clock menu more-arrow no yes mobile

Filed under:

Finally, an answer to the question: AI — what is it good for?

Got a protein? This AI will tell you what it looks like.

AlphaFold’s prediction for the structure of protein F20H23.2
AlphaFold’s prediction for the structure of protein F20H23.2.
Bryan Walsh is the editor of Vox’s Future Perfect section, which covers the policies, people, and forces that could make the future a better place for everyone. He worked at Time magazine for 15 years as a foreign correspondent in Asia, a climate writer, and an international editor — and wrote a book on existential risk.

That headline might seem a bit churlish, given the tremendous amount of energy, investment, and hype in the AI space, as well as undeniable evidence of technological progress. After all, AI today can beat any human in games ranging from chess to Starcraft (DeepMind’s AlphaZero and AlphaStar); it can write a B- college history essay in seconds with a few prompts (OpenAI’s GPT-3); it can draw on-demand illustrations of surprising creativity and quality (OpenAI’s DALL-E 2).

For AI proponents like Sam Altman, OpenAI’s CEO, these advances herald an era where “AI creative tools are going to be the biggest impact on creative work flows since the computer itself,” as he tweeted last month. That may turn out to be true. But in the here and now, I’m still left somewhat underwhelmed.

Not by what these AI tools can do, exactly. Typing a short prompt into DALL-E 2 and getting back, say, “a medieval painting where the wifi isn’t working” feels close to magic. Still, human beings can write essays and human beings can draw illustrations, and while GPT-3 and DALL-E 2 can do those tasks faster, they can’t really do them better. They’re superhuman in velocity, not quality. (The exception in the above group is DeepMind’s game-playing model, which really is superhuman — just ask poor defeated Go master Lee Se-dol — but until those AI skills can be employed in the much more complex real world, it’s mostly an interesting research project.)

So AI can be fascinating and cool and even be a little bit scary, but what it isn’t yet is truly able to play a vital role in solving important problems — something that can be seen in the fact that all of these advances have yet to boost America’s sluggish productivity numbers.

That’s why the recent news about AlphaFold, an AI model from DeepMind that can predict the three-dimensional structure of proteins, seems genuinely monumental — heralding not just a new era in artificial intelligence but a new era in useful, important science.

A “grand challenge” solved

For decades, molecular biologists have been trying to crack what’s known as “the protein-folding problem.”

Proteins are the biological drivers of everything from viruses to human beings. They begin as strings of chemical compounds before they fold into unique 3D shapes. The nature of those shapes — as much as the amino acids that make them up — define what proteins can do, and how they can be used.

Predicting what shape a protein will take based on its amino acid sequence would allow biologists to better understand its function and how it relates to other molecular processes. Pharmaceuticals are often designed using protein structural information, and predicting protein folding could greatly accelerate drug discovery, among other areas of science.

However, the issue in the protein-folding problem is that identifying a protein’s eventual structure has generally taken scientists years of strenuous lab work. What researchers needed was an AI algorithm that could quickly identify the eventual shape of a protein, just as computer vision systems today can identify human faces with astounding accuracy. Up until just a few years ago, the best computational biology approaches to protein-folding prediction were still far below the accuracy scientists could expect from experimental work.

Enter AlphaFold. Another product of DeepMind, the London-based AI company that was bought by Google (which later became Alphabet) in 2014, AlphaFold is an AI model designed to predict the three-dimensional structure of proteins. AlphaFold blew away the competition in a biennial protein-structure prediction challenge in late 2020, performing almost as well as gold-standard experimental work, but far faster.

AlphaFold predicts protein structures through a deep learning neural network that was trained on thousands of known proteins and their structures. The model used those known connections to learn to rapidly predict the shape of other proteins, in much the same way that other deep learning models can ingest vast quantities of data — in the case of GPT-3, about 45 terabytes of text data — to predict what comes next.

AlphaFold was recognized by the journal Science as 2021’s Breakthrough of the Year, beating out candidates like Covid-19 antiviral pills and the application of CRISPR gene editing in the human body. One expert even wondered if AlphaFold would become the first AI to win a Nobel Prize.

“A new era of digital biology”

The breakthroughs have kept coming.

Last week, DeepMind announced that researchers from around the world have used AlphaFold to predict the structures of some 200 million proteins from 1 million species, covering just about every protein known to human beings. All of that data is being made freely available on a database set up by DeepMind and its partner, the European Molecular Biology Laboratory’s European Bioinformatics Institute.

“Essentially you can think of it as covering the entire protein universe,” DeepMind CEO Demis Hassabis said at a press briefing last week. “We are at the beginning of a new era of digital biology.”

The database basically works as a Google search for protein structures. Researchers can type in a known protein and get back its predicted structure, saving them weeks or more of work in the lab. The system is already being used to accelerate drug discovery, in part through an Alphabet sister company called Isomorphic Laboratories, while other researchers are tapping AlphaFold to identify enzymes that could break down plastics.

The sheer speed enabled by AlphaFold should also help cut the cost of research. Kathryn Tunyasuvunakool, a DeepMind research scientist, told reporters that AlphaFold required only about 10 to 20 seconds to make each protein prediction. That could be especially useful for researchers laboring on neglected diseases like leishmaniasis and Chagas disease, which are perennially underfunded because they mostly strike the desperately poor.

“AlphaFold is the singular and momentous advance in life science that demonstrates the power of AI,” tweeted Eric Topol, the director of the Scripps Research Translational Institute.

AI that’s useful — now

It may well be that AI models like GPT-3 that deal in general language are ultimately more influential than a more narrow application like AlphaFold. Language is still our greatest signal of intelligence and potentially even consciousness — just witness the recent controversy over whether another advanced language model, Google’s LaMDA, had become sentient.

But for all their advances, such models are still far from that level, and far even from being truly reliable for ordinary users. Companies like Apple and Amazon have labored to develop voice assistant AIs that are worthy of the name. Such models also struggle with bias and fairness, as Sigal Samuel wrote earlier this year, which is a problem to be solved with politics rather than technology.

DeepMind’s AlphaFold model isn’t without its risks. As Kelsey Piper wrote earlier this year about AI and its applications in biology, “Any system that is powerful and accurate enough to identify drugs that are safe for humans is inherently a system that will also be good at identifying drugs that are incredibly dangerous for humans.” An AI capable of predicting protein structures could theoretically be put to malign uses by someone looking to engineer biological weapons or toxins.

To its credit, DeepMind says it weighed the potential dangers of opening up its database to the public, consulting with more than 30 experts in biosecurity and ethics, and concluded that the benefits — including in speeding the development of effective defenses against biological threats — outweighed any risks. “The accumulation of human knowledge is just a massive benefit,” Ewen Birney, director of the European Bioinformatics Institute, told reporters at the press briefing. “And the entities which could be risky are likely to be a very small handful.”

AlphaFold — which DeepMind has said is the most complex AI system it has ever built — is a highly effective tool that can do things humans can’t do easily. In the process, it can make those human biologists even more effective at their jobs. And in the age of Covid, those jobs are more important than ever, as is their new AI assistant.

A version of this story was initially published in the Future Perfect newsletter. Sign up here to subscribe!

Sign up for the newsletter Sign up for Vox Recommends

Get curated picks of the best Vox journalism to read, watch, and listen to every week, from our editors.