Combining human thinking with computer tactics: the future of medicinal chemistry?
In this interview, Willem van Hoorn, Head of Chemoinformatics at Exscientia, discusses artificial intelligence and how Exscientia is applying this technology to drug discovery.
Can you tell us a bit about yourself and your role at Exscientia (Dundee and Oxford, UK)?
At secondary school my favourite subject was chemistry. Around that time home computers became affordable and I got into programming. Computers and chemistry have since then been a common thread. I signed up for a chemical engineering course at university, as they offered a minor in computer science, and I naively thought chemical engineering was similar to chemistry – it wasn't, however, I have no regrets. This was followed by a PhD in computational chemistry, where I was doing transition state calculations, then a postdoc in the US doing Monte Carlo computer simulations.
In 1999, I joined Pfizer in Sandwich (UK) as computational chemist supporting medicinal chemists. Soon after library chemistry and HTS took off, and new tools were needed to deal with the resulting much larger but noisier data sets. Machine learning and data pipelining tools came to the rescue and slowly I became a chemoinformatician. Andrew Hopkins, who is now the CEO of Exscientia, was my colleague at Pfizer and we had lots of discussions on how machine learning and chemoinformatics could transform drug discovery. When he started Exscientia, I was really keen to join and I am now part of a like-minded group of people, who are all full of ideas on how to transform drug discovery using AI. My role is focused on developing our active learning algorithms, which help in performing smarter experiments.
What is artificial intelligence?
There is a lot of overlap with machine learning, but in the context of drug discovery we see it as going beyond ranking (human-generated) compound ideas by machine models to having a system that suggests what to make next.
How is Exscientia applying artificial intelligence to drug discovery?
We have built a platform called Centaur, inspired by centaur chess, which came about after Garry Kasparov was defeated by Deep Blue in 1997: this was the first time a computer beat the world champion in a non-trivial game. Computers may beat humans but it soon became clear that humans using computers were better than computers alone. In centaur chess, a human/computer team plays against other human/computer teams. Our Centaur platform enables the same synergy between human strategic thinking and computer tactics to dramatically speed up drug discovery.
Want to learn more about computational chemistry? You may be interested in:
You recently spoke at the 29th Symposium of Medicinal Chemistry in Eastern England – can you give us a quick overview of the work you presented here?
I gave a high level overview of the Centaur platform, including two examples of how it performed in small molecule drug discovery projects that were run in collaboration with two pharma companies. In the last few years we have gained experience in how to best use the platform and we now have a track record of finding a clinical candidate in less than 500 compounds. This is about a fivefold reduction of the long-term industry average.
How do you expect the continued development of artificial intelligence techniques to affect drug discovery in the coming years?
To paraphrase a quote I recently read on Twitter: AI won't replace medicinal chemists but medicinal chemists who use AI will replace the ones who don't. This is less ominous than it sounds; previously, chemists who use computers in the form of Scifinder, Reaxys, etc. have replaced chemists who don't, and the chemist who uses AI can very well be the same chemist who previously didn’t use AI.
The quote can also be seen in a larger context: companies that discover drugs using AI will replace companies that don't.