An evaluation of automated model-building procedures for protein crystallography

The computer programs ARP/wARP, MAID and RESOLVE are designed to build protein structures into experimentally phased electron-density maps without any user intervention, requiring only diffraction data and sequence information. However, the MAID and RESOLVE systems, which seek to extend the range of automated model-building to approximately 3 A resolution, have yet to receive significant testing outside the small numbers of data sets used in their development. Since these two systems employ a large number of scoring functions and decision-making heuristics, additional tests are required to establish their usefulness to the crystallographic community. To independently evaluate these programs, their performance was tested using a database containing 41 experimentally phased maps between 1.3 and 2.9 A resolution from a diverse set of protein structures. At resolutions higher than 2.3 A the most successful program was ARP/wARP 6.0, which accurately built an average of 90% of the main chain. This system builds somewhat larger fractions of the model than the previous version ARP/wARP 5.1, which accurately built an average of 87% of the main chain. Although not specifically designed for model building into high-resolution maps, MAID and RESOLVE were also quite successful in this resolution regime, typically building approximately 80% of the main chain. At 2.3-2.7 A resolution the MAID and RESOLVE programs automatically built approximately 75% of the main-chain atoms in the protein structures used in these tests, which would significantly accelerate the model-building process. Data sets at lower resolution proved more problematic for these programs, although many of the secondary-structure elements were correctly identified and fitted.