A new multidimensional potential is described that encodes for the relative spatial arrangement of the peptidyl backbone units as observed within a large database of high-resolution X-ray structures. The detailed description afforded by such an analysis provides an opportunity to study the atomic details of hydrogen bonding in proteins. The specification of the corresponding potential of mean force (PMF) is based on a defined set of physical principles and optimized to yield the maximum advantage when applied to protein structure refinement. The observed intricate differences between hydrogen-bonding geometries within various patterns of secondary structure allow application of the PMF to both validation of protein structures and their refinement. A pronounced improvement of several aspects of structural quality is observed following the application of such a potential to a variety of NMR-derived models, including a noticeable decrease in backbone coordinate root-mean-square deviation relative to the X-ray structures and a considerable improvement in the Ramachandran map statistics.