Python Programming for Biology: Bioinformatics and Beyond

Comparative modelling

When we do not have direct experimental evidence for the structure of a protein, we can

sometimes still come up with a good guess called a model if we know the structure of a

closely  related  protein.  This  method  is  known  as  comparative  modelling  or  homology

---

modelling.  Strictly  speaking  even  the  structures  of  proteins  determined  by  X-ray

crystallography  and  NMR  can  also  be  thought  of  as  models,  as  prior  information  about

normal  molecular  geometries  is  used  and  there  is  always  some  uncertainty.  However,

direct experimental data constrains the models much more (and generally crystallography

more so than NMR) and the more data you have the closer the model will be to the native

conformation.

Comparative  modelling  relies  on  the  observation  that  when  proteins  evolve  their

structures  change  more  slowly  than  their  amino  acid  sequence  does.  Hence,  if  we  can

detect  two  proteins  that  have  sufficiently  similar  amino  acid  sequences,  and  thus  infer  a

common ancestry, then we can be confident that they have structural similarities. Also, the

closer  the  sequence  similarity  between  two  such  homologous  proteins,  the  closer  their

structural  similarity  will  be.  There  are  two  basic  steps  when  building  a  structural  model

based  upon  the  structure  of  a  protein  homologue:  find  a  homologue  of  known  structure

and then use the homologue’s structure to guide the building of a model.

For  the  query  protein  of  unknown  structure  we  use  its  sequence  to  find  potential

homologues which do have a known structure, to act as the structural template.  Template

detection  uses  a  special  kind  of  sequence  alignment,  which  is  especially  sensitive  and

accurate.  Rather  than  using  the  regular,  general  substitution  matrices  like  BLOSUM  or

PAM,  comparative  modelling  tends  to  use  family-specific  scoring  matrices  or,  for  even

better  homologue  detection,  substitution  tables  that  are  specific  to  the  structural

environment.  Using  environment-specific  substitution  data  allows  an  alignment  to  be

sensitive  to  the  way  that  amino  acid  changes  in  evolution  depend  on  structure.  For

example,  serine  swapping  for  proline  is  more  common  in  turns  than  in  alpha-helices,

because  proline  tends  to  disrupt  helices.  We  know  the  structural  environment  for  each

position in a sequence alignment because we know the structure of the template, and the

best  guess  is  that  the  query  sequence  has  the  same  structural  environment,  even  if  the

residues  differ.  Such  structural  environments  are  typically  defined  by  combining  side-

chain hydrogen bonding, solvent exposure

6

and secondary-structure categories. Thus, for

example, you would have substitution matrices for exposed alpha-helix; buried, side-chain

hydrogen-bonding alpha-helix; exposed beta-sheet, to name but a few.

Once a template is selected, and the sequence-structure alignment tells us which query

residues are equivalent to which template residues, the next step is to build the computer

model. Generally the backbone of the model is built first, then the side chains, given that

these  may  vary  significantly  between  the  query  and  template,  and  finally  loops  are

modelled  in  the  regions  that  were  not  aligned,  i.e.  where  there  were  gaps.  The  initial

model  may  be  built  by  assembling  fragments  of  the  query  structure  using  the

conformations borrowed from one or more templates. Alternatively, it may be built by the

application of spatial restraints, derived from the templates, on to the model of the query

polypeptide.  The  model  is  then  subjected  to  a  minimisation  procedure  to  find  the

conformation  that  best  satisfies  these  restraints.  A  popular  program  for  such  restraint-

based modelling is MODELLER.

7