Comparison of Kernels for Protein Secondary Structure Prediction With a Support Vector Machine

Two years ago, Ryan Weber developed a simple support vector machine for protein secondary structure prediction to investigate the complexity of kernel and algorithm necessary to perform adequate classification [1]. For my project, I have resurrected Ryan‘s project, tested his conclusions, and extended the support vector machine to use the spectrum string kernel designed by Leslie, Eskin, and Noble [2], as well as a modified version of the polynomial kernel. My results for the classic Gaussian and polynomial kernels were not initially up to the specifications Ryan Weber had reported (55% accuracy, with 59% reported), which I suspect is due to many differences in our implementation (especially in how we combine predictions). Adding window weighting to the kernels improved the accuracy of these results, bringing them back up to the reported accuracy. The new spectrum kernel performed very well, about the same as the polynomial and Gaussian kernels without window weighting (54%); however, the polynomial and Gaussian kernels were still the best, generating predictions of approximately 59% accuracy at best. The p-sequential kernel performed terribly in this implementation, with only ∼ 45% accuracy. These results are still very far from the state-of-the-art in protein structure prediction (current accuracy is around 80%).