Transcript Pre-processing the PDB

Being a binding site: Characterizing
Residue-Composition of Binding Sites
on Proteins
Vince Grolmusz
joint work with
Zoltán Szabadka and Gábor Iván,
Protein Information Technology Group
Department of Computer Science, Eötvös University
Budapest, Hungary
The Protein Data Bank

It is a collection of the experimentally
determined 3D structures of biopolymers and
their complexes, today it contains more than
45 ,000 entries

Experimental methods include



X-Ray Diffraction
Nuclear magnetic resonance (NMR) spectroscopy
PDB file formats



pdb format
mmCIF format
XML format
The graph model of molecules

The molecule is modelled with a graph where the
vertices are the atoms and the edges are the covalent
bonds

Each atom has an atomic number and a formal charge

Each bond has an order that can be
 0 for coordinated covalent bonds
 1,2 or 3 for single, double and triple bonds
respectively

Aromatic ring systems are modelled with alternating
single and double bonds

A steric model is a graph model plus 3D coordinates for
the atoms
Main problems

Given a pdb file, find the steric model of each
molecule in it

Find the molecules which have unrealistic
steric models

Make a searchable database of different
protein-ligand complexes which fulfil certain
additional quality requirements
Our solution: The RS-PDB Database
(RS stands for Rich-Structure)
Difficulties and solutions

The two main difficulties with these problems

the basic units of a pdb entry are the residues and
HET groups, and not the molecules

there are atoms, whose coordinates could not be
determined, and these are simply missing from the
files

Therefore the problem can not be solved for every
entries

We developed a method to automatically process
the PDB mmCIF files and created a database with
an approximate solution and marked the places,
where there are errors or ambiguities
HET Group Dictionary

The basic units of a pdb entry are the residues and HET
groups, these will be called monomers

A monomer can be a molecule or a molecule fragment

Each monomer has a unique code: ASN, C, MG, NAD, …

The covalent structure of these monomers are in a
separate part of the PDB, the “PDB Chemical
Component Dictionary'‘, formerly called the HET Group
Dictionary (HGD)

We converted the structure descriptions of these
monomers to the graph model and put them in our HGD
database
Processing of an mmCIF file (1)
Polymers

We read all the so called entities from the file,
each of them containing one ore more
monomers

Each entity has a type, that can be polymer,
non-polymer or water, and each polymer
entity has a polymer type

Next we build the polymers from the
monomers, one-by-one, for example in the
case of proteins:
Constructing Polypeptide chains – the
peptide bond
A
O
O
R
R
HXT
HXT
HA
H
N
N
HN2
CA
C
C
OXT
HN2
...
C
OXT
C
OXT
HN2
CA
N
H
HA
OXT
CA
HN2
CA
N
H
HA
HXT
HXT
R
R
2
n
n-1
O
1
A
O
O
H
R
HA
R
HA
N
H
N
...
CA
C
CA
HN2
CA
C
C
N
R
2
OXT
HXT
H
H
O
1
C
HA
HA
R
CA
N
n-1
n
O
In the case of amino acid PRO, we
remove both HT1 and HT2; if, in the
case of a non-standard amino acid
(i.e., protein monomer), the above
mentioned atoms are not present, we
refuse to make chain.
HA
H
O
When a new amino acid (i.e., a
monomer) is added we remove
the atoms OXT and HXT from the
end of the chain, and the atom
HN2 from the new monomer, and
add a covalent bond between
the atoms C and N.


After the polymers are built, we define three
types of polymer molecules

Polypeptide chains (P) : >10 monomers long

DNA/RNA chains (N)
: >5 monomers long

Polysaccharides (S)
: >5 monomers long
The sequence of these polymers will give the
graph model of the molecules
Processing of an mmCIF file (2)
Ligands and their bond graph

Initially all monomers not belonging to a polymer are
distinct ligands, their graph model taken from the HGD

We read all the available atomic coordinates from the
mmCIF file to create the (partial) steric models

We find all pairs of atoms with distance less then 6 Å,
building a kd-tree for this purpose

If two atoms from different molecules are within
covalent distance, we try to combine their graphs

If this fails, or the atoms are too close, we record this in
a separate database table containing bond errors

Next, crystallization artefacts and “junk” ligands are
removed (Similarly as in the PDBBind database).
Database of protein-ligand complexes
and binding sites

A protein-ligand complex consists of a ligand and one or
more protein chains that have atoms in van der Waals
distance from the ligand; these atoms are painted red in
the figure:
Getting rid of redundancies





PDB is strongly biased in the direction of
“popular” or “important” proteins; some
chains (e.g., bovine trypsin) are present in
more than 100 PDB entries.
When mapping binding sites in the PDB,
redundancies must be dealt with;
If to the chain A ligand X is bound to the same
place in different PDB id’s -> counted once;
If to the chain A ligand X is bound at distinct
places -> counted twice or more
Result: 25,000 binding sites -> 19,000 B.S.
Residues in binding sites
•Next, those residues are collected from
protein chains, that are close to the ligands:
•We go through the ligand atoms oneby-one and find those protein atoms
which were closer to them than 1.05
times the sum of the Van der Waals radii
of the two atoms scanned;
•We do not have covalently bound ligands; they were already filtered out .
•Next we identify the residues containing these atoms: for every
binding site a subset of the 20 amino acids were created.
If the same residue appeared more than once, we inserted
it only once into the residue-set: we are interested in the plain
appearance of the residue at the binding site.
Binding site residue frequencies
Association rules in residue-sets


We are interested in implication-like rules such as:
(ALA,LEU)
(ILE,VAL)
that is, if a binding site contains amino acids leucine and
alanine, it will ``likely'' contain also valine and isoleucine.
Main attributes of the rules are:
support:
Prob(ALA,LEU,ILE,VAL)
confidence: Prob((ILE,VAL) | (ALA,LEU))
lift: Prob(ALA,LEU,ILE,VAL)/(Prob(ILE,VAL)Prob(ALA,LEU))
What is interesting?

Association rules X
Y, where Y is a very frequently
appearing residue-subset, are not interesting generally.

On the other hand, if Y is infrequent, then the support and
the confidence generally will not reach the thresholds to be
included in our results.

For example, Y=GLY appears very frequently, while Y=CYS
or Y=TRP appears rarely.

Association rules of unusually high and unusually low lifts
and rules of form X
Y with high confidence and not-toohigh support for Y are of particular interest. Our next
figures here visualize such remarkable data.
Our first figure…
…was created by deleting all X



GLY association
rules for clarity, and including only those rules
which satisfy that
their support is at least 7.15% and
their confidence is at least 0.5 and
at least one of the following conditions hold:




a) their confidence is at least 0.8 or
b) their lift is at least 1.8 or
c) their lift is at most 0.97 or
d) their support is at least 24%.
High-confidence area
Low-lift area
High-support area
Figure 2 contains rules, where…

all X
GLY association rules are deleted
for clarity, and

the support is at least 7.15% and

the confidence is at least 0.55 and

the lift is at least 1.7.
Here, ALA, the sixth most frequent
residue, is present in almost all bases;
and THR (threonine), the tenth most
frequent residue appears in the center;
all bases have 3 or 4 elements.
All large fan-in stars
contains GLY
Conclusions

We believe that by the analysis of the
residue-composition of the binding sites in
a really large and reliable data set, one
can identify pretty interesting data
patterns, applicable in inhibitor and drug
design;

We think that this work is just one of the
first steps in that direction.
Thank you very much!