skip to primary navigationskip to content


Computational drug discovery by combining physics with deep learning

 Supervisor: Dr Alpha Lee

Predicting the affinity of unknown molecules towards a target is a significant challenge limiting the drug discovery pipeline. In the status quo, drug discovery is still an art that relies of chemists’ intuitions, which can often be incorrect. A reason why those intuitions might be flawed is that chemical space is high dimensional - there are many ways in which two molecules might be different to each other. Moreover, even if chemical intuitions are qualitatively correct, a human chemist often struggles to predict the uncertainty of his/her prediction. The aim of this project is to use deep learning to tease out subtle correlations within chemical datasets that a human chemist would have missed, and develop scalable Bayesian method to predict the uncertainty of the predictions.


The first stage of the project will focus on developing representations of chemical molecules to predict protein-ligand affinity. The second stage of the project will focus on developing scalable Bayesian deep learning architectures to estimate the prediction uncertainty. A particular focus of our modeling approach will be integrating physical insights into the design of machine learning models. We will obtain training data from collaborators in the industry as well as from mining data in the public domain.