Hybrid approach-I This combined amino acid and dipeptide composition-based SVM classifier was able to achieve an overall accuracy of 82.53% which was about 1% superior over the amino acid composition based SVM method (kernel=RBF, g=100, c=6, j=2). The individual accuracy obtained for four types of subcellular localization are shown below:

Hybrid approach-II Secondly, we developed another hybrid module by combining amino acid composition and PSSM based matrix. The SVM input vector pattern was 420 (20 for amino acid and 400 for PSSM). Best results were obtained with RBF kernel (g=45, c=2, j=4) with an overall accuracy of 84.84% which was about 2% superior over the hybrid approach - I based SVM method. The individual accuracy obtained for four types of subcellular localization are shown below:                                                              [ HOME ]   [ SUBMIT ]   [ TOP ]

Hybrid approach-III Further, we attempted another hybrid module by combining amino acid composition, dipeptide composition and PSSM matrix. The SVM input vector pattern increased to 820 (20 for amino acid, 400 for dipeptide composition and 400 for PSSM). Best results were obtained with RBF kernel (g=35, c=4, j=2) with an overall accuracy of 84.51% which was almost at par with hybrid approach-II. However, the overall accuracy of all the hybrid approaches attempted could not exceed the overall accuracy obtained with PSSM based SVM module alone. The individual accuracy obtained for four types of subcellular localization are shown below:

N-terminal amino acid composition Most proteins have sorting signals that relies on the presence of an N-terminal targeting sequence that is recognized by a translocation machinery. These signals are responsible for targeting proteins to various subcellular localizations in the cell. Therefore, we developed a SVM module based on the N-terminal amino acid composition of each protein having 20 SVM input vector pattern. The SVM module was developed at various levels of N-terminal residue length (10, 15, 20, 25 and 30 amino acids) in order to achieve maximum accuracy. Best results were obtained at 25 residue length with RBF kernel (g=30, c=1, j=3) with an overall accuracy of 70.88%. The individual accuracy obtained for four types of subcellular localization are shown below:                                                                                                                [ HOME ]   [ SUBMIT ]   [ TOP ]

C-terminal amino acid composition We also developed a SVM module based on the C-terminal amino acid composition of each protein having 20 SVM input vector pattern. Here, we also altered the C-terminal residue length (10, 15, 20, 25 and 30 amino acids) in order to achieve maximum accuracy. Best results were obtained with RBF kernel (g=40, c=1, j=3) with an overall accuracy of 64.18%. This accuracy was far lower than the N-terminal based accuracy indicating the presence of sorting signals in the N-terminus region of a protein sequence. The individual accuracy obtained for four types of subcellular localization are shown below:

Splitted Amino Acid Composition (SAAC) Best results were obtained with RBF kernel (g=6, c=2, j=2) with an overall accuracy of 79.78%. Within the terminal based SVM approaches, this method had the highest accuracy as compared to the accuracy of other terminal based SVM modules. The individual accuracy obtained for four types of subcellular localization are shown below:                                                           [ HOME ]   [ SUBMIT ]   [ TOP ]

N-terminal + Remaining part amino acid composition We also attempted a SVM module based on the division of protein into two parts viz. N-terminus and the remaining part of the sequence. The amino acid composition was calculated separately for both the parts so that it gave 40 (2 x 20) SVM vector pattern. Maximum overall accuarcy of 79.23% was obtained with RBF kernel (g=18, c=2, j=3). The individual accuracy obtained for four types of subcellular localization are shown below:

C-terminal + Remaining part amino acid composition Similarly, we attempted a SVM module based on the division of protein into two parts viz. C-terminus and the remaining part of the sequence. The amino acid composition was calculated separately for both the parts so that it gave 40 (2 x 20) SVM vector pattern. Maximum overall accuarcy of 76.48% was obtained with RBF kernel (g=15, c=2, j=2). This accuracy was lower than the N-terminal+remaining part composition based accuracy. The individual accuracy obtained for four types of subcellular localization are shown below: