Technical Report STP: The Sample-Train-Predict Algorithm and Its Application to Protein Structure Meta-Selection

The importance and the difficulty of the folding problem have led scientists to develop several computational methods for protein structure prediction. Despite the abundance of protein structure prediction methods, these approaches have two major limitations. First, the top ranked model reported by a server is not necessarily the best predicted model. The correct predicted model may be ranked within the top 10 predictions after some false positives. Second, no single method can give correct predictions for all proteins. To attempt to remedy these limitations, protein structure prediction “meta” approaches have been developed. A metaserver can select a set of candidate models by ranking models obtained from other servers. In this article we present the Sample-Train-Predict algorithm and its application to implement a new model quality assessment program (MQAP) based on a consensus of five MQAP’s then we discuss the application of our MQAP as a meta-selector. STP depends on the clustered nature of the training data and it can dynamically handle constantly growing training data. STP selects clusters which are similar to the input data, then trains a model on these clusters, and finally uses the trained model to get predictions for the input data. Our experimental results show that a hierarchical model trained using STP outperforms any tested model quality assessment program by 7%-8%. When selecting from predictions made by humans in a standard benchmark CASP7, our meta-selector achieves about 3% improvement above the best human predictor.