The Application of multiobjective optimisation to protein-ligand docking

Despite the intense efforts that have been devoted to the development of scoring
functions for protein-ligand docking, they are still limited in their ability to identify
the correct binding pose of a ligand within a protein binding site. A deeper
understanding of the intricacies of scoring functions is therefore essential in order to
develop these effectively. The aim of the work described in this thesis is to analyse
the individual interaction energy types which form the individual components of a
force field-based scoring function.
To do this, & protein-ligand docking algorithm that is based on multiobjective
optimisation has been developed. Multiobjective optimisation allows for the
optimisation of several objectives simultaneously and this has been applied to the
individual interaction energy types of the GRID scoring function. Traditionally these
interaction energy types are summed together and the total energy is used to guide the
search. By using individual energy types during optimisation, their roles can be better
understood. The interaction energy types that have been used here are the electrostatic
and hydrogen bond interactions combined, and van der Waals interactions.
The algorithm is first tested on two datasets containing twenty complexes. The results
show that the different interaction energy types have varying influences when it
comes to successfully docking certain complexes, and that it is important to fmd the
right balance of interaction energy types so as to find correct solutions. Ofthe twenty
complexes, the algorithm found correct solutions for fifteen.
To improve the performance of the algorithm, a few enhancements were introduced.
This includes a simplex minimisation process with a Lamarckian element. The
algorithm was retested on the twenty complexes, and the newer version was found to
outperform the original version, finding correct solutions for seventeen of the twenty
complexes. To extensively study the capabilities of the algorithm, it was tested on varied datasets,
including the FlexX dataset. The algoritlun's performance was also compared to a
single-objective docking tool, Q-fit. The comparison betw~en the multiobjective and
single-objective methodologies revealed that single-objective methods can sometimes
fail at finding correct docked solutions because they are unable to correctly balance
the interaction energy types comprising a scoring function. The study also showed
that a multiobjective optimisation method can reveal the reasons why a given docking
algorithm may fail at fmding a correct solution.
Finally, the algorithm was extended to incorporate desolvation energy as a third
objective. Though these results are preliminary, they revealed some interesting
relationships between the different objectives.