Google DeepMind expanded its AlphaFold protein database to over 200 million predicted 3D structures, covering nearly every sequenced organism. While the massive catalog accelerates research in life sciences, structural biologists warn that predictions cannot replace painstaking experimental structure determination.
Mapping the Protein Universe With Artificial Intelligence
Biologists gained glimpses into the molecular world when an artificial intelligence system rendered the predicted 3-D shapes of more than 200 million proteins. The July release expanded a public database initially launched in 2021, which had debuted with 350,000 structures, including nearly all known human proteins. Created by London-based DeepMind, AlphaFold combines deep-learning techniques to predict how long chains of amino acids fold into complex architectural shapes by recognizing patterns in structures previously solved through decades of experimental work using electron microscopes and other methods.
Demis Hassabis, cofounder and CEO of DeepMind, described the scope during a July 26 news briefing by stating, You can think of it as covering the entire protein universe.
Hassabis also noted that the database now encompasses almost every organism on the planet that has had its genome sequenced
and highlighted the system’s accessibility, pointing out that You can look up a 3-D structure of a protein almost as easily as doing a key word Google search.
DeepMind partnered with the European Bioinformatics Institute of the European Molecular Biology Laboratory to host the public library and make the structures available in a public database.
“You can think of it as covering the entire protein universe.”
Demis Hassabis, cofounder and CEO of DeepMind
Translating Predictions Into Real-World Biological Insights
Knowing a protein’s architecture carries direct implications for human health. A protein’s shape dictates how enzymes capture small molecules for chemical reactions and how multi-protein complexes snap into formation. Researchers have already utilized some of the 2021 predictions to develop potential new malaria vaccines, improve understanding of Parkinson’s disease, work out how to protect honeybee health, gain insight into human evolution and more. DeepMind has also focused AlphaFold on neglected tropical diseases, including Chagas disease and leishmaniasis, which can be debilitating or lethal if left untreated.
Building on AlphaFold 2, Google DeepMind developed a separate system called AlphaMissense to predict whether a missense genetic variant is likely to be pathogenic or benign. Missense variants are the most common type of genetic variant, involving a single change in the DNA sequence that results in a substitution of one amino acid for another in a protein. Analyzing a massive dataset of variants that includes each variant’s frequency in the human population and its location in the protein sequence—alongside structural context derived from AlphaFold—AlphaMissense has already provided high-confidence predictions for most human missense variants (Cheng et al., 2023). Meanwhile, AlphaFold 3 extends beyond protein structure prediction to model intricate biological assemblies involving DNA, RNA, ligands, ions, and diverse chemical modifications.
Addressing the Limits of Computational Modeling
Despite the enthusiasm surrounding computational structural biology, the release of the vast dataset was greeted with excitement by many scientists, while others worry that researchers will take the predicted structures as the true shapes of proteins. Decades of slow-going experiments have revealed the structure of more than 194,000 proteins, all housed in the Protein Data Bank. There are still things AlphaFold can’t do — and it wasn’t designed to do.
Furthermore, AlphaFold 3 (AF3) uses a diffusion generative model, a powerful class of AI also used for image generation, for structure prediction. The diffusion generative model first creates an initial state in which the atoms of the protein are scattered at random by noise, and then removes that noise, moving the atoms toward positions of higher probability corresponding to positions of lower energy. However, conformational changes in proteins are vital to their function yet remain challenging for state-of-the-art artificial intelligence, such as AlphaFold3, to predict. While proteins function by switching between multiple conformational states, AlphaFold is known to predict only a single conformation for many proteins, thereby limiting its applicability to the life sciences, including drug design.
To overcome this, researchers at the Institute for Molecular Science (IMS), National Institutes of Natural Sciences, and the Graduate University for Advanced Studies, SOKENDAI—specifically the research group of Jun Ohnuki and Kei-ichi Okazaki—set out to develop a novel AlphaFold-based method for sampling conformational changes. The group introduced a repulsive force between predicted structures in AlphaFold, allowing AlphaFold3 to sample the multiple conformational states that its default settings rarely capture, and succeeded in predicting protein conformational changes that had been difficult for AlphaFold with its default settings. These results are published online in JACS Au.
