Methods that can automatically assess the quality of computationally predicted protein structures are important, as they enable the selection of the most accurate structure from an ensemble of predictions. Assessment methods that determine the quality of a protein's structure by comparing it against the various structures predicted by different servers have been shown to outperform approaches that rely on the intrinsic characteristics of the structure itself.
We developed an algorithm to estimate the quality of a predicted protein structure using a consensus approach. Our method uses LGA to align the structure in question to the structures for the same protein predicted by different servers and estimates the per-residue error by averaging the distances across these alignments. On a dataset containing 892,299 positions from 4,969 CASP7 submissions, our method achieved a root mean squared error (RMSE) of 6.69angstroms, which is significantly better than the 8.21angstroms achieved by the winning scheme in CASP7 for this problem (Pcons). We further improved these results to 6.51angstroms by developing a scheme that carefully selects which distances to average based on the predicted quality of the overall structure. We also examined the use of machine learning approaches to learn an appropriate aggregation scheme, which led to a simple weight learning approach achieving a 2.61angstroms RMSE on a reduced dataset.
Our results show that the use of LGA alignments and aggregation of raw distances is the primary reason for its performance advantage. In addition, our results show that beyond a binary inclusion/exclusion decision, learning from the data a set of weights by which the structures of the different servers can be aggregated can further improve performance.