Last month, Alphabet’s artificial intelligence subsidiary, DeepMind, stunned the world of science by presenting something truly spectacular: a snapshot of nearly every existing protein on Earth — 200 million of them.
AlphaFold, DeepMind’s protein structure program, is impressive because it reveals so much fundamental information about living organisms.
Proteins are the building blocks of life, after all, and as such they are essential to life and to the development of medicines. Proteins can be drug targets, and they can themselves be drugs. In either case, it is important to know the intricate ways in which they fold into various shapes. Their coils, floppy bits, hidden pockets and sticky patches can control, for example, when a signal is sent between cells or if a process is turned on or off.
Until now, capturing an image of a protein has required painstaking work lasting anywhere from days to months to years — work that has sometimes never come to fruition.
Since the early 1990s, scientists have been trying to train computers to predict a protein’s structure based on its genetic sequence. AlphaFold had the first taste of success in 2020, when it correctly predicted the structures of a handful of proteins. The next year, DeepMind put on its server about 365,000 proteins.
Now, it’s put the entire universe of proteins up for grabs — in animals, plants, bacteria, fungi and other living things. All 200 million of them.
Much as the gene-editing tool Crispr revolutionized the study of human disease and the design of drugs to target genetic errors, AlphaFold’s feat is fundamentally changing the way new medicines can be invented.
“Anybody who could have thought that machine learning was not yet relevant for drug hunting surely must feel different,” said Jay Bradner, president of the Novartis Institutes for BioMedical Research, the pharma company’s research arm. “I’m on it more than Spotify.”
Count me as one of the former skeptics. I hadn’t discounted the possibility that AI might have an impact on the drug industry, but I was weary of the many biotech firms hyping often ill-defined machine-learning capabilities. Companies often claimed that they could use AI to invent a new drug without acknowledging that the starting point — a protein structure — still needed to be worked out by a human. And so far, people have had to first invent drugs for the computer to improve upon them.
Producing the full compendium of proteins is something entirely different — and outside the usual hype cycle. It’s little wonder that executives at biotech and pharma companies are widely adopting AlphaFold’s revelations.
Rosana Kapeller, the chief executive officer of Rome Therapeutics, offers an example from her company’s labs. Rome is probing the “dark genome,” the repetitive portion of the human genetic code that is believed to be largely a relic of ancient viruses. Rome’s team spent more than six months refining its first image of one protein embedded in that dark genome. Just one day after they captured an initial snapshot of a second protein, DeepMind dropped its complete load of images. Within 24 hours, Rome’s scientists had perfected their picture. “So you see,” she said, “that’s amazing.”
None of this is to say that AlphaFold will solve every problem in drug discovery, or even that its 200 million protein images are perfect. They’re not. Some need more work, and others are more akin to a child’s scribbles than the fleshed out images researchers hope for.
Scientists in industry and academic labs tell me that even when the snapshots are imperfect, they contain enough information to provide a rough sense of where the important bits are. David Liu, a professor at the Broad Institute of MIT and Harvard and founder of multiple biotech companies, said the technology still allows researchers in his lab to “achieve that Zen-like understanding state” to decide where to tinker with a protein to change its properties.
But proteins also don’t exist as still snapshots. Depending on the job they’re performing at a given moment, they yawn and jiggle and twist inside the swamp of a cell. In other words, AlphaFold gives us protein Instagram; scientists would love to have protein TikTok or, eventually, protein YouTube.
Even if that becomes possible one day, this addresses just one step in the process of creating new drugs. The most expensive part is testing that new medicine in humans.
Nevertheless, AlphaFold’s pictures can help drugmakers get to the testing stage faster. DeepMind’s feat may have taken several years of scientific exploration, but it produced something with truly monumental consequences. And it made that work freely available. (Of course, it has also launched its own biotech company, Isomorphic Labs, to take a stab at capitalizing on the advances.)
Finally, we are getting a glimpse of AI’s potential to transform the drug industry. And now it’s possible to consider which problems machine-learning might solve next for science and medicine.
More from other writers at Bloomberg Opinion:
• AI’s Next Big Thing Is Fake Data: Parmy Olson
• Maybe AI Technology Isn’t as Scary as We Thought: Tyler Cowen
• AI Needs a Babysitter, Just Like the Rest of Us: Parmy Olson
This column does not necessarily reflect the opinion of the editorial board or Bloomberg LP and its owners.
Lisa Jarvis is a Bloomberg Opinion columnist covering biotech, health care and the pharmaceutical industry. Previously, she was executive editor of Chemical & Engineering News.
More stories like this are available on bloomberg.com/opinion
©2022 Bloomberg L.P.