QSAR Modeling Of Bioactive Peptide Activity Based On Geostatistics And SVR

Bioactive peptide is a short peptide molecule that plays a very important role in life, which have potential applied value in many fields. Predicting the activity of bioactive peptides precisely is the theoretical basis for finding or designing of highly active bioactive peptides. Quantitative sequence-activity relationship, QSAR, is a popular method for the prediction of the activity of bioactive peptides. Generally the generating of a QSAR model contains three steps:descriptor extraction, descriptor selection, regression model. Hence, this article improved the generating of a QSAR model from these three aspects. We used two datasets to test the performance of our new QSAR model. One dataset is angiotensin converting enzyme inhibitor peptides dataset and the other is antimicrobial peptides dataset. This content mainly include:(1) In the extracting descriptors, geostatistics can capture the correlation of spatial, so we utilized this characteristic and extracted association descriptor based on geostatistics and the physical and chemical properties of amino acids. In consideration of the amino acid residue effect of based on specific physical and chemical properties of bioactive peptide activity, at the same time considering the different amino acid residues between interactions. Through statistical calculation, the magnitude of association can be reflected in the descriptor. The result showed association descriptors extracted by geostatistics and the physical and chemical properties of amino acids can express the information of sequence better, especially when the activity of bioactive peptides was more affected by the interaction effect between amino acids, association descriptor extracted by geostatistics had better forecasting results than direct extracted by physical and chemical properties.(2) In the selection of descriptors, we absorbed the advanced idea of Minimum Redundancy Maximum Relevance, which was a popular feature selection method in solving classification problems. The improved MRMR descriptor selection method can be used in QSAR models built, which is a regression problem that composed of continuous variable, at the same time, we introduced descriptors one by one for a second selection. The results showed our improved MRMR descriptor selection method eliminated redundant descriptors in the initial descriptors effectively, and it reduced the complexity, improved the prediction performance, and enhanced the interpretative nature of QSAR model. Furthermore, the retained descriptors which selected by our improved MRMR descriptor selection method had good interpretation and biological significance. Through the analysis of those retained descriptors, we explored the residue position and high frequency descriptors which affected the activity of angiotensin converting enzyme inhibitor peptides, and we summarized the general rules of how the physical and chemical properties of amino acids affect the activity of antimicrobial peptides.(3) In the construction of regression model, we used support vector regression as a basic egression model, and we applied geostatistics to select nearest neighbor samples for a private prediction. The results showed private prediction based on geostatistics further improved the prediction performance of our QSAR model, our prediction performance were all better than other models.Our new method has good prediction ability and interpretation in the prediction of the activity of bioactive peptide, which would have widespread application prospect in the domain of high dimensional regression and other QSAR researches.