We’re delighted to announce that we have appointed a couple of senior industry figures to be deepmirror’s new scientific advisors. Today, we sit down with Jérémy Besnard, a co-founder of Exscientia, to find out why he’s joined the deepmirror journey:
Tell us about your background and expertise in drug discovery
I’m an organic chemist, and studied mainly in France before moving to the UK in 2008 to do my PhD at the University of Dundee. That’s where I started working around the themes of de novo design and how to apply AI and machine learning to generate compounds.
It was during this time that I co-founded Exscientia, first serving as Chief Technology Officer, and then as Chief Designer. In my time at Exscientia, I led the computational chemistry and the AI machine learning groups. I was involved in our drug discovery and technology strategies, as well as leading discovery projects with compounds that entered the clinic.
After Exscientia merged with Recursion, I switched to consultancy full-time. I worked closely with the team at Medicines for Malaria Venture (MMV) as a scientific advisor, and also began advising startups in the drug discovery space.
Throughout my consulting work, one question kept coming up: how can startups ensure that learnings from one AI discovery project are appropriately captured and used in the next project for the next customer?
Why did you decide to join deepmirror?
I first encountered deepmirror’s work through MMV, and the Drug Design for Global Health (dd4gh) project that the two organisations ran together. The focus on accessibility was incredibly compelling for me; not just that this was a cutting-edge technology being deployed to labs around the world for crucial global health challenges, but that the entire platform was built to be used by the chemists actually doing this work.
Scientists really love going through data. And while many labs have their own information, sometimes they may not know how to use it individually in the best possible way. dd4gh made it possible to ease the access to important data in a way that benefits more people. In a way, it helps democratise AI and machine learning data.
As well as the ease-of-use, I was impressed by deepmirror’s commitment to building the data infrastructure for drug discovery AI models. As with a lot of the pharma industry, drug discovery is hampered by the data being siloed and protected above all else. deepmirror’s innovative approach to building models that can learn from this data while keeping it private and secure presents a very exciting opportunity for AI to finally fulfil its potential in drug discovery.
How has drug discovery changed since you were building Exscientia?
It has changed so much and I think it's still changing, especially when we talk about AI. The potential impact that AI could have has become well accepted, and people are now more often asking why it’s not being used. There’s been a bit of a ‘coming of age’ process over the past few years; AI is now everywhere, and it will keep evolving.
When we were building Exscientia, everything was a bit less structured and planned. We were moving so fast, we’d build a prototype and it would either work or it wouldn’t, but we’d be moving on and building something else. With deepmirror, everything is so deliberate, and there’s an incredibly clear plan for where the technology is going, and how it’s going to help its customers.
Building AI tools has become a lot easier. I can easily ask Claude to build me an interface where I can input data and then get something that looks like a working template. In short, a lot of very basic functions will become easy to reproduce. But the expertise in the model and data curation will be missing, which means you may not get the output you want. To use AI successfully in drug discovery, there is a lot of science that needs to happen behind the scenes, and that is where generalist LLMs can’t help.
How do you see the value of AI in the drug discovery process?
AI in drug discovery suffers from a data problem more than a software problem. If you want to build a model just on your own data set you can do that, but your outputs will always be limited. We speak to lots of chemists in small molecule discovery, and they often say generative compounds are just “too silly”. That’s not really a failure in the generative model, rather a lack of scale in the training data.
Building that massive data set is crucial to the rest of the infrastructure deepmirror is building. There are no shortcuts here; to paraphrase Sun Tzu, ‘AI provides the method for building fast, but human expertise provides the method for building right’. deepmirror’s commitment to generating feedback, learning and iterating is what powers its ability to deliver the best possible results for its users.
