Next generation software for placing atoms into electron density maps

X-ray crystallography is the most common method used to determine the three-dimensional structure of biological macromolecules, with proteins being of particular interest. Both the amplitudes and phases of the diffracted X-ray waves are needed to construct an electron density map, which is then interpreted by building an atomic model. However, the phases cannot be directly measured and must be estimated using either experimental phasing or molecular replacement. BUCCANEER is a program for automatically building protein models into electron density maps, which is iterated in a pipeline with global refinement to refine the model and update the map. The amount of time-consuming manual completion required depends on the success of automated building, which may fail in difficult cases with low resolution data or poorly estimated phases. The aim of this work was to improve automated building and therefore make structure solution quicker and easier. Potential developments to BUCCANEER were explored, but it was changes to the pipeline that proved to be most effective. The pipeline control system was updated and the following steps were added: shift-field refinement, classical density modification, addition of water and dummy atoms, pruning, and final rebuilding of side chains. The new pruning steps delete chains, residues and side chains using two neural networks, which were trained to predict main-chain and side-chain correctness by combining many validation metrics. The set of 54 experimental phasing cases previously used for testing BUCCANEER was expanded to 202 experimental phasing and 1351 molecular replacement cases. The combined pipeline changes substantially improved performance, increasing the mean completeness of the experimental phasing cases from 85% to 91% and the molecular replacement cases from 40% to 74%. The updated pipeline was released as a new program called ModelCraft.