Was kann ich per Knopfdruck über eine PDB-Struktur lernen?

4. Vorlesung WS 2004/05

Softwarewerkzeuge 1

Was kann ich per Knopfdruck über eine PDB-Struktur lernen?

PdbSum webseite:

http://www.ebi.ac.uk/thornton-srv/databases/cgi-bin/pdbsum/


Softwarewerkzeuge 2

Klassifizierung in CATH



Softwarewerkzeuge 3

Darstellung der Sekundärstruktur



Softwarewerkzeuge 4

Konservierung innerhalb Proteinfamilie

Oberfläche entsprechend Konservierung

eingefärbt.



Softwarewerkzeuge 5

Multiples Sequenzalignment



Softwarewerkzeuge 6

Ramachandran-Diagramm



Softwarewerkzeuge 7

Oberfläche

Spalten (clefts) auf Oberfläche sind

mögliche Bindungstaschen!



Softwarewerkzeuge 8

Sekundärstrukturvorhersage: PSIPRED

D.T. Jones, J Mol Biol 292, 195 (1999); http://bioinf.cs.ucl.ac.uk/psipred/

Enge, sehr polare Bindungstasche auf Proteinoberfläche.


Softwarewerkzeuge 9

Qualität von PSIRED-Vorhersagen

D.T. Jones, J Mol Biol 292, 195 (1999); http://bioinf.cs.ucl.ac.uk/psipred/

Ergebnis für 187 Testproteine mit unterschiedlichen Faltungen.

Genauigkeit von PSIPRED:Ca. 75%


Softwarewerkzeuge 10

Vorhersage von TM-Helices

http://darwin.nmsu.edu/~molb470/fall2003/Projects/koul/tmhmm.html

Residuen in Transmembranhelices

sind fast ausschließlich hydrophob.

Länge einer TM-Helix ≥ 20 Residuen.

HMMs sind sehr erfolgreich um TM-

Helices vorherzusagen (>90%

Genauigkeit).



Analyse der Oberfläche: elektrostatisches Potential

Sheinerman, Honig, J Mol Biol 318, 161 (2002)

Proteinoberflächen an Protein-

Protein-Bindungsstellen sind

häufig elektrostatisch

komplementär.

Surface representation of the

electrostatic potential of unbound

monomers of 4 protein-protein

complexes. Open book view of the

protein–protein interfaces is shown. Color

range from deep red to deep blue

corresponds to the range in the values of

electrostatic potential from −10 to

+10kT/e, where k is the Boltzmann

constant, T is the absolute temperature

and e is a proton's charge.



PROCHECK: Qualitätscheck für ProteinstrukturenThe Ramachandran plot shows the phi-psi torsion angles for all residues in the structure (except those at the chain termini). Glycines are separately identified by triangles as these are not restricted to the regions of the plot appropriate to the other sidechain types. Colouring/shading scheme: the darkest areas (here shown in red) correspond to the "core" regions representing the most favourable combinations of phi-psi values. The regions are labelled as follows:

A - Core alpha B - Core beta L - Core left-handed alpha p - Allowed epsilon a - Allowed alpha b - Allowed beta l - Allowed left-handed alpha ~a - Generous alpha ~p - Generous epsilon ~l - Generous left-handed alpha ~b - Generous beta

The different regions were taken from the observed phi-psi distribution for 121,870 residues from 463 known X-ray protein structures. The two most favoured regions are the "core" and "allowed" regions which correspond to 10° x 10° pixels having more than 100 and 8 residues in them, respectively. The "generous" regions were defined by Morris et al. (1992) by extending out by 20° (two pixels) all round the "allowed" regions. In fact, the authors found very few residues in these "generous" regions, so they can probably be treated much like the "disallowed" region and any residues in them investigated more closely.

Ideally, one would hope to have over 90% of the residues in the "core" regions. The percentage of residues in the "core" regions is one of the

better guides to the stereochemical quality of a protein structure. http://www.biochem.ucl.ac.uk/~roman/procheck



PROCHECK

The plot shows separate Ramachandran plots

are shown for each of the 20 different amino

acid types.

The darker the shaded area on each plot, the

more favourable the region. The data on

which the shading is based has come from a

data set of 163 non-homologous, high-

resolution protein chains chosen from

structures solved by X-ray crystallography to a

resolution of 2.0Å or better and an R-factor no

greater than 20%.

The numbers in brackets, following each

residue name, show the total number of data

points on that graph. The red numbers above

the data points are the reside-numbers of the

residues in question (ie showing those

residues lying in unfavourable regions of the

plot).

http://www.biochem.ucl.ac.uk/~roman/procheck



PROCHECK: analysis of side chain angles




PROCHECKThe 6 graphs show how the structure (represented by the solid square) compares with well-refined structures at a similar resolution. The dark band in each graph represents the results from the well-refined structures; the central line is a least-squares fit to the mean trend as a function of resolution, while the width of the band on either side of it corresponds to a variation of one standard deviation about the mean. In some cases, the trend is dependent on the resolution, and in other cases it is not. The 6 properties plotted are: a. Ramachandran plot quality. This property is measured by the percentage of the protein's residues that are in the most favoured, or core, regions of the Ramachandran plot. For a good model structure, obtained at high resolution, one would expect this percentage to be over 90%. However, as the resolution gets poorer, so this figure decreases - as might be expected. The shaded region reflects this expected decrease with worsening resolution. b. Peptide bond planarity. This property is measured by calculating the standard deviation of the protein structure's omega torsion angles. The smaller the value the tighter the clustering around the ideal of 180 degrees (which represents a perfectly planar peptide bond). c. Bad non-bonded interactions. This property is measured by the number of bad contacts per 100 residues. Bad contacts are selected from the list of non-bonded interactions and are defined as contacts where the distance of closest approach is less than or equal to 2.6Å. d. Calpha tetrahedral distortion. This property is measured by calculating the standard deviation of the zeta torsion angle. This is a notional torsion angle in that it is not defined about any actual bond in the structure. Rather, it is defined by the following four atoms within a given residue: Calpha, N, C, and Cbeta. e. Main-chain hydrogen bond energy. This property is measured by the standard deviation of the hydrogen bond energies for main-chain hydrogen bonds. The energies are calculated using the method of Kabsch & Sander (1983). f. Overall G-factor. The overall G-factor is a measure of the overall normality of the structure. The overall value is obtained from an average of all the different G-factors for each residue in the structure.




The 5 properties plotted are:

a. Standard deviation of the chi-1

gauche minus torsion angles.

b. Standard deviation of the chi-1

trans torsion angles.

c. Standard deviation of the chi-1

gauche plus torsion angles.

d. Pooled standard deviation of all

chi-1 torsion angles.

e. Standard deviation of the chi-2

trans torsion angles.

PROCHECK




PROCHECK


Distributions of each of the different

main-chain bond lengths in the structure.

The solid line in the centre of each plot

corresponds to the small-molecule mean

value, while the dashed lines either side

show the small-molecule standard

deviation, the data coming from Engh

& Huber (1991).

Highlighted bars correspond to values

more than 2.0 standard deviations from

the mean, though the value of 2.0 can be

changed by editing the procheck.prm file.



PROCHECK


Distributions of each of the different

main-chain bond angles in the

structure. The solid line in the centre

of each plot corresponds to the small-

molecule mean value, while the

dashed lines either side show the

small-molecule standard deviation,

the data coming from Engh

& Huber (1991).

If any of the histogram bars lie off the

graph, to the left or to the right, a large

arrow indicates the number of these

outliers (as in the CA-C-O and CB-CA-

C plots above).



PROCHECK

RMS distances from planarity for the

different planar groups in the structure.

The dashed lines indicate different ideal

values for aromatic rings (Phe, Tyr, Trp,

His) and for planar end-groups (Arg, Asn,

Asp, Gln, Glu).

The default values are 0.03Å and 0.02Å,

respectively.




Wie kann man 2 Proteinstrukturen vergleichen?

Paarweise Sequenzvergleiche

Paarweise Strukturvergleiche?



Partitioning protein space into homologous families

Protein architecture. The tramtrack protein

[Protein Data Bank entry 2drp (30)] is a

small protein (525 heavy atoms,

63 residues, and 6 elements of secondary

structure), yet it exhibits typical modular

protein architecture with two compact

structural domains, the so-called zinc

fingers.

(A) The most detailed description of

atomic positions is required to understand

the function of the tramtrack protein (gray

and black, running left to right), which

involves binding to a specific base

sequence of DNA (white).

Holm, Sander Science 273, 5275 (1996)




(B) The complicated 3D shape of proteins

is encoded in their linear sequence of

amino acids. Side chains stripped off, the

polypeptide backbone (thick) can be seen

meandering from the bottom left to the

upper right. Regular patterns of hydrogen

bonding (thin lines) between amide and

carbonyl groups of the polypeptide

backbone give rise to secondary structure,

shown schematically in (C) as arrows for

strands and cylinders for helices (with

zinc atoms as spheres).




Meaning of structural equivalenceShape comparison aims at the 1:1 enumeration of

equivalent polymer units in 2 protein molecules.

The problem and solution can be represented in

3D, as a rigid-body superimposition; in 2D, as

similar patterns in distance matrices; or in 1D, as

an alignment of amino acid sequences.

Here, the comparison of the tramtrack protein with

another zinc finger protein, the human enhancer-

binding protein MBP-1 [PDB entry 1bbo], is used

as an example.

(A) In the 3D comparison, the problem is to find a

translation and rotation of one molecule (red:

1bbo) onto the other (blue: 2drpA). The 3D

superimposition (residue centers only, green lines

join equivalenced residue centers, zinc atoms as

spheres) is not exact because of an internal

rotation of the two zinc finger domains relative to

one another. Holm, Sander Science 273, 5275 (1996)



Ranges of similarity between proteins

Holm et al. Prot Sci 1, 1691 (1992)



Surprising similarities












Partitioning protein space into homologous families(B) The 2D distance matrices reveal

the conserved structure of the zinc

fingers (left: distance matrices of the

whole structures; black dots are

intramolecular distances less than

12 Å, 1bbo at bottom and 2drpA on

top; right: distance matrices brought

into register by keeping only rows or

columns corresponding to

structurally equivalent residues).

(C) One-dimensional alignment of

amino acid strings. Evolutionary

comparison aligns the histidine (H)

residues involved in zinc binding

(bold; helices and strands of

secondary structure are underlined).




2 Algorithms for structural alignment(A) The 3D lookup is a fast heuristic algorithm that catches easy-to-find

structural similarities and is part of the Dali 3D search server. The idea is

that in favorable cases, 3D superimposition of only a pair of secondary

structure elements (SSEs) leads to superimposition of the entire

structures.

Top: Structure comparison of an SH3 domain of c-Src kinase [1cskA,

query structure] with the enzyme papain [1ppn, target structure] reveals

similar domain folds, although there is no sequence relation between the

proteins and one is much larger. The appropriate orientation of the

molecules is found by exhaustive comparison of internal coordinate

frames of each protein. An internal coordinate frame is defined by an

ordered pair of SSEs (centering one SSE at the origin, aligning it with the

y axis, and rotating the molecule around this axis so that the center of a

second SSE is in the positive x-y plane).

Bottom left: Target structure, papain, loaded onto the SSE lookup grid.

Each pair of SSEs where the segment midpoints are within 12 Å defines a

coordinate frame relative to the grid axes. The figure shows the

transformed positions of the 12 SSEs of papain (dotted lines) in each of

the 100 different coordinate frames defined by different pairs of SSEs.

Bottom right: The target lookup grid is probed with the SH3 domain, which

has four SSEs (thick continuous lines). The coordinate frames shown are

the ones yielding the best 3D match of four segments. Iterative extension

of a residue-wise alignment starting from the preorientation defined by the

SSE match shown here leads to the equivalence of 43 C atoms with

1.7 Å root-mean-square positional deviation on an optimal least-squares

superimposition. Holm, Sander Science 273, 5275 (1996)



Branch-and-bound algorithm(B) A branch-and-bound algorithm is guaranteed to yield the global optimum but may, in the

worst case, need an exponential number of steps to do so. An implementation of this

algorithm is an essential part of the Dali 3D search server.

First, protein structures A and B are represented by distance matrices (bottom left and right;

each point in a matrix is a residue-residue distance; an internal square is a set of contacts

made by two segments; the secondary structure segments are ,, and ). The problem of

shape comparison becomes one of finding a best subset of residues in each matrix (subsets

of rows and columns) such that the set of residues in protein A has a similar pattern of

intramolecular distances as the set in protein A, as in Fig. 2B. A single solution to the problem

is given in terms of the two sets of equivalent residues (an alignment), as shown in Fig. 2C.

The solution space consists of all possible placements of residues in protein B relative to the

segments of residues of protein A. The key algorithmic idea is to recursively split the solution

subspace (schematically shown as a circle at upper left, in which each point is a solution to

the problem and the lines divide subsets of solutions) that yields the highest upper bound

until there is a single alignment trace left: start with the entire circle; calculate the upper

bound for the left (9) and right (17) half; choose the right half and split it into top (upper bound

10) and bottom (upper bound 16) quarters; choose the bottom part and split it (left: 14; right:

12); choose the right part; and so on until the area of solution space has shrunk to a single

solution (shown as the residue-residue alignment matrix enlarged at right). The upper bound

for each part of the solution space is estimated in terms of a simplified subproblem that asks

for the best match of residues in protein B onto a predefined set of residues in protein A (the

match is illustrated by the circle-ended line connecting the single square in matrix A with a

set of candidate squares in matrix B). The best match is the one with the maximal pair score

(sum of similarities of distances between the square in A and the square in B). The

predefined set corresponds to residues in secondary structure elements ( , ). The upper

bound for each of the segment-segment submatrices of matrix A is found by calculating the

similarity scores between the submatrix in A and all accessible submatrices in B. An upper

bound of the total similarity score (sum over all segment-segment submatrices in A) for one

set of solutions is given by the sum of separately calculated upper bounds for each segment-

segment pair of matrix A. The method for choosing constraints that define a set of solutions

works in terms of defining allowed residue ranges at each stage of the iteration and is not

illustrated. Holm, Sander Science 273, 5275 (1996)



Recurrent folds

(A) A small number of frequently occurring

domains (folds) covers a large fraction of all

known protein structures. The 287 structurally

unique protein domains (folds) are ranked in

descending order of occurrence in the

representative set of 740 proteins. Domains

ranked 1 through 16 occur 10 or more times

each. Domains ranked 1 through 26 cover 50%

of all known structures that is, the essential parts

of these structures can be constructed from

these domains or described in terms of these

domains (within the limits of similarity within a

domain class). Domains ranked about 170 or

higher occur only once in the current database

(singlets).




Partitioning protein space into homologous families(B) Examples of frequently observed fold classes, with one class

from each of the attractor regions in Fig. 5 (each attractor region

contains several classes, where the term "class" is defined in the

text). Color coding indicates which parts of the fold are present in

more or fewer members of the class. The color changes from

light blue (regions present in 100% of members of the fold class)

to red (0% occupancy). The representative classes are defined

as follows (attractor, class name, and number of recurrences in

sequence-unique set of 740 structures): attractor I: parallel :

COOH-terminal domain of succinyl-CoA synthetase chain

(126); attractor II: -meander: mouse opg2 immunoglobulin

heavy chain variable domain (52); attractor III: -helical:

myoglobin; attractor IV: -zigzag: COOH-terminal domain of

pertussis toxin; and attractor V: meander: COOH-terminal

domain of phosphoglycerate dehydrogenase. Note that other fold

classes in the same attractor region are not shown, but the most

frequently occurring are shown in Fig. 5B.




(C) Growth and redundancy of protein 3D structures in the

Protein Data Bank.

Entry: one of currently more than 4000 sets of protein

coordinates in the PDB.

Family: collection of proteins set as equivalent if pairwise

sequence identity exceeds 25%.

Fold: fold class as defined above.

The number of new structure entries grows rapidly in time (note

logarithmic scale). Redundancy is defined in terms of sequence

similarity (sequence families) or structure similarity (fold

classes). Currently, there are about 6.4 entries per sequence

family and 2.4 families per fold class, for a total of 15 entries

per fold. One may expect that in the near future a new fold will

appear for about every 15 new entries. The curve of new folds

lags behind the curve of sequence-unique families, which

indicates the increasing frequency of recurrent folds in newly

solved structures (although this may be the result of bias in

experimental work). There is no indication that the growth in

new fold classes is slowing down at present.





Partitioning protein space into homologous families(B) 40% of all known domains (protein substructures) are covered by 16 fold

classes (shown as topology diagrams; , -helix segment; , -strand segment;

thick bar, parallel chain connection between segments; thin bars, antiparallel

connection; arc, helices crossing at roughly right angles). Although each fold

class has individual features, most fold classes map to five attractor regions

(peaks I through V).

All folds with sheets of mainly parallel strands map to attractor I. The parallel

folds contain a x unit, where the intervening segment (x) is required to

reverse chain direction so that the strands are parallel. The unit has a

preferred handedness determined by polymer physics and the natural twist of

strands. Attractor II contains a variety of helical folds. The connectivity of

elements in the folds of attractors III and IV contains meander motifs

suggestive of the collapse of a long hairpin, either of strands only or of

strands alternating with a helical pair, ()2. The zigzag motif of attractor V is

simply a series of antiparallel hairpin connections between sequentially

adjacent strands. Elementary polymer physics indicates that interactions in

space between regions of the chain that are close in sequence are much more

probable than those between sequence-distant regions. The zigzag motif

occurs both in flat sheets and barrels, and there is considerable variation in the

length of strands (about 4 residues in propeller blades, about 13 in porin

barrels). Holm, Sander Science 273, 5275 (1996)



Evolutionary adaptation of enzyme function

(A) Discovery of an essential structure-function feature

by shape comparison. A structure database search with

DNA polymerase detects kanamycin

nucleotidyltransferase (rather than other known DNA or

RNA polymerases) as the nearest neighbor in fold space

and reveals conserved residues and structural features

supporting the active site.

Following up the lead provided by structure database

searching with profile searches in sequence databases

resulted in the identification of the same characteristics

in a large superfamily of nucleotidyltransferases.

The biological functions of member families range from

DNA repair to regulation of biosynthetic pathways and

antibiotic resistance.





(B) Variety of substrate specificity of a

common chemical reaction on an essential

protein substructure is the remarkable result

of biological evolution. All member enzymes

of this extended family unified as a result of

shape comparison catalyze a common

chemical reaction, the coupling of nucleoside

triphosphates (black squares and dots) to a

free hydroxyl group by means of elimination

of pyrophosphate [top row: DNA polymerase

, DNA nucleotidyl exotransferase; middle

row: polyadenylate polymerase, (2‘-5‘)

oligoadenylate synthetase, kanamycin

nucleotidyltransferase; bottom row: protein PII

uridylyltransferase, glutamine synthetase

adenylyltransferase, and streptomycin 3‘-

adenylyltransferase].




Partitioning protein space into homologous familiesa, All-against-all structure alignment by DALI reveals a hierarchical

organization of fold space. The method is sensitive enough to recognize

similarities of general folding pattern — e.g., the -sandwich topology of

superoxide dismutase and immunoglobulin domains — and selective

enough to give higher scores to pairs of structures with more closely

superimposable C traces — e.g., any two globins score higher than any

globin–phycocyanin pair. Structure similarity alone yields an operational

definition of 'folds'. The thick circles denoting folds (left) are defined using a

uniform radius for clusters of structural neighbors. The vertical bar (right)

denotes cutting the fold dendrogram at a uniform value of structural

similarity. However, the level of structural similarity, or degree of structural

divergence, varies between different families, and we need other criteria to

delineate superfamilies.

b, Divergent evolution from a common ancestor retains not only the fold but

also many functional features. This means that homologs remain in a

structural neighborhood and can be delineated by similar functional

attributes (marked here by similar color) in the map of fold space.

Functional convergence (from independent evolutionary origins) would

appear as blotches of similar color in disconnected regions of the map of

fold space and in disjoint branches of the fold dendrogram. Partitioning the

fold dendrogram in terms of functional similarities yields family-specific

thresholds in terms of structural similarity (nodes that partition the fold

dendrogram into functionally conserved superfamilies are circled on the

right). This combination of structural and functional similarity measures

results in an automatically generated hierarchical classification m_n at the

fold (m) and superfamily (n) levels. Dietmann & Holm, Nat Struct Biol 8, 953 (2001)



Proteinstruktur-Analyse

c, The principles are illustrated on a branch of the fold dendrogram

consisting of aminopeptidases (1xjo and 1amp), carboxypeptidase

(1aye), purine nucleoside phosphorylases (1b8oA, 1cb0A and

1ecpA), pyrrolidone carboxyl peptidase (1a2zA), peptidyl–tRNA

hydrolase (2pth) and hydrogenase maturating endopeptidase

(1cfzA). The functional similarity between all pairs of structures is

evaluated using a neural network with output in the range 0

(analogous)-1 (homologous) — for example, (1cb0A, 1b8oA) =

0.91, (1amp, 1aye) = 0.74, (1cfzA, 2pth) = 0.59, (1xjo, 1amp) =

0.30 and (1a2zA, 2pth) = 0.13. Here, line thickness indicates the

magnitude of the term (i,j) - (Eq. 1) with color-coding for positive

(red) or negative (blue) values. The threshold parameter was

arbitrarily set to 0.30 in this numerical example.

d, The protein set is partitioned into superfamilies in the context of

the fold dendrogram. Node scores s(C) are computed for each node,

with = 0.30. For example, each structure is homologous to itself;

therefore, leaf nodes get a score s(leaf) = 1.00 - = 0.70, whereas

s(1cfzA, 2pth) = (1.00 + 1.00 + (2 0.59)) / 4 - = 1.98. The optimal

partition (circled nodes) maximizes the sum of node scores over

selected nodes (underlined scores). This optimal partition is stable

for threshold values 0.09 < < 0.53.

Dietmann & Holm, Nat Struct Biol 8, 953 (2001)







Proteinstruktur-Vergleich durch Feature-Vector

Input für Neuronales Netzwerk ist ein Feature-Vector.


„Keyword similarity“: Vektorprodukt für Häufigkeiten von Swissprot-Keywörter innerhalb der beiden

Sequenzfamilien.

„Functional preference“ is pro Aminosäure definiert und wird über alle Residuen in einem 3D-Cluster von

konservierten Residuen summiert.



Funktionszuordnung per Strukturvergleich




Zusammenfassung

Viele, sehr bequeme Tools verfügbar, mit denen man schnell einen guten

Überblick über bestimmte Proteinstrukturen erhalten kann.

Proteinstruktur ist evolutionär wesentlich länger konserviert als Sequenz

Strukturvergleiche erlauben es, wesentlich entferntere Verwandtschaften

aufzudecken.

Numerische Klassifizierung erlaubt (nun erstmals) eine robuste, automatische

evolutionäre Klassifikation von Proteinstrukturen.


Documents

Was kann ich per Knopfdruck über eine PDB-Struktur lernen?