Comparative Study of Particle Swarm Approaches for the Prediction of Functionally Important Residues in Protein Structures

Abstract

Prediction of functionally important amino acids in protein structures is challenging problem in the area of protein function prediction. In the quest of looking for better machine learning approaches to address this problem, we have compared a support vector machine and a neural network trained with a particle swarm algorithm (PSO) to nonlinearly combine a subset of features selected from a set of 314 features describing catalytic residues in protein structures. We compare this approach against three other approaches. Results show trade-offs for two of the approaches on the precision-recall curves. While no approach surpassed the performance of the linear kernel support vector machine (SVM) classifier, the performance of the PSO was comparable to that of a feature selected SVM.