Adjusting Local Conformational Sampling For Fragment Assembly Protein Structure Prediction Based On Secondary Structure Complexity

Fragment assembly protein structure prediction is one of the most successful methods whenever reliable templates (for homology-based approaches) and/or massive computational resources (for physics-based approaches) are not available. However, it suffers from important limitations: tremendous search space, energy scores inaccuracy, and consequently the large number of decoys which are needed to be generated. Taking advantage of the different protein sequence-structure complexity shown by the various types of secondary structure, - using Rosetta - we propose to customize the diversity of fragments for each region of the conformation being built. By eventually reducing the size of search space, this approach permits better exploitation of promising areas. Experiments demonstrate the value of the proposed strategy: compared to standard Rosetta's performance in terms of first model, accuracy improves significantly (~6%), respectively drastically (~24%), when using 20,000, resp. 2,000, decoy-based predictions. Furthermore, performance using 2,000 decoys is equivalent to that of standard Rosetta using 20,000 decoys, which means that predictions can be executed on a standard PC instead of a highperformance computing system.