DeepMind faces criticism from scientists skeptical of breakthrough

• Google’s DeepMind, a London-based artificial intelligence start-up, is facing backlash from the scientific community after announcing that its AlphaFold program is a “solution” to the problem of protein folding.
• DeepMind researchers said on Monday the startup has successfully modeled complex protein structures – and suggested its findings could revolutionize drug discovery and medicine.
• While a number of independent academics hailed the breakthrough, several said it was not clear how well AlphaFold would perform in the real world.
• John Moult, the chairman of the academic competition that announced DeepMind’s breakthrough, defended the results, telling Business Insider he had looked at the results “very carefully.”
• Visit the Business Insider homepage for more stories.

Academics and researchers are debating claims by Google-owned artificial intelligence firm DeepMind that it has solved one of biology’s most difficult problems.

On Monday, DeepMind said it broke new ground in understanding the behavior of microscopic proteins. He said his artificial intelligence program AlphaFold could reliably predict their shape, effectively solving a problem that has plagued scientists for decades.

Professor Venki Ramakrishnan, winner of the Nobel Prize in Chemistry, hailed the results as an “astonishing” achievement.

And the DeepMind team wrote in a blog on Monday: “This breakthrough demonstrates the impact AI can have on scientific discovery and its potential to dramatically accelerate progress in some of the most fundamental areas that explain and shape our world.”

But some academics are skeptical of how DeepMind’s claims should be presented as a “solution” to the protein folding problem. They asked DeepMind to put AlphaFold’s code in the public domain and said it was not clear how the program would work outside of a restricted framework.

DeepMind’s breakthrough was part of the Critical Assessment of Structure Prediction (CASP), a global competition specially set up to test research teams on their ability to predict the shape of a protein from its sequence. of amino acids.

Max Little, an associate professor and lecturer in computer science at the University of Birmingham, told Business Insider that DeepMind’s AI has only shown its potential “in the context of the CASP database challenge.”

He said: “We can’t really be sure how well AlphaFold works in the face of the much richer and more diverse protein spectrum found in the real world of living organisms.”

Here’s what DeepMind did

DeepMind

Protein is a key part of life, found in humans, animals, plants, and microscopic organisms. They are invisible to the human eye and constantly rearrange themselves, making it difficult to study and predict their behavior.

The way proteins move (or ‘fold’) around your body – turning from a chain of amino acids into more complex 3D structures – has big implications for your health and is tied to everything. , from Alzheimer’s disease to the flu. This is why scientists have spent most of 50 years trying to predict their movement.

If you already know how a protein will behave, they say, you could theoretically change its behavior. For example, by stopping a misfolded protein in its tracks, you could prevent its host from contracting a neurodegenerative disease like Parkinson’s disease. You could also target medical treatment better, as you would have a better idea of ​​how a person’s body is reacting, avoiding unpleasant side effects in advance.

In 2018, DeepMind first introduced AlphaFold in CASP, but the results were not deemed concrete enough to be medically useful.

This year, the latest version of AlphaFold was trained on a “protein database” made up of around 170,000 structures and matched to predictions made by scientists in the lab – a much longer and more expensive process. – with high accuracy within two thirds. cases.

Experts say DeepMind’s research may not apply outside of a narrow framework

Professor Michael Thompson, an expert in structural biology at the University of California, Merced, called the idea that protein folding had been solved “laughable.”

“Frankly, the hype doesn’t serve anyone,” he wrote on Twitter, adding that the company could “never keep the promise that was made”.

He then said: “Until DeepMind shares their code, nobody on the ground cares and they’re just the ones patting each other on the back.”

Thompson said that “the progress of the prediction is impressive.” He added: “However, taking a big step forward is not the same as ‘solving’ a decades-old problem in biology and chemical physics.”

“Although this is the most important thing that has happened in protein folding, the problem is not yet resolved,” said Vishal Gulati, a seasoned venture capitalist specializing in deep tech and healthcare startups. , in a blog post published on Tuesday.

He said DeepMind’s research was “not a minor achievement” but added: “Compared to the protein folding problem, CASP is a game. It is a very difficult game but it is a set of problems. reduced that helps us train tools and standardize performance … It’s a necessary step but it’s not enough. “

Lior Pachter, professor of computational biology at the California Institute of Technology, agreed, writing on Twitter that the protein folding problem was not “well defined”.

“I don’t mind that Google has promoted this,” he wrote. “It bothers me that many biologists (computer scientists) who should be better informed are shouting, ‘The protein folding is solved!’

“Have respect for yourself.”

But the CASP president said he had reviewed the results and that code sharing was not the norm

The results may be expected to spark vigorous public debate.

In an email exchange with Business Insider, CASP President John Moult dismissed the criticisms, writing: “CASP is not a game, it is a science experiment designed to test folding methods in similar situations. of reality … What is missing? “

He added, “Maybe the people promoting the results actually watched them? [It] seems like this person didn’t. ”

Responding to criticism of DeepMind’s refusal to share AlphaFold’s code, Moult wrote, “It’s an old chestnut in the field. While code sharing is obviously desirable, and some groups do, it doesn’t. has never been the norm. [It is] not clear why DeepMind should be held to a higher standard than others. “

He then said: “Fifty years of listening to false statements about this problem have made me the biggest skeptic in the world. But I have examined these results very carefully … Obviously, this is only the beginning. of what DeepMind and others will achieve with these kinds of approaches. “

A DeepMind spokesperson declined to comment when approached by Business Insider.