DeepMind’s AI could accelerate drug discovery

A new study suggests that AlphaFold — an AI trained by Google sister company DeepMind to predict protein structures — could be useful for drug discovery after all.

Protein folding 101: Practically everything about your body is determined by proteins, complex molecules made up of long chains of amino acids. Just one of these chains can contain thousands of amino acids, and a single protein can be made up of several chains. 

A protein’s function is determined by how these amino acid chains fold up into a three-dimensional shape. If we know a protein’s structure, we can potentially alter its behavior — by introducing a drug that binds to the protein, for example.

Unfortunately, traditional methods for determining a protein’s structure are extremely slow, expensive, and complicated. As a result, researchers had only mapped the structure of 17% of the human body’s 20,000 proteins, as of 2021.

The AI solution? AlphaFold — an AI trained to predict a protein’s structure based only on the sequence of amino acids in its chains — aims to solve this “protein folding problem.”

In 2021, DeepMind published AlphaFold-predicted structures for all of the proteins in the human body, but when researchers plugged those predictions into their drug discovery models, which analyze a protein’s structure to identify chemicals that will bind to it, they found that chemicals that they already knew met the criteria weren’t always flagged by their models. 

That has led some to question how useful AlphaFold’s predictions will actually be for accelerating drug discovery.

What’s new? A new study, led by Brian Shoichet, a pharmaceutical chemist at UC San Francisco, and Bryan Roth, a structural biologist at the University of North Carolina at Chapel Hill, suggests that AlphaFold could be a useful drug discovery tool after all.

They had a theory that, while the slight differences between AlphaFold’s predictions and traditionally identified structures could cause certain bindable chemicals to be overlooked, they might still lead to the identification of new, equally promising compounds.

“There were no two molecules that were the same.”

Brian Shoichet

To test the theory, they screened a database containing hundreds of millions of chemicals for ones that would bind to two proteins — sigma-2 and 5-HT2A — based on protein structures that had been determined using traditional methods. 

They then did the same thing using AlphaFold’s predicted structures for the same proteins, and the list of flagged chemicals was completely different.

“There were no two molecules that were the same,” Shoichet told Nature. “They didn’t even resemble each other.”

Next, they identified hundreds of promising compounds from each list and headed into the lab to test whether the chemicals could actually bind to the proteins and meaningfully alter their function — and amazingly, the “hit rates” were about the same for each group.

In other words, the traditional and AI-generated ways of finding new drug candidates were both highly useful, even though their lists didn’t overlap.

Looking ahead: The researchers have shared an early draft of their study on the preprint server bioRXiv, so the results still need to be peer reviewed and replicated by other labs — but if they hold up, they suggest that AlphaFold could be more useful for drug discovery than early research suggested.

“Compared to actually going out and getting a new structure, you could advance the project by a couple of years and that’s huge,” said Shoichet.

We’d love to hear from you! If you have a comment about this article or if you have a tip for a future Freethink story, please email us at tips@freethink.com.