Research

Publications

For an updated list of my research publications, please visit my Google Scholar page. For my thoughts on my own papers, please see here.

Current work (as of 2024)

Disclaimer: I have signed a confidentiality agreement as part of my current position and therefore will not be able to disclose everything I am working on.

I am generally interested in how machine learning can enhance scientific discovery (for positive applications). Due to issues of robustness, generalization, and small dataset sizes I generally do not believe that replacing existing systems with deep neural networks is the best strategy for most practical problems (at least with current technology). Instead, I think that the way forward is to identify small parts of larger systems in science where machine learning could have an advantage over existing approaches, then try to develop methods that are well-suited to these niches. Currently the niche I am focusing on is early-stage drug candidate generation.

Some specific problems I am working on (or have worked on recently) are:

• Applying Bayesian optimization to drug screening (and making it work in practice)
• Molecular property prediction with quantified uncertainty
• Retrosynthetic planning (how to synthesize novel molecules)
• Meaningful evaluation of ML algorithms in chemistry (average loss on a test set can be misleading)

General research interests

• Sequential decision making (e.g. Bayesian optimization)
• Gaussian processes and other kernel methods
• Learning on small datasets
• How can large language models help discovery? Until recently I thought they had limited potential for this, but I have started to change my mind.
• Model robustness and reliability
• Uncertainty quantification
• Evaluation of machine learning algorithms in a way that reflects their practical usage
• AI safety/alignment