Protein Structure and Function: Application of Bioinformatics Methods - John Rigden 2014

Fold Recognition
Remote Homology Detection Without Alignment
Consensus Approaches

Recent CASP experiments have demonstrated that consensus Methods, which combine data from multiple Structure/29.html">Fold Recognition servers into a unified prediction, offer a significant advantage. These "metaservers" consistently outperform many of the individual methods upon which they are built. Such individual methods include sequence-to-profile alignment, Hidden Markov Models, profile-profile alignment, and threading.

Among the most popular techniques that combine predictions within metaservers are Peons (Wallner and Elofsson 2005), 3D-Shotgun (Fischer 2003), and 3D-Jury (Ginalski et al. 2003). The simplest yet remarkably efficient approach is 3D-Jury. This method compares 3D models generated by different servers through structural alignment. The models are then re-evaluated based on their structural similarity to the other models in the group. Consequently, if several relatively independent fold recognition systems select similar templates and subsequently produce comparable alignments, these models will receive higher scores compared to other, less typical models. The Peons method combines this 3D-Jury approach with a neural network trained to distinguish models that possess properties common to all protein structures from those that lack them (much like the Empirical energy function used in threading). Finally, the 3D-Shotgun method calculates a 3D-Jury score for each residue in every model, after which a new model is assembled from the most common, or "consensus," parts. While this can lead to severe model fragmentation—and a series of experiments have been aimed at mitigating this drawback—the issue remains unresolved.

An extensive investigation into the reasons behind the high performance of metaservers was conducted by Bennett-Lovsey et al. (2008). The authors concluded that the improvements stem largely not from identifying distant homologs per se, but from enhanced accuracy, i.e., the elimination of false positives. This phenomenon occurs because when multiple diverse structure prediction systems are combined, the probability of all of them making the same error is significantly lower than the probability of obtaining a concordant result. Any sequence feature that might cause one or two prediction methods to fail is unlikely to have the same impact on the majority of methods. Combining classifiers and prediction algorithms into ensembles to boost performance is a well-established research field bridging statistical pattern recognition and machine learning (Jain et al. 2000; Kuncheva and Whitaker 2003). Unfortunately, even after decades of research, fundamental theory still does not provide a blueprint for constructing optimal ensembles. As a result, trial and error remains the primary guiding principle in metaserver design.