Last year DeepMind beat its rivals in a protein-folding competition. As we all know, proteins are the complex molecules that underpin all biology, and predicting their structure is difficult. But DeepMind's AlphaFold2 made a huge leap in its capabilities, producing results that matched experimental data to a few atomic resolutions.
In July, the company published a paper describing AlphaFold2, made the code public, and abandoned a library of 350,000 protein structures, promising 100 million more.
This week, Alphabet announced that it would build on DeepMind AlphaFold2's breakthrough by creating a new company called Isomorphic Labs, to apply artificial intelligence to drug discovery.
"You can think of Isomorphic as a sister company to DeepMind," Hassabis told Stat. Our idea is to advance the potential of computational AI approaches and reimagine the entire drug discovery process."
While AlphaFold2's success has spurred efforts, protein folding is just one step — arguably easier than the others — in the arduous process of drug discovery.
Hassabis has bigger ideas. Although details are unclear, the new company appears to be building a range of AI models to ease key bottlenecks in the process. Instead of identifying and developing drugs themselves, they will sell a model platform to pharmaceutical companies as a service. These may address how proteins interact, the design of small molecules, how well the molecules bind, and the prediction of toxicity, Hassabis told Stat.
Most of the company's not insignificant costs are spent on pure research. DeepMind made its first profit in 2020, but its customers are mostly Alphabet. Some wonder whether it will come under more pressure to focus on commercial products. The decision to create an independent business based on DeepMind's research suggests otherwise. If it can advance the field as a whole, perhaps it makes sense to fund a new organization — or be one seeded by future breakthroughs — rather than divert resources from DeepMind's more fundamental research.
In a way, Isomorphic will build its business from scratch. AlphaFold2 is certainly a big deal, but protein modeling is just the tip of the iceberg for drug discovery. Moreover, while AlphaFold2 has access to hundreds of thousands of freely available, already modeled protein structures, most of the molecular data relevant to drugs is stored in proprietary databases owned by pharmaceutical companies.
There are also big differences between models and predictions, and even between promising animal trials and drugs approved for human consumption. Most drugs will fail to some extent. Despite a surge in investment in recent years, ai approaches have yet to yield significant results. Exscientia has two drugs in phase I clinical trials, and Valo Health has two drugs expected to enter phase II trials later this year or early next year. The full promise of AI in drug discovery may still be years away.