The AI-based software AlphaFold2 has revolutionized the study of 3D protein structures. Players from Northern Germany are in the middle of this protein folding revolution and are involved in improving applications of the software.
What is AlphaFold2 and how does it work?
For structural biologists around the world, July 2021 marked a turning point. London-based firm DeepMind – part of Google’s parent company Alphabet – released an artificial intelligence (AI) tool called AlphaFold2. The deep learning algorithm can predict the 3D shape of proteins from their genetic sequence with almost pinpoint accuracy. This has transformed the studies of thousands of biologists worldwide. Researchers in the North of Germany are in the middle of this AI protein-folding revolution. One of them is Jan Kosinski, a structural bioinformatician who leads a research group at the Hamburg branch of the European Molecular Biology Laboratory (EMBL). He is also a member of the Centre for Structural Systems Biology (CSSB) at DESY. Kosinski’s passion is to solve tricky biomolecular puzzles – which means elucidating the accurate 3D architecture of proteins.
Figuring out the 3D shape of a giant
One structural biology masterpiece was the human nuclear pore complex which Kosinski‘s team solved together with Martin Beck’s team from the Max Planck Institute of Biophysics and other collaborators. The nuclear pore complex (NPC) is a true molecular giant which sits on the membrane separating the nucleus from the cytoplasm. It is doughnut-shaped and works as both a gateway and a checkpoint for molecules that travel between the cytoplasm and the nucleus. For structural biologists, the human NPC is a challenging yet exciting 3D puzzle with around 30 different proteins each present in multiple copies. This amounts to around 1000 puzzle pieces.
The key to figuring out the complex‘s architecture was to combine several experimental and computational methods. “AlphaFold2 was a breakthrough moment for us,” says Agnieszka Obarska-Kosińska, a postdoc in Beck’s and Kosinski’s labs who performed the molecular modelling. “Before, we didn’t know anything about the structure of many of the proteins within the NPC. You cannot assemble a puzzle when you don’t know what the pieces look like. But AlphaFold2, combined with other approaches, enabled us to predict those shapes.”
Rapid progress thanks to AlphaFold2
The resulting model, published in top journal “Science”, was so complete and detailed that it enabled the researchers to create time-resolved molecular simulations that shed more light on how the NPC works.
Jan Kosinski sees three aspects that have made AlphaFold2 a game changer for his and his team’s way of doing protein science.
- “It speeds up our work – we get a good impression of what the protein we are working on looks like within minutes instead of months or years,” he says.
- “It makes our work easier,
- and finally, AlphaFold2 makes our work more fun – many technical and tedious steps are now automated so we can focus more on our biological questions!”
The further development of AlphaFold2
Since its sensational debut, AlphaFold2 has been developed further.
Kosinski points out that scientists now have a better understanding of where the software works best, and in which aspects it cannot replace experimental methods. For example, while AlphaFold2 gives an idea about the structure of a protein, X-ray and high-resolution cryo-electron microscopy (cryo-EM) methods still usually give the more precise structures needed for understanding enzymatic reactions or designing drugs. DeepMind and EMBL’s European Bioinformatics Institute (EMBL-EBI) have partnered to create a database for making the predictions freely available to the scientific community. The database now includes over 200 million protein models.
“This means we have models for almost every catalogued protein known to science,” Kosinski says. After having cracked the structural code of the NPC, his group at EMBL Hamburg is now focusing on even more difficult molecular puzzles in infection biology. “We are now using AlphaFold2 in combination with our own software to model the structures of viruses such as influenza and Lassa to map their interactions with human proteins. We have also just started a bigger project on the malaria parasite funded by the CSSB.”
AlphaFold2 was a breakthrough moment for us.
Dr Agnieszka Obarska-Kosińska
EMBL Hamburg and MPI of Biophysics
EMBL and DESY create a hotspot for structural biology
Jan Kosinski is convinced that Hamburg has the potential to develop into a hotspot for AI-based structural biology. At EMBL Hamburg and CSSB, there are already a number of AI-based projects in structural biology. There is also a growing AI community at the DESY campus including researchers from other disciplines such as physics or instrumentation. “I think our potential to attract AI experts or students lies in tackling important problems in fundamental biology and questions relevant to human health.”
Text: Philipp Graf
Featured image: © Agnieszka Obarska-Kosińska
