An assessment of the accuracy of methods for predicting hydrogen positions in protein structures

The addition of hydrogen atoms to models or experimental structures of proteins that contain only non-hydrogen atoms is a common step in crystallographic structure refinement, in theoretical studies of proteins, and in protein structure prediction. Accurate prediction of the hydrogen positions is essential, since they constitute around half of the atoms in proteins and hence contribute significantly to their energetics. Many computational tools exist for predicting hydrogen positions, although to date no quantitative comparison has been made of their accuracy or efficiency. Here we take advantage of the recent increase in ultra-high-resolution X-ray crystal structures (< 0.9 angstrom resolution), as well as of a number of relatively high-resolution neutron diffraction structures (< 1.8 angstrom resolution), to compare the quality of the predictions generated by a large set of commonly used methods. These include CHARMM, CNS, GROMACS, MCCE, MolProbity, WHAT IF, and X-PLOR. The hydrogen atoms that lack a rotational degree of freedom are mostly, but not always, accurately predicted. For hydrogens with a rotational degree of freedom, all the methods give much less accurate predictions. The predictions for the hydroxyl hydrogens are analyzed in detail, particularly those buried within the protein, and some explanation is provided for the errors observed. The results provide a means to make informed decisions regarding the choice and implementation of methodologies for placing hydrogens on structures of proteins. They also point to shortcomings in current force fields and suggest the need for improved descriptions of hydrogen bonding energetics.