Analysis and assessment of ab initio three-dimensional prediction, secondary structure, and contacts prediction

CASP3 saw a substantial increase in the volume of ab initio 3D prediction data, with 507 datasets for fifteen selected targets and sixty-one groups participating. As with CASP2, methods ranged from computationally intensive strategies that attempt to recreate the physical and chemical forces involved in protein folding to the more recent knowledge-based approaches. These exploit information from the structure databases, extracting potentially similar fragments and/or distance constraints derived from multiple sequence alignments. The knowledge-based approaches generally gave more consistently successful predictions across the range of targets, particularly that of the Baker group (Bystroff and Baker, J Mol Biol 1998;281:565-577; Simons et al. Proteins Suppl 1999;3:171-176), which used a fragment library. In the secondary structure prediction category, the most successful approaches built on the concepts used in PHD (Rost et al. Comput Appl Biosci 1994;10:53-60), an accepted standard in this field. Like PHD, they exploit neural networks but have different strategies for incorporating multiple sequence data or position-dependent weight matrices for training the networks. Analysis of the contact data, for which only six groups participated, suggested that as yet this data provides a rather weak signal. However, in combination with other types of prediction data it can sometimes be a useful constraint for identifying the correct structure.