Proteins are the molecular workhorses of cells. To understand their functions, scientists soon realized that they had to figure out proteins’ shapes. But solving a protein structure is a challenging task: It requires purifying the protein of interest, finding the ideal conditions for it to crystalize (when necessary), and then using methods such as X-ray crystallography to determine the exact arrangement of atoms in the macromolecule.
Understanding a single protein structure might not tell researchers much about its function, though, because many proteins work as complexes with these protein-protein interactions dictating the pace of virtually every cellular process.
“Biology really happens when things come together,” said Martin Steinegger, a computational biologist at Seoul National University. “[Proteins] function together as units. They have their interaction partners. They have their conformational states. They have their stoichiometries. All of these things are somehow only visible once you go from one unit to multi-units,” he explained.
Steinegger’s team is part of an ambitious collaboration between Google DeepMind, NVIDIA, and the European Bioinformatics Institute at the European Molecular Biology Laboratory (EMBL-EBI) that aims to make the prediction of protein complex structures at scale possible using the AI-powered AlphaFold system.
In a recent announcement and preprint, the researchers shared the initial fruits of their labor by adding almost two million high-confidence protein complex structure predictions from different taxonomic groups to the AlphaFold Protein Structure Database.1
Charting the Path to Protein Complex Prediction
In 2021, the release of AlphaFold2 showed that it was possible to predict protein structures with almost 90 percent accuracy—an unprecedented feat as previous systems, including AlphaFold’s first version, could only predict the three-dimensional structure of amino acid sequences with up to 60 percent accuracy.2,3 To improve AlphaFold’s prediction capabilities, Google’s DeepMind partnered with Steinegger, who suggested using a database of metagenomic-sourced proteins, now known as the Big Fantastic Database (BFD), as the basis for AlphaFold’s multiple sequence alignment (MSA). In an MSA, similar protein sequences from different organisms are arranged next to each other, and AlphaFold compares and analyzes these sequences, looking for similarities and differences to better understand the evolutionary relationship among these proteins. AlphaFold uses these relationships to produce protein structure predictions.
AlphaFold2’s highly accurate 3D protein structure predictions fueled the tool’s popularity among researchers; yet, the system had its limitations, including the fact it could not predict protein complexes, especially at scale.
AlphaFold 3.0: Structures of Protein Complexes
To take AlphaFold from single protein structure predictions to protein complex predictions, the researchers in the multi-team collaboration brainstormed ideas to identify key aspects to optimize. First, they realized that they needed a more efficient method to generate MSAs. Second, to be able to scale up the AlphaFold pipeline to make millions of predictions fairly quickly, they needed more computational power. Finally, they had to figure out how to analyze the data, ensure its accuracy, and display it in a way that scientists could use it.
To overcome some of these limitations, Steinegger collaborated with NVIDIA to develop the MMseqs2-GPU tool, which sped up the search for protein homologs in large databases and the subsequent protein structure predictions without compromising accuracy.4 Then, scientists led by structural bioinformatician Sameer Velankar at the EMBL-EBI tested metrics to ensure that AlphaFold had high certainty in its protein structure predictions. Velankar and his team also redesigned the interface of the AlphaFold Database, which they manage and through which AlphaFold’s predictions are freely available, to offer scientists access to high-quality information about the protein complexes.
While recent efforts to predict protein complex structures have focused on model organisms, including humans, Steinegger explained they wanted to create a more comprehensive database that also included species that are not as well-studied.5
“This is really critical from the perspective of comparative genomics, [to] look at genomes [of] different organisms,” he said. “But you can also now think of comparative structure-omics, where you take complexes and compare these complexes across other organisms at scale, so you can actually get structural insights, structural differences between humans and other organisms. That's really where a lot of interesting biology comes out.”
Using the latest AlphaFold pipeline, the team predicted almost 31 million protein complex structures. Of those, they deemed 1.8 million homodimers—protein complexes consisting of two identical proteins—high-confidence and made them available to the scientific community in the AlphaFold Database. They classified the remaining complexes either as still under analysis or low-confidence, with the latter made available as a list for scientists to download. The team expects to add more high-confidence predictions to the AlphaFold Database in the near future.
Even though the researchers took several steps to ensure that the structural information added to the database was sound, Velankar cautioned they are still predictions, meaning that they might not exactly reflect how two proteins interact. That is where the community of users comes into play, he explained. Researchers using the AlphaFold Database can flag errors in the protein complex predictions and communicate them to the team, helping to keep the database as accurate as possible.
“Understanding protein function means understanding how it makes complexes, understanding what the dynamics is,” explained Velankar. “We are taking steps in that direction. We started with single chains, and now we have gone to having homodimer or heterodimer prediction. It still means there are many more things to do, but this is a next step in terms of how we can help researchers.”
- Han Y, et al. AlphaFold Database expands to proteome-scale quaternary structures. Preprint. 2026.
- Senior AW, et al. Improved protein structure prediction using potentials from deep learning. Nature. 2020;577(7792):706-710.
- Jumper J, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596(7873):583-589.
- Kallenborn F, et al. GPU-accelerated homology search with MMseqs2. Nat Methods. 2025;22(10):2024-2027.
- Zhang J, et al. Predicting protein-protein interactions in the human proteome. Science. 2025;390(6771):eadt1630.
