Estimating the Class Posterior Probabilities in Protein Secondary Structure Prediction

Abstract

Abstract. Support vector machines, let them be bi-class or multi-class, have proved efficient for protein secondary structure prediction. They can be used either as sequence-to-structure classifier, structure-to-structure classifier, or both. Compared to the classifier most commonly found in the main prediction methods, the multi-layer perceptron, they exhibit one single drawback: their outputs are not class posterior probability estimates. This paper addresses the problem of post-processing the outputs of multi-class support vector machines used as sequence-to-structure classifiers with a structure-to-structure classifier estimating the class posterior probabilities. The aim of this comparative study is to obtain improved performance with respect to both criteria: prediction accuracy and quality of the estimates.

Effect of simple ensemble methods on protein secondary structure prediction

“…Different post-processing can be applied to map SVMs and MSVMs outputs into class posterior probabilities (Platt 2000). The quality of posterior probability estimates is subject to many recent studies (Zhang and Jordan 2006;Guermeur and Thomarat 2011;Wallace 2012). Here, we used the softmax function which is the most common mapping.…”