Sequential search leads to faster, more efficient fragment-basedde novoprotein structure prediction

Abstract

MotivationMost current de novo structure prediction methods randomly sample protein conformations and thus require large amounts of computational resource. Here, we consider a sequential sampling strategy, building on ideas from recent experimental work which shows that many proteins fold cotranslationally.ResultsWe have investigated whether a pseudo-greedy search approach, which begins sequentially from one of the termini, can improve the performance and accuracy of de novo protein structure prediction. We observed that our sequential approach converges when fewer than 20 000 decoys have been produced, fewer than commonly expected. Using our software, SAINT2, we also compared the run time and quality of models produced in a sequential fashion against a standard, non-sequential approach. Sequential prediction produces an individual decoy 1.5–2.5 times faster than non-sequential prediction. When considering the quality of the best model, sequential prediction led to a better model being produced for 31 out of 41 soluble protein validation cases and for 18 out of 24 transmembrane protein cases. Correct models (TM-Score > 0.5) were produced for 29 of these cases by the sequential mode and for only 22 by the non-sequential mode. Our comparison reveals that a sequential search strategy can be used to drastically reduce computational time of de novo protein structure prediction and improve accuracy.Availability and implementationData are available for download from: http://opig.stats.ox.ac.uk/resources. SAINT2 is available for download from: https://github.com/sauloho/SAINT2.Supplementary information Supplementary data are available at Bioinformatics online.

RFQAmodel: Random Forest Quality Assessment to identify a predicted protein structure in the correct fold

“…by EigenTHREADER for each model as a feature. 60% of the targets in each set, in line with numbers reported previously [1].…”

“…In particular, proteins in the Validation set with B eff <100 102 tended to be longer than proteins on the Training set with B eff <100, which suggests 103 that the Validation set may be more challenging for protein structure prediction. 104 Protein Structure Prediction 105 To produce models for all targets in our Training and Validation sets, we used our 106 fragment-assembly protocol SAINT2 [1] (for details, see SI Section 4 and [1]) with the 107 parameters given in the original publication, with the exception of secondary structure 108 prediction. We used DeepCNF Q8 to predict secondary structure, as DeepCNF Q8 had 109 a slightly higher precision for targets with large B eff values, and results in marginal calculated across all models produced for each target.…”

“…Contact-based features (2): The contact component of the SAINT2 score 162 (see [1] for more details) and the proportion of satisfied predicted contacts (positive 163 predictive value, PPV). Here, we considered a predicted contact to be a satisfied if the 164 C-β atoms (C-α in the case of glycine) of the two residues predicted to be in contact 165 were less than 8Å apart in the model output by SAINT2.…”

RFQAmodel: Random Forest Quality Assessment to identify a predicted protein structure in the correct fold

“…Template-free protein structure prediction protocols routinely produce hundreds to thousands of models for a given target [1]. Users need to be able to identify if a good model exists in this ensemble.…”

“…We ensured that these sets were well-balanced in terms of protein length, number of effective sequences [7], SCOP class [13], and other properties that are known to have an effect on modelling success. We used our sequential protein structure prediction protocol SAINT2 [1] to generate 500 models for each of the 488 protein domains. Using the Training set, we show that predicted contact map alignment scores are as effective for ranking models as existing state-of-the-art quality assessment scores.…”