YOU'LL FIND HERE
  1. The presentation of the course
  2. Theoretical background
  3. Practical sections
 

                This sections aims to deal with all the theoretical an practical course of 1D predicitions.

PROTEIN 1D PREDICTIONS [BACK TO MAIN]

 

                There are not methods available to predict the protein 3D structure (folding) from the amino acid sequence. However, there are well established methods allowing to predict more simple structural features. These features can provide with useful indormation.
            Due to the genome sequencing projects the number of sequences has increased exponentially. Nonetheless, the crystal structures are not growing at the same rate (see figures below in the table). Then we have about 25.000 crystal structures (year 2004) whereas at the same time there are about 22 millions (year 2002). Why such a shift? Technically, is easier to obtain a DNA sequence (and then translate it) than obtaining a X-ray or NMR structure. We can obtain a DNA sequence in 2-3 days whereas the experimental requirements to obtain crystal might varie between 1-3 years.


24444 entries updated 24th February 2004

http://www.rcsb.org/pdb		 22,318,883 in year 2002 (February 2003)

http://www.ncbi.nlm.nih.gov/Genbank/genbankstats.html

    It has been proposed (and most of the times proven) that the protein 3D structure is a consequence of its primary amino acid composition. However there are many other factors that could influence the folding of a given protein. For instance chaperone proteins are essential for folding. But generally speaking it is assumed that the final structure is represented by the form in its minimum free energy state. That's why the native structure is coded by its primary amino acid sequence. There are some limits to this assumption. For example, there is not accuracy to establish the basic parameters improtant for folding, and also the computing resources are not well developed in these terms so far. There is not possible to predict structure from primary sequence yet as inferred by CASP experiments.
    We can, however, assign secondary structure elements to a primary sequence. In a 3 state condition, there are several algorithms capable to conduct such assignations. This can be useful for further attempts to thread. In fact, the most advanced methods are those regarding predicting secondary structure. This has happened because of the combiantion of such methods and evolutionary information available in the databases.

    The 1D structural predictions even if partially wrong, are very useful to obtain information regarding function of a given protein. Sometimes we can extract information related to the active site of a protein and even perform more complex predictions

 

Residue properties

The primary data we can extract from a protein are its physico-chemical characteristics like: hydrophobicity, polarity, etc.
We can then generate for instance the hydrophobicity plot of a protein to extract particular hydrophobic regions. This can be of help for further secondary predicitions.
There are several tools performing this kind of calculations from the primary sequence. Most of them provide with user-friendly WWW interfaces, some examples:

This server inputs a query and allows to select within 54 parameters.
The returned output is a graphic representation about the variation of the selected parameter along the sequence. A text file is also provided .

Secondary structure

There are three state elements of secondary structure: alpha-helix, beta-strand, and loops. This is to predict the arregment of those elements protein using its primary sequence. Secondary structure is assigned based on the hydrogen bonds profiles between the carboxyl and NH at the backbone. Most of the programs make use of neural network to train with known secondary structure proteins. Most of them also use additional information, extracted for instance from multiple alignments.

About secondary structure predictions methods

  • 1951 Pauling y Corey suggest there are local conformation patterns in proteins, such as helices and beta strands.
  • 1957 Szent-Györgyi & Cohen try to correlate the presence of certain residues with a particular secondary structure element. (i.e.: proline)
  • 1960 Blout, Fasman et al. & 1962 Blout, an extension of the previous one. They correlate this time the total number of amino acids with the total number of secondary structure elements.
  • 1960. Kendrew et al. & Perutz et al, Spectrophotometric determination the Myoglobin and hemoglobine proteins.
There are several ways to describe the secondary structure prediction methods. 1.- Probably the most descriptive one is to refer to them in a time-basis frame. (Eisenhaber, Persson and Argo, 1995) describe them.
The msot important contributions can be summarized in four main groups:
  • 1974. Chou y Fasman proposed a statistical method. This was based on the propensity of amino acids to adopt secondary structures. Those observations were made on 15 structures obtained by X-ray difraction. Obviously, these propensities were generated by the physico-chemical properties of the amino acids (special cases are glycine and proline). This method has improved increasing the number of proteins. Calculations were made using 5-6 residue windowas. This method reached an accuracy of ~50% (when 62 proteins were used as a benchmark)
  • 1978. Garnier improved this method including sifnificant interaction pairs. This method is ~60% reliable.
  • 1993. Levin improved the predictions including information from multiple alignments. Conserved regions within the alignment are indicative of importance in function. Moreover, those regions tend to have a more conserved structure. This strengthen the predicitions. The realiability for this method is ~69%.
  • 1994. Rost y Sander combined neural network with multiple alignments. The realiability improved to ~72%.

2.- Another way to describe the methods are based on the algorithm:

  • First generation methods:

    •

Statistical methods analzying amino acid propensity to adopt certain secondary structures.

The firt one (Chou & Fasman, 1975). Inference of data coming from 15 X-ray structures.

The reliability for this method is ~ 50% (when 62 proteins are used to infer statistics)

  • Second generation methods

The main improvement relies on the combination of using the major databases and calculating statistics based on neighbour segments (11-21 residues). The statistics evaluates the propensity of the central residue to be in a given secondary structure.

First and second generation methods show obvious problems

  • Reliability (3 state predictions) <70%

  • Reliability for B-strands 28-48% (~random)

  • alpha and beta too short

This is due to:

  • Experimental structures vary even within crystals of the same protein

  • Secondary structure depends on long range interactions (more than 11-21 resideues). This effect is stronger for betas. .

  • Third generation methods

The inclusion of evolutionary information allowed to improve the reliability of the predictions.The profiles obtained from multiple alignments are indicative of structural constrains. Moreover, these profiles don't contain local information as the evolutive preassure takes place at the 3D level and not at the priamry sequence level.

Therefore PSI-BLAST and Hidden Markov models improve remarkably the predictions.

Pros:

  • relaibility (3-state predictions)  70%

  • reliability for alpha ~ reliability for alpha ~ loops

Caveats:

  • Bad alignments => bad predicitions!

  • whne long-range interactions are taking place, this might lead to confusion between alphas and betas.

  • Caution to evaluate results of special proteins!

 

Kyte-Doolittle Hydropathy Scale

Scheme for PHD Protein Predictor Methods

Example: output of 3 secondary prediction servers. The query is a SH3 domain. The structure was assigned with DSSP. Reliability levels are: C+F=59%, GORIII=65% and PHD=72%. The RELIABILITY INDEX are 0-9. For > 4 predicitions were correct.

Example: Reliability ((3-states/residue) for different servers

First generation: Chou & Fasman, Lim, GORI

Second generation: Schneider, ALB, GORIII

Third generation: LPAG, COMBINE, S83, NSSP, PHD

 

Available public servers are:

  • PHDsec neural network for secondary structure prediction, accesibility, and trans-membrane helix. Uses multiple alignments. Reliability ~70%.

  • Jpred2  uses 2 neural networks and includes evolutive information (PSI-BLAST). The version 2 evaluates the results of 4 networks (JNet, NSSP, Predator, PHD) to improve reliability.

  • PROF Based on multiple alignments and other residue properties obtained from databases. Relaibility ~70%.

  • PSIpred  Based on filtered psi-blast profiles and neural networks (combines reults from different methods). relaibility is >76%.

  • SAM-T99  A neural network and multiple alignments profiles using hidden markov models.

  • SSpro Uses bi-directional recursive neural networks of fixed window size. Allwos to use the whole protein as an input.


Accesibility

Aims to predict the residue exposition to a solvent. The most detailed and fast method calculates accesibility estimating the exposed volumen to the solvent of each residue envolved in a connolly structure. (and lately implemented in DSSP). A simplification of this will be to go from normalized values (observed values divided by the maximum value) to a 2 state values, being those "buried" (relative accessibility < 16%) and "exposed" (relative accesibility ≥ 16%).
A classic method assigns one of these two values "buried/exposed" depending on the hidrophobicity of a particular residue. Accroding to this, highly hydrophobic areas are predicted as buried. However, like in secondary structure methods, there are some improvements due to the use of neural networks and/or training of sets of known structure.

Dickerson´s Dodecamer

Accessibility to solvent (Fig. from B. Rost).   Solvent accesibility is measure basically roling a spheric water molecule over the protein surface. The total value is the sum of the surface corresponding to each residue (normally (normalmente 0-300 Å2). In order to comapare amino acids you can calculate relative values (the percentage oa accessible area). Simpler descriptions only distinguish between two states: buried (in figure residues 1-3 and 10-12) and exposed (residues 4-9).


Accessibility for each position of a protein 3D structure is evolutionary conserved within each sequence family. Therefore, the information derived from multiple alignments has been used to develop new methods reaching a reliability from ~75% to ~79%.

Some public servers:

PHD y PROFphd (available trough the server) PredictProtein) Neural networks including multiple alignment derived information. These 2 servers are the only ones predicting real values for relative accessibility (matrix with 0, 1, 4, 9, 16, 25, 36, 49, 64, 81). JPred2 uses Psi_Blast based profiles as input to a neural network and gives back two states: "buried/exposed".

Hydropathicity plot, Kyte-Doolitle

 

Trans-membrane Helices

One of the most difficult aspects of proteomics is the determination of X-ray structure from trans-membranar proteins. They are almost impossible to cristalize and are difficult to approach by NMR (although some improvements have been developed). They are two main classes of transmembranar proteins: The ones introducing helix in the lipid membrane, and the other ones forming pores built by beta-strands (porin like). (Figure) . Up to date there are not public servers capable to predict this second group proteins, due to the lack of experimental evidence. However, for the first group (helical trans-membrane) the exact location of the helix can be predicted with high accuracy, exploring all the possible conformations.

Trans-membranar helices (Fig. from B. Rost).   For certain classes of membranar proteins, the hydrophobic segments are inside the lipid layer, they are located perpendicular to the membrne surface. helices are considered like rigid cylinders. The helix orientation can be defined by the first N-terminal region. Topology is defined as "outside" when N-terminal region is located at the extra-cellular region (A protein), and as "inside" when N-terminal region is located at the cytoplasm (proteínas B y C). The lower part describes the "inside-out-rule".  

These proteins prsent strong structural constrains because the lipd layer restricts the freedom. These helices can be predicted from observations: (a) they are mostly hydorphobic and average 12-35 residues, (b) the blobular regions between helices are typically less than 60 residues, (c) the bigger of the trans-membrane helices has a characteristic distribution of positive R and K (defined in the inside-out rule by Gunnar von Heijne) ins such a way that internal loops have more positive charges than loops in external regions, (d) Long globular regions (> 60 residuos) differ in composition from those subjected to the inside-out rule.

The vast majority of methods are based on neuronal-betworks or similar algorithms training benchmarks of known structure proteins. This increase the reliability because those helices have usually very well specific patterns. Including evolutive information improves the trans-membrane predicitions.

Available public servers are: (listado):

  • TopPred2, one of the classics

  • MEMSAT Introduces a dynamic programming to optimize the predictions based on statistical preferences.

  • TMAP Uses statistics and alignment profiles.

  • PHD combines neural networks, multiple alignments and dynamic progrmming to optimize predictions.

  • DAS use profiles hidrofobitos

  • SOSUI  combines hydrophobic propensities

  • TMHMM The most advanced one. The most reliable one. A similar one is HMMTOP). 

The reliability for the best methods (HMMTOP2, PHDhtm, y TMHMM2) predict correctly all the helices for a 70% of the studied proteins. In a 60% it also predicts correctly the right topology. The best reliability is PHDhtm at 70%. Generally speaking the methods tend to under-estimate the predictions ina ~ 86% and to mix signal peptides with trans-membrane helices. Moreover, the majority of methods, especially those based on hydropathicity over-estimate the helices predicitions in globular proteins in a range of ~90%. The error rate is 25% for PHDhtm and 34% for TMHMM2. This can be somehow fixed because the prediction methods for signal peptides are very reliable and the majority of helices incorrectly predicted start before the 10th residue from the N-terminal region MET. allowing for corrections. Despite all these problems all these predictions are useful for screening in whole genomes.

 

Transmembrane Helices (TMHMM Plot)

 

Post-transcripcional Modifications

"ExPASy Proteomics tools" (http://www.expasy.ch/tools/)

  • PSORT signal peptide prediction

  • TargetP Subcelullar location

  • SignalP Signal peptide

  • ChloroP Chloroplast peptides

  • MITOPROT mitochondrial

  • Predotar plastids and mitochonria

  • NetOGlyc O-glicosilation in mammal proteins

  • NDictyOGlic GlcNAc O-glicosilation in "Dictyostelium"

  • YinOYang Bindig O-beta-GlcNAc in eukarya

  • big-PI Predictor GPI (glicosil-fosfatidil inositol)

  • DGPI anchoring and breaking of GPI sites

  • NetPhos Eukaryotic phosphorialtion (Ser, Thr, Tyr)

  • NetPicoRNA predicción de sitos de ruptura para proteasas en proteínas de picornavirus

  • NMT N-myristoil N-terminal

  • Sulfinator

 

(SignalP)

http://www.cbs.dtu.dk/services/SignalP/

 

 

 

 

PRACTICAL MODULES
 [BACK TO MAIN MENU]


1D FEATURES

 

1D properties are those represented by secondary structure elements : helix (H), sheet (E) and loops;for accessibility (buried o exposed; or accesibility precentage);   for hidrophobicity etc. 1D are useful to predict 3D.
 
 

AA :        Residues for the sequence
OBSsec: Observed secondary structure (E: sheet, H: helice)
OBSacc: Observed accessibility(e: exposed, b: buried)
PHDsec: Predicted secondary structure
PHDacc: Predicted accessibility

 

Servers and programas

 

SECONDARY STRUCTURE PREDICTION:

Evaluation:  EVA
      PredictProtein
      (PHDsec)      		 PHDsec predicts the secondary structure multiple alignments. Neural networks (reliability = 72%, Rost & Sander, PNAS, 1993 , 90, 7558-7562; Rost & Sander, JMB, 1993 , 232, 584-599; and Rost & Sander, Proteins, 1994, 19, 55-72).  example [Output]
      JPred Jpred e   In the case of one query: it generates an automatic alignment with clustal, which is filtered with SCANPS.  Ejemplo [Output]
      PsiPred PSIPRED integrates neural networks "two feed-forward" which analyses the Psi-Blast output(Altschul et al., 1997). Reliability Q3 = 77%. Example [Output]

Training

¿What can you say about the secondary structure of this query?

Use the urldescribed in theory section.

1. Send these sequences to secondary structure predictions servers. (PHD, JPred, PsiPred). Compare the results.

Then send it to the signal peptide... (SignalP)

- Optional: generate a multiple alignment and send it to any server acepting this input. (JPred). Compare the results with the ones obtained when a unique sequence is sent.

>1_T0112 Ketose Reductase / Sorbitol Dehydrogenase, Bemisia argentifolii
MASDNLSAVL YKQNDLRLEQ RPIPEPKEDE VLLQMAYVGI CGSDVHYYEH GRIADFIVKD PMVIGHEASG TVVKVGKNVK HLKKGDRVAV EPGVPCRRCQ FCKEGKYNLC PDLTFCATPP DDGNLARYYV HAADFCHKLP DNVSLEEGAL LEPLSVGVHA CRRAGVQLGT TVLVIGAGPI GLVSVLAAKA YGAFVVCTAR SPRRLEVAKN CGADVTLVVD PAKEEESSII ERIRSAIGDL PNVTIDCSGN EKCITIGINI TRTGGTLMLV GMGSQMVTVP LVNACAREID IKSVFRYCND YPIALEMVAS GRCNVKQLVT HSFKLEQTVD AFEAARKKAD NTIKVMISCR QG

>APTE_DROME
MGVCTEERPVMHWQQSARFLGPGAREKSPTPPVAHQGSNQCGSAAGANNNHPLFRACSSSSCPDICDHST

>AREA_EMENI
MSGIAQLRLSDRVSNTPTTTADTVSDAMNLDDFIIPFSPSDHPSPSTTKASEATTGAIPIKARRDQSASE

>ARG1_YEAST
MTSNSDGSSTSPVEKPITGDVETNEPTKPIRRLSTPSPEQDQEGDFEEEDDDDKFSVSTSTPTPTITKTK

 

2. Send this sequence to TM-preodiction servers(TMHMM),

>2_636 AA
MEGPAFSKPL KDKINPWGPL IILGILIRAG VSVQHDSPHQ VFNVTWRVTN LMTGQTANVT SLLGTMTDAF PKLYFDLCDL IGDDWDETGL GCRTPGGRKR ARTFDFYVCP GHTVPTGCGG PREGYCGKWG CETTGQAYWK PSSSWDLISL KRGNTPRNQG PCYDSSAVSS NIKGATPGGR CNPLVLEFTD AGKKASWDGP KVWGLRLYRS TGIDPVTRFS LTRQVLNIGP RVSIGPNPVI TDQLPPSRPV QIMLPRPPQP PPPGAASIVP ETAPPSQQPG TGDRLLNLVD GAYRALNLTS PDKTQECWLC LVAGPPYYEG VAILGTYSNH TSAPANCSVA SQHKLTLSEV TGQGLCVGAV PKTHQALCNT TQTSSRGSYY LVAPTGTMWA CSTGLTPCIS TTILNLTTDY CVLVELWPRV TYHSPSYVYG LFERSNRHKR EPVSLTLALL LGGLTMGGIA AGIGTGTTAL MATQQFQQLQ AAVQDDLREV EKSISNLEKS LTSLSEVVLQ NRRGLDLLFL KEGGLCAALK EECCFYADHT GLVRDSMAKL RERLNQRQKL FESTQGWFEG LFNRSPWFTT LISTIMGPLI VLLMILLFGP CILNRLVQFV KDRISVVQAL VLTQQYHQLK PIEYEP

PHD_TM output

 

3. Send this sequence to(NetOGly, NetPhos)

>3_41 AA
ASYDGHKLVAGYDFTPPSTPSTDDPNVCREYSYKLGTYGAP

NetOGlyc output

>4_153 AA
ASQKRPSQRHGSKYLATASTMDHARHGFLPRHRDTGILDSIGRFFGGDRGAPKNMYKDSHHPARTAHYGSLPQKSHGRTQ DENPVVHFFKNIVTPRTPPPSQGKGRKSAHKGFKGVDAQGTLSKIFKLGGRDSRSGSPKPELVISALIVESRR

NetPhos output

 [BACK TO MAIN MENU]