AI-enabled drug discovery is a nascent field with many breakthroughs still to come, but the promise of real impact is on the way. While machine learning (ML) for drug discovery isn’t a new field, over the last few years progress has accelerated, particularly in understanding and modelling proteins through structure prediction models like AlphaFold2 and protein language models. This is just the start, with a huge opportunity to apply all the lessons and successes of deep learning from other domains to approach fundamental problems in chemistry and biology.
Drug discovery takes place in the microscopic world, invisible to the human eye, where molecules interact in vast networks of biomolecular machines. There is a huge amount of complexity here, making it intractable to describe these systems using traditional mathematical and physics-based approaches. This is exactly where machine learning and AI can excel, becoming the perfect description language for these biological systems by learning from data. But to do this effectively we have to move past simple models that draw shallow statistical inferences from hand-crafted descriptions of the molecular world, to deep neural networks that are built and trained to capture the underlying mechanics of these systems, thereby generalising to uncharted territories.
By training models to represent the biomolecular world, we create a digital version of these hidden systems that are controllable and malleable, giving us a platform to question hypotheses. We can then search and explore the biomolecular space represented in these models: just like DeepMind’s algorithm AlphaGo was able to search the space of ‘Go’ intelligently to find the next move, the algorithms we develop at Iso are able to search the landscape of molecules and disease to create new therapeutics in collaboration with chemists and biologists.
Since the start of my career - which coincided with the so-called “deep learning” renaissance - I’ve been driven by the opportunity to develop breakthrough capabilities using machine learning. I’ve worked on a huge range of different methods and domains, from computer vision and language modelling, to core deep learning, neural network design, generative models, and large-scale deep reinforcement learning. Something I’ve seen again and again is how it is possible to move the needle from 0 to 100 on seemingly impossible problems when deep learning is applied from first principles.
But you cannot work on challenges like this alone: at Isomorphic Labs, we have hired world class ML researchers and engineers, brilliant biologists and chemists, all of whom work alongside our great people, legal, and operational teams (known internally as ‘Isos’). We believe that coming up with breakthrough models in this space will require not only diverse scientific and engineering skills, but also a wide array of professional experience.
Some Isos come with backgrounds rooted in drug discovery, but many come with deep experience in other fields such as core machine learning or machine learning applied to other industries. For us to achieve our mission, our whole team needs to come together, and work creatively and collaboratively to pave the way for this burgeoning field. We have to immerse ourselves in the depths of chemistry, biology, and physics to understand where machine learning can really make a step-change in modelling capability, and crucially have real-world impact on the most challenging bottlenecks for medicinal chemists.
We know this is a hard problem, and there will be many bumps and twists in the road: that’s the nature of truly ambitious research. Still, it’s enormously promising that even in the first year we are meeting and exceeding our ambitious goals on seemingly impossible problems. The opportunity to apply deep learning to reimagine drug discovery and model the hidden world within us is very much within reach.
Interested in joining Max's team? View our open roles here.
.png)
.svg)
