]]
 
|}
 
|}
   
Line 370: Line 373:
 
</code>
 
</code>
   
The sequence differs from the one of the GBA sequence, as DSSP uses the PDB-file (1ogs), containing the sequence without the signaling peptide and both domains of the protein, as input. In the following table, the secondary structure of domain A, as assigned with DSSP, is compared to the secondary structure listed in Uniprot. It can be seen, that the result of DSSP is quite good. Most helices are assigned correctly and the beta sheets are determined most of the time as bends. Only rarely, helices and sheets get mixed up with turns.
+
The sequence differs from the one of the GBA sequence, as DSSP uses the PDB-file (1ogs), containing the sequence without the signaling peptide and both chains of the protein, as input. In the following table, the secondary structure of chain A, as assigned with DSSP, is compared to the secondary structure listed in Uniprot. It can be seen, that the result of DSSP is quite good. Most helices are assigned correctly and the beta sheets are determined most of the time as bends. Only rarely, helices and sheets get mixed up with turns.
 
<br/>
 
<br/>
   
Line 428: Line 431:
   
 
<br/>
 
<br/>
  +
  +
=== Discussion ===
  +
  +
If we compare PSIPRED, Jpred3 and DSSP you can see, that the last two tools show the best results. The reason for the good result of Jpred3 may be, that it uses a lot of sequences with an E-value of 0. These are examples, where the structure is the same. So the prediction must be good.
  +
DSSP also shows a very good result which is highly consistent with the sequence in Uniprot. So the method works quite well.
   
 
== Prediction of disordered regions ==
 
== Prediction of disordered regions ==
Line 433: Line 441:
 
=== General ===
 
=== General ===
   
Disordered regions are regions with no fixed secondary and therefore tertiary structure. The Ramachandran angles of these regions are very flexible, which indicates, that they can have several different secondary structures, often depending on the binding to a substrate.<br/>
+
Disordered regions are regions with no fixed secondary and therefore no fixed tertiary structure. The Ramachandran angles of these regions are very flexible, which indicates, that they can have several different secondary structures, often depending on the binding to a substrate.<br/>
There are two types of disordered regions: the extended or random coil like ones and the collapsed or molten gloguble like ones. They have a lot of different functions, for example they are responsible for the DNA/RNA/protein recognition or the specificity and affinity of the binding. They are often involved in regulatory functions.<ref>http://www.pondr.com/pondr-tut1.html</ref>
+
There are two types of disordered regions: the extended or random coil like ones and the collapsed or molten globule like ones. They have a lot of different functions, for example they are responsible for the DNA/RNA/protein recognition or the specificity and affinity of the binding. They are often involved in regulatory functions.<ref>http://www.pondr.com/pondr-tut1.html</ref>
   
 
=== DISOPRED ===
 
=== DISOPRED ===
Line 448: Line 456:
 
'''Results'''
 
'''Results'''
   
DISOPRED predictions for a false positive rate threshold of 5% are shown in the table below. Asterisks (*) represent disorder predictions and dots (.) predictions of order. The confidence estimates give a rough indication of the probability that each residue is disordered. The signal peptide is marked as a disordered region with a very high confidence which is quite interesting. In addition, a part of a beta sheet was predicted as a disordered regions which may indicate that the prediction might be wrong.
+
DISOPRED predictions for a false positive rate threshold of 5% are shown in the table below. Asterisks (*) represent disorder predictions and dots (.) predictions of order. The estimated confidence gives a rough indication of the probability that each residue is disordered. <br/>
  +
The signal peptide is marked as a disordered region with a very high confidence which is quite interesting. In addition, a part of a beta sheet was predicted as a disordered regions which may indicate that the prediction might be wrong at least in this part.
   
 
{| border="0" border-spacing:0;"
 
{| border="0" border-spacing:0;"
Line 523: Line 532:
   
 
==== POODLE-I ====
 
==== POODLE-I ====
[[File:glucocerebrosidase_poodle-i.png|thumb|right|result of POODLE-I (POODLE series only)]]
+
[[File:glucocerebrosidase_poodle-i.png|thumb|right|'''Figure 6:''' Result of POODLE-I (POODLE series only)]]
The amino acids predicted as disordered are listed in the table below. Furthermore a visualization of the disorder probability of the amino acids is shown in the graphic to the right. Two regions have been predicted as disordered: the first region once again is part of the signal peptide and the second regions is part of a beta sheet.
+
The amino acids predicted as disordered are listed in the table below. Furthermore a visualization of the disorder probability of the amino acids is shown in Figure 6 to the right. Two regions have been predicted as disordered: the first region once again is part of the signal peptide and the second region is part of a beta sheet.
   
 
{| border="0" style="border-spacing:3;"
 
{| border="0" style="border-spacing:3;"
Line 580: Line 589:
 
==== POODLE-S ====
 
==== POODLE-S ====
   
The residues predicted as disordered are listed in the table below. To the right, the different disorder probabilities are illustrated. POODLE-S predicts a part of the signal peptide as disordered as well. The prediction is based on looking for regions that are missing in the X-ray structure. Based on a high B-factor region some disordered regions are found in the rear part of the protein. The first one is located in no defined secondary structure element in Uniprot, the second one is part of a helix.
+
The residues predicted as disordered are listed in the table below, where ''D'' indicates a disordered region and ''0'' and ordered region. In the figures (Figure 7 and 8) to the right, the different disorder probabilities are illustrated. POODLE-S predicts a part of the signal peptide as disordered as well. The prediction is based on looking for regions that are missing in the X-ray structure. Based on a high B-factor region some disordered regions are found in the rear part of the protein. The first one is located in no defined secondary structure element in Uniprot, the second one is part of a helix.
[[File:glucocerebrosidase_poodle-s.png|thumb|right|result of POODLE-S (based on missing residues)]]
+
[[File:glucocerebrosidase_poodle-s.png|thumb|right|'''Figure 7:''' Result of POODLE-S (based on missing residues)]]
[[File:glucocerebrosidase_poodle-s-bfactor.png|thumb|right|result of POODLE-S (based on B-factor region)]]
+
[[File:glucocerebrosidase_poodle-s-bfactor.png|thumb|right|'''Figure 8:''' Result of POODLE-S (based on B-factor region)]]
 
{| border="0" style="border-spacing:3;"
 
{| border="0" style="border-spacing:3;"
 
|'''no.'''||'''AA&nbsp;'''||'''ORD/DIS/xray&nbsp;'''||'''Prob./xray&nbsp;'''||'''ORD/DIS/B-factor&nbsp;'''||'''Prob./B-factor&nbsp;'''
 
|'''no.'''||'''AA&nbsp;'''||'''ORD/DIS/xray&nbsp;'''||'''Prob./xray&nbsp;'''||'''ORD/DIS/B-factor&nbsp;'''||'''Prob./B-factor&nbsp;'''
Line 644: Line 653:
   
 
==== POODLE-L ====
 
==== POODLE-L ====
[[File:glucocerebrosidase_poodle-l.png|thumb|right|result of POODLE-L]]
+
[[File:glucocerebrosidase_poodle-l.png|thumb|right|'''Figure 9:''' Result of POODLE-L]]
   
POODLE-L did not predict any disordered regions, which indicates, that there are no disordered regions with a length of at least 40 residues in the protein.
+
POODLE-L did not predict any disordered regions (cf. Figure 9), which indicates, that there are no disordered regions with a length of at least 40 residues in the protein.
 
<br clear="all"/>
 
<br clear="all"/>
   
Line 667: Line 676:
 
==== IUPred: short disordered regions ====
 
==== IUPred: short disordered regions ====
   
The results of IUPred for short disordered regions are listed in the table below and illustrated in the graphic to the right. The tool predicts disordered regions with a high probability at the beginning of the sequence. This may be explained by the presence of a signal peptide at this position. It is hard to say, if the other regions are disordered as well, because the probabilities are not really high. A disordered region of a length two is really short and therefore not very probable. At the end there is no secondary structure given at Uniprot, which indicates that a disordered region my be present.
+
The results of IUPred for short disordered regions are listed in the table below and illustrated in Figure 10 to the right. The tool predicts disordered regions with a high probability at the beginning of the sequence. This may be explained by the presence of a signal peptide at this position. It is hard to say, if the other regions are disordered as well, because the probabilities are not really high. A disordered region of a length of two residues is really short and therefore not very probable. At the end there is no secondary structure given at Uniprot, which indicates that a disordered region could be present.
   
[[File:Iupred_short_disorder.png||thumb|right|Diagram of IUPred for short disorders]]
+
[[File:Iupred_short_disorder.png||thumb|right|'''Figure 10:''' Diagram of IUPred for short disorders]]
   
 
{| border="0" style="border-spacing:3;"
 
{| border="0" style="border-spacing:3;"
Line 734: Line 743:
 
|536||Q||0.6474
 
|536||Q||0.6474
 
|}
 
|}
 
 
   
 
==== IUPred: long disordered regions ====
 
==== IUPred: long disordered regions ====
   
The results of IUPred for long disordered regions are listed in the table below and illustrated in the graphic to the right. It predicts some very short disordered regions with a very low probability. Therefore it is not very probable, that there are really disordered regions.
+
The results of IUPred for long disordered regions are listed in the table below and are illustrated in Figure 11 to the right. It predicts some very short disordered regions with a very low probability. Thus it is not very probable, that there are really disordered regions.
   
[[File:Iupred_long_disorder.png||thumb|right|Diagram of IUPred for long disorders]]
+
[[File:Iupred_long_disorder.png||thumb|right|'''Figure 11:''' Diagram of IUPred for long disorders]]
   
 
{| border="0" style="border-spacing:3;"
 
{| border="0" style="border-spacing:3;"
Line 782: Line 789:
 
|236 ||L||0.5139
 
|236 ||L||0.5139
 
|}
 
|}
 
   
 
==== IUPred: structured regions ====
 
==== IUPred: structured regions ====
   
[[File:Iupred_structured_regions.png||thumb|right|Diagram of IUPred for structured regions]]
+
[[File:Iupred_structured_regions.png||thumb|right|'''Figure 12''': Diagram of IUPred for structured regions]]
   
IUPred for structured regions predicts globular domains for position 4 to 536. So all except the first four residues are predicted as structured. An illustration of the prediction can be seen to the right.
+
IUPred for structured regions predicts globular domains for position 4 to 536. So all except the first four residues are predicted as structured. An illustration of the prediction can be seen in Figure 12.
 
<br clear="all"/>
 
<br clear="all"/>
   
Line 801: Line 807:
 
''' Results '''
 
''' Results '''
   
The results of META-Disorder are shown in the table below and in the graphic to the right. The tool predicts a disordered region ranging from amino acid 1 to amino acid 6, which is part of the signal peptide. It is hard to say whether it is really a disordered region, as the beginning of a sequence is often mispredicted as a disordered region.
+
The results of META-Disorder are shown in the table below and in Figure 13 to the right. The tool predicts a disordered region ranging from amino acid 1 to amino acid 6, which is part of the signal peptide. It is hard to say whether it is really a disordered region, as the beginning of a sequence is often mispredicted as a disordered region.
   
   
[[File:Glucocerebrosidase_meta_disorder.png||thumb|right|Result of PredictProtein for disordered regions.]]
+
[[File:Glucocerebrosidase_meta_disorder.png||thumb|right|'''Figure 13: '''Result of PredictProtein for disordered regions.]]
   
 
{| border="0" border-spacing:3;"
 
{| border="0" border-spacing:3;"
Line 826: Line 832:
 
=== Discussion ===
 
=== Discussion ===
   
Glucocerebrosidase (Uniprot-ID: P04062) is not listed in the Database of Protein Disorder Disport <ref>http://www.disprot.org/index.php</ref>, which indicates that there are no disordered regions in the protein. The different prediction methods predict probable disordered regions in different parts of the protein which are most of the time located in a certain structural element and therefore may be wrong. Furthermore the different predictions have no real overlaps apart from the regions in the signal peptide. The latter is not present in the mature protein and therefore describes no disordered region in the final structure.
+
Glucocerebrosidase (Uniprot-ID: P04062) is not listed in the Database of Protein Disorder Disport <ref>http://www.disprot.org/index.php</ref>, which indicates that there are no disordered regions in the protein. The different prediction methods predict probable disordered regions in different parts of the protein which are most of the time located in a certain structural element and therefore may be wrong. Furthermore the different predictions have no real overlaps apart from the regions in the signal peptide, where almost all tools show at least a few disordered residues. But the latter is not present in the mature protein and therefore describes no disordered region in the final structure.
  +
  +
== Prediction of coiled coils ==
  +
  +
Our additional method to test another structural feature was the test of coiled coils. A coiled coil is a motif, which is often part of regulation of gene expression. Several alpha helices together build this motif. They usually contain a heptad repeat, which is a pattern hxxhcxc, where h are hydrophobic residues and c charged amino acid residues. This structure allows the special folding of coiled coils.<ref>http://en.wikipedia.org/wiki/Coiled_coil</ref>
  +
In our case we do not expect any coiled coils because Glucocerebrosidase is not a transcription factor and also does not show such helix structures. We used three tools to verify this assumption.
  +
  +
=== COILS ===
  +
  +
[[File:Glucocerebrosidase_Coils.gif||thumb||'''Figure 14:''' Result of COILS for Glucocerebrosidase.]]
  +
  +
[http://www.ch.embnet.org/software/COILS_form.html COILS] is a method which compares the given sequence to a database of known coiled coils. With that it calculates a similarity score and so gets the probability of a coiled coil motif.<ref>Lupas, A., Van Dyke, M., and Stock, J. (1991), Predicting Coled Coils from Protein Sequences, Science 252:1162-1164</ref> The result can be seen in Figure 14 to the right.
  +
  +
''' Usage '''
  +
* Webserver: http://www.ch.embnet.org/software/COILS_form.html
  +
* Input: protein sequence in fasta format, for matrix and window width we used the standard parameter
  +
  +
<br clear = "all"/>
  +
=== MultiCoil ===
  +
  +
[[File:Glucocerebrosidase_MultiCoil.gif||thumb||'''Figure 15:''' Result of MultiCoil for Glucocerebrosidase]]
  +
  +
[http://groups.csail.mit.edu/cb/multicoil/cgi-bin/multicoil.cgi MultiCoil] is based on Parcoil, which predicts coiled coils by pairwise residue correlations.<ref>Bonnie Berger, David B. Wilson, Ethan Wolf, Theodore Tonchev, Mari Milla, and Peter S. Kim, "Predicting Coiled Coils by Use of Pairwise Residue Correlations", Proceedings of the National Academy of Science USA, vol 92, aug 1995, pp. 8259-8263.</ref>. It is a better version for two- and three-stranded coiled coils.<ref>Ethan Wolf, Peter S. Kim, and Bonnie Berger, "MultiCoil: A Program for Predicting Two- and Three-Stranded Coiled Coils", Protein Science 6:1179-1189. June 1997.</ref>. The result is illustrated in Figure 15.
  +
  +
''' Usage '''
  +
* Webserver: http://groups.csail.mit.edu/cb/multicoil/cgi-bin/multicoil.cgi
  +
* Input: protein sequence in fasta format
  +
<br clear = "all"/>
  +
=== Parcoil2 ===
  +
  +
[[File:Glucocerebrosidase_Parcoil2.gif||thumb||'''Figure 16:''' Result of Parcoil2 for Glucocerebrosidase]]
  +
  +
[http://groups.csail.mit.edu/cb/paircoil2/ Parcoil2] works like Parcoil or Multicoil, but is an approved version of 2006, which uses again pairwise residue probabilities and an updated database.<ref>A.V. McDonnell, T. Jiang, A.E. Keating, B. Berger, "Paircoil2: Improved prediction of coiled coils from sequence", Bioinformatics Vol. 22(3) (2006)</ref> The result can be seen in Figure 16 to the right.
  +
  +
''' Usage '''
  +
* Webserver: http://groups.csail.mit.edu/cb/paircoil2/paircoil2.html
  +
* Input: protein sequence in fasta format, minimum search window and p-score as standard
  +
<br clear = "all"/>
  +
=== Discussion ===
  +
  +
None of the three tools, COILS, MultiCoil and Parcoil2 predicted any coiled coil motif. This is not surprising as Glucocerebrosidase does not contain any coiled coils.
   
 
== Prediction of transmembrane alpha-helices and signal peptides ==
 
== Prediction of transmembrane alpha-helices and signal peptides ==
Line 846: Line 892:
 
The high similarity between the hydrophobic region of a transmembrane helix and the one of a signal peptide leads to cross-predictions when conventional transmembrane topology and signal peptide predictors as TMHMM and SignalP are used. Predictors which are based on submodels for both make less errors coming from cross-predictions and help to discriminate against false positives. Furthermore, a predicted signal peptide indicates that the N-terminus of the protein is non-cytoplasmic and is therefore helpful to assign the orientation of the protein. <ref>Käll L. Krogh A, & Sonnhammer, E. L. (2007) Advantages of combined tranasmembrane topology and signal peptide prediction - the Phobius web server. Nucleic Acids Res., Vol. 35, Web server issue, S.429-32</ref><br/>
 
The high similarity between the hydrophobic region of a transmembrane helix and the one of a signal peptide leads to cross-predictions when conventional transmembrane topology and signal peptide predictors as TMHMM and SignalP are used. Predictors which are based on submodels for both make less errors coming from cross-predictions and help to discriminate against false positives. Furthermore, a predicted signal peptide indicates that the N-terminus of the protein is non-cytoplasmic and is therefore helpful to assign the orientation of the protein. <ref>Käll L. Krogh A, & Sonnhammer, E. L. (2007) Advantages of combined tranasmembrane topology and signal peptide prediction - the Phobius web server. Nucleic Acids Res., Vol. 35, Web server issue, S.429-32</ref><br/>
 
Predictors: [[#Phobius and PolyPhobius|Phobius]], [[#Phobius and PolyPhobius|Polyphobius]], [[#OCTOPUS and SPOCTOPUS|SPOCTOPUS]].
 
Predictors: [[#Phobius and PolyPhobius|Phobius]], [[#Phobius and PolyPhobius|Polyphobius]], [[#OCTOPUS and SPOCTOPUS|SPOCTOPUS]].
  +
  +
==== Topology of Glucocerebrosidase ====
  +
  +
Glucocerebrosidase is located in the lysosome and does not contain any transmembrane regions. Residues 1 to 39 are part of a signal peptide <ref>http://www.uniprot.org/uniprot/P04062</ref>. Therefore the prediction tools should place the protein in the non-cytoplasm and not find any transmembrane helices, but instead a signal peptide.
   
 
=== TMHMM ===
 
=== TMHMM ===
TMHMM is a method to predict transmembrane topology of membrane-spanning proteins. It is based on a hidden Markoc model with an architecture of 7 types of states (helix core, helic caps on both sides, one loop on the cytoplasmic side, two loops on the non-cytoplasmic side and a globular domain in the middle of each loop) which correspond to the biological system. The method was established by Sonnhammer et al. in 1998 <ref>Sonnhammer EL, von Heijne G, Krogh A. A hidden Markov model for predicting transmembrane helices in protein sequences. Proc Int Conf Intell Syst Mol Biol. 1998;6:175–182.</ref>.
+
TMHMM is a method to predict transmembrane topology of membrane-spanning proteins. It is based on a hidden Markov model with an architecture of 7 types of states (helix core, helic caps on both sides, one loop on the cytoplasmic side, two loops on the non-cytoplasmic side and a globular domain in the middle of each loop) which correspond to the biological system. The method was established by Sonnhammer et al. in 1998 <ref>Sonnhammer EL, von Heijne G, Krogh A. A hidden Markov model for predicting transmembrane helices in protein sequences. Proc Int Conf Intell Syst Mol Biol. 1998;6:175–182.</ref>.
   
 
''' Usage '''
 
''' Usage '''
Line 857: Line 907:
 
''' Results '''
 
''' Results '''
   
[[Image: Tmhmm_posterior_probabilities_glucocerebrosidase.jpg |thumb|right| TMHMM posterior probabilities for glucocerebrosidase]]
+
[[Image: Tmhmm_posterior_probabilities_glucocerebrosidase.jpg |thumb|right|'''Figure 17:''' TMHMM posterior probabilities for glucocerebrosidase]]
 
<code>sp|P04062|GLCM_HUMAN Length: 536<br/>
 
<code>sp|P04062|GLCM_HUMAN Length: 536<br/>
 
sp|P04062|GLCM_HUMAN Number of predicted TMHs: 0<br/>
 
sp|P04062|GLCM_HUMAN Number of predicted TMHs: 0<br/>
Line 865: Line 915:
 
sp|P04062|GLCM_HUMAN TMHMM2.0 outside 1 536</code>
 
sp|P04062|GLCM_HUMAN TMHMM2.0 outside 1 536</code>
   
TMHMM predicts no transmembrane segments for the sequence of glucocerebrosidase which is correct.
+
TMHMM predicts no transmembrane segments (cf. output and Figure 17) for the sequence of glucocerebrosidase which is correct.
   
   
Line 872: Line 922:
 
=== Phobius and PolyPhobius ===
 
=== Phobius and PolyPhobius ===
   
Phobius is based on a hidden Markoc model which contains submodels for both transmembrane helices and signal peptides and therefore obtains a better discrimination between the two segments than predictors for only one of them. This method was presented in 2004 by Käll et al. <ref>Käll L., et al. A combined transmembrane topology and signal peptide prediction method. J. Mol. Biol. 2004;338:1027–1036.</ref>
+
Phobius is based on a hidden Markov model which contains submodels for both transmembrane helices and signal peptides and therefore obtains a better discrimination between the two segments than predictors for only one of them. This method was presented in 2004 by Käll et al. <ref>Käll L., et al. A combined transmembrane topology and signal peptide prediction method. J. Mol. Biol. 2004;338:1027–1036.</ref>
Polyphobius is a pradiction server that additionally uses an algorithm to include homology information. The performance of transmembrane topology and signal peptide prediction is increased by incorporating extra support from homologs. <ref>Käll L., et al. An HMM posterior decoder for sequence feature prediction that includes homology information
+
Polyphobius is a prediction server that additionally uses an algorithm to include homology information. The performance of transmembrane topology and signal peptide prediction is increased by incorporating extra support from homolougs. <ref>Käll L., et al. An HMM posterior decoder for sequence feature prediction that includes homology information
 
Bioinformatics, 21 (Suppl 1):i251-i257, June 2005.</ref>
 
Bioinformatics, 21 (Suppl 1):i251-i257, June 2005.</ref>
   
Line 881: Line 931:
   
 
''' Results - Phobius '''
 
''' Results - Phobius '''
[[Image: Phobius_posterior_probabilities_glucocerebrosidase.jpg |thumb|right| Phobius posterior probabilities for glucocerebrosidase]]
+
[[Image: Phobius_posterior_probabilities_glucocerebrosidase.jpg |thumb|right|'''Figure 18:''' Phobius posterior probabilities for glucocerebrosidase]]
 
<code>ID sp|P04062|GLCM_HUMAN<br/>
 
<code>ID sp|P04062|GLCM_HUMAN<br/>
 
FT SIGNAL 1 39 <br/>
 
FT SIGNAL 1 39 <br/>
Line 889: Line 939:
 
FT TOPO_DOM 40 536 NON CYTOPLASMIC.</code>
 
FT TOPO_DOM 40 536 NON CYTOPLASMIC.</code>
   
Phobius predics a signal peptide ranging from amino acid 1 to amino acid 39 of the sequence of glucocerebrosidase. This goes along with the information given on Uniprot <ref>http://www.uniprot.org/uniprot/P04062</ref> that the protein has a 39 residue signal sequence. The presence of a signal peptide explains the differences between the sequences of sp|P04062|GLCM_HUMAN and its corresponding PDB structure 1OGS: the sequence of 1OGS has 39 amino acids less than the sequence of sp|P04062|GLCM_HUMAN as the signal peptide is missing in the mature structure. The prediction, that the protein is non-cytoplasmic is true, as glucocerebrosidase is located in the lysosome.
+
Phobius predicts a signal peptide ranging from amino acid 1 to amino acid 39 of the sequence of glucocerebrosidase. This goes along with the information given on Uniprot <ref>http://www.uniprot.org/uniprot/P04062</ref> that the protein has a 39 residue signal sequence. The presence of a signal peptide explains the differences between the sequences of sp|P04062|GLCM_HUMAN and its corresponding PDB structure 1OGS: the sequence of 1OGS has 39 amino acids less than the sequence of sp|P04062|GLCM_HUMAN as the signal peptide is missing in the mature structure. The prediction, that the protein is non-cytoplasmic is true, as glucocerebrosidase is located in the lysosome.
 
<br clear="all"/>
 
<br clear="all"/>
   
Line 898: Line 948:
 
''' Results - PolyPhobius '''
 
''' Results - PolyPhobius '''
   
[[Image: Polyphobius_posterior_probabilities_glucocerebrosidase.jpg |thumb|right| Phobius posterior probabilities for glucocerebrosidase]]
+
[[Image: Polyphobius_posterior_probabilities_glucocerebrosidase.jpg |thumb|right|'''Figure 19:''' Phobius posterior probabilities for glucocerebrosidase]]
 
<code>ID sp|P04062|GLCM_HUMAN<br/>
 
<code>ID sp|P04062|GLCM_HUMAN<br/>
 
FT SIGNAL 1 39 <br/>
 
FT SIGNAL 1 39 <br/>
Line 912: Line 962:
   
 
OCTOPUS (obtainer of correct topologies for uncharacterized sequences) was developed in 2007 by Viklund et al. The method combines hidden markov models and artificial neural networks. Furthermore, OCTOPUS is the first method that integrates the modelation of reentrant-, membrane dip-, and TM hairpin regions. <ref>Viklund, H., and A. Elofsson. 2008. OCTOPUS: improving topology prediction by two-track ANN-based preference scores and an extended topological grammar. Bioinformatics 24:1662-1668.</ref>
 
OCTOPUS (obtainer of correct topologies for uncharacterized sequences) was developed in 2007 by Viklund et al. The method combines hidden markov models and artificial neural networks. Furthermore, OCTOPUS is the first method that integrates the modelation of reentrant-, membrane dip-, and TM hairpin regions. <ref>Viklund, H., and A. Elofsson. 2008. OCTOPUS: improving topology prediction by two-track ANN-based preference scores and an extended topological grammar. Bioinformatics 24:1662-1668.</ref>
SPOCTOPUS is an extention of the OCTOPUS algorithm which additionally predictes signal peptides for reducing predictions of transmembrane regions as signal peptides and the other way round. The method was first mentiond by Viklund et al. in 2008. <ref>Viklund, H., et al. 2008. SPOCTOPUS: a combined predictor of signal peptides and membrane protein topology Bioinformatics 24:2928-2929.</ref>
+
SPOCTOPUS is an extension of the OCTOPUS algorithm which additionally predicts signal peptides for reducing predictions of transmembrane regions as signal peptides and the other way round. The method was first mentioned by Viklund et al. in 2008. <ref>Viklund, H., et al. 2008. SPOCTOPUS: a combined predictor of signal peptides and membrane protein topology Bioinformatics 24:2928-2929.</ref>
   
 
''' Usage '''
 
''' Usage '''
Line 939: Line 989:
   
 
''' Results '''
 
''' Results '''
[[Image: Signalp_prediction_glucocerebrosidase.jpg |thumb|right| Prediction of signal peptide with SignalP 3.0 (image taken from the web server<ref>http://www.cbs.dtu.dk/services/SignalP/</ref>)]]
+
[[Image: Signalp_prediction_glucocerebrosidase.jpg |thumb|right|'''Figure 20:''' Prediction of signal peptide with SignalP 3.0 (image taken from the web server<ref>http://www.cbs.dtu.dk/services/SignalP/</ref>)]]
 
<code>>sp|P04062|GLCM_HUMAN<br/>
 
<code>>sp|P04062|GLCM_HUMAN<br/>
 
Prediction: Signal peptide<br/>
 
Prediction: Signal peptide<br/>
Line 950: Line 1,000:
   
 
=== TargetP ===
 
=== TargetP ===
TargetP, a neural network-based tool, predicts the subcellular location of eukaryotic proteins based on the predicted presence of N-terminal presequences like mitochodraial targeting peptides, chloroplast trasit peptides and pathway signal peptides. For the latter two potential cleavage sites can be predicted as TargetP uses ChloroP and SignalP. This method, which only uses N-terminal sequence information, was published in 2000 by Emanuelsson et al. <ref>Emanuelsson O., et al. Predicting Subcellular Localization of Proteins Based on their N-terminal Amino Acid Sequence. J.Mol.Biol. (2000) 300, 1005-1016.</ref>
+
TargetP, a neural network-based tool, predicts the subcellular location of eukaryotic proteins based on the predicted presence of N-terminal presequences like mitochondrial targeting peptides, chloroplast trasit peptides and pathway signal peptides. For the latter two potential cleavage sites can be predicted as TargetP uses ChloroP and SignalP. This method, which only uses N-terminal sequence information, was published in 2000 by Emanuelsson et al. <ref>Emanuelsson O., et al. Predicting Subcellular Localization of Proteins Based on their N-terminal Amino Acid Sequence. J.Mol.Biol. (2000) 300, 1005-1016.</ref>
   
 
''' Usage '''
 
''' Usage '''
Line 965: Line 1,015:
   
 
TargetP does not predict the signal peptide of glucocerebrosidase. Although the score for a signal peptide is much higher than the one for a mitochondrial targeting peptide, it locates the protein with a very low reliability class rather at any other location than in chloroplast, mitochondrion or the secretory pathway.
 
TargetP does not predict the signal peptide of glucocerebrosidase. Although the score for a signal peptide is much higher than the one for a mitochondrial targeting peptide, it locates the protein with a very low reliability class rather at any other location than in chloroplast, mitochondrion or the secretory pathway.
 
   
 
=== Further Examples of Use ===
 
=== Further Examples of Use ===
Line 993: Line 1,042:
   
 
==== TMHMM ====
 
==== TMHMM ====
The topology predictions of TMHMM for the different proteins are illustrated in the graphic below. TMHMM predicts 6 different transmembrane segments for BACR_HASLA, but indicates, that a signal peptide is possible. According to Uniprot, BACR_HASLA consists of 7 different transmembrane segments and no signal peptide. The last transmembrane helix, ranging from amino acid 217 to 236 was not predicted by TMHMM. The prediction that RET4_HUMAN and INSL5_Human do not contain any transmembrane segments goes along with the corresponding Uniprot entries. TMHMM finds 2 transmembrane segments in LAMP1_HUMAN. The first transmembrane segment is in reality a signal paptide, as indicated in the Uniprot entry and is therefore another example for the confusion of transmembrane segments and signal peptides. The single transmembrane helix of A4_HUMAN was predicted correctly. <br/> <br/>
+
The topology predictions of TMHMM for the different proteins are illustrated in the graphic below. TMHMM predicts 6 different transmembrane segments for BACR_HASLA, but indicates, that a signal peptide is possible. According to Uniprot, BACR_HASLA consists of 7 different transmembrane segments and no signal peptide. The last transmembrane helix, ranging from amino acid 217 to 236 was not predicted by TMHMM. The prediction that RET4_HUMAN and INSL5_Human do not contain any transmembrane segments goes along with the corresponding Uniprot entries. TMHMM finds 2 transmembrane segments in LAMP1_HUMAN. The first transmembrane segment is in reality a signal peptide, as indicated in the Uniprot entry and is therefore another example for the confusion of transmembrane segments and signal peptides. The single transmembrane helix of A4_HUMAN was predicted correctly. <br/> <br/>
   
 
[[Image: Tmhmm_membrane_topologies_further_examples.jpg |center| Membrane topologies predicted with TMHMM for different proteins]]
 
[[Image: Tmhmm_membrane_topologies_further_examples.jpg |center| Membrane topologies predicted with TMHMM for different proteins]]
Line 1,056: Line 1,105:
   
 
=== General ===
 
=== General ===
  +
[[Image: Go_tree_example.jpg |thumb|right|'''Figure 21:''' GO-tree with a reference annotation and a prediction.)]]
 
 
The Gene Ontology Consortium tries to unify the terminology of gene and gene product attributes across all species to decrease non-consistent descriptions in different databases. It therefore has developed three different ontologies: cellular component, molecular function and biological process. <ref>http://www.geneontology.org/GO.doc.shtml</ref>
 
The Gene Ontology Consortium tries to unify the terminology of gene and gene product attributes across all species to decrease non-consistent descriptions in different databases. It therefore has developed three different ontologies: cellular component, molecular function and biological process. <ref>http://www.geneontology.org/GO.doc.shtml</ref>
For each prediction, precision and recall are calculated by comparing the paths to the root in the GO-tree of the predicted and the reference terms. Comparing the paths to the root, allows one to not count a predicted parent of a reference GO-term as false positive, but instead as true positive and the resulting values are therefore more reliable than the ones obtained by just comparing the predicted with thre reference GO-terms. For the GO-tree, the transitive closure<ref>http://www.geneontology.org/scratch/transitive_closure/go_transitive_closure.links</ref> of june 2008 is used. This file is not up to date and therefore does not contain each GO-term, but still helps to give an impression of how good the predictions are.
+
For each prediction, precision (tp/(tp+fp)) and recall (tp/(tp+fn)) are calculated by comparing the paths to the root in the GO-tree of the predicted and the reference terms. Comparing the paths to the root allows one to not count a predicted child of a reference GO-term as false positive, but instead as true positive and the resulting values are therefore more reliable than the ones obtained by just comparing the predicted with thre reference GO-terms. Example: In Figure 21, one can see an exemplary GO tree with a reference annotation and a prediction. If one would just compare the GO terms without taking the whole tree into account, this would result in one false positive (term 2), one false negative (term 3) and zero true positives. In contrast, if one uses the paths to the root of the tree of both reference annotation and prediction, one gets three true positives (domain, term 1, term 2) and one false negative (term 3). <br/>
  +
For the GO-tree, the transitive closure<ref>http://www.geneontology.org/scratch/transitive_closure/go_transitive_closure.links</ref> of june 2008 is used. This file is not up to date and therefore does not contain each GO-term, but still helps to give an impression of how good the predictions are.
   
   
Line 1,153: Line 1,203:
   
 
ProtFun is an ''ab initio'' prediction server for protein function which is based on sequence derived protein features as predicted post translational modifications, protein sorting signals and phisical/chemical properties that have been calculated from the amino acid composition of the input sequence. <ref>Jehnsen L., et al. Ab initio prediction of human orphan protein function from post-translational modifications and localization features. J. Mol. Biol., 319:1257-1265, 2002</ref>
 
ProtFun is an ''ab initio'' prediction server for protein function which is based on sequence derived protein features as predicted post translational modifications, protein sorting signals and phisical/chemical properties that have been calculated from the amino acid composition of the input sequence. <ref>Jehnsen L., et al. Ab initio prediction of human orphan protein function from post-translational modifications and localization features. J. Mol. Biol., 319:1257-1265, 2002</ref>
 
   
 
''' Usage'''
 
''' Usage'''
Line 1,236: Line 1,285:
 
<br/>
 
<br/>
 
PDF-File containing the results of the different methods and the correct GO-term annotations: [[File:GO_term_prediction_for_several_different_proteins.pdf]]
 
PDF-File containing the results of the different methods and the correct GO-term annotations: [[File:GO_term_prediction_for_several_different_proteins.pdf]]
  +
  +
=== Discussion ===
  +
  +
GOPET predicts GO terms corresponding to molecular function very well. The fact, that the recall values are quite low for the further examples of use is due to the fact, that all GO-terms and not only the ones for molecular function have been used as reference.
  +
The GO-terms retrieved with Pfam are very accurate: in each prediction, only correct GO-terms have been predicted which results in a precision of 1.0.
  +
ProtFun2.2 differs from the other two prediction methods, as it only assigns the protein to one gene ontology category. This resuls in very low recall values and it happend several times, that the prediction was not correct.
  +
  +
None of the three different methods was able to predict each of the annotated GO-terms. But this is not surprising, as the Gene Ontology consists of a very large number of terms. GOPET and Pfam are therefore a good solution to get a broad overview of the different functions and localizations of your protein of interest.
   
 
== References ==
 
== References ==
   
 
<references />
 
<references />
  +
  +
  +
[[Category:Gaucher_Disease]]

Latest revision as of 17:52, 28 August 2011

In this section several different sequence based predictions are applied to the sequence of glucocerebrosidase. For each prediction tool, the webserver and the required input is indicated. As not all results presented in this website were obtained by using the webservers, but instead by running the tools locally (e.g. PSIPRED, DISOPRED, TMHMM and SignalP), there might be differences in the results when trying to reproduce them.


Secondary structure prediction

General

Figure 1: Beta sheets, α-helices and stabilizing hydrogen bonds. <ref>http://www.nature.com/horizon/proteinfolding/background/images/importance_f3.gif</ref>

The secondary structure of a protein describes the local conformation of its polypeptide chain, which is limited by the peptide bond and hydrogen bonding considerations. Two types of secondary structures are dominating the local conformations of a polypeptide chain: alpha (α) helices and beta (β) sheets (cf. Figure 1). Some helix structures occuring less frequent are e.g. π-helices and 3,10-helices. Helices and sheets are stabilized by hydrogen bond interactions between the backbone atoms of the corresponding residues and are the only regular secondary structural elements present in proteins (cf. Figure 1). Loops or coils are examples for irregular structural elements. <ref>http://en.wikipedia.org/wiki/Biomolecular_structure#Secondary_structure</ref>



PSIPRED

PSIPRED is a method by David T. Jones, published 1999 in JMB with "Protein Secondary Structure Prediction Based on Position-specific Scoring Matrices". PSIPRED works with a two-stage neural network that analyses the output of PSI-BLAST to predict the secondary structure of a protein. <ref>David T. Jones, Protein Secondary Structure Prediction Based on Position-specific Scoring Matrices, JMB, 1999</ref>

Usage

Results

The online, as well as the local version of PSIPRED were applied to the sequence of glucocerebrosidase. Both runs resulted in different results which are compared to the secondary structure given in Uniprot. <ref>http://www.uniprot.org/uniprot/P04062</ref>. As one can see in the table below, the results differ a lot which may be traced back to the fact that different parameters were used in both versions. The comparison to the structure listed in Uniprot shows however, that the results of both versions differ in many regions considerably from the reference structure.


Conf: 988898954488887622315999999999998641038968865325999649995388
online: CCCCCCCCCCCCCCCCCHHHHHHHHHHHHHHHHHHHHCCCCCCCCCCCCCCCEEEEECCC
Conf: 987898955489988742200466888998986410038977877777863169974474
local: CCCCCCCCCCCCCCCCCEEEEHHHHHHHHHHHHHHHCCCCCCCCCCCCCCCEEEEEECCC
uniprot: ------------------------------------------------EEEE-EEEEEE-
AA: MEFSSPSREECPKPLSRVSIMAGSLTGLLLLQAVSWASGARPCIPKSFGYSSVVCVCNAT
 
Conf: 558889998889992599996377885421237645688875108995378301079247
online: CCCCCCCCCCCCCCCEEEEEECCCCCCCCCCCCCCCCCCCCCCCEEEECCCCCEEEEEEE
Conf: 148899998788875431100345640022100111177897107840966454557422
local: CCCCCCCCCCCCCCCCCCCCCCCCCCCEEEEEEEEECCCCCCCCEEEECCCCCCCEEEEE
uniprot: -------------EEEEEEEE-----EEEEEEE-EEE----EEEEEEEEEEEEEE--EEE
AA: YCDSFDPPTFPALGTFSRYESTRSGRRMELSMGPIQANHTGTGLLLTLQPEQKFQKVKGF
 
Conf: 300233899997249999999999860597882105999750588999986666899999
online: EECCCHHHHHHHHCCCHHHHHHHHHHCCCCCCCEEEEEEEEECCCCCCCCCCCCCCCCCC
Conf: 011335889987508927898998851396893001358621344677653324799999
local: CCCCCHHHHHHHHHCCHHHHHHHHHHHCCCCCCEEEEEEEECCCCCCCCCCCCCCCCCCC
uniprot: EE--HHHHHHH----HHHHHHHHHHHH-CCCC---EEEEEEE--EEEEE------EEE--
AA: GGAMTDAAALNILALSPPAQNLLLKSYFSEEGIGYNIIRVPMASCDFSIRTYTYADTPDD
 
Conf: 689999994100245289999999971999389971377785612147247999889999
online: CCCCCCCCCHHCCCCCHHHHHHHHHHCCCCCEEEECCCCCCCCCEECCCCCCCCCCCCCC
Conf: 721111368543220024799998733999689957899974220056347854325899
local: CCCCCCCCCCCCCCCHHHHHHHHHHHCCCCCEEEECCCCCCCCCCCCCCCCCCCCCCCCC
uniprot: --------HHHH--HHHHHHHHHHH-----EEEEEEE---HHH----EEEEE-EEEE---
AA: FQLHNFSLPEEDTKLKIPLIHRALQLAQRPVSLLASPWTSPTWLKTNGAVNGKGSLKGQP
 
Conf: 922699999999999999975490786872012579899999999986349999999999
online: CCHHHHHHHHHHHHHHHHHHHCCEEEEEEECCCCCCCCCCCCCCCCCCCCCHHHHHHHHH
Conf: 971468799999999967663395143898112789787678873222114422121122
local: CCHHHHHHHHHHHHHHHHHHHCCCCEEEEEEECCCCCCCCCCCCCCCCCCCCCCCCHHHH
uniprot: -HHHHHHHHHHHHHHHHHHH-----EEEEE-----HHH------------HHHHHHHHHH
AA: GDIYHQTWARYFVKFLDAYAEHKLQFWAVTAENEPSAGLLSGYPFQCLGFTPEHQRDFIA
 
Conf: 955799851689972999944888873334664149955640224689831699998033
online: HHHHHHHHCCCCCCEEEEEECCCCCCHHHHHHHHCCCHHHHCCCCEEEEEECCCCCCHHH
Conf: 111332310577410134212544556520222238976651151878702212236320
local: HHHHHHHHCCCCCCCEEEEECCCCCCCCCCHHHHCCCHHHHHCCEEEEEECCCCCCCCCC
uniprot: -HHHHHH--CCCCEEEEEEEEEHHH--HHHHHHH--HHHH----EEEEEEE------HHH
AA: RDLGPTLANSTHHNVRLLMLDDQRLLLPHWAKVVLTDPEAAKYVHGIAVHWYLDFLAPAK
 
Conf: 412688750999509994343699998866567831444255999999996402335772
online: HHHHHHHHCCCCCEEEEECCCCCCCCCCCCCCCCCHHHHHHHHHHHHHHHHHCCEEEEEE
Conf: 011111126898001101210389653344579861210143212566552001100000
local: CCCCCCCCCCCCCCEEHHHHCCCCCCCCCCCCCCCHHHHCCCCHHHHHHHHHHHHHHEEE
uniprot: HHHHHHHH---EEEEEEEEE--------------HHHHHHHHHHHHHHHH--EEEEEEEE
AA: ATLGETHRLFPNTMLFASEACVGSKFWEQSVRLGSWDRGMQYSHSIITNLLYHVVGWTDW
 
Conf: 000169999986689878535895679769986202333102244469939999542389
online: EECCCCCCCCCCCCCCCCCCEEEECCCCEEEECCHHHHHHHHCCCCCCCCEEEEEEECCC
Conf: 023699999860001325228998208703226821232123444679927984435078
local: CCCCCCCCCCCCEECCCCCCEEEEECCCCEEECCCEEEECCCCCCCCCCCEEEEEEEECC
uniprot: E-----------------EEEEEHHH-EEEE-HHHHHHHHHH-------EEEEEEEEE--
AA: NLALNPEGGPNWVRNFVDSPIIVDITKDTFYKQPMFYHLGHFSKFIPEGSQRVGLVASQK
 
Conf: 99028999928998999999099993779999099304998518951899999309
online: CCCEEEEEECCCCCEEEEEECCCCCCEEEEEEECCCCEEEEECCCCEEEEEEEEEC
Conf: 99538995649997799999146898214741998642000389842568774139
local: CCCCEEEEECCCCCEEEEEEECCCCCEEEEECCCCCCCCCCCCCCCEEEEEEEECC
uniprot: EEEEEEEE-----EEEEEEE-EEE-EEEEEEECCCEEEEEEE---EEEEEEE----
AA: NDLDAVALMHPDGSAVVVVLNRSSKDVPLTIKDPAVGFLETISPGYSIHTYLWRRQ

Jpred3

Jpred3 was published in 2008 by Christian Cole, Jonathan D. Barber and Geoffrey J. Barton as "The Jpred 3 secondary structure prediction server" in Nucl. Acids Res. The Jnet algorithm predicts the secondary structure and the solvent accessibility of a protein with the help of alignment profiles. Therefore it uses the position-specific scoring matrix (PSSM) created by PSI-BLAST and a hidden Markov model. The final prediction is made with a neural network.<ref>http://nar.oxfordjournals.org/content/36/suppl_2/W197.full</ref>
As input, either the protein sequence or a multiple sequence alignment is taken. We decided to take the protein sequence.

Usage

Results

Jpred3 predicts the majority of the secondary structure elements correctly: Sheets and helices get mixed up rarely. The same applies to both secondary structures being overlooked and certain regions being assigned to a certain structure by mistake. For the prediction a lot of BLAST-Hits with an E-value of 0 and one Hit with an E-Value of 2e-52 were used: 2wkl, 3keh, 3ke0, 3gxm, 3gxi, 3gxf, 3gxd, 2wcg, 2vt0, 2v3f, 2v3e, 2v3d, 2nt1, 2nt0, 2nsx, 2j25, 2f61, 1y7v, 1ogs (self-hit), 2wnw. Of these proteins, all but 2wnw are glucocerebrosidase proteins of Homo Sapiens. The latter is a hydrolase activated by a transcription factor from salmonella typhimurium, which seems to be a glucocerebrosidase as well. Due to the high identity the prediction is very good. Some examples of the used proteins are shown in Figures 2 to 5.


OrigSeq MEFSSPSREECPKPLSRVSIMAGSLTGLLLLQAVSWASGARPCIPKSFGYSSVVCVCNATYCDSFDPPTFPALGTFSRYESTRSGRRMELSMGPIQANH
Jnet -----------------HHHHHHHHHHHHHHHHHHHH---------------EEEEE-----------------EEEEEEE------------------
jhmm ----------------HHHHHHHHHHHHHHHHHHHHH---------------EEEEE-----------------EEEEEEE------------------
jpssm ------------------HHHHHHHHHHHHHHHHHH---------------EEEEEE-----------------EEEEEEE------------------
uniprot ------------------------------------------------EEEE-EEEEEE--------------EEEEEEEE-----EEEEEEE-EEE--
 
OrigSeq TGTGLLLTLQPEQKFQKVKGFGGAMTDAAALNILALSPPAQNLLLKSYFSEEGIGYNIIRVPMASCDFSIRTYTYADTPDDFQLHNFSLPEEDTKLKIP
Jnet ----EEEEEE----EEEEEEEEEEHHHHHHHHH----HHHHHHHHHHH-------EEEEEE----------------------------------HHHH
jhmm ----EEEEE-----EEEEEEEEEEEHHHHHHHH----HHHHHHHHHHH-------EEEEEE----------------------------------HHHH
jpssm ----EEEEEE----EEEEEEEE-HHHHHHHHHH----HHHHHHHHHHH------EEEEEEEE--------------------------------HHHHH
uniprot --EEEEEEEEEEEEEE--EEEEE--HHHHHHH----HHHHHHHHHHHH-CCCC---EEEEEEE--EEEEE------EEE----------HHHH--HHHH
 
OrigSeq LIHRALQLAQRPVSLLASPWTSPTWLKTNGAVNGKGSLKGQPGDIYHQTWARYFVKFLDAYAEHKLQFWAVTAENEPSAGLLSGYPFQCLGFTPEHQRD
Jnet HHHHHHHH----EEEEE--------EE-------------------HHHHHHHHHHHHHHHHH----EEEEE---------------------HHHHHH
jhmm HHHHHHHH----EEEEE-----------------------------HHHHHHHHHHHHHHHHH----EEEEE---------------------HHHHHH
jpssm HHHHHHHHH---EEEEE--------EEE-----------------HHHHHHHHHHHHHHHHHH----EEEEEE--------------------HHHHHH
uniprot HHHHHHH-----EEEEEEE---HHH----EEEEE-EEEE----HHHHHHHHHHHHHHHHHHH-----EEEEE-----HHH------------HHHHHHH
 
OrigSeq FIARDLGPTLANSTHHNVRLLMLDDQRLLLPHWAKVVLTDPEAAKYVHGIAVHWYLDFLAPAKATLGETHRLFPNTMLFASEACVGSKFWEQSVRLGSW
Jnet HHHHHHHHHHHH-----EEEEEE--------HHHHHHH--HHHHHHHHEEEE----------HHHHHHHHHH-----EEEEEEEE--------------
jhmm HHHHHHHHHHHH------EEEEE-------HHHHHHHH--H-HHHHHHEEEE----------HHHHHHHHHH-----EEEEEEEE--------------
jpssm HHHHHHHHHHHH-----EEEEEEE-------HHHHHH----HHHHHH--EE-----------HHHHHHHHHH-----EEEEEEE--------------H
uniprot HHH-HHHHHH--CCCCEEEEEEEEEHHH--HHHHHHH--HHHH----EEEEEEE------HHHHHHHHHHH---EEEEEEEEE--------------HH
 
OrigSeq DRGMQYSHSIITNLLYHVVGWTDWNLALNPEGGPNWVRNFVDSPIIVDITKDTFYKQPMFYHLGHFSKFIPEGSQRVGLVASQKNDLDAVALMHPDGSA
Jnet HHHHHHHHHHHHHHHHHHHHHHHHHHH----------------EEEEE----EEEE---HHHHHHHH-------EEEEE-------EEEEEEE-----E
jhmm -HHHHHHHHHHHHHHHHHHHHHHHHHH----------------EEEEE----EEEE---HHHHHHHH-------EEEE--------EEEEEEE-----E
jpssm HHHHHHHHHHHHHHHHHHHHHHHHHHE----------------EEEEE----EEEE--HHHHHHHH--------EEEEEE------EEEEEEEE----E
uniprot HHHHHHHHHHHHHH--EEEEEEEEE-----------------EEEEEHHH-EEEE-HHHHHHHHHH-------EEEEEEEEE--EEEEEEEE-----EE
 
OrigSeq VVVVLNRSSKDVPLTIKDPAVGFLETISPGYSIHTYLWRRQ
Jnet EEEEEE-----EEEEEEE---EEEEEEE----EEEEEEE--
jhmm EEEEEE-----EEEEEEE---EEEEEEE----EEEEEEE--
jpssm EEEEEE----EEEEEEEE---EEEEEEE---EEEEEEE---
uniprot EEEEE-EEE-EEEEEEECCCEEEEEEE---EEEEEEE----


Figure 2: 2WKL
Figure 3: 3KEH
Figure 4: 1OGS
Figure 5: 2WNW


Comparison with DSSP

DSSP, which stands for Define Secondary Structure of Proteins, was published by Wolfgang Kabsch and Chris Sander in 1983 with the title "Dictionary of protein secondary structure: pattern recognition of hydrogen-bonded and geometrical features".<ref>http://swift.cmbi.ru.nl/gv/dssp/</ref>
The DSSP algorithm recognises the secondary structure of a protein by defining hydrogen bonds with an electrostatic definition. The different patterns of these hydrogen bonds constitute one of the eight possible secondary structure types. DSSP is therefore a secondary structure assignment tool, rather than a prediction tool. <ref>http://en.wikipedia.org/wiki/DSSP_%28protein%29</ref>

Usage

Results

The result of DSSP consists of four lines: The first line describes the amino acid sequence, the second line the secondary structure, the third the residues which are involved in symmetry contacts (marked with an asterix (*)) and the fourth solvent accessible residues, which are marked with an A. The secondary structure consists of the elements H, T, 3 and S, where H indicates an alpha helix, T an hydrogen bonded turn, 3 a residue in an isolated beta-bridge and S a bend, which is a region of high curvature.

                    10        20        30        40        50        60
                     |         |         |         |         |         |
   1 -   60 ARPCIPKSFGYSSVVCVCNATYCDSFDPPTFPALGTFSRYESTRSGRRMELSMGPIQANH
   1 -   60      SSS TTTTSSSSSSTT            TTSSSSSSSSTTT  TSSSSSS  TT
   1 -   60                     *                                      *
   1 -   60 AAA AAAAAAAA        A     AA  A  A    A    AA A  AA   A AAAA
                    70        80        90       100       110       120
                     |         |         |         |         |         |
  61 -  120 TGTGLLLTLQPEQKFQKVKGFGGAMTDAAALNILALSPPAQNLLLKSYFSEEGIGYNIIR
  61 -  120    TSSSSSSSSSSSSS  SSSSS  HHHHHHHTTT HHHHHHHHHHHHTTTTT   SSS
  61 -  120   *    * *
  61 -  120 A A  A   A AAAA A             A  A  AAA  A   A    AA
                   130       140       150       160       170       180
                     |         |         |         |         |         |
 121 -  180 VPMASCDFSIRTYTYADTPDDFQLHNFSLPEEDTKLKIPLIHRALQLAQRPVSLLASPWT
 121 -  180 SSST  TTTTT   T  TTT TT TT    HHHHTTHHHHHHHHHHH TT  SSSSSST
 121 -  180          **       *           **
 121 -  180          AAA    AAAA AA AA    A  AA   A  AA AAA A A
                   190       200       210       220       230       240
                     |         |         |         |         |         |
 181 -  240 SPTWLKTNGAVNGKGSLKGQPGDIYHQTWARYFVKFLDAYAEHKLQFWAVTAENEPSAGL
 181 -  240   333 TT TTTTT   TT TTTHHHHHHHHHHHHHHHHHHHTT   TSSST TTTT333
 181 -  240        ***** *
 181 -  240       AAAAAA A   AAA  AAA A   A   A  A   AAA A             A
                   250       260       270       280       290       300
                     |         |         |         |         |         |
 241 -  300 LSGYPFQCLGFTPEHQRDFIARDLGPTLANSTHHNVRLLMLDDQRLLLPHWAKVVLTDPE
 241 -  300 TTT  T      HHHHHHHHHHTHHHHHHTTTTTTTSSSSSSSS333TTHHHHHHHTTHH
 241 -  300  ** *                                        *
 241 -  300 AAAAA      A AA  A   A   A  AA A AA A      A A   A  A   A AA
                   310       320       330       340       350       360
                     |         |         |         |         |         |
 301 -  360 AAKYVHGIAVHWYLDFLAPAKATLGETHRLFPNTMLFASEACVGSKFWEQSVRLGSWDRG
 301 -  360 HHTT  SSSSSSSTTT   HHHHHHHHHHH TTTSSSSSSSS    TTT T  TT HHHH
 301 -  360                **           *   *             **
 301 -  360   AA        A  AA A AA   A AA  AA            AAA A  A    A
                   370       380       390       400       410       420
                     |         |         |         |         |         |
 361 -  420 MQYSHSIITNLLYHVVGWTDWNLALNPEGGPNWVRNFVDSPIIVDITKDTFYKQPMFYHL
 361 -  420 HHHHHHHHHHHHTTSSSSSSSST   TTT   TT      TSSSS333TSSSS HHHHHH
 361 -  420                               *** ***        **
 361 -  420     A      AA            AAA    A A A        AAA
                   430       440       450       460       470       480
                     |         |         |         |         |         |
 421 -  480 GHFSKFIPEGSQRVGLVASQKNDLDAVALMHPDGSAVVVVLNRSSKDVPLTIKDPAVGFL
 421 -  480 HHHHTT  TT SSSSSSSTT  TSSSSSSS TTT SSSSSSS TTT SSSSSSSTTTSSS
 421 -  480                 *****                               *     *
 421 -  480         A       AAAAAAA       AAA         A AAA A   A AA  A
                   490
                     |
 481 -  497 ETISPGYSIHTYLWHRQ
 481 -  497 SSSS TTSSSSSSS
 481 -  497                 *
 481 -  497 A A A          AA

            500       510       520       530       540       550
              |         |         |         |         |         |
 498 -  557 ARPCIPKSFGYSSVVCVCNATYCDSFDPPTFPALGTFSRYESTRSGRRMELSMGPIQANH
 498 -  557      SS  TTTT SSSS TT      T     TTSSSSSSSSTTT  TSSSSSS  T
 498 -  557   * *                           *  *             ****** *
 498 -  557 AAA AAAAAAAA        A    AAAAAA AA    A    AA A   A A A AAAA
            560       570       580       590       600       610
              |         |         |         |         |         |
 558 -  617 TGTGLLLTLQPEQKFQKVKGFGGAMTDAAALNILALSPPAQNLLLKSYFSEEGIGYNIIR
 558 -  617   TTSSSSSSSSSSSSS  SSSSS  HHHHHHHHTT HHHHHHHHHHHHTTTTT   SSS
 558 -  617                               *     *
 558 -  617 AAAA A   A AAAA A             A  AA AAA  A   A    A
            620       630       640       650       660       670
              |         |         |         |         |         |
 618 -  677 VPMASCDFSIRTYTYADTPDDFQLHNFSLPEEDTKLKIPLIHRALQLAQRPVSLLASPWT
 618 -  677 SSST  TTTTT   T  TTT TT TT    HHHHTTHHHHHHHHHHH TT  SSSSSST
 618 -  677          ***    *             *
 618 -  677          AAA    AAAA  A AA A  A  AAA     AA AAA AAA
            680       690       700       710       720       730
              |         |         |         |         |         |
 678 -  737 SPTWLKTNGAVNGKGSLKGQPGDIYHQTWARYFVKFLDAYAEHKLQFWAVTAENEPSAGL
 678 -  737   333 TT TTTTT   TT TTTHHHHHHHHHHHHHHHHHHHTT   TSSST TT33333
 678 -  737   *
 678 -  737   A   AAA  A A   AAA      A   A   A  A   AAA A             A
            740       750       760       770       780       790
              |         |         |         |         |         |
 738 -  797 LSGYPFQCLGFTPEHQRDFIARDLGPTLANSTHHNVRLLMLDDQRLLLPHWAKVVLTDPE
 738 -  797 TTT  T      HHHHHHHHHHTHHHHHHTTTTTTTSSSSSSSS333TTHHHHHHHTTHH
 738 -  797   ***                                        **
 738 -  797 AAA A      A A  AA   A   A  AA A AA A      A A   A  A   A AA
            800       810       820       830       840       850
              |         |         |         |         |         |
 798 -  857 AAKYVHGIAVHWYLDFLAPAKATLGETHRLFPNTMLFASEACVGSKFWEQSVRLGSWDRG
 798 -  857 HHTT  SSSSSSSTTT   HHHHHHHHHHH TTTSSSSSSSST  TTTT T  TT HHHH
 798 -  857                **   *       ** *              **
 798 -  857   A         A  AA A AA   A AAA AA            AAA A  A    A
            860       870       880       890       900       910
              |         |         |         |         |         |
 858 -  917 MQYSHSIITNLLYHVVGWTDWNLALNPEGGPNWVRNFVDSPIIVDITKDTFYKQPMFYHL
 858 -  917 HHHHHHHHHHHHTTSSSSSSSST   TTT   TT      TSSSS333TSSSS HHHHHH
 858 -  917                               *   **         **
 858 -  917     A      AA             AA    AAAA         AAA
            920       930       940       950       960       970
              |         |         |         |         |         |
 918 -  977 GHFSKFIPEGSQRVGLVASQKNDLDAVALMHPDGSAVVVVLNRSSKDVPLTIKDPAVGFL
 918 -  977 HHHHTT  TT SSSSSSSTT  TSSSSSSS TTT SSSSSSS TTT SSSSSSSTTTSSS
 918 -  977
 918 -  977         A       AAAAAAA        AA         A AAA A   A A   A
            980       990
              |         |
 978 -  994 ETISPGYSIHTYLWHRQ
 978 -  994 SSSS TTSSSSSSS
 978 -  994
 978 -  994 A A            AA

The sequence differs from the one of the GBA sequence, as DSSP uses the PDB-file (1ogs), containing the sequence without the signaling peptide and both chains of the protein, as input. In the following table, the secondary structure of chain A, as assigned with DSSP, is compared to the secondary structure listed in Uniprot. It can be seen, that the result of DSSP is quite good. Most helices are assigned correctly and the beta sheets are determined most of the time as bends. Only rarely, helices and sheets get mixed up with turns.

uniprot ----------EEEE-EEEEEE--------------EEEEEEEE-----EEEEEEE-EEE-
DSSP -----SSS-TTTTSSSSSSTT------------TTSSSSSSSSTTT--TSSSSSS--TT-
 
uniprot ---EEEEEEEEEEEEEE--EEEEE--HHHHHHH----HHHHHHHHHHHH-CCCC---EEE
DSSP TSSSSSSSSSSSSS--SSSSS--HHHHHHHTTT-HHHHHHHHHHHHTTTTT---SSS
 
uniprot EEEE--EEEEE------EEE----------HHHH--HHHHHHHHHHH-----EEEEEEE-
DSSP SSST--TTTTT---T--TTT-TT-TT----HHHHTTHHHHHHHHHHH-TT--SSSSSST-
 
uniprot --HHH----EEEEE-EEEE----HHHHHHHHHHHHHHHHHHH-----EEEEE-----HHH
DSSP --333-TT-TTTTT---TT-TTTHHHHHHHHHHHHHHHHHHHTT---TSSST-TTTT333
 
uniprot ------------HHHHHHHHHH-HHHHHH--CCCCEEEEEEEEEHHH--HHHHHHH--HH
DSSP TTT--T------HHHHHHHHHHTHHHHHHTTTTTTTSSSSSSSS333TTHHHHHHHTTHH
 
uniprot HH----EEEEEEE------HHHHHHHHHHH---EEEEEEEEE--------------HHHH
DSSP HHTT--SSSSSSSTTT---HHHHHHHHHHH-TTTSSSSSSSS----TTT-T--TT-HHHH
 
uniprot HHHHHHHHHHHH--EEEEEEEEE-----------------EEEEEHHH-EEEE-HHHHHH
DSSP HHHHHHHHHHHHTTSSSSSSSST---TTT---TT------TSSSS333TSSSS-HHHHHH
 
uniprot HHHH-------EEEEEEEEE--EEEEEEEE-----EEEEEEE-EEE-EEEEEEECCCEEE
DSSP HHHHTT--TT-SSSSSSSTT--TSSSSSSS-TTT-SSSSSSS-TTT-SSSSSSSTTTSSS
 
uniprot EEEE---EEEEEEE---
DSSP SSSS-TTSSSSSSS---


Discussion

If we compare PSIPRED, Jpred3 and DSSP you can see, that the last two tools show the best results. The reason for the good result of Jpred3 may be, that it uses a lot of sequences with an E-value of 0. These are examples, where the structure is the same. So the prediction must be good. DSSP also shows a very good result which is highly consistent with the sequence in Uniprot. So the method works quite well.

Prediction of disordered regions

General

Disordered regions are regions with no fixed secondary and therefore no fixed tertiary structure. The Ramachandran angles of these regions are very flexible, which indicates, that they can have several different secondary structures, often depending on the binding to a substrate.
There are two types of disordered regions: the extended or random coil like ones and the collapsed or molten globule like ones. They have a lot of different functions, for example they are responsible for the DNA/RNA/protein recognition or the specificity and affinity of the binding. They are often involved in regulatory functions.<ref>http://www.pondr.com/pondr-tut1.html</ref>

DISOPRED

DISOPRED was published by Ward JJ, Sodhi JS, McGuffin LJ, Buxton BF and Jones DT in 2004 in the Journal of Molecular Biology with the title "Prediction and functional analysis of native disorder in proteins from the three kingdoms of life".<ref>http://bioinf.cs.ucl.ac.uk/disopred/</ref>
The method is based on a neural network which was trained with the SVMlight support vector machine package. DISOPRED first uses PSI-BLAST with a filtered sequence database and uses the position-specific scoring matrix at the final iteration to generate inputs for DISOPRED.<ref>http://cms.cs.ucl.ac.uk/typo3/fileadmin/bioinf/Disopred/disopred_help.html</ref>

Usage


Results

DISOPRED predictions for a false positive rate threshold of 5% are shown in the table below. Asterisks (*) represent disorder predictions and dots (.) predictions of order. The estimated confidence gives a rough indication of the probability that each residue is disordered.
The signal peptide is marked as a disordered region with a very high confidence which is quite interesting. In addition, a part of a beta sheet was predicted as a disordered regions which may indicate that the prediction might be wrong at least in this part.

conf 999999999988777630000000000000000000000000000000000000000000
pred ****************............................................
AA MEFSSPSREECPKPLSRVSIMAGSLTGLLLLQAVSWASGARPCIPKSFGYSSVVCVCNAT
conf 000000000000000000000001456676677776543210000000000000000000
pred .........................************.......................
AA YCDSFDPPTFPALGTFSRYESTRSGRRMELSMGPIQANHTGTGLLLTLQPEQKFQKVKGF
conf 000000000000000000000000000000000000000000000000000000000000
pred ............................................................
AA GGAMTDAAALNILALSPPAQNLLLKSYFSEEGIGYNIIRVPMASCDFSIRTYTYADTPDD
conf 000000000000000000000000000000000000000000000000000021000000
pred ............................................................
AA FQLHNFSLPEEDTKLKIPLIHRALQLAQRPVSLLASPWTSPTWLKTNGAVNGKGSLKGQP
conf 000000000000000000000000000000000000000000000000000000000000
pred ............................................................
AA GDIYHQTWARYFVKFLDAYAEHKLQFWAVTAENEPSAGLLSGYPFQCLGFTPEHQRDFIA
conf 000000000000000000000000000000000000000000000000000000000000
pred ............................................................
AA RDLGPTLANSTHHNVRLLMLDDQRLLLPHWAKVVLTDPEAAKYVHGIAVHWYLDFLAPAK
conf 000000000000000000000000000000000000000000000000000000000000
pred ............................................................
AA ATLGETHRLFPNTMLFASEACVGSKFWEQSVRLGSWDRGMQYSHSIITNLLYHVVGWTDW
conf 000000000000000000000000000000000000000000000000000000022331
pred ............................................................
AA NLALNPEGGPNWVRNFVDSPIIVDITKDTFYKQPMFYHLGHFSKFIPEGSQRVGLVASQK
conf 00000000000000000000000000000000000000000000000000000004
pred ........................................................
AA NDLDAVALMHPDGSAVVVVLNRSSKDVPLTIKDPAVGFLETISPGYSIHTYLWRRQ

POODLE

POODLE stands for Prediction Of Order and Disorder by machine LEarning. It is by S. Hirose, K. Shimizu, N. Inoue, S. Kanai and T. Noguchi and was published in 2008 in CASP8 Proceedings as "Disordered region prediction by integrating POODLE series".

It consists of three predictions:

  • short disorder regions prediction (POODLE-S: short disorder regions, missing region in X ray structure or high B-factor region)
  • long disorder regions prediction (POODLE-L: mainly longer than 40 consecutive amino acids)
  • unfolded protein prediction

POODLE-I is based on a work-flow approach. POODLE uses a machine learning approach and only needs the amino acid sequence for prediction.<ref>http://mbs.cbrc.jp/poodle/help.html</ref>

Usage


POODLE-I

Figure 6: Result of POODLE-I (POODLE series only)

The amino acids predicted as disordered are listed in the table below. Furthermore a visualization of the disorder probability of the amino acids is shown in Figure 6 to the right. Two regions have been predicted as disordered: the first region once again is part of the signal peptide and the second region is part of a beta sheet.

no. AA ORD/DIS Prob.
1 M D 0.851
2 E D 0.813
3 F D 0.781
4 S D 0.753
5 S D 0.757
6 P D 0.813
7 S D 0.777
8 R D 0.743
9 E D 0.699
10 E D 0.707
11 C D 0.711
12 P D 0.656
13 K D 0.633
14 P D 0.609
15 L D 0.581
16 S D 0.553
17 R D 0.535
18 V D 0.51
...
95 I D 0.58
96 Q D 0.812
97 A D 0.892
98 N D 0.744
99 H D 0.551

POODLE-S

The residues predicted as disordered are listed in the table below, where D indicates a disordered region and 0 and ordered region. In the figures (Figure 7 and 8) to the right, the different disorder probabilities are illustrated. POODLE-S predicts a part of the signal peptide as disordered as well. The prediction is based on looking for regions that are missing in the X-ray structure. Based on a high B-factor region some disordered regions are found in the rear part of the protein. The first one is located in no defined secondary structure element in Uniprot, the second one is part of a helix.

Figure 7: Result of POODLE-S (based on missing residues)
Figure 8: Result of POODLE-S (based on B-factor region)
no. AA  ORD/DIS/xray  Prob./xray  ORD/DIS/B-factor  Prob./B-factor 
1 M D 0.764 D 1
2 E D 0.707 D 0.901
3 F D 0.731 D 0.882
4 S D 0.759 D 0.875
5 S D 0.764 D 0.847
6 P D 0.75 D 0.825
7 S D 0.782 D 0.811
8 R D 0.751 D 0.791
9 E D 0.707 D 0.711
10 E D 0.692 D 0.661
11 C D 0.679 D 0.568
12 P D 0.66 D 0.596
13 K D 0.633 D 0.557
14 P D 0.607 D 0.517
15 L D 0.588 O 0.493
16 S D 0.55 O 0.489
17 R D 0.523 O 0.424
18 V D 0.534 O 0.391
...
100 T O 0.368 D 0.571
101 G O 0.29 D 0.54
102 T O 0.243 D 0.555
...
360 K D 0.527 0 0.133
361 A D 0.534 0 0.138
362 T D 0.577 0 0.162
363 L D 0.594 0 0.162
364 G D 0.513 0 0.163

POODLE-L

Figure 9: Result of POODLE-L

POODLE-L did not predict any disordered regions (cf. Figure 9), which indicates, that there are no disordered regions with a length of at least 40 residues in the protein.

IUPred

IUPred is a method established by Zsuzsanna Dosztányi, Veronika Csizmók, Péter Tompa and István Simon which was published 2005 in Bioinformatics with the title "IUPred: web server for the prediction of intrinsically unstructured regions of proteins based on estimated energy content".

The idea of IUPred is to estimate the ability of polypeptides to form stabilizing contacts. It depends on the surrounding sequence and its chemical properties. Intrinsically unstructured regions (IUs) cannot form such contacts. With a 20 by 20 energy predictor matrix the energies can be calculated and with that the probability of IUs.<ref>http://iupred.enzim.hu/Theory.html</ref>

For IUPred one of the three following options can be chosen<ref>http://iupred.enzim.hu/Help.html</ref>:

  • long disorder: predicts context-independent global disorder with at least 30 consecutive residues of predicted disorder (neighbourhood of 100 residues is considered)
  • short disorder: predicts short, probably context-dependent, disordered regions, e.g. missing residues in the X-ray structure (neighbourhood of 25 residues is considered)
  • structured domains: predicts putative structured domains

Usage


IUPred: short disordered regions

The results of IUPred for short disordered regions are listed in the table below and illustrated in Figure 10 to the right. The tool predicts disordered regions with a high probability at the beginning of the sequence. This may be explained by the presence of a signal peptide at this position. It is hard to say, if the other regions are disordered as well, because the probabilities are not really high. A disordered region of a length of two residues is really short and therefore not very probable. At the end there is no secondary structure given at Uniprot, which indicates that a disordered region could be present.

Figure 10: Diagram of IUPred for short disorders
Position  Residue  Disorder Tendency
1 M 0.9753
2 E 0.9369
3 F 0.9280
4 S 0.9009
5 S 0.8857
6 P 0.7869
7 S 0.7418
8 R 0.7034
9 E 0.5992
10 E 0.5549
...
85 G 0.5126
86 R 0.4458
87 R 0.5412
88 M 0.5173
89 E 0.5173
90 L 0.5992
91 S 0.5846
92 M 0.5900
93 G 0.5900
94 P 0.5900
95 I 0.5374
...
103 G 0.5173
104 L 0.5084
...
533 W 0.5514
534 R 0.5992
535 R 0.6124
536 Q 0.6474

IUPred: long disordered regions

The results of IUPred for long disordered regions are listed in the table below and are illustrated in Figure 11 to the right. It predicts some very short disordered regions with a very low probability. Thus it is not very probable, that there are really disordered regions.

Figure 11: Diagram of IUPred for long disorders
Position  Residue  Disorder Tendency
1 M 0.4864
2 E 0.5017
3 F 0.5707
...
87 R 0.4979
88 M 0.4979
89 E 0.4979
90 L 0.6136
91 S 0.5901
92 M 0.5992
93 G 0.5017
...
229 A 0.5055
230 V 0.5055
231 N 0.5211
...
235 S 0.5055
236 L 0.5139

IUPred: structured regions

Figure 12: Diagram of IUPred for structured regions

IUPred for structured regions predicts globular domains for position 4 to 536. So all except the first four residues are predicted as structured. An illustration of the prediction can be seen in Figure 12.

META-Disorder

META-Disorder was published in 2009 by Avner Schlessinger, Marco Punta, Guy Yachdav, Laszlo Kajan and Burkhard Rost in PLoS ONE.<ref>http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0004433</ref> It is a method which combines NORSnet<ref>https://www.rostlab.org/owiki/index.php/Norsnet</ref>, PROFbval<ref>https://rostlab.org/owiki/index.php/Profbval</ref> and Ucon<ref>https://www.rostlab.org/owiki/index.php/UCON</ref> to predict disordered regions. As input only the amino acid sequence is needed.<ref>https://www.rostlab.org/owiki/index.php/Metadisorder</ref>

Usage

Results

The results of META-Disorder are shown in the table below and in Figure 13 to the right. The tool predicts a disordered region ranging from amino acid 1 to amino acid 6, which is part of the signal peptide. It is hard to say whether it is really a disordered region, as the beginning of a sequence is often mispredicted as a disordered region.


Figure 13: Result of PredictProtein for disordered regions.
position residue score
1 M 0.636
2 E 0.596
3 F 0.596
4 S 0.586
5 S 0.561
6 P 0.551


Discussion

Glucocerebrosidase (Uniprot-ID: P04062) is not listed in the Database of Protein Disorder Disport <ref>http://www.disprot.org/index.php</ref>, which indicates that there are no disordered regions in the protein. The different prediction methods predict probable disordered regions in different parts of the protein which are most of the time located in a certain structural element and therefore may be wrong. Furthermore the different predictions have no real overlaps apart from the regions in the signal peptide, where almost all tools show at least a few disordered residues. But the latter is not present in the mature protein and therefore describes no disordered region in the final structure.

Prediction of coiled coils

Our additional method to test another structural feature was the test of coiled coils. A coiled coil is a motif, which is often part of regulation of gene expression. Several alpha helices together build this motif. They usually contain a heptad repeat, which is a pattern hxxhcxc, where h are hydrophobic residues and c charged amino acid residues. This structure allows the special folding of coiled coils.<ref>http://en.wikipedia.org/wiki/Coiled_coil</ref> In our case we do not expect any coiled coils because Glucocerebrosidase is not a transcription factor and also does not show such helix structures. We used three tools to verify this assumption.

COILS

Figure 14: Result of COILS for Glucocerebrosidase.

COILS is a method which compares the given sequence to a database of known coiled coils. With that it calculates a similarity score and so gets the probability of a coiled coil motif.<ref>Lupas, A., Van Dyke, M., and Stock, J. (1991), Predicting Coled Coils from Protein Sequences, Science 252:1162-1164</ref> The result can be seen in Figure 14 to the right.

Usage


MultiCoil

Figure 15: Result of MultiCoil for Glucocerebrosidase

MultiCoil is based on Parcoil, which predicts coiled coils by pairwise residue correlations.<ref>Bonnie Berger, David B. Wilson, Ethan Wolf, Theodore Tonchev, Mari Milla, and Peter S. Kim, "Predicting Coiled Coils by Use of Pairwise Residue Correlations", Proceedings of the National Academy of Science USA, vol 92, aug 1995, pp. 8259-8263.</ref>. It is a better version for two- and three-stranded coiled coils.<ref>Ethan Wolf, Peter S. Kim, and Bonnie Berger, "MultiCoil: A Program for Predicting Two- and Three-Stranded Coiled Coils", Protein Science 6:1179-1189. June 1997.</ref>. The result is illustrated in Figure 15.

Usage


Parcoil2

Figure 16: Result of Parcoil2 for Glucocerebrosidase

Parcoil2 works like Parcoil or Multicoil, but is an approved version of 2006, which uses again pairwise residue probabilities and an updated database.<ref>A.V. McDonnell, T. Jiang, A.E. Keating, B. Berger, "Paircoil2: Improved prediction of coiled coils from sequence", Bioinformatics Vol. 22(3) (2006)</ref> The result can be seen in Figure 16 to the right.

Usage


Discussion

None of the three tools, COILS, MultiCoil and Parcoil2 predicted any coiled coil motif. This is not surprising as Glucocerebrosidase does not contain any coiled coils.

Prediction of transmembrane alpha-helices and signal peptides

General

Transmembrane topology

The topology of a membrane protein is characterized by the number of membrane spanning segments in the protein. The transmembrane regions of the protein are hydrophobic and have a length of aproximately 15-30 residues which is enough to cross the lipid bilayer of the membrane once. The different transmembrane regions are connected by hydrophilic loops which are located outside the membrane. These attributes can be used to predict the transmembrane topology of a protein.
Predictors: TMHMM, OCTOPUS

Signal peptides

Signal Peptides are located at the N-terminus of a protein sequence and direct the transport of a protein to its correct location. Signal Sequences have a typical size of 20-30 residues and have a tripartite structure: a hydrophobic core region (h-region) which is flanked by a basic n-region and a slightly polar c-region. The sequence variation among signal sequences affects ER targeting, translocation and signal peptidase cleavage. <ref>Hegde R.S. and Bernstein H.D. (2006) The surprising complexity of signal sequences. Trends Biochem Sci 31(10), 563-71</ref>
Predictors: SignalP, TargetP

Combined transmembrane and signal peptide prediction

The high similarity between the hydrophobic region of a transmembrane helix and the one of a signal peptide leads to cross-predictions when conventional transmembrane topology and signal peptide predictors as TMHMM and SignalP are used. Predictors which are based on submodels for both make less errors coming from cross-predictions and help to discriminate against false positives. Furthermore, a predicted signal peptide indicates that the N-terminus of the protein is non-cytoplasmic and is therefore helpful to assign the orientation of the protein. <ref>Käll L. Krogh A, & Sonnhammer, E. L. (2007) Advantages of combined tranasmembrane topology and signal peptide prediction - the Phobius web server. Nucleic Acids Res., Vol. 35, Web server issue, S.429-32</ref>
Predictors: Phobius, Polyphobius, SPOCTOPUS.

Topology of Glucocerebrosidase

Glucocerebrosidase is located in the lysosome and does not contain any transmembrane regions. Residues 1 to 39 are part of a signal peptide <ref>http://www.uniprot.org/uniprot/P04062</ref>. Therefore the prediction tools should place the protein in the non-cytoplasm and not find any transmembrane helices, but instead a signal peptide.

TMHMM

TMHMM is a method to predict transmembrane topology of membrane-spanning proteins. It is based on a hidden Markov model with an architecture of 7 types of states (helix core, helic caps on both sides, one loop on the cytoplasmic side, two loops on the non-cytoplasmic side and a globular domain in the middle of each loop) which correspond to the biological system. The method was established by Sonnhammer et al. in 1998 <ref>Sonnhammer EL, von Heijne G, Krogh A. A hidden Markov model for predicting transmembrane helices in protein sequences. Proc Int Conf Intell Syst Mol Biol. 1998;6:175–182.</ref>.

Usage


Results

Figure 17: TMHMM posterior probabilities for glucocerebrosidase

sp|P04062|GLCM_HUMAN Length: 536
sp|P04062|GLCM_HUMAN Number of predicted TMHs: 0
sp|P04062|GLCM_HUMAN Exp number of AAs in TMHs: 1.77867
sp|P04062|GLCM_HUMAN Exp number, first 60 AAs: 1.62607
sp|P04062|GLCM_HUMAN Total prob of N-in: 0.06840
sp|P04062|GLCM_HUMAN TMHMM2.0 outside 1 536

TMHMM predicts no transmembrane segments (cf. output and Figure 17) for the sequence of glucocerebrosidase which is correct.



Phobius and PolyPhobius

Phobius is based on a hidden Markov model which contains submodels for both transmembrane helices and signal peptides and therefore obtains a better discrimination between the two segments than predictors for only one of them. This method was presented in 2004 by Käll et al. <ref>Käll L., et al. A combined transmembrane topology and signal peptide prediction method. J. Mol. Biol. 2004;338:1027–1036.</ref> Polyphobius is a prediction server that additionally uses an algorithm to include homology information. The performance of transmembrane topology and signal peptide prediction is increased by incorporating extra support from homolougs. <ref>Käll L., et al. An HMM posterior decoder for sequence feature prediction that includes homology information Bioinformatics, 21 (Suppl 1):i251-i257, June 2005.</ref>

Usage - Phobius

Results - Phobius

Figure 18: Phobius posterior probabilities for glucocerebrosidase

ID sp|P04062|GLCM_HUMAN
FT SIGNAL 1 39
FT REGION 1 19 N-REGION.
FT REGION 20 31 H-REGION.
FT REGION 32 39 C-REGION.
FT TOPO_DOM 40 536 NON CYTOPLASMIC.

Phobius predicts a signal peptide ranging from amino acid 1 to amino acid 39 of the sequence of glucocerebrosidase. This goes along with the information given on Uniprot <ref>http://www.uniprot.org/uniprot/P04062</ref> that the protein has a 39 residue signal sequence. The presence of a signal peptide explains the differences between the sequences of sp|P04062|GLCM_HUMAN and its corresponding PDB structure 1OGS: the sequence of 1OGS has 39 amino acids less than the sequence of sp|P04062|GLCM_HUMAN as the signal peptide is missing in the mature structure. The prediction, that the protein is non-cytoplasmic is true, as glucocerebrosidase is located in the lysosome.

Usage - PolyPhobius

Results - PolyPhobius

Figure 19: Phobius posterior probabilities for glucocerebrosidase

ID sp|P04062|GLCM_HUMAN
FT SIGNAL 1 39
FT REGION 1 23 N-REGION.
FT REGION 24 34 H-REGION.
FT REGION 35 39 C-REGION.
FT TOPO_DOM 40 536 NON CYTOPLASMIC.

Polyphobius returns the same predictions as Phobius: a 39 residue signal sequence with slightly different region ranges.

OCTOPUS and SPOCTOPUS

OCTOPUS (obtainer of correct topologies for uncharacterized sequences) was developed in 2007 by Viklund et al. The method combines hidden markov models and artificial neural networks. Furthermore, OCTOPUS is the first method that integrates the modelation of reentrant-, membrane dip-, and TM hairpin regions. <ref>Viklund, H., and A. Elofsson. 2008. OCTOPUS: improving topology prediction by two-track ANN-based preference scores and an extended topological grammar. Bioinformatics 24:1662-1668.</ref> SPOCTOPUS is an extension of the OCTOPUS algorithm which additionally predicts signal peptides for reducing predictions of transmembrane regions as signal peptides and the other way round. The method was first mentioned by Viklund et al. in 2008. <ref>Viklund, H., et al. 2008. SPOCTOPUS: a combined predictor of signal peptides and membrane protein topology Bioinformatics 24:2928-2929.</ref>

Usage

OCTOPUS - Results

Results of the topology prediction with Octopus

OCTOPUS predicts a transmembrane segment ranging from amino acid 16 to amino acid 36. The region from amino acid 1 to 15 is predicted to be cytoplasmic and the remaining amino acids (from amino acid 37) are indicated to be non cytoplasmic. This is an example for a misclassification between transmembrane segments and signal peptides when a prediction tool for only transmembrane segments is used.

SPOCTOPUS - Results

Results of the topology prediction with Spoctopus

SPOCTOPUS correctly identifies the regions from amino acid 1 to amino acid 39 as a signal peptide. The combination of a signal peptide and transmembrane segment prediction helps to eliminate the misclassification made by a transmembrane segment predictor (cf. results of OCTOPUS).

SignalP

SignalP uses a combination of several artificial neural networks and hidden Markov models to predict the presence and location of signal peptide cleavage sites of three different organism groups: eukaryotes, Gram-negative and Gram-positive bacteria. The current version 3.0 was published in 2004 by Bendtsen et al. <ref>Bendtsen JD, Nielsen H, von Heijne G, Brunak S. Improved prediction of signal peptides: SignalP 3.0. J Mol Biol. 2004;340:783–795.</ref>

Usage

Results

Figure 20: Prediction of signal peptide with SignalP 3.0 (image taken from the web server<ref>http://www.cbs.dtu.dk/services/SignalP/</ref>)

>sp|P04062|GLCM_HUMAN
Prediction: Signal peptide
Signal peptide probability: 0.516
Signal anchor probability: 0.001
Max cleavage site probability: 0.423 between pos. 39 and 40

SignalP predicts the signal peptide of glucocerebrosidase correctly. Even the cleavage site is located correctly between amino acid 39 and amino acid 40. The location of the different regions of the signal peptide can be seen in the illustration to the right.

TargetP

TargetP, a neural network-based tool, predicts the subcellular location of eukaryotic proteins based on the predicted presence of N-terminal presequences like mitochondrial targeting peptides, chloroplast trasit peptides and pathway signal peptides. For the latter two potential cleavage sites can be predicted as TargetP uses ChloroP and SignalP. This method, which only uses N-terminal sequence information, was published in 2000 by Emanuelsson et al. <ref>Emanuelsson O., et al. Predicting Subcellular Localization of Proteins Based on their N-terminal Amino Acid Sequence. J.Mol.Biol. (2000) 300, 1005-1016.</ref>

Usage

Results

Name Len mTP SP other Loc RC TPlen
sp_P04062_GLCM_HUMAN 536 0.091 0.364 0.612 _ 4 -
cutoff 0.000 0.000 0.000

TargetP does not predict the signal peptide of glucocerebrosidase. Although the score for a signal peptide is much higher than the one for a mitochondrial targeting peptide, it locates the protein with a very low reliability class rather at any other location than in chloroplast, mitochondrion or the secretory pathway.

Further Examples of Use

As the protein glucocerebrosidase does not contain any transmembrane segments, the application of the different tools mentioned above will further be demonstrated with some other proteins: Bacteriorhodopsin (BACR_HALSA)<ref>http://www.uniprot.org/uniprot/P02945</ref>, Retinol-binding protein 4 (RET4_HUMAN)<ref>http://www.uniprot.org/uniprot/P02753</ref>, Insulin-like peptide INSL5 (INSL5_HUMAN)<ref>http://www.uniprot.org/uniprot/Q9Y5Q6</ref>, Lysosome-associated membrane glycoprotein 1 (LAMP1_HUMAN)<ref>http://www.uniprot.org/uniprot/P11279</ref> and Amyloid beta A4 protein (A4_HUMAN)<ref>http://www.uniprot.org/uniprot/P05067</ref>.

Topology

The table below indicates the correct number of transmembrane segments, the localization of the N-terminus and presence or absence of a signal peptide in the five proteins mentioned above. A more detailed description with the exact positions of the membrane segments and the cleavage sites can be seen in the corresponding Uniprot entries.

Protein # Transmembrane Segments Signal Peptide Localization of N-terminus
BACR_HALSA 7 0 extracellular
RET4_HUMAN 0 1 extracellular
INSL5_HUMAN 0 1 extracellular
LAMP1_HUMAN 1 1 lumenal
A4_HUMAN 1 1 extracellular


TMHMM

The topology predictions of TMHMM for the different proteins are illustrated in the graphic below. TMHMM predicts 6 different transmembrane segments for BACR_HASLA, but indicates, that a signal peptide is possible. According to Uniprot, BACR_HASLA consists of 7 different transmembrane segments and no signal peptide. The last transmembrane helix, ranging from amino acid 217 to 236 was not predicted by TMHMM. The prediction that RET4_HUMAN and INSL5_Human do not contain any transmembrane segments goes along with the corresponding Uniprot entries. TMHMM finds 2 transmembrane segments in LAMP1_HUMAN. The first transmembrane segment is in reality a signal peptide, as indicated in the Uniprot entry and is therefore another example for the confusion of transmembrane segments and signal peptides. The single transmembrane helix of A4_HUMAN was predicted correctly.

Membrane topologies predicted with TMHMM for different proteins

Phobius and PolyPhobius

The results of Phobius and PolyPhobius are identical. There are only minor differences in the length of the different regions of a signal peptide. The topology predictions of both methods for the different proteins are illustrated in the graphic below. The predictions for all of the proteins were made correctly by Phobius and PolyPhobius. Sometimes the transmembrane segments are shifted some amino acid positions to the right or to the left. The ortientation of the proteins, the presence of signal peptides and the overall topology of the proteins go along with the corresponding Uniprot entries.

Membrane topologies predicted with Phobius and PolyPhobius for different proteins

OCTOPUS

OCTOPUS makes a lot of false predictions resulting from a confusion between signal peptides and transmembrane regions. Each signal peptide was predicted as either transmembrane segment or reentrant/dip region. BACR_HASLA, which does not have a signal peptide, was predicted correctly.

Membrane topologies predicted with OCTOPUS

SPOCTOPUS

SPOCTOPUS, which expends the OCTOPUS algorithm with a signal peptide prediction, predicts the overall topology, orientation of the protein and presenence of signal peptides in each case correctly. This is a very good example, that a combined signal peptide and transmembrane prediction is more reliable and makes less errors than a single transmembrane prediction.

Membrane topologies predicted with SPOCTOPUS

SignalP

If a signal peptide is present in the protein, SignalP predicts it with a very high confidence, both with the neural networks and the hidden Markov models. The presence of a signal peptide is predicted correctly for the proteins RET4_HUMAN, INSL5_HUMAN, LAMP1_HUMAN and A4_HUMAN. The results of the neural network and the hidden Markov models differed for the protein BACT_HALSA which does not have a signal peptide. The S-Score of the neural networks indicates that there is a signal peptide and that the corresponding cleavage site is between position 38 and 39. In contrast, the hidden Marcov models predict a signal anchor.


Protein Signal Anchor Prob. Signal Peptide Prob. Cleavage Site Prob. Cleavage Site
BACR_HALSA 0.86 0.02 0.00 15-16
RET4_HUMAN 0.00 1.00 0.98 18-19
INSL5_HUMAN 0.00 1.00 0.91 22-23
LAMP1_HUMAN 0.00 1.00 0.85 28-29
A4_HUMAN 0.00 1.00 0.99 17-18


TargetP

TargetP predicts present signal peptides (for proteins RET4_HUMAN, INSL5_HUMAN, LAMP1_HUMAN and A4_HUMAN) correctly and with a very high (1-2) reliability class in each case. The method predicts a signal peptide for BACR_HALSA as well, but the reliablility class is in this case very low (4), which indicates that the prediction is not very safe.

Name Len mTP SP other Loc RC TPlen
sp_P02945_BACR_HALSA 262 0.019 0.897 0.562 S 4 116
sp_P02753_RET4_HUMAN 201 0.242 0.928 0.020 S 2 18
sp_Q9Y5Q6_INSL5_HUMA 135 0.074 0.899 0.037 S 1 22
sp_P11279_LAMP1_HUMA 417 0.043 0.953 0.017 S 1 28
sp_P05067_A4_HUMAN 770 0.035 0.937 0.084 S 1 17
cutoff 0.000 0.000 0.000

Discussion

The application of the different tools for transmembrane region and signal peptide prediction to a variety of proteins shows that predictors which combine both elements are more reliable and make less errors coming from misclassifications than single predictors. Therefore, if possible, a predictor for both, transmembrane segment and signal peptide like SPOCTOPUS, Polyphobius or Phobius, should be used.

Prediction of GO terms

General

Figure 21: GO-tree with a reference annotation and a prediction.)

The Gene Ontology Consortium tries to unify the terminology of gene and gene product attributes across all species to decrease non-consistent descriptions in different databases. It therefore has developed three different ontologies: cellular component, molecular function and biological process. <ref>http://www.geneontology.org/GO.doc.shtml</ref> For each prediction, precision (tp/(tp+fp)) and recall (tp/(tp+fn)) are calculated by comparing the paths to the root in the GO-tree of the predicted and the reference terms. Comparing the paths to the root allows one to not count a predicted child of a reference GO-term as false positive, but instead as true positive and the resulting values are therefore more reliable than the ones obtained by just comparing the predicted with thre reference GO-terms. Example: In Figure 21, one can see an exemplary GO tree with a reference annotation and a prediction. If one would just compare the GO terms without taking the whole tree into account, this would result in one false positive (term 2), one false negative (term 3) and zero true positives. In contrast, if one uses the paths to the root of the tree of both reference annotation and prediction, one gets three true positives (domain, term 1, term 2) and one false negative (term 3).
For the GO-tree, the transitive closure<ref>http://www.geneontology.org/scratch/transitive_closure/go_transitive_closure.links</ref> of june 2008 is used. This file is not up to date and therefore does not contain each GO-term, but still helps to give an impression of how good the predictions are.


Annotation of Glucocerebrosidase

The following GO-term annotations are taken from Uniprot<ref>http://www.ebi.ac.uk/QuickGO/GProtein?ac=P04062</ref> and are used as reference for a comparison with the different prediction tools.


Accession Term Ontology
GO:0005975 carbohydrate metabolic process biological process
GO:0008219 cell death biological process
GO:0006629 lipid metabolic process biological process
GO:0007040 lysosome organization biological process
GO:0006665 sphingolipid metabolic process biological process
GO:0008152 metabolic process biological process
GO:0005765 lysosomal membrane cellular component
GO:0016020 membrane cellular component
GO:0005764 lysosome cellular component
GO:0043169 cation binding molecular function
GO:0003824 catalytic activity molecular function
GO:0004348 glucosylceramidase activity molecular function
GO:0016798 hydrolase activity, acting on glycosyl bonds molecular function
GO:0016787 hydrolase activity molecular function
GO:0005515 protein binding molecular function

GOPET

GOPET, a Gene Ontology term Prediction and Evaluation Tool, uses homology searches and Support Vector Machines to predict the molecular function GO-terms for sequences of any organism. It was made public in 2006 by Vinayagam et al. <ref>Vinayagam A., et al. GOPET: A tool for automated predictions of Gene Ontology terms. BMC Bioinformatics. 2006; 7: 161.</ref>

Usage

Results

GOPET predicts 3 different GO-terms for glucocerebrosidase. The annotations made by GOPET are correct: each predicted GO-term is listed in the corresponding Uniprot entry of glucocerebrosidase. GOPET does not predict two molecular function GO-terms: protein binding and cation binding. In total, GOPET achieves a recall of 0.67 and a precision of 1.00 if only the GO-terms corresponding to molecular function are taken into account.


GOid Aspect Confidence GO term
GO-ID:0016787 F 98% hydrolase activity
GO-ID:0004348 F 97% glucosylceramidase activity
GO-ID:0016798 F 97% hydrolase activity acting on glycosyl bonds

Pfam

The database Pfam contains protein families and domains and is based on hidden Markov models. It consists of two parts: Pfam-A is curated and therefore contains high quality data whereas Pfam-B is generated automatically. Pfam was presented by Sonnhammer et al. in 1997. <ref>Sonnhammer E., et al., Pfam: A Comprehensive Database of Protein Domain Families Based on Seed Alignments. PROTEINS: Structure, Function, and Genetics 28:405-420(1997)</ref>

Usage

Results
Pfam assigns glucocerebrosidase to the "O-Glycosyl hydrolase family 30". To retrieve the GO-annotations for this family, the pfam2go file <ref>http://www.geneontology.org/external2go/pfam2go</ref> of the Gene Ontology website had to be used. Each GO-term listed in Pfam is listed in the corresponding Uniprot site as well. This results in a precision of 1.00 and a recall of 0.65.


Accession Term Ontology
GO:0004348 glucosylceramidase activity molecular function
GO:0006665 sphingolipid metabolic process biological process
GO:0007040 lysosome organization biological process
GO:0005764 lysosome cellular component

ProtFun 2.2

ProtFun is an ab initio prediction server for protein function which is based on sequence derived protein features as predicted post translational modifications, protein sorting signals and phisical/chemical properties that have been calculated from the amino acid composition of the input sequence. <ref>Jehnsen L., et al. Ab initio prediction of human orphan protein function from post-translational modifications and localization features. J. Mol. Biol., 319:1257-1265, 2002</ref>

Usage

Results
ProtFun assigns glucocerebrosidase to immune response (GO:0006955) which is not tue. Therefore the results of ProtFun have a recall and a precission of 0.00.

 Functional category                  Prob     Odds
 Amino_acid_biosynthesis              0.035    1.593
 Biosynthesis_of_cofactors            0.182    2.528
 Cell_envelope                     => 0.504    8.262
 Cellular_processes                   0.032    0.438
 Central_intermediary_metabolism      0.382    6.063
 Energy_metabolism                    0.067    0.740
 Fatty_acid_metabolism                0.027    2.088
 Purines_and_pyrimidines              0.538    2.213
 Regulatory_functions                 0.031    0.191
 Replication_and_transcription        0.126    0.471
 Translation                          0.082    1.863
 Transport_and_binding                0.560    1.365
 Enzyme/nonenzyme                     Prob     Odds
 Enzyme                            => 0.773    2.698
 Nonenzyme                            0.227    0.318
 Enzyme class                         Prob     Odds
 Oxidoreductase (EC 1.-.-.-)          0.083    0.399
 Transferase    (EC 2.-.-.-)          0.228    0.660
 Hydrolase      (EC 3.-.-.-)          0.272    0.859
 Lyase          (EC 4.-.-.-)          0.045    0.961
 Isomerase      (EC 5.-.-.-)          0.011    0.345
 Ligase         (EC 6.-.-.-)          0.017    0.332
 Gene Ontology category               Prob     Odds
 Signal_transducer                    0.054    0.251
 Receptor                             0.027    0.158
 Hormone                              0.001    0.206
 Structural_protein                   0.002    0.087
 Transporter                          0.024    0.222
 Ion_channel                          0.018    0.307
 Voltage-gated_ion_channel            0.004    0.195
 Cation_channel                       0.012    0.268
 Transcription                        0.070    0.550
 Transcription_regulation             0.030    0.237
 Stress_response                      0.085    0.962
 Immune_response                   => 0.153    1.804
 Growth_factor                        0.005    0.376
 Metal_ion_transport                  0.009    0.020

Further Examples of Use

The different prediction tools have also been applied to the additional proteins already used in the transmembrane and signal peptide prediction section. The reference annotations and the predictions are listed in the following pdf-File: File:GO term prediction for several different proteins.pdf. The resulting precision and recall values are listed in the table below. This time, recall and precision of GOPet have been calculated by a comparison to all annotated GO-terms of the corresponding proteins, not only to the ones being a molecular function.
The Pfam database search is the only method which returns only correct GO-terms for each of the 5 proteins, although the coverage is only high for INSL5. GOPet, as well as ProtFun do not predict any correct GO-terms for LAMP1, which is quite interesting, as the precision for the other proteins is quite high (apart from A4_HUMAN: in this case, the predicted gene ontology category of ProtFun could not be mapped to a certain GO-term, and therefore neither precision or recall could be calculated). The low recall values of ProtFun and GOPet may have reasons: GOPet only predicts GO-terms of the ontology "molecular function" and ProtFun only classifies the protein to one gene ontology category.

Bacteriorhodopsin Retinol-binding protein 4 Insulin-like peptide INSL5 Lysosome-associated
membrane glycoprotein 1
Amyloid beta A4 protein
GOPet
Precision 0.62 0.62 1.00 0.00 0.70
Recall 0.21 0.04 0.57 0.00 0.05
Pfam
Precision 1.00 1.00 1.00 1.00 1.00
Recall 0.26 0.01 0.71 0.07 0.03
ProtFun 2.2
Precision 1.00 1.00 1.00 0.00 -
Recall 0.02 0.02 0.57 0.00 -


PDF-File containing the results of the different methods and the correct GO-term annotations: File:GO term prediction for several different proteins.pdf

Discussion

GOPET predicts GO terms corresponding to molecular function very well. The fact, that the recall values are quite low for the further examples of use is due to the fact, that all GO-terms and not only the ones for molecular function have been used as reference. The GO-terms retrieved with Pfam are very accurate: in each prediction, only correct GO-terms have been predicted which results in a precision of 1.0. ProtFun2.2 differs from the other two prediction methods, as it only assigns the protein to one gene ontology category. This resuls in very low recall values and it happend several times, that the prediction was not correct.

None of the three different methods was able to predict each of the annotated GO-terms. But this is not surprising, as the Gene Ontology consists of a very large number of terms. GOPET and Pfam are therefore a good solution to get a broad overview of the different functions and localizations of your protein of interest.

References

<references />