AlphaFold from DeepMind could be the future of science – and artificial intelligence

Here’s an idea: Artificial intelligence – what’s the point of it?

This might sound crude, given the huge amount of energy, investment, and hype in AI, as well as undeniable evidence of technological progress. After all, today’s artificial intelligence can beat any human being in games ranging from chess to me starcraft (DeepMind’s AlphaZero and AlphaStar); can write B- An essay on the history of the college in seconds with a few prompts (GPT-3 from OpenAI); can draw Illustrations upon request Of creativity and amazing quality (OpenAI’s DALL-E 2).

For AI proponents such as Sam Altman, CEO of OpenAI, these developments herald an era “when creative AI tools will have the greatest impact on creative workflows since the computer itself,” he said. chirp Last month. This may be true. But here and now, I still feel somewhat frustrated.

Not by what these AI tools can do, exactly. Write a short prompt in DALL-E 2 and return, say, “Medieval board where wifi doesn’t work“A feeling close to magic. However, humans can write articles and humans can draw illustrations, while GPT-3 and DALL-E 2 can do these tasks.” Faster, they really can’t do it better. They are superior in speed, not quality. (The exception in the group above is the DeepMind gameplay model, which Really superhuman Just ask the poor lost man Go Mister Lee Si Doll — but until these AI skills are put into practice in a more complex real world, it’s mostly an interesting research project.)

So AI can be cool and cool and even a little intimidating, but unless it’s really capable of playing 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 3D structure of proteins, is really huge — heralding not only a new era in AI, but a new era in useful and important science.

‘Great Challenge’ solved

For decades, molecular biologists have attempted to discover what is known asProtein folding problem. “

Proteins are the biological driver of everything from viruses to humans. They begin as chains of chemical compounds before being folded into unique 3D shapes. The nature of these forms – as far as the amino acids they are made of – determine what proteins can do and how they can be used.

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

However, the problem with the protein folding problem is that determining the final structure of a protein has generally taken scientists years of hard laboratory work. What the researchers needed was an artificial intelligence algorithm that could quickly determine the final shape of a protein, just as today’s computer vision systems can identify human faces with amazing accuracy. Until a few years ago, the best computational biology approaches for predicting the folding of proteins were still much less The accuracy that scientists can expect from experimental work.

Enter AlphaFold. Another product of DeepMind, the London-based artificial intelligence company Purchased by Google (later Alphabet) in 2014AlphaFold is a model of artificial intelligence Designer To predict the three-dimensional structure of proteins. AlphaFold blew up the competition In a challenge to predict protein structure every two years in late 2020, experimental work performed nearly identical to the gold standard, but much faster.

AlphaFold predicts protein structures through a deep learning neural network that has been trained on thousands of known proteins and their structures. The model used those known connections to learn to quickly predict the shape of other proteins, in the same way that other deep learning models can ingest massive amounts of data — in the case of GPT-3, about 45 TB of text data – To predict what will come next.

Alpha Fold was aware from the magazine Sciences As the best breakthrough of 2021, it outperformed candidates such as the Covid-19 antiviral pill and the human body’s CRISPR gene-editing application. one expert even wondered If AlphaFold becomes the first artificial intelligence to win a Nobel Prize.

A new era of digital biology

Breakouts continued to appear.

Last week, DeepMind announce That researchers from around the world have used AlphaFold to predict the structures of about 200 million proteins from a million species, covering just about every protein known to humans. All this data is provided free of charge at Database Created by DeepMind and its partner, the European Bioinformatics Institute of the European Molecular Biology Laboratory.

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

The database basically acts as a Google search for protein structures. Researchers can write down a known protein and restore its expected structure, saving them weeks or more of lab work. The system is already in use Accelerate drug discoverypartly through Alphabet’s sister company called Isomorphic Laboratories, while other researchers are bugging AlphaFold to identify enzymes that can break down plastic.

The sheer speed that AlphaFold enables should also help lower your search cost. DeepMind Research Scientist Catherine Tunyasofunakul told reporters that AlphaFold only takes about 10 to 20 seconds to predict each protein. This could be particularly useful for researchers working on neglected diseases such as leishmaniasis and Chagas disease, which are perennially underfunded because they often afflict the extreme poor.

“AlphaFold is a unique and significant advance in the life sciences that demonstrates the power of artificial intelligence,” chirp Eric Topol, director of the Scripps Research Translational Institute.

Artificial intelligence is useful – now

AI models such as GPT-3 that deal with a general language may be more effective than a narrower application such as AlphaFold. Language is still our greatest sign of intelligence and perhaps even consciousness – just a witness The last controversy About whether another advanced language model, Google’s LaMDA, has become conscious.

But despite all the progress, these paradigms persist Far from this level, and far from being really reliable for regular users. Companies like Apple and Amazon he strived To develop voice assistant artificial intelligence systems worthy of the name. Such models also struggle with Bias and fairnessAs Segal Samuel wrote earlier this year, this is a problem that must be solved with politics, not technology.

The AlphaFold model from DeepMind is not without risks. Like Kelsey Piper wrote earlier This year on artificial intelligence and its applications in biology, “Any system that is robust 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 very dangerous to humans.” An AI capable of predicting protein structures could theoretically be put to malicious uses by someone looking to engineer bioweapons or toxins.

DeepMind is credited with having assessed the potential risks of opening its database to the public, consulting with more than 30 biosecurity and ethics experts, and concluding that the benefits — including speeding up the development of effective defenses against biological threats — outweigh any risks. “The accumulation of human knowledge is just a huge benefit,” Ewen Bernie, director of the European Institute of Bioinformatics, told reporters at a press conference. “And the potentially risky entities are likely to be a very small handful.”

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

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

Leave a Comment