New bulk-solvent models improves model-to-data fit and facilitates map interpretation

Bio-macromolecular crystals contain between 10 and 90% of solvent. This solvent is mostly disordered so it cannot be interpreted in terms of an atomic model. Owing to its simplicity and yet relatively good modeling power, flat bulk-solvent model is the most commonly used model to account for disordered solvent in modern crystallographic software packages such as CNS, CCP4 or Phenix. This model assumes electron density is constant anywhere in the unit cell where there is no atomic model placed. While this may be a reasonable approximation for some crystal structures or at initial stages of structure determination, it may be less accurate at final stages. Major deviations from the assumption of a flat model include: 1) local concentration of solvent component of specific types, such as lipid belts in membrane proteins, 2) unmodeled ligands, 3) partial occupancy of solvent in small isolated regions (between or inside macromolecules), and 4) lack of solvent in certain regions (hydrophobic cores). These deviations manifest themselves as elevated R factors in the lowest resolution shells as well as residual features in difference maps: positive in cases when the flat solvent model is inadequate in accounting for distinct features, or negative when solvent model is used in regions with no solvent. To overcome the limitations of the existing bulk solvent model we have proposed a non-uniform bulk-solvent model that allows for solvent variation across the unit cell volume. The new model splits initially binary (0/1) solvent masks into several masks by applying connectivity analysis. These masks are then split further into more masks based on analysis of difference maps. This final set of solvent masks is used to compute the individual bulk solvent contributions to the total model structure factor. Tests on all deposited structures in PDB that have diffraction data and cross-validation flags available indicate systematic improvement of model-to-data fit with no signs of over-fitting as judged by the Rfree factor. Tests on selected models demonstrate notable improvements in map quality especially for weak features such as ligands or solvent molecules. All described tools will be available in Phenix.