Evaluation of comparative protein structure modeling by MODELLER-3

We evaluate homology-derived 3D models of dihydrofolate reductase (DFR1), phosphotransferase enzyme IIA domain (PTE2A3), and mouse/human UBC9 protein (UBC9(24)) which were submitted to the second Meeting on the Critical Assessment of Techniques for Protein Structure Prediction (CASP). The DFR1 and PTE2A3 models, based on alignments without large errors, were slightly closer to their corresponding X-ray structures than the closest template structures. By contrast, the UBC9(24) model was slightly worse than the best template due to a misalignment of the N-terminal helix. Although the current models appear to be more accurate than the models submitted to the CASP meeting in 1994, the four major types of errors in side chain packing, position and conformation of aligned segments, position and conformation of inserted segments, and in alignment still occur to almost the same degree. The modest improvement probably originates from the careful manual selection of the templates and editing of the alignment, as well as from the iterative realignment and model building guided by various model evaluation techniques. This iterative approach to comparative modeling is likely to overcome at least some initial alignment errors, as demonstrated by the correct final alignment of the C terminus of DFR.