Analysing arid predicting the detailed three dimensional conformation of protein structures is a critical and important task within structural bioinformatics with impact on other fields, e.g.. drug design and delivery, sensing technologies, etc. Unfortunately, it is hard to identify one methodology that will give the best prediction of the three-dimensional structure for any sequence. That is, some predictors are best suited for some sequences and not for others. In trying to address this drawback of current prediction algorithms the research community introduced the concept of protein prediction metaservers. In this paper we propose a new metaserver method called 3D-Judge that uses an artificial neural network (ANN) to select the best model from among models produced by individual servers. The fundamental innovation we introduce is that the AXN is not only used to decide which models and servers to use as good predictions but, crucially, it is also used to analyse and "remember" the past performances of the servers it has access to. Thus, our method acts as both a kind of majority voting algorithm, by selecting models arising from different servers based on their mutual similarity, and also a reinforced learning method that takes cues from historical data of previously solved structures. We train and evaluate our metaserver based on previous GASP results and we compare SD-Judge with a popular and effective metaserver, namely. 3D-Jury. The obtained results indicate that 3D-Judge is competitive with 3D-Jury, outperforming it on many cases. We also present a discussion on future extensions to 3D-Judge.
JavaScript jest wyłączony w Twojej przeglądarce internetowej. Włącz go, a następnie odśwież stronę, aby móc w pełni z niej korzystać.