Explorations Of The Influence Factors On Accuracies Of Protein Secondary Structure Prediction

One of the most persistent problems in bioinformatics has been the unraveling of the protein primary structure to their unique tertiary structure. Most current protein secondary structure prediction programs employ multiple sequence alignments to capture local sequence patterns as input information for machine learning techniques. However, such local sequence patterns ignore the amino acids'intrinsic propensities for three states of the secondary structure, namely, n-helices,β-strands, or others (often referred to as coils). For this reason, we propose an approach to integrate the multiple sequence alignment profiles with amino acid propensities for machine learning input coding schemes. The position specific scoring matrices (PSSM) from PSI-BLAST were integrated with amino acid conformation parameters and hydrophobicity properties for protein secondary structure prediction with support vector machines (SVMs).The paper described SVM-based method with hydrohpobicity and HEC propensity with PSSM to predict protein secondary structure, which also used a two-layer SVM. The result analysis with correlative coefficient showed that the hydrophobicity and the HEC propensity had little relationship with the Q3 results but they had obviously relationship with their SOV results. The two-layer SVM technique showed improvement on both Q3 and SOV. The integrated method increased Q3 and SOV by 2.76% and 1.25% respectively.