Improving amino-acid identification, fit and C(alpha) prediction using the Simplex method in automated model building

Automated methods for protein model building in X-ray crystallography typically use a two-phased approach that involves first modeling the protein backbone followed by building in the side chains. The latter phase requires the identification of the amino-acid side-chain type as well as fitting of the side-chain model into the observed electron density. While mistakes in identification of individual side chains are common for a number of reasons, sequence alignment can sometimes be used to correct errors by mapping fragments into the true (expected) amino-acid sequence and exploiting contiguity constraints among neighbors. However, side chains cannot always be confidently aligned; this depends on having sufficient accuracy in the initial calls. The recognition of amino-acid side-chains based on the surrounding pattern of electron density, whether by features, density correlation or free atoms, can be sensitive to inaccuracies in the coordinates of the predicted backbone C(alpha) atoms to which they are anchored. By incorporating a Nelder-Mead Simplex search into the side-chain identification and model-building routines of TEXTAL, it is demonstrated that this form of residue-by-residue rigid-body real-space refinement (in which the C(alpha) itself is allowed to shift) can improve the initial accuracy of side-chain selection by over 25% on average (from 25% average identity to 32% on a test set of five representative proteins, without corrections by sequence alignment). This improvement in amino-acid selection accuracy in TEXTAL is often sufficient to bring the pairwise amino-acid identity of chains in the model out of the so-called ;twilight zone' for sequence-alignment methods. When coupled with sequence alignment, use of the Simplex search yielded improvements in side-chain accuracy on average by over 13 percentage points (from 64 to 77%) and up to 38 percentage points (from 40 to 78%) in one case compared with using sequence alignment alone.