In silico Secondary Structure Prediction Method (Kalasalingam University Structure Prediction Method) using Comparative Analysis

MATERIALS AND METHODS

This study was conducted at Kalasalingam University, Bioinformatics Laboratory in 2007-2009.

HNN
Hierarchial neural network prediction method can be seen as an improvement on the famous classifier developed by QIAN and SEJNOWSKI derived from the system NET-TALK (Sivan et al., 2007). This is mainly made up of two types of network; they are sequence to structural network and structure to structure network. This improvement mainly deals with two points. They are many technical tricks have been used to increase the content on which the prediction is made and concominatantly decrease by two orders of magnitude of the number of parameters. Many physico-chemical data have been explicity incorporated in the predictors used by the structure to structure network.

SOPMA
Self optimized prediction method is based on the homologue method of Levin et al. (1993). This method correctly predicts 69.5% of amino acids for a three description of the secondary structure (α-helix, β-sheet and random coil) in a whole database containing 126 chains of non-homologous proteins on combination of SOPMA and PHD methods correctly predicts 82.2% of residues.

ISPBD
In my earlier methods, ISPBD (Innovative Structure Prediction using Bioinformatics Databases), was developed to predict the secondary structure of the proteins from amino acid sequences using the generated structure prediction parameters. In the above method β-sheet is much better prediction than SSPDP (Mugilan and Veluraja, 2000) PHD (King et al., 1997; Rost et al., 1994) DSC (Rost and Sander, 1993), NNSSP (Salamov and Solovyev, 1995) and NNPREDICT (Kneller et al., 1990) methods.

PHD
The prediction (King et al., 1997; Rost et al., 1994) is well balanced between alpha-helix, beta-strand and the loop-65% of the observed strand residues are predicted correctly (Rost et al., 1994). The accuracy in predicting the content of three secondary structure types is comparable to that of circular dichroism spectroscopy.

SSPRO
The SSPRO belongs to the scratch protein predictor and are shown in many versions. It is a server of protein secondary structure prediction (Pollastri and Mclysaght, 2005). The SSPRO includes the direct incorporation of homologous proteins secondary structure and probabilistic methods to improve the accuracy. Its output is different from that of all the other methods.

METHOD OF CALCULATION

We have retrieved the sequential information from selected 377 non-homologous proteins PDB databases for this analysis. Collected primary sequences of the selected protein were submitted to the various structure prediction servers like HNN, SOPMA, ISPBD, SSPRO and PHD. Secondary structural information of the selected proteins were carried out from the output of the various secondary structure methods using PYTHON programming. In other way secondary structural information were collected from DSSP output for the selected proteins. Prediction Accuracy (PA) for various methods were obtained from the following formula.

Using the prediction accuracy we find out a new method KLUSP. Prediction Accuracy (PA) calculated chart for the various methods are shown in Fig. 1.

Fig. 1:	Flow chart for Prediction accuracy calculation

Fig. 2:	Flow chart for Prediction accuracy calculation

Flow chart for the in silico secondary structure prediction method indicates that, KLUSP was rendered based on comparative analysis (Fig. 2). Chart explains that on comparison of all prediction accuracy the higher prediction accuracy (High PA) for α-helix and β-sheet were found out and combine together to carried out KLUSP.

RESULTS AND DISCUSSION

Sequential information for 377 non-homologous proteins were collected from the PDB databases. The collected sequences were submitted to the various secondary structure methods server page and found out the structural information for the above sequences. Structural information for selected proteins (377) were also collected from the DSSP out put (Fig. 3).

Output of DSSP and various secondary structure prediction methods of model sequence were plotted. Prediction accuracy for various methods were calculated based on DSSP results using the above formula. α -helix and β-sheet Prediction accuracy for selected proteins were calculated using python programming. Prediction accuracy results of 10 proteins were calculated (Table 1).

Average prediction accuracy for α-helix and β-sheet for selected proteins (377) were calculated. α-helix average prediction accuracy for HNN, SOPMA, ISPBD, SSPRO and PHD are 64.12, 78.79, 79.62, 77.59 and 80.39, respectively and β-sheet average prediction accuracy for HNN, SOPMA, ISPBD, SSPRO and PHD are 50.65, 57.95, 78.21, 67.26 and 64.56, respectively.

Among the methods used for our comparison, SSPRO was known to be of highest prediction accuracy for helix and ISPBD has highest Prediction accuracy for sheet.

We found out the in silco method (KLUSP) from the output of secondary structure prediction methods SSPRO and ISPBD. In the first step, we have to assign sequence for structure prediction. It was submitted to SSPRO and ISPBD methods. α-helical structural information from SSPRO output and β sheet structural information from ISPBD output were alone taken from among all the α-helix and β-sheet structural elements. Thus, combining these two methods we can generate a new method known as KLUSP. Now by comparing KLUSP with various secondary structure prediction methods and then prediction accuracy was calculated for the above methods. Structural information of the various methods and DSSP output for 1G3P protein were analysed. In the Table 2 sequence number 6-10 favour the α-helical structure in SSPRO and KLUSP methods, but the sequence number 18-20 favour the β-sheet in ISPBD and KLUSP methods. For the above information overall prediction accuracy for KLUSP is much better than ISPBD and SSPRO methods. The result clearly indicate that KLUSP has higher accuracy than the SSPRO and PHD method.

Fig. 3:	Comparison of secondary structures

Table 1:	Prediction accuracy results of randomly selected 10 proteins
HNN: Hierarchial neural network, SOPMA: Seif optimized prediction method, ISPBD: Innovative structure prediction using bioinformatics databases, PHD: Prediction method, SSPRO: Scratch protein predictor

The following proteins are randomly selected and compared with new method KLUSP. Prediction accuracy of proteins 1LIT, 1XNB, 1A7S, 1G3P, 1RHS and 3SEB is shown in Table 3. Prediction accuracy on analyzing these proteins with KLUSP, it is on the whole good. Average prediction accuracy of α-helix for various methods HNN, SOPMA, ISPBD, PHD, SSPRO and KLUSP is 67.18, 85.15, 78.5, 71.42, 88.77 and 88.7. In the result KLUSP is comparable to SSPRO.

Similarly, Table 4 indicates the average prediction accuracy of β-sheet for various methods HNN, SOPMA, ISPBD, PHD, SSPRO and KLUSP is 50.72, 56.23, 73.23, 70.57, 63.25 and 72.9. From the result KLUSP is comparable to ISPBD.

Overall (α-helix + β-Sheet) prediction accuracy of various methods for the above selected proteins are shown in Fig. 4. From overall prediction accuracy of KLUSP for the protein 1LIT and 1XNB has been comparable to SSPRO. But our method has higher prediction accuracy than other methods like HNN, SOPMA, ISPBD, PHD and SSPRO.

Average prediction accuracy(overall) for various methods HNN, SOPMA, ISPBD, PHD, SSPRO and KLUSP is 58.95, 70.69, 75.87, 70.99, 76.01 and 80.80, respectively. The above result clearly indicates that our method (KLUSP) is better than other methods. So predictions made by KLUSP better than a prediction using for instance of all methods.

Table 2:	Structural information of the various methods and DSSP output for 1 G3P protein
H: alpha helix, E: beta sheet, SEQ: sequence, Amino acid: A,E,T,V,E,S, HNN: Hierarchial neural network, SOPMA: Seif optimized prediction method, ISPBD: Innovative structure prediction using bioinformatics databases, PHD: Prediction method, SSPRO: Scratch protein predictor

Table 3:	Average prediction accuracy for alpha
HNN: Hierarchial neural network, SOPMA: Seif optimized prediction method, ISPBD: Innovative structure prediction using bioinformatics databases, PHD: Prediction method, SSPRO: Scratch protein predictor

Table 4:	Average prediction accuracy for beta
HNN: Hierarchial neural network, SOPMA: Seif optimized prediction method, ISPBD: Innovative structure prediction using bioinformatics databases, PHD: Prediction method, SSPRO: Scratch protein predictor

Fig. 4:	Overall (α-helix + β-sheet) prediction accuracy of various methods

DISCUSSION

Protein Secondary structure prediction is an important step towards understanding how proteins fold in three dimensions. Our recent analysis of many proteins with the databases HNN, SOPMA, GORIV, PHD, SSPRO is done for keeping the target for improvement in protein secondary structure prediction. These Bioinformatics Databases are libraries of life sciences information, collected from scientific experiments, published literature, high throughput experiment technology and computational analysis. They contain information from research areas including Genomics, Proteomics, Metabolomics, Microarray gene expression, Phylogenetics. Information contained in biological databases includes gene function, structure, localization (both cellular and chromosomal), clinical effects of mutations as well as similarities of biological sequences and structures. Biological database design, development and long-term management is a core area of the discipline of Bioinformatics. So predictions made by KLUSP, better than a prediction using for instance of all methods. Some of the significance of our KLUSP method is overall accuracy of 81% and Beta sheets predict with an accuracy of 84%.

CONCLUSION

Our new method completely revolutionized protein secondary structure predictions, KLUSP, taking it into an area where, it actually is very useful. For instance you will achieve higher accuracy in secondary structure prediction from the modern methods than other methods so far we compared. And the accuracy is so high that it is often the first method used when trying to predict the structure of a protein. In future, by using this higher predicting method, we can improve the design of drug and molecular modeling.

ACKNOWLEDGMENT

The author thanks Bioinformatics, KLU, who provided the computational facility for this study. The author would also like to thank the funding agency DST, Newdelhi.