String Kernels and High-Quality Data Set for Improved Prediction of Kinked Helices in α-Helical Membrane Proteins

The reasons for distortions from optimal α-helical geometry are widely unknown, but their influences on structural changes of proteins are significant. Hence, their prediction is a crucial problem in structural bioinformatics. For the particular case of kink prediction, we generated a data set of 132 membrane proteins containing 1014 manually labeled helices and examined the environment of kinks. Our sequence analysis confirms the great relevance of proline and reveals disproportionately high occurrences of glycine and serine at kink positions. The structural analysis shows significantly different solvent accessible surface area mean values for kinked and nonkinked helices. More important, we used this data set to validate string kernels for support vector machines as a new kink prediction method. Applying the new predictor, about 80% of all helices could be correctly predicted as kinked or nonkinked even when focusing on small helical fragments. The results exceed recently reported accuracies of alternative approaches and are a consequence of both the method and the data set.