The algorithm used by WAM, essentially an update on that used by Oxford Molecular's AbM, is detailed as follows. Improvements to AbM are indicated.
Framework and canonicals
Build and fit together the frameworks. The most sequence-homologous frameworks to the sequence being modelled are chosen from a database of known antibody structures. The light and heavy chains are chosen separately and fitted together using certain conserved interface residues on a mean beta-barrel of the known structures. Sidechains are resequenced using a maximum-overlap method.
Build the canonical loops. The most sequence-homologous known loops of the same canonical class are used. Sidechains which are conserved in chi-1 only are built at this stage (note: canonical structures conserve only some of the sidechain positions, and normally chi-1 only) ; like those of the framework, they are resequenced by maximum-overlap. A new stage involves 5 rounds of steepest-descent minimisation to smooth the joins.
Non-canonical loops: original AbM algorithm
Several changes have been made to the modelling algorithm for non-canonical loops in WAM compared to AbM. Therefore the original AbM algorithm will be described first, followed by the changes implemented in WAM.
For non-canonical loops, a different method is used. This is the CAMAL (Combined Antibody Modelling Algorithm) of Martin, Cheetham and Rees (1989), and consists of a combined database/conformational search.
Database search. A set of c-alpha to c-alpha distance constraints, derived from the known structures of that length of the current CDR, is used to extract, from the entire Brookhaven Protein Data Bank, a number of putative models. For short loops (7 residues or less), we stop here as the database search is considered to saturate conformational space, but for longer loops, the centre of the loop is rebuilt using the CONGEN conformational search of Bruccoleri and Karplus (see below).
Conformational search. A 5-residue segment in the centre of each database loop is deleted, and the conformational space of the two end residues of this segment are sampled. The optimum conformation of the middle three residues is then calculated using the Go and Scheraga chain closure algorithm.
Energy screen. The VFF potential of Dauber-Osguthorpe et al is used to energy minimise, and screen the set of conformations generated by CAMAL. This consists of van der Waals, bond, angle and torsion energies. Electrostatics have been found to favour "false positive" conformations so have been turned off.
Structural Determining Residue filter. Finally, the final five conformations by energy are ranked using this filtering method, which compares the torsion angles of particular residue types in the models with those observed in known antibody structures.
Non-canonical non-H3 loops: modification in WAM
CAMAL takes advantage of the increased saturation of conformational space of CONGEN but minimises its disadvantage: time. However, it is still relatively time consuming, when one considers that each conformation also needs to be energy screened (up to a day for CAMAL + energy screen per CDR). Modelling can be accelerated by using the database search only, and furthermore, tests across a number of cases have shown that this does not, in non-H3 loops, significantly decrease the accuracy of the model. Therefore, in WAM, this is the default, and recommended, method for non-canonical, non-H3 loops.
Modelling the H3 loop: modifications in WAM
Modelling CDR-H3 loops is the area where most of the updates in WAM have taken place, largely resulting from greater knowledge about H3 structure . The CAMAL method described above is used, with the following changes:
a) Modifications to database search. For structures predicted to be kinked, the C-alpha to C-alpha distance constraints are derived only from kinked loops. Also, the loop/ framework join is energy minimised subsequent to grafting as bad clashes have been persistently found to raise the energy of otherwise good structures here.
b) Alteration of CONGEN rebuild range. The most exposed region of H3 is towards the N-terminus; hence, the rebuild region has been moved there from the centre.
c) Clustering the models after VFF minimisation. A large number of the generated models are very similar in conformation, so a simple RMSD-based clustering is used on the lowest energy conformations to cut down the number.
d) Final screen. The SDR filtering has been replaced by a number of improved methods. These screens are amongst the key improvements in WAM and merit a good deal of detail, and are therefore described separately here. This page also shows the relative success of the alternative screens, together with recommendations for the optimum screen to use for differing length CDR-H3 loops.
Sidechain modelling
As has been seen, only the framework and conserved canonical sidechains are built early on. Those remaining are the canonical non-conserved sidechains, and the non-canonical sidechains, which require special, theoretical methods to build. This is also a key development in WAM compared to AbM and as such is discussed separately here .
Last updated 22/11/01