Breakthrough: AI and better medicines

Ask the doctor

Q. Do you think artificial intelligence will ever make a scientific breakthrough that leads to better medicines?

A. In fact, it's already happened. A remarkable recent example is called the protein structure breakthrough. The word "breakthrough" is often overused, but not in this case.

Proteins are the workhorses of every cell. How each protein does its work depends on its shape. Many diseases occur because of defects in particular proteins. To develop a drug that targets a defective protein, the scientist benefits from knowing that protein's shape.

The techniques traditionally used to determine a protein's shape are slow and difficult. For 60 years, however, scientists knew that there was a much faster solution: if you knew the structure of the specific gene that makes a specific protein, you should, in theory, be able to predict the shape of that protein. Unfortunately, the structure of very few genes was known.

However, beginning in the 1970s, scientists developed techniques for determining the structure of genes. By the early 21st century the structure of every human gene, and many animal genes, had been determined. But scientists still struggled with translating theory into practice, with predicting the shape of a protein from the structure of its gene.

In the 1990s, scientists began a competition to use mathematical models and computers to achieve that goal. Over time, predictions improved somewhat, but they still were pretty bad. In 2018, an artificial intelligence company owned by Google called DeepMind joined the competition. It did surprisingly well, but still not well enough to be useful. By the end of 2020, though, DeepMind's predictions had become accurate. The next question was whether its software was fast enough to determine the structure of very many proteins.

Indeed, very few protein structures were known. As of July 1, 2021, thousands of scientists working for nearly a century had determined the shape of only about 30% of the 20,000 human proteins, and the shape of only 0.01% of the 280 million nonhuman proteins. Could DeepMind really improve those dismal numbers?

On July 22, 2021, DeepMind published its answer. It reported a predicted structure for 98% of the human proteins, up from 30% on July 1. By the end of 2021, it says it will publish the shape of 50% of the 280 million nonhuman proteins, up from 0.01%. Now, that's a breakthrough.

Experts differ as to how rapidly this remarkable advance will lead to new treatments. Time will tell. For now, though, one thing is clear: the marriage of information technology and modern biology led to this breakthrough. These two fields may be the most important intellectual achievements of the past century. And they happened, and could only have happened, because societies invested in science.

Image: © sergunt/Getty Images