Nobel Laureate John Jumper: AI is Revolutionizing Scientific Discovery

Jumper describes moving from physics into computational biology, then studying biophysics, chemistry, and machine learning. At DeepMind, he worked on applying AI to scientific problems. He says the goal is to give scientists tools that help them make discoveries, including advances that could support medicine.

Proteins fold into complex shapes that let them perform many functions in cells, and those structures can help scientists understand disease and develop drugs. Determining a structure experimentally can take substantial effort: one paper Jumper cites reported that crystals formed after more than a year. The Protein Data Bank held about 200,000 known structures, while billions of protein sequences were available.

Jumper describes AlphaFold as mapping a protein sequence to a predicted structure. Its ingredients included public data, compute, and research; the final model used 128 TPU v3 cores for two weeks. He argues that research ideas were central: AlphaFold 2 trained on 1% of the data matched or exceeded AlphaFold 1, and blind assessment found about a third of the error of competing groups.

DeepMind released AlphaFold's code and a prediction database that grew from 300,000 to 200 million structures. The database let biologists compare predictions with proteins they had not solved themselves, helping build trust. Researchers also found uses the team had not anticipated, including protein interaction prediction.

Jumper describes researchers at MIT's Zhang Lab using an AlphaFold prediction of a molecular syringe to identify parts that might control cell targeting. They replaced those parts with a designed protein and demonstrated delivery of fluorescent proteins to selected cells in a mouse brain. Jumper says tools like AlphaFold can amplify experimental work, and expects AI for science to extend into broader areas.