Evaluating protein structure-prediction schemes using energy landscape theory

Protein structure prediction is beginning to be, at least partially, successful. Evaluating predictions, however, has many elements of subjectivity, making it difficult to determine the nature and extent of improvements that are most needed. We describe how the funnel-like nature of energy functions used for protein structure prediction determines their quality and can be quantified using landscape theory and multiple histogram sampling methods. Prediction algorithms exhibit a "caldera"-like landscape rather than a perfectly funneled one. Estimates are made of the expected number of effectively distinct structures produced by a prediction algorithm.