Python Tools for Computational Molecular Biology
Python Tools for Computational Molecular Biology
João Rodrigues anaryin@gmail.com
Mentors
Eric Talevich
Diana Jaunzeikare
Peter Cock
Biopython is a very popular library in Bioinformatics and Computational
Biology. Its Bio.PDB module, originally developed by Thomas Hamelryck,
is a simple yet powerful tool for structural biologists. Although it
provides a reliable PDB parser feature and it allows several
calculations (Neighbour Search, RMS) to be made on macromolecules, it
still lacks a number of features that are part of a researcher’s daily
routine. Probing for disulphide bridges in a structure and adding polar
hydrogen atoms accordingly are two examples that can be incorporated in
Bio.PDB, given the module’s clever structure and good overall
organisation. Cosmetic operations such as chain removal and residue
renaming – to account for the different existing nomenclatures – and
renumbering would also be greatly appreciated by the community.
Another aspect that can be improved for Bio.PDB is a smooth
integration/interaction layer for heavy-weights in macromolecule
simulation such as MODELLER, GROMACS, AutoDock, HADDOCK. It could be
argued that the easiest solution would be to code hooks to these
packages’ functions and routines. However, projects such as the recently
developed edPDB or the more complete Biskit library render,
in my opinion, such interfacing efforts redundant. Instead, I believe it
to be more advantageous to include these software’ input/output formats
in Biopython’s SeqIO and AlignIO modules. This, together with the
creation of interfaces for model validation/structure checking
services/software would allow Biopython to be used as a pre- and
post-simulation tool. Eventually, it would pave the way for its
inclusion in pipelines and workflows for structure modelling, molecular
dynamics, and docking simulations.
The schedule below was organised to be flexible, which means that some features will likely be done early. Also, the weeks include documentation and unit testing efforts for the features, with extended periods for reviewing these efforts at the two points during the project (halfway, final week).
Getting familiar with development environment (GitHub account, Git, Biopython’s repository, Bug tracking system, etc)
Gather scientific literature and discuss some of the to-be-implemented methods.
NeighbourSearch classTesting and consolidating the features thoroughly.
Write documentation & examples for each feature, to be included in
Biopython’s Wiki and Bio.PDB’s FAQ.
Mid-term Evaluations. Discussing with mentors current state of project and adjust following schedule to comply with project’s needs.
SeqIOAlignIOBio.PDB.Polypeptide functionprotein.find_homoseq() )Hosted at this GitHub branch
Since I’m adding some methods that are useful/logical only for proteins,
having them exposed in Structure.py for every molecule could be
misleading. We decided then to add a as_protein() method that allows
protein-specific methods to be accessed. The following example
demonstrates how this call works. Note how the search_ss_bonds
method is absent from dir(s) but not from dir(prot).
from Bio.PDB import PDBParser
p = PDBParser()
s = p.get_structure('example', '4PTI.pdb')
dir(s)
# Cut for viewing purposes
['__doc__', ... , 'renumber_residues', 'set_parent', 'xtra']
prot = s.as_protein()
dir(prot)
['__doc__', ... , 'renumber_residues', 'search_ss_bonds', 'set_parent', 'xtra']
Since parse_pdb_header is far from optimal and is likely to change in
the future, I opted to forfeit reading SEQREQ records to account for
gaps. However, ignoring this information and renumbering based on ATOM
records would make us lose information on gaps. I opted to subtract the
first residue number-1 to all residues thus making the numbering start
in 1 and still keep gaps. I also added an argument (start) to allow the
user to set which number to start the counting from.
Example:
from Bio.PDB import PDBParser
p = PDBParser()
s = p.get_structure('example', '1IHM.pdb')
print list(s.get_residues())[0]
<Residue ASP het= resseq=1029 icode= >
s.renumber_residues()
print list(s.get_residues())[0]
<Residue ASP het= resseq=1 icode= >
The same rationale from SEQRES applies for the exclusion of looking up
SSBOND. Also, instead of using NeighborSearch to look for pairs of
cysteins in bond distance, I instead used the minus operator since it
has been overloaded to return the distance between two atoms (Page 10 of
the FAQ).
The average distance cited in the literature is 2.05A but other software
packages and my own tests set 3.0A as a good threshold. Still, the user
can set his own threshold manually.
The function returns an iterator with tuples of pairs of residues.
from Bio.PDB import PDBParser
p = PDBParser()
s = p.get_structure('example', '4PTI.pdb')
prot = s.as_protein()
for bond in prot.search_ss_bonds():
print bond
(<Residue CYS het= resseq=5 icode= >, <Residue CYS het= resseq=55 icode= >)
(<Residue CYS het= resseq=14 icode= >, <Residue CYS het= resseq=38 icode= >)
(<Residue CYS het= resseq=30 icode= >, <Residue CYS het= resseq=51 icode= >)
Added parsing for REMARK350 to parse_pdb_header since there was
already a bit written for another REMARK section. This extracts the
transformation matrices and the translation vector from the header, that
is then fed to the Structure function. Each new rotated structure is
created as a new MODEL. I chose this because crystal structures very
rarely have more than one MODEL instance and also because NMR models
don’t have REMARK 350 that often (at least to my knowledge).
from Bio.PDB import PDBParser
p = PDBParser()
s1 = p.get_structure('a', '4PTI.pdb')
s1.build_biological_unit()
'Processed 0 transformations on the structure.' # Identity matrix is ignored.
s2 = p.get_structure('b', 'homol_1bd8.pdb') # A homology model
s2.build_biological_unit()
'PDB File lacks appropriate REMARK 350 entries to build Biological Unit.'
s3 = p.get_structure('c', '1IHM.pdb')
s3.build_biological_unit()
'Processed 59 transformations on the structure.'
Following discussion between the mentors and me, we decided that maybe it was better to not only include a webserver for this purpose but also a local algorithm. This would not limit the user when there he/she lacks an internet connection.
The interface for the WHATIF Protonation service has been implemented, although it should be regarded as highly experimental for now. Interfacing this server included writing a small parser for a PDBXML-like format, which is expected to have serious bugs in its initial versions. I ran some simple tests and it works. It doesn’t support water molecules yet, nor any other molecules other than proteins. Such issues will be hopefully solved later on..
For those brave enough to want to test it (and help me debug it), here’s an example usage.
from Bio.Struct.WWW import WHATIF
from Bio import Struct
server = WHATIF.WHATIF() # Performs a sort of PING to the server. Gracefully exits if the servers are down.
# Get the protein structure
structure = Struct.read('4PTI.pdb')
protein = structure.as_protein() # This excludes water molecules
# Upload the structure to the WHATIF server
# This should convert the structure from a Structure object to a string via tempfile and PDBIO
# I was having some issues uploading structures...
id = server.UploadPDB(protein)
# Protonate
# Returns a Structure Object / WARNING! Bug prone for now.
protein_h = server.PDBasXMLwithSymwithPolarH(id)
Regarding the local implementation, after much reading I settled on
using PyMol’s algorithm. It seems to allow for protonation of any
structure, regardless of its nature (protein, DNA, etc). Its vectorial
and matrix operations can likely be optimized with Numpy and Biopython’s
Vector module. This first implementation works for proteins only.
I’ll add general molecule support later.
from Bio import Struct
from Bio.Struct import Hydrogenate as H
s = Struct.read('1ctf.pdb')
p = s.as_protein()
prot = H.Hydrogenate_Protein()
prot.add_hydrogens(p)
A Center of Mass function was developed first as part of a new module
Bio.Struct.Geometry. It allows for calculation of the center of geometry
(all masses are equal) and center of mass (taking into account elemental
masses for the atoms). The masses are a new Atom object feature derived
from this list and from PyMol.
Essentially, all atoms of a structure now get their mass defined when
the structure is created (check Atom.py and this
thread
for details). This is obviously experimental.
To calculate the center of mass of any Entity (Structure, Model, Chain, Residue) or a List of Atoms:
from Bio.Struct.Geometry import center_of_mass
from Bio import Struct
s = Struct.read('4PTI.pdb')
print center_of_mass.__doc__
Returns gravitic or geometric center of mass of an Entity.
Geometric assumes all masses are equal (geometric=True)
Defaults to Gravitic.
print center_of_mass(s)
[14.833301303933874, 21.431581746366263, 4.1218478418007134]
print center_of_mass(s, geometric=True)
[14.805324902127458, 21.365571977563405, 4.1108949403803985]
As of now, 3 CG models are supported.
1) CA-Trace 2) ENCAD 3-point model (CA, O, Side Chain bead) 3) MARTINI protein model (BB, Side Chain points [S1 to S4])
An example, picking up the s Structure from above:
p = s.as_protein() # To expose the CG method
ca_trace = p.coarse_grain()
# One atom per residue
print (len(list(p.get_residues())) == len(list(ca_trace.get_atoms())) )
True
cg_encad = p.coarse_grain('ENCAD_3P')
for residue in cg_encad.get_residues():
print residue.resname, residue.child_list
ARG [<Atom CA>, <Atom O>, <Atom CMA>]
PRO [<Atom CA>, <Atom O>, <Atom CMA>]
ASP [<Atom CA>, <Atom O>, <Atom CMA>]
PHE [<Atom CA>, <Atom O>, <Atom CMA>]
CYS [<Atom CA>, <Atom O>, <Atom CMA>]
LEU [<Atom CA>, <Atom O>, <Atom CMA>]
GLU [<Atom CA>, <Atom O>, <Atom CMA>]
PRO [<Atom CA>, <Atom O>, <Atom CMA>]
PRO [<Atom CA>, <Atom O>, <Atom CMA>]
TYR [<Atom CA>, <Atom O>, <Atom CMA>]
.....
CYS [<Atom CA>, <Atom O>, <Atom CMA>]
GLY [<Atom CA>, <Atom O>]
GLY [<Atom CA>, <Atom O>]
ALA [<Atom CA>, <Atom O>, <Atom CMA>]
cg_martini = p.coarse_grain('MARTINI')
for residue in cg_martini.get_residues():
print residue.resname, residue.child_list
ARG [<Atom BB>, <Atom S1>, <Atom S2>]
PRO [<Atom BB>, <Atom S1>]
ASP [<Atom BB>, <Atom S1>]
PHE [<Atom BB>, <Atom S1>, <Atom S2>, <Atom S3>]
CYS [<Atom BB>, <Atom S1>]
LEU [<Atom BB>, <Atom S1>]
GLU [<Atom BB>, <Atom S1>]
PRO [<Atom BB>, <Atom S1>]
PRO [<Atom BB>, <Atom S1>]
TYR [<Atom BB>, <Atom S1>, <Atom S2>, <Atom S3>]
......
CYS [<Atom BB>, <Atom S1>]
GLY [<Atom BB>]
GLY [<Atom BB>]
ALA [<Atom BB>]
Implement as part of Structure.py and based loosely on the contribution
of Ramon
Crehuet. The
DisorderedAtom objects are removed from the residue and a single Atom
object is added corresponding to the location of the user’s choice
(keep_loc argument) which defaults to A.
An example, still keeping s from above:
s = s.remove_disordered_atoms(verbose=True)
0 residues were modified
# Now if we load a structure with disordered atoms
ds = Struct.read('1MC2.pdb')
ds.remove_disordered_atoms(verbose=True)
Residue TRP:1010 has 8 disordered atoms: CD1/CD2/NE1/CE2/CE3/CZ2/CZ3/CH2
Residue VAL:1018 has 3 disordered atoms: CB/CG1/CG2
Residue LEU:1024 has 4 disordered atoms: CB/CG/CD1/CD2
Residue ARG:1043 has 7 disordered atoms: CB/CG/CD/NE/CZ/NH1/NH2
Residue MET:1092 has 4 disordered atoms: CB/CG/SD/CE
Residue ARG:1107 has 7 disordered atoms: CB/CG/CD/NE/CZ/NH1/NH2
Residue GLU:1108 has 4 disordered atoms: CG/CD/OE1/OE2
Residue ASP:1111 has 4 disordered atoms: CB/CG/OD1/OD2
Residue SER:1116 has 1 disordered atoms: OG
Residue SER:1131 has 1 disordered atoms: O
10 residues were modified
Biopython supports BLAST (local and remote through NCBI servers). We
bridged both Bio.PDB and Bio.Blast modules to allow an easier search for
sequence homologues. For now, it supports remote BLAST through
Bio.Blast.NCBIWWW and functions as a blackbox - i.e. users cannot change
any search parameter. If one wants to fully use BLAST he/she should use
the regular BLAST module. This is just a convenience function.
It is accessible only to Protein objects. It queries the PDB subset
database of NCBI BLAST servers with the Structure object’s sequence,
auto-adjusting parameters for short sequences (less than 15 residues).
It returns a list ranked by Expectation Value with some informational values (e-value, identities, positives, gaps), the PDB code of the match, and the alignment.
from Bio import Struct
s = Struct.read('1A8O.pdb')
p = s.as_protein()
seq_homologues = p.find_seq_homologues()
for homologues in seq_homologues:
print homologues[0], homologues[1]
print homologues[-1]
print
2BUO 1.82482e-31
DIRQGPKEPFRDYVDRFYKTLRAEQASQEVKNW-TETLLVQNANPDCKTILKALGPGATLEE--TACQG
DIRQGPKEPFRDYVDRFYKTLRAEQASQEVKNW TETLLVQNANPDCKTILKALGPGATLEE TACQG
DIRQGPKEPFRDYVDRFYKTLRAEQASQEVKNWMTETLLVQNANPDCKTILKALGPGATLEEMMTACQG
1AUM 1.82482e-31
DIRQGPKEPFRDYVDRFYKTLRAEQASQEVKNW-TETLLVQNANPDCKTILKALGPGATLEE--TACQG
DIRQGPKEPFRDYVDRFYKTLRAEQASQEVKNW TETLLVQNANPDCKTILKALGPGATLEE TACQG
DIRQGPKEPFRDYVDRFYKTLRAEQASQEVKNWMTETLLVQNANPDCKTILKALGPGATLEEMMTACQG
.....
MODELLER PIR format support was added to SeqIO as ‘pir-modeller’.
Currently, the format can be read but not written. An example of the
format follows, as well as an example of the parser’s usage.
>P1;5fd1
structureX:5fd1:1 :A:106 :A:ferredoxin:Azotobacter vinelandii: 1.90: 0.19
AFVVTDNCIKCKYTDCVEVCPVDCFYEGPNFLVIHPDECIDCALCEPECPAQAIFSEDEVPEDMQEFIQLNAELA
EVWPNITEKKDPLPDAEDWDGVKGKLQHLER*
from Bio import SeqIO
handle = open('test_pir.txt')
records = SeqIO.parse(handle, 'pir-modeller')
for i in records:
print i
ID: 5fd1
Name: 5fd1
Description: ferredoxin
Number of features: 0
/r_factor= 0.19
/end_residue=106
/initial_chain=a
/end_chain=a
/record_type=X-Ray Structure
/initial_residue=1
/resolution= 1.90
/source_organism=Azotobacter vinelandii
Seq('AFVVTDNCIKCKYTDCVEVCPVDCFYEGPNFLVIHPDECIDCALCEPECPAQAI...LER', ProteinAlphabet())