DeepMind’s AlphaFold Is Rewriting the Rules of Protein Research

Last updated: October 15, 2026

DeepMind has released free, public predictions for the structure of roughly 350,000 proteins across 20 organisms—including 98.5 percent of the human proteome, roughly 20,000 distinct structures. The data, generated by the lab's AlphaFold AI, is hosted by the European Molecular Biology Laboratory and available for immediate download.

The release includes human proteins alongside those of mice, fruit flies, and bacteria such as E. Coli. AlphaFold technical lead John Jumper describes the full human set as "HumanProteome.zip"—about 50 gigabytes on disk.

Why protein structure matters

Proteins build tissue, ferry signals, and fight infection. Their function follows from their shape, and that shape has always been difficult to determine. Experimental methods like X-ray crystallography or cryo-electron microscopy can take months per protein; some structures never yielded at all. AlphaFold was built to remove that bottleneck.

What the release contains

The current database holds roughly 350,000 predictions. By year-end, DeepMind aims to publish 100 million more, covering organisms across the tree of life. All data is free for any use with attribution, hosted permanently by EMBL.

What scientists are saying

DeepMind CEO Demis Hassabis called the release "the culmination of DeepMind's entire 10-year journey"—from game-playing AI toward real-world impact. EMBL Director-General Edith Heard called the eventual database "transformative for our understanding of how life works."

Outside reactions are cautiously optimistic. Marcelo C. Sousa, a biochemistry professor at the University of Colorado, said the data will not change patient care this year but will have a "huge impact for the scientific community." Scientists must recalibrate workflows when an entire reference dataset lands at once.

"Even 20,000 structures is far more than any biologist can manually review," said DeepMind senior researcher Kathryn Tunyasuvunakool. "Analyzing hundreds of thousands—it is a different world."

Our analysis

This is the Human Genome Project of structural biology: a public baseline that shifts the default starting point for research.

The strategic move is the open license. Anyone can use the data for anything with attribution—a direct challenge to closed prediction services and proprietary databases. A small biotech in Bangalore and a pharma giant start from the same footing. The bottleneck moves from "can we get the structure" to "what do we do with it."

Who wins:** Researchers in underfunded labs, drug hunters working on neglected diseases, structural biologists pivoting to applied work. **Who faces pressure:** Commercial prediction services, labs whose value depended on solving structures others couldn't. **The open question:** How well predictions hold up when tested experimentally—and whether the research community can develop tools to use this data at scale.

What to watch next

• Year-end expansion:** Will DeepMind hit its 100-million-structure target?
• Commercial adoption:** Watch for preclinical pipelines citing AlphaFold data.
• Accuracy debates:** The program assigns confidence scores; researchers must learn to interpret them.
• Attribution tracking:** How attribution is monitored across commercial use remains unresolved.

Key takeaways

• DeepMind released AlphaFold predictions for ~350,000 proteins across 20 organisms, including 98.5 percent of the human proteome.
• The data is free for any use with attribution, hosted permanently by EMBL, with 100 million structures targeted by year-end.
• The release is being compared to the Human Genome Project for its scope and downstream impact.
• Treatments are further off than headlines suggest, but early-stage discovery and target validation should accelerate dramatically.