Abstract Details


Poster 2: A Multi-label Approach to Multi-Target Prediction.

M. Avid Afzal1, Hamse Y. Mussa1, Andreas Bender1, Robert C. Glen1
1The Unilever Centre for Molecular Science Informatics, Cambridge University
In conventional approaches to Cheminformatics, multi-class classification deals with learning from a given training set consisting of small compounds (usually in the form of chemical structures) and their associated biological/chemical activities which are generally represented by a single label per chemical structure. In other words a chemical structure is associated with a set of disjoint labels and it is assumed that a chemical structure can only have one label at a time. If there are two labels, the classification is commonly referred to as a binary classification problem; for more classes, it is called a multi-class classification problem. In machine learning, these classifications are referred to as single-label classifications.

In real-world situations a given small molecule (ligand) can simultaneously interact with a set of different biological targets. In this case, the ligand should hence be associated with different labels and the conventional single label multi-class classification approach should be modified accordingly. This leads to (what is defined in machine learning) as a multi-label multi-class approach to classification. It is a new paradigm that has, in recent years, received increased attention in a wide variety of research areas such as text classification, and scene and video classification. We wish to extend this approach to classification problems in Cheminformatics, namely the prediction of the polypharmacology of compounds.

In this contribution we present multi-target target prediction results obtained from a number of multi-label multi-class approaches, and hence provide a chemically relevant application of these recent developments in the machine learning field.

Return to Programme