Recently there has been great interest in applying state-of-the-art machine learning methods to problems in Bio- and Chemoinformatics. The success of kernel methods and the development of kernels which operate over discrete structures such as strings and trees opens up many possibilites for progress in these fields.
This project would focusses on using kernel methods for cheminformatic tasks, such as, in the first instance, developing predictive models which generalise well across chemical series. We are investigating among other aspects the construction of new kernels which can represent molecular structures in a more structured way than the traditional vector-of-statistics approach, which in turn would hopefully lead to greater accuracy on predictive tasks.