University of Copenhagen Algorithms for Protein Structure Prediction

The problem of predicting the three-dimensional structure of a protein given its amino acid sequence is one of the most important open problems in bioinformatics. One of the carbon atoms in amino acids is the Cα-atom and the overall structure of a protein is often represented by a so-called Cα-trace. Here we present three different approaches for reconstruction of Cα-traces from predictable measures. In our first approach [63, 62], the Cα-trace is positioned on a lattice and a tabu-search algorithm is applied to find minimum energy structures. The energy function is based on half-sphere-exposure (HSE) and contact number (CN) measures only. We show that the HSE measure is much more information-rich than CN, nevertheless, HSE does not appear to provide enough information to reconstruct the Cα-traces of real-sized proteins. Our experiments also show that using tabu search (with our novel tabu definition) is more robust than standard Monte Carlo search. In the second approach for reconstruction of Cα-traces, an exact branch and bound algorithm has been developed [67, 65]. The model is discrete and makes use of secondary structure predictions, HSE, CN and radius of gyration. We show how to compute good lower bounds for partial structures very fast. Using these lower bounds, we are able to find global minimum structures in a huge conformational space in reasonable time. We show that many of these global minimum structures are of good quality compared to the native structure. Our branch and bound algorithm is competitive in quality and speed with other state-of-the-art decoy generation algorithms. Our third Cα-trace reconstruction approach is based on bee-colony optimization [24]. We demonstrate why this algorithm has some important properties that makes it suitable for protein structure prediction. Our approach for model quality assessment (MQA) [64] makes use of distance constraints extracted from alignments to templates. We show how to use CN probabilities in an optimization algorithm for selecting good distance constraints and we introduce the concept of non-contacts. When comparing our algorithm with state-of-the-art MQA algorithms on the CASP7 benchmark, our algorithm is among the top-ranked algorithms. We are currently participating in CASP8 MQA with this algorithm.