Sequence-based predictions TSD
- 1 Secondary structure
- 2 Disorder
- 3 Transmembrane helices
- 4 Signal peptides
- 5 GO terms
- 6 Pfam
- 7 References
Proteins: Ribonuclease inhibitor P10775 , Divalent-cation tolerance protein CutA (CutA) Q9X0E6 , Serine/threonine-protein phosphatase 2B catalytic subunit alpha isoform (CAM-PRP catalytic subunit) Q08209.
Ribonuclease inhibitor and CutA are located in the cytoplasm whereas the CAM-PRP catalytic subunit is located in the nucleus.
DSSP and handling of differing sequences
DSSP builds upon 3D structures, therefore a PDB entry has to be selected for every given Uniprot entry. The chosen mapping is 2bnh for P10775, 1kr4 for Q9X0E6, 1aui for Q08209 and 2gjx for P06865. This creates an additional problem. All other resources base their predictions on the Uniprot sequence. The sequence used by DSSP, inferred from the PDB file might be significantly different due to changes of the experimentalists solving the structure. There are automated ways to resolve this, using PDBs new mmCIF files, which provide a residue-level mapping between the atom recored inferred sequence and the SEQRES record sequence. From there one could use SIFTS which provides a residue-level mapping between SEQRES and Uniprot. However both resources, mmCIF mapping and SIFTS, are automated and, while surely developed with great care, looking at the sequence might be considered favorable and will also directly point out interesting parts. Therefore manual alignments were performed and, if applicable, special cases noted in the text.
Comparison of predictions and annotation
In the following the secondary structure prediction of PSIPred and reprof is compared to the 3D-structure based annotation of DSSP. Uniprot annotation was not included, since its source is DSSP <ref name="uniprotsecstruct">http://www.uniprot.org/manual/helix</ref>. To compare the methods a common alphabet had to be established. All outputs were normalized to only contain the residues types H (Helix), E (Beta-strand) or C (Coil). Details on the parsing can be found in the protocol. The figures were created using the Latex package cpssp <ref name="cpssp">http://www.ctan.org/pkg/cpssp</ref>, details on these can also be found in the protocol.
<xr id="fig:ssP06865"/> shows the comparison of assigned secondary structure for the Hex A alpha subunit. Several things are noticable here. Firstly, there are several short helices in the DSSP annotation that neither PSIPred nor reprof predicted. These short regions are not, as one might first assume, less common helix types, but in fact, as a check in the original DSSP output confirms 'normal' (i+4->i) alpha-helices. This issue cannnot be observed for beta-sheets where the length does not show an immediate impact on prediction performance. Comparing PSIPred and reprof it can be easily observed that PSIPred shows very good correlation with the DSSP annotation while reprof often differs from these two. The disagreement is noticeably stronger in the catalytic domain of the protein (beginning at position 167, cf. <xr id="fig:pfam2gjx"/>) and mostly appears in the prediction of beta-sheets instead of alpha-helices. This is surprising but also means that both prediction methods at least correctly predict the central TIM barrel in this domain.
<xr id="fig:ssP10775"/> shows the comparison of assigned secondary structure for the ribonuclease inhibitor. It is apparent that both prediction methods largely agree with the annotation in DSSP, which is not entirely surprising given the peculiar horseshoe like secondary structure of alternating alpha-helices and beta-sheets that hardly leaves place for exchange or removal of a secondary structure element (c.f. <xr id="fig:2bnh"/>). Nonetheless, there are differences present, most noticeably the behaviour of reprof to either elongate helices or shift them to the left or right compared to the DSSP and PSIPred annotations. This leads to the fact that some of the beta strand-turn-alpha helix motifs are mistaken by reprof for something that would be more accurately described as alpha helix-turn-alpha helix motif. Given that these inhibitor are known to very tightly interact with the ribonucleases <ref name="Dickson2010">Dickson,K. and Haigis,M. (2005) Ribonuclease inhibitor: structure and function. Progress in nucleic acid research and, 6603, 1-23.</ref> it would be surprising to see that this does not significantly impair function and is therefore a very important finding.
<xr id="fig:ssQ9X0E6"/> shows the comparison of assigned secondary structure for 'cation tolerance protein' CutA which reinforces the behaviours observed so far. While PSIPred's prediction is very close to the DSSP annotation, reprof does not predict some of the beta-sheets and tends to extend helices. Again this is important to note, since the beta-sheet assembly is thought to be essential for assumption of the trimeric biological assembly <ref name="Savchenko2009">Savchenko,A. and Skarina,T. (2004) X-Ray Crystal Structure of CutA From Thermotoga maritima at 1.4 Å Resolution. Proteins: Structure,, 54, 162-165.</ref>.
Finally, <xr id="fig:ssQ08209"/> shows the comparison of assigned secondary structure for the phosphatase. This protein seems comparably hard to predict, both PSIPred and reprof make several errors. PSIPred does not predict any structural elements where there are none annotated according to DSSP, however some are mistaken for the opposite and serveral ones simply missed. On the other hand reprof correctly predicts some of these, but misses others and in addition again mistakes several long alpha helices for beta-sheets.
In conclusion it can be seen that a very general agreement with the annotation by DSSP can always be achieved at least by PSIPred. Mostly this prediction methods' errors remain low and could be considered minor.
Reprof shows much more detrimental errors, namely mistaking alpha-helices for beta-sheets or elongating and shifting predicted alpha-helices which can lead to a failure of predicting beta sheets. Most of these errors can be shown to be severe in terms of the known function of the particular structural element.
While the last entry hinted at the fact that on some proteins both methods have almost equally large problems, PSIPred was much more convincing overall.
The protein disorder prediction was performed with IUPred for the same proteins as in section 1. The option long was chosen as prediction type as this is most suitable to find any disordered regions in a protein, which are long enough (>30 residues) to have an impact on protein structure.
In the plots in <xr id="tbl:iupred"/> the calculated disorder tendency is displayed for every residue.
All predictions for the proteins express a fluctuating tendency which is overall lower than 0.5.
The prediction of CutA (<xr id="tbl:iupred"/>, bottom left) has the lowest and least fluctuating curve. Ribonuclease inhibitor (<xr id="tbl:iupred"/>, top right) and Hex A alpha subunit (<xr id="tbl:iupred"/>, top left) have very similar profiles. None of these 3 proteins can be assigned a disordered region in accordance with the prediction. This complies with the known resolution of the structures of these proteins: For the ribonuclease inhibitor, the leucine rich repeats leave very little room for disorder and CutA is very small where the only coiled regions seem to only directly connect the ordered regions (c.f. <xr id="fig:1kr4"/>).
The only protein exeeding the 0.5 cutoff is the CAM-PRP catalytic subunit (<xr id="tbl:iupred"/>, bottom right) which shows signs of disorder at the beginning of the sequence and towards the end. The first region is about 10 residues long and the latter begins roughly at residue 425 and spans 100 amino acids. This prediction can be validated by the annotations of CAM-PRP in Disprot<ref name="disprot">Sickmeier,M. et al. (2007) DisProt: the Database of Disordered Proteins. Nucleic acids research, 35, D786-93.</ref>. The assigned disordered regions are 1 - 13, 390 - 414 (CaM-binding domain), 374 - 468, 469 - 486 (Autoinhibitory region) and 487 - 521. All were detected by X-ray crystallography.
The first disordered region was well detected by IUPred. The region starting at position 374 is hinted at by a peak in the prediction, however the signal is too strong to assume a disordered region of the size as annotated in Disprot. Furthermore, the IUPred results hint at the distinction between the disordered region till position 486 and the disorder starting at 487 as the curve expresses a steep rise in this region.
In conclusion, IUPred supplies a very accurate and reliable prediction for the given protein set.
Differing sequences and mapping of gold standard
Since OPM and PDBTM rely on 3D structures to provide a TMH annotation, every Uniprot entry that is to be evaluated needs to be assigned a PDB entry. The mapping chosen is 3pbl for P35462, 1orq for Q9YDF8 and 2d57 for P47863, as well as 2gjx for P06865. As for the secondary structure prediction a problem lies in the mapping needed between the PDB ATOM record and a Uniprot sequence. For this task an automated approach was used: The PDB offers mmCIF files which contain a per residue level mapping of the ATOM record and the SEQRES sequence. In addition SIFTS <ref name="SIFTS">Velankar,S. et al. (2005) E-MSD: an integrated data resource for bioinformatics. Nucleic acids research, 33, D262-5.</ref> provides a residue level mapping between a PDB SEQRES sequence and Uniprot sequences. This allows a transfer of the annotation in OPM and PDBTM onto the Uniprot sequences used.
Comparison of predictions and annotation
Note that in the following PDBTM residue types 'H' and 'L' were considered as transmembrane helices. The figures were created using code from T. Nugent <ref name="drawtmh">http://www.cs.ucl.ac.uk/staff/T.Nugent/code.html</ref>. It should be noted that from the given data one cannot decide which side is extracellular and which intracellular, the distinction is simply due to the way the module works. The consensus only considers the OPM and PDBTM annotations.
Generally it can be seen in all of the following comparisons, that OPM and PDBTM usually agree on the presence of transmembrane helices, but the exact length and residue level position of the helices differs. This is to be expected, given that even provided with a 3D structure, the annotation of a helix is not a trivial task. Lipids or solvents used are too small and agile to be part of the resolved structure and even if they were present, it is a matter of the tool's author to decide at which part of the region building the transition between the membrane inside and extracellular surface the cut in the 2D annotation should be made. Owing to this problem, scores used to assess the performance of transmembrane helix prediction, having been proposed for a long time, do not penalize a prediction that is not 100% aligned with the 3D annotation but qualitatively count a helix as correctly predicted if there at least three overlapping residues and no other helix shares an overlap <ref name="chen2002">Chen,C.P. et al. (2002) Transmembrane helix predictions revisited. Protein Science, 11, 2774–2791.</ref>.
<xr id="fig:tmhP35462"/> shows the comparison between OPM and PDBTM annotation and PolyPhobius prediction for the dopamine receptor. There is high agreement between OPM and PDBTM and by the above mentioned scoring system, PolyPhobius correctly identifies all transmembrane helices. It seems from the figure that there are two additional helices towards the end of the sequence, that are overpredicted, however the hydrophobicity plot already hints that this might not be an error of PolyPhobius. Indeed manually checking the entries in OPM and PDBTM reveals, that these two helices do exist, and PolyPhobius correctly predicted them. They are note displayed because 3pbl is annotated as a chimera in SIFTS. The mapping to P35462 only extends up to residue 230 and then switches to P00720.
<xr id="fig:tmhQ9YDF82"/> shows the comparison between OPM and PDBTM annotation and PolyPhobius prediction for the potassium channel. While all three methods agree on three C-terminal transmembrane helices, there are two N-terminal ones, that are predicted by PolyPhobius and present in PDBTM, but not annotated by OPM. Checking the OPM entry for 1orq it is revelaed that three N-terminal helices were actually explicitly excluded during creation of the database due to possible misalignments. The reference given for this decision in OPM <ref name="voltgate">Mackinnon,R. (2004) Structural biology. Voltage sensor meets lipid membrane. Science, 306, 1304-5.</ref> discusses the different theories of how the sensor works and in what way the helices are arranged in the open and closed formation. Indeed the literature agrees that there are in total six transmembrane helices in each monomer. Manual observation of a more recent structure 2r9r<ref name="">Long,S.B. et al. (2007) Atomic structure of a voltage-dependent K+ channel in a lipid membrane-like environment. Nature, 450, 376-82.</ref> supports this finding and also reveals that there is an additional re-entrant helix that does not cross the membrane and is oriented towards the inside of the tetramer. This is also recognized in the annotation by OPM and PDBTM (which explicitly mentions the re-entrant helix by residue type 'L'), both annotating a total of seven helices.
This suggests that the prediction performed by PolyPhobius could in fact be correct and a logical next step would be to assess the prediction performance using, not 1orq but 2r9r as a reference structure. However, a simple pairwise sequence alignment between 1orq:C and 2r9r:H reveals that the sequences are fairly dissimilar (Alignment length 201, similarity 49.8%) which is also the reason why there is no mapping available in SIFTS. This is due to the fact that 1orq describes the first voltage gate potassium channel found in Aeropyrum pernix<ref name="1orq">Jiang,Y. et al. (2003) X-ray structure of a voltage-dependent K+ channel. Nature, 423, 33-41.</ref>, while 2r9r is a chimera of Kv1.2 and Kv2.1, both from mammals. In conclusion a satisfying assessment of this would require more time than available for now, but it should definitely be noted that PolyPhobius performance might not be as bad as it seems at first glance.
Finally, <xr id="fig:tmhP47863"/> shows the comparison between OPM and PDBTM annotation and PolyPhobius prediction for aquaporin. Aquaporin is one of the archetypal structures for re-entrant helices, where two opposing ones form part of the central pore's surface <ref name="reentrant">Viklund,H. et al. (2006) Structural classification and prediction of reentrant regions in alpha-helical transmembrane proteins: application to complete genomes. Journal of molecular biology, 361, 591-603.</ref>. These are both annotated in OPM and PDBTM, however PolyPhobius misses them in the otherwise correct prediction. This could be due to the fact that these helices are comparably new and have not been known for a long time. While the re-entrant helix in the potassium channel was almost as long as a transmembrane helix, the two ones in aquaporin are much shorter making it hard to a method like PolyPhobius that was not created with these regions in mind, to identify them. Indeed, applying the newer MEMSAT-SVM <ref name="msvm">Nugent,T. and Jones,D.T. (2009) Transmembrane protein topology prediction using support vector machines. BMC bioinformatics, 10, 159.</ref> that was specifically trained for re-entrant helices shows that this method can identify all helices correctly.
PolyPhobius did not predict any transmembrane helices on the human Hex A alpha subunit which is correct. As can be seen in <xr id="fig:hexapp"/> there where no ambiguities and the soluble nature of the protein has been clearly identified. PolyPhobius does however find a signal peptide, which will be further discussed in the next section.
In conclusion, PolyPhobius showed very good performance on the set of four proteins. Residue-level accuracy is not achieved but actually cannot be achieved even among the 'gold-standards' OPM and PDBTM and is therefore not an issue. A problem is presented though by the difficulties to recognize the recently discovered structural elements of re-entrant helices.
Proteins: Serum albumin P02768, Lysosome-associated membrane glycoprotein 1 (LAMP-1) P11279, Aquaporin-4 (AQP-4) P47863.
According to Uniprot, HEXA, LAMP-1 and Serum albumin contain a signal peptide. LAMP-1 is a membrane protein which passes the membrane with one helix. Serum albumin, the main protein of plasma, is a secreted extracellular protein. AQP-4 is a multi-pass membrane protein which forms a water-specific channel.
The prediction of the displayed results was performed with SignalP version 4.0.
SignalP employs 3 main scores for the prediction of signal peptides: C, S and Y. The S-score stands for the actual signal peptide prediction, with high scores indicating that the corresponding amino acid is part of a signal peptide, and low scores indicating that the amino acid is part of a mature protein. The C-score is the cleavage score, which indicates the best cleavage site when significantly high. (When a cleavage site position is referred to by a single number, the number indicates the first residue in the mature protein.) Y-max is a derivative of the C-score, combined with the S-score calculated to give a better cleavage site prediction than the raw C-score alone <ref name="sigpref">
Bendtsen, J. D. et. al. (2004) Improved prediction of signal peptides: SignalP 3.0.
J. Mol. Biol., 340:783-795</ref>. For non-secretory proteins all scores are supposed to be very low.
<xr id="tbl:signalp"/> shows the results of the SignalP predictions and <xr id="tab:signals"/> gives a comparison of the predicted signal peptide positions and the validation from the Signal Peptide Website <ref name="signalpwebsite">http://www.signalpeptide.de</ref>. Additional scores can be viewed here.
HEXA, LAMP-1 and Serum albumin are correctly predicted as having one signal peptide at the beginning of the sequence and AQP-4 is identified as a mature protein. Even the exact positions of the peptides are predicted accurately and thus the performance of SignalP turns out exceptionally satisfactory.
Cross check with PolyPhobius
PolyPhobius was specifically built to account for false predictions of signal peptides as transmembrane helices and should therefore be able to distinguish them. Indeed PolyPhobius perfectly predicts the signal peptides in Serum Albumin and LAMP-1 (where it also finds the single transmembrane helix at the C-terminus). For the Hex A alpha subunit, only a signal peptide is predicted, however the cleavage region already ends at position 19, instead of 22. For aquaporin only the transmembrane helices are predicted an no signal peptide is found.
In conclusion PolyPhobius performs very well. While the prediction of signal peptides is not on par with SignalPv4, this is not what PolyPhobius was designed to do. The actual discrimination between signal peptide and transmembrane helices worked flawlessly for the test proteins.
<xr id="tab:gopetgo"/> depicts the prediction results for the Hex A alpha subunit from GOpet. The predictions are all given with a high confidence of at least 61%. To validate the results QuickGO and AmiGO were employed. As the GO annotations from these tools were all mostly inferred from electronic annotation (IEA) the predictions were additionally validated manually.
Hexosaminidase A is involved in the hydrolysis of terminal N-acetyl-D-hexosamine residues where only the alpha subunit is able to hydrolyse GM2 gangliosides. In presence of the cofactor GM2-activator protein (GM2AP) the alpha subunit of Hex A catalyses the removal of β-D-GalNAc from GM2. Thus the first 5 GO terms are considered correctly assigned. "Hydrolase activity hydrolyzing O-glycosyl compounds" is falsely predicted, but as it is not completely incorrect it should be viewed as merely a little shift away from the exact function. Hit number 6, the hydrolization of N-glycosyl compounds, is a very convincing assignment because it is not only true but also fairly specific. The prediction with the lowest confidence again proves correct as it is known that the Hex A alpha subunit forms a heterodimer with the Hex A beta-subunit. Altogether hexosaminidase activity and hydrolase activity, as part of more or less specific descriptions, dominate this GO term prediction. If the protein had been unknown, the GOpet prediction would have revealed many helpful functions which depict the protein accurately and already quite detailed.
|GO-Term ID||Type||Confidence||GO-Term description||Validation|
|GO:0003824||Molecular function||97%||catalytic activity||true||false||true|
|GO:0004563||Molecular function||96%||beta-N-acetylhexosaminidase activity||true||true||true|
|GO:0015929||Molecular function||96%||hexosaminidase activity||false||false||true|
|GO:0016787||Molecular function||96%||hydrolase activity||true||false||true|
|GO:0016798||Molecular function||96%||hydrolase activity acting on glycosyl bonds||true||false||true|
|GO:0004553||Molecular function||96%||hydrolase activity hydrolyzing O-glycosyl compounds||true||false||false|
|GO:0016799||Molecular function||77%||hydrolase activity hydrolyzing N-glycosyl compounds||false||false||true|
|GO:0046982||Molecular function||61%||protein heterodimerization activity||true||false||true|
ProtFun2.2 employes various tools for the ab initio protein function prediction. A large number of feature prediction servers such as SignalP are queried to obtain information about the submitted protein sequence. These are integrated into final predictions of the cellular role, enzyme class, and selected Gene Ontology categories. The classifications are based on two scores: The first score (influenced by the prior probability of that class) is the estimated probability that the entry belongs to the class in question. The second number (independent of the prior probability) is the odds that the sequence belongs to that class. The class with the highest information gain is chosen as prediction with the exception that ProtFun refrains from marking GO categories if the score with the highest information content has odds lower than 1 <ref name="protfun"> http://www.cbs.dtu.dk/services/ProtFun-2.2/output.php</ref>.
In the following the ProtFun2.2 prediction for the hexosaminidase alpha-subunit are analysed. As supplementary information, there is a detailed depiction of the ProtFun2.2 output available.
The Gene Ontology categories are displayed in <xr id="tab:protfun"/> (left). For the GO categories there is no single prediction above 10% and no entry receives an odds ratio greater than 1, thus Hex A is not attributed to any of the categories. With a closer examination of these classes it becomes clear that neither of them matches the function of our protein.
Further on, there is a prediction of cellular role provided which classifies the protein into 12 different functional categories based on a scheme developed by Monica Riley for E. coli in 1993 <ref name="protfunn">Riley, M. (1993) Functions of the gene products of Escherichia coli: Microbiol Rev. </ref>.
Here the Hex A alpha subunit is assigned to "Cell_envelope" with a probability of over 80% (see <xr id="tab:protfun"/>, right) . This prediction seems to be the most accurate.
In addition, the Hex A subunit is classified to be an enzyme, more specifically a ligase (EC 6.-.-.-) with the probability of 8,5%. This assignment is surprising as there is an EC number, (EC 3.-.-.-), which receives a much higher probability of 32.9% but it is neglected due to its lower odds ratio. Here the selection according to the highest information content has clearly failed as the hexosaminidase indeed belongs to the hydrolases, enzyme class 3. The exact EC classification is 220.127.116.11.
The classification of the cellular role could pose as a hint in the right direction of the Hex A subunit function. In addition the GO classification is not actually false, the categories apparently just do not cover the whole range of protein functions and therefore the Hex A alpha subunit can not be assigned. Apart from that, the GO category and the enzyme number predictions can be disregarded for Hex A and therefore the performance of ProtFun2.2 can be assessed as rather unsatisfying in this particular case.
The Pfam-A sequence search reveals two significant Pfam-A domains within the Hex A alpha subunit: The Glycosyl hydrolase family 20, domain 2 and the Glycosyl hydrolase family 20, catalytic domain (see <xr id="fig:pfam"/>).
HEXA is almost completely spanned by these two domains. The Glyco hydro 20b domain (green) reaches from position 35 to position 165 and the Glyco hydro 20 domain (red) directly follows up, occupying a region from position 167 to 488. Left unmapped are a mostly coiled region at the beginning of the subunit and a helix followed by another coiled region at the end of the sequence. <xr id="fig:pfam2gjx"/> visualizes the Pfam annotation in the 3D structure of the Hex A alpha subunit.
Both domain annotations are correct, albeit they don't seem very specific. The catalytic domain also belongs to the Pfam clan Glyco_hydro_tim, consisting of glycosyl hydrolases that contain a TIM barrel fold. The TIM barrel can be seen in the middle of catalytic domain in <xr id="fig:pfam2gjx"/>. The clan is very large <ref name="pfamclan">PFAM clan statistics</ref> with 41 members and amongst others also contains a domain associated with Fabry disease.
Interestingly, Pfam also infers an active site residue E323 which is indeed thought to be important for catalytic activity as already outlined in the introduction.
In conclusion Pfam of course cannot provide the wealth of information the prediction methods claim to deliver, however its manual curation and high quality data in combination with the recently introduced step towards crowd sourcing Wikipedia <ref name="Punta2012a">Punta,M. et al. (2012) The Pfam protein families database. Nucleic acids research, 40, D290-301.</ref> make it an at least equally valuable resource.