Kernel methods for integrating biological data

Dick de Ridder and Marc Hulsman

Integrative bioinformatics focuses on the construction of approximate models of biological phenomena, such as gene regulation, protein interaction and complex formation, or protein function. Such models are based on a wealth of prior knowledge (databases, literature) and high-throughput measurement data available. A major challenge is how to combine these various sources of information, which often differ in data type, bias, coverage etc.

Over the last decade, kernel methods have been increasingly employed to tackle such problems. Kernels can be used in many algorithms, including classification and regression (the support vector machine), dimensionality reduction and statistics. A large number of kernels specifically tailored for certain types of (biological) data are now available, and various methods have been proposed to combine kernels.

In this tutorial, we will introduce kernel-based predictive algorithms, discuss a number of kernels relevant to biological modeling and methods to integrate various kernels for prediction. We will end by discussing some applications of kernel combination to biological problems.