News

The Hopkins Lab has joined a new EPSRC-funded collaboration whose aim to apply “meta-learning” to the ChEMBL database of bioactivity to find patterns in what type of machine learning works best for what type of problem.

QSAR (Quantitative Structure-Activity Relationship) models are regression or classification models which relate a set of "predictor" variables (X) to a response variable (Y). They can be used to predict bioactivity (e.g. kinase inhibition) from a chemical structure and thereby guide drug discoverers to design more potent inhibitors. The predictors consist of physico-chemical properties or molecular descriptors of compounds while the response variables would typically be some kind of biological activity.

The aim of the collaboration is to use newly available public domain chemoinformatics databases to systematically run extensive comparative QSAR experiments. We will then generalize these results to learn which target-type/ compound-type/compound-representation/learning-method combinations work best together. This is known as “meta-learning” - using machine learning to learn about QSAR leaning. This knowledge will be made publically available to guide future QSAR learning and hopefully decrease the time and cost of developing new drugs.

The other members of the Meta-QSAR collaboration come from the labs of Professor Ross King from the University of Manchester and Dr Laris Soldatova from Brunel University in London.