Protein-Protein Interface Prediction based on a Novel SVM Speedup

Protein-protein interactions play a crucial role in many cellular processes. Prediction of amino acid residues that appear in interaction sites helps decipher protein functions. Since a significant number of complexes have large enough interfaces, we hypothesize that the complex formation follows the induced-fit mechanism rather than the lock-and-key mechanism. Therefore, one should be able to characterize interface regions by frequent appearances of unstructured or flexible amino acid residues in those regions. For this residue prediction problem, we designed a novel method called “tree decomposition support vector machine” (TDSVM) that can handle large samples. Previously, the sizes of protein chains used as training data were generally in the scope of hundreds, whereas TDSVM extends the number to thousands (4,064 in our case), which yields more than a million samples, represented as feature vectors. Using TDSVM to speed up the training of kernel-based support vector machines (SVMs), at a factor of nearly 300, we were able to perform numerous experiments efficiently to optimize the parameters and feature selection that would otherwise take months. As a result, we achieved prediction outcomes with substantially high scores in F1-measure and Matthews correlation coefficient (MCC) using only protein-sequence information.