Research article

Benchmarking pK_a prediction

Matthew N Davies^1* , Christopher P Toseland^1* , David S Moss² and Darren R Flower¹

¹  Edward Jenner Institute for Vaccine Research, Compton, Berkshire, RG20 7NN, UK

²  School of Crystallography, Birkbeck College, Malet Street, London WC1E 7HX, UK

author email corresponding author email* Contributed equally

BMC Biochemistry 2006, 7:18doi:10.1186/1471-2091-7-18

Published:	2 June 2006

Abstract

Background

pK_a values are a measure of the protonation of ionizable groups in proteins. Ionizable groups are involved in intra-protein, protein-solvent and protein-ligand interactions as well as solubility, protein folding and catalytic activity. The pK_a shift of a group from its intrinsic value is determined by the perturbation of the residue by the environment and can be calculated from three-dimensional structural data.

Results

Here we use a large dataset of experimentally-determined pK_as to analyse the performance of different prediction techniques. Our work provides a benchmark of available software implementations: MCCE, MEAD, PROPKA and UHBD. Combinatorial and regression analysis is also used in an attempt to find a consensus approach towards pK_a prediction. The tendency of individual programs to over- or underpredict the pK_a value is related to the underlying methodology of the individual programs.

Conclusion

Overall, PROPKA is more accurate than the other three programs. Key to developing accurate predictive software will be a complete sampling of conformations accessible to protein structures.