Secundary Structure of Physicochemical Clustered Proteins

Diverse methods have been proposed for protein secondary structure prediction. However, such task still presents a challenge in bioinformatics. In this article various of these methods are implemented and analysed. First, a baseline using Support Vector Machine. Then a convolutional neural network (CNN), a Long Short-Term Memory (LSTM) and a strategy of Ensembling both of these methods. Lastly, a novel technique Secundary Structure of Physicochemical Clustered Proteins (SSPCP) is proposed, which combines multiple CNNs trained accordingly to a protein feature clustering and combined using a neural network. The rationale behind SSPCP is that amino acids from proteins which have similar physicochemical characteristics should have the same secondary structure prediction for similar amino acids, but amino acids from differing proteins might have different structures. All of these methods use as features PSSM matrices extracted from PSIBLAST. For performance evaluation, 25pdb dataset was split into training and validation and the same subsets were used on all these methods achieving the Q3 score of CNN: 70.11%, LSTM: 69.25%, Ensemble: 70.71%, SSPCP: 70.91%. The experimental results show that the features extracted from clustering of physicochemical properties of proteins seem to improve the accuracy of highly specific CNN models for accurate protein secondary structure prediction.