AI tools, like ChatGPT, are revolutionising how the world generates text, images and code. In a similar way machine learning algorithms and generative AI are upending traditional processes in life sciences and collapsing time frames in drug discovery and materials development.
AlphaFold by DeepMind is probably the most well-known Machine Learning model in this space. It predicts the 3D structure of a protein from its amino acid sequence and has been used by more than 1 million researchers over the 18 months in which it has been publicly available. Since then, a plethora of other AI tools have emerged, such as the recently open-sourced RFDiffusion, a Machine Learning model which allows researchers to generate computational protein designs using just their laptop.
As AI makes rapid progress in the world of bits, translating those computational designs into physical, functioning proteins remains challenging. Adaptyv Bio has set out to address this with its next-generation protein foundry. Combining advanced robotics, microfluidics and synthetic biology techniques, Adaptyv Bio is building a full-stack platform to allow protein engineers to validate their AI-generated protein designs.
“Proteins are at the core of the biorevolution, whether in the form of new medicines, better enzymes for research and industrial applications or as materials with novel properties,” said Julian Englert, CEO and Co-founder at Adaptyv Bio. “As a protein designer, you now have access to incredible new AI tools such as AlphaFold or RFDiffusion. But validating your protein designs in the lab to see if they work remains a huge pain. Imagine every time you used Github Copilot to generate some code you had to wait 10 weeks for it to execute or to tell you that it had a bug. And imagine each execution costs US$1000. That’s pretty much the situation for protein designers today.”
The thing that AI models need the most is data – for training them and for improving the predictions they make. By making it easier to generate data about how well the designed proteins work, Adaptyv Bio allows protein engineers and AI models to get more feedback about their designs and helps them steer towards better-performing proteins.
“Think of the AI in a self-driving car,” added Englert. “To keep the car on the road and drive it to the destination, an AI model needs to have a tight feedback loop by getting lots of high-quality data from the car’s camera sensors. It’s pretty much the same principle for an AI model designing new proteins, just that the feedback mechanism here is making the proteins in our lab and testing how they perform.”