Automatic modeling of protein backbones in electron-density maps via prediction of C^α coordinates

Most crystallographers today solve protein structures by first building as much of the protein backbone as possible and then modeling the side chains. Automating the determination of backbone coordinates by computer-based interpretation of the electron density would enhance the speed and possibly improve the accuracy of the structure-solution process. In this paper, a new computational procedure called CAPRA is described that predicts coordinates of C^α atoms in density maps and outputs chains of C^α atoms representing the backbone of the protein. The result constitutes a significant step beyond tracing the density, because there is ideally a one-to-one correspondence between atoms predicted in the chains output by CAPRA and C^α atoms in the true structure (refined model). CAPRA is based on pattern-recognition techniques, including extraction of rotation-invariant numeric features to represent patterns in the density and use of a neural network to predict which pseudo-atoms in the trace are closest to true C^α atoms. Experiments with several MAD and MIR electron-density maps of 2.4–2.8 Å resolution reveal that CAPRA is capable of building ∼90% of the backbone of a protein molecule, with an r.m.s. error for C^α coordinates of around 0.9 Å.