Hierarchical graph kernels for classification of molecules

The central problem of pharmaceutical research is to understand the effect of a certain molecule on human or animal biology. Many machine learning models have been developed in recent years and show great efficiency and accuracy for this kind of predictions. However, a problem common to many of the best algorithms is that they require huge amounts of data to be sure the models are well trained and make accurate predictions. This is a serious issue when such data volume is not available. We propose to adapt a family of methods called hierarchically defined kernels we belive is well suited to tackle the problem of molecule classification for small datasets. This method works by decomposing the objects of interest in a tree-like structure of sub-objects (like a picture can be split up into sub-pictures, each sub-pictures be decomposed into smaller sub-pictures and so on) and constructing iteratively a similarity measure starting with a similarity on the smallest sub-objects.

Faculty Supervisor:

Guillaume Rabusseau

Student:

Partner:

Valence Discovery Inc

Discipline:

Computer science

Sector:

Biotechnology; Pharmaceuticals; Information and Communications Technology

University:

Université de Montréal

Program: