The tests are almost entirely automated, requiring a range of high-end equipment to prepare the samples and take them through the various stages of the testing process: antibodies are grown based on their genetic sequence and then used in biological tests – samples – tested the diseased tissue for which they were designed. Humans monitor the process, but their job is mainly to transport samples from one machine to the next.
“Once you have the experimental results of this first set of 700 molecules, that information is fed back to the model and used to refine the model’s understanding of space,” says Field. In other words, the algorithm begins to paint a picture of how different antibody designs change the effectiveness of the treatment – getting better with each subsequent round of antibody designs, carefully balancing the use of potentially fruitful designs with exploring new areas.
“One challenge with conventional protein engineering is that once you find something that works reasonably well, you tend to make a lot of very small changes to that molecule to see if you can refine it further,” says Field . These tweaks can improve one property—such as the ease of large-scale production of the antibody—but have a disastrous impact on the many other properties required, such as selectivity, toxicity, potency, and more. The traditional approach means you could end up going the wrong way, or missing the wood for the trees – endlessly tweaking something that works a little, while there might be far better options in a completely different part of the map.
You’re also limited by the number of tests you can do, or how many “shots on goal,” as Field puts it. That means human protein engineers tend to look for things they know work. “This creates all of these heuristics, or rules of thumb, that human protein engineers use to find safe spaces,” Field says. “But this quickly leads to an accumulation of dogmas.”
The LabGenius approach delivers unexpected solutions that people might not have thought of, and finds them faster: from problem definition to completion of the first batch, it takes just six weeks, all driven by machine learning models. LabGenius has raised $28 million from companies like Atomico and Kindred and is beginning to partner with pharmaceutical companies and offer its services like a consulting firm. According to Field, the automated approach could also be extended to other forms of drug discovery, thereby streamlining the long, “artisanal” process of drug discovery.
Ultimately, Field says, it’s a prescription for better care: antibody treatments that are more effective or have fewer side effects than existing human-developed treatments. “You find molecules that you would never have found with conventional methods,” he says. “They are very different and often counter-intuitive to designs that a human would come up with — which should allow us to find molecules with better properties, ultimately leading to better outcomes for patients.”
This article appears in the September/October 2023 issue of WIRED UK magazine.