| Summary: | Fragment-based approaches are the current standard for de novo protein structure prediction. These approaches rely on accurate and reliable fragment libraries to generate good structural models. We demonstrate that fragments presenting different predominant predicted secondary structures should be treated differently during fragment library generation. Using this information, we developed Flib and shown that it generates fragment libraries with higher precision and coverage than two other methods. We explored co-evolution to identify pairs of residues that are in contact, which were then used to improve model generation. We performed a comparative analysis of nine methods in terms of their precision and their usefulness to de novo structure prediction. Our results show that metaPSICOV stage 2 produces the most accurate predictions and that metaPSICOV stage 1 generates the best modelling results. In general, contact predictors are good at identifying contacts between β-strands and bad at identifying contacts between a-helices. We also show that the ratio of satisfied predicted contacts can be used to assess whether correct models were generated for a given target. We also investigated whether the biological process of cotranslational protein folding, the notion that proteins fold as they are being synthesized, can be used to improve de novo protein structure prediction. Our tool for this investigation is SAINT2. SAINT2 differs from conventional fragment-assembly approaches as it is able to perform predictions sequentially from N to C-terminus, starting with a small peptide that is extended as the simulation progresses (SAINT2 Cotranslational). SAINT2 is also able to generate decoys in a standard non-sequential fashion (SAINT2 In Vitro). We compared SAINT2 Cotranslational to SAINT2 In Vitro and shown that SAINT2 Cotranslational generally produces better answers, generating an individual decoy between 1.5 to 2.5 times faster than SAINT2 In Vitro. Our results suggest that biologically inspired structure prediction can improve search heuristics and final model quality.
|
|---|
