Gram-Schmidt Kernels Applied to Structure Activity Analysis for Drug Design

We introduce a novel approach for structure activity relationship analysis based on the use of a special kernel. The kernel eÆciently performs Gram-Schmidt orthogonalisation in a kernel de ned feature space. We show that support vector machines in conjunction with Gram-Schmidt kernel, a recent method to extract features, can be adopted successfully to predict the inhibition of dihydrofolate reductase by pyrimidines. We show that the feature space generated by Gram-Schmidt kernel does capture enough information, to the extent that it can outperform state-of-the-art systems. We also show that support vector machines in conjunction with Gram-Schmidt kernels can be applied successfully to predict the inhibition of dihydrofolate reductase by triazines. A preliminary experimental comparison of performance of the kernel with standard kernels and decision trees is made showing encouraging results.