Scientists show how artificial intelligence can help to study protein complexes

Researchers at Martin Luther University Halle-Wittenberg (MLU) and the European Molecular Biology Laboratory in Hamburg have developed a new method for studying proteins. The team has succeeded in developing an AI-based method for analysing cryo-electron microscopy data. In future, this will enable several protein complexes to be simultaneously examined directly in cells. The team presents its work in the scientific journal "Structure".

Cryo-electron microscopy is a relatively new method that can be used to study the structure of materials, cells or proteins. In principle it works like this: Samples are instantly frozen and bombarded with electrons. The microscope initially creates two-dimensional images that can then be compiled into a 3D model. In the case of proteins, this allows extremely detailed images of their structure to be made at the single-atom level. Knowledge about the structure of proteins is crucial. "Proteins are the workhorses of the cell; they control all of the important processes - from bone growth to metabolism. However, the role a protein plays and how it functions cannot be precisely understood until its structure is known," says Assistant Professor Panagiotis Kastritis from the Centre for Innovation Competence HALOmem at MLU. This knowledge can, in turn, be used to create treatments for a number of diseases such as Alzheimer’s or cancer. 

However, cryo-electron microscopy only provides three-dimensional structures and cannot determine which protein is which. The data must be interpreted by the researchers themselves, and that can lead to potential errors, as Ioannis Skalidis, a doctoral student in Kastritis’ research group, explains: "The 3D structure of some proteins is nearly identical even though they have completely different functions in the body. When researchers use their experience to decide how to interpret the data and identify the proteins, this can greatly impact the results of their experiments." Therefore, the team developed a new method that automatically interprets the raw data from the electron microscope. To do this, the scientists used several of open-source programmes, some of which they developed themselves, as well as DeepMind’s AlphaFold. In July 2021 its developers reported that they were able precisely predict the structure of most human proteins. Until then, this was considered one of molecular biology’s trickiest puzzles. 

In the new study, the team from Halle and Hamburg examined protein complexes as they occur in cells rather than looking at isolated proteins. "Proteins generally do not function on their own, but in larger groups. The structure of their binding sites changes depending on what the proteins are working alongside. That is why it is so important to study them under conditions that are as realistic as possible," explains Kastritis. The new combination of methods used by the team will also enable other research groups to more easily analyse the structures of such protein complexes in the future. 

The work was funded by the Federal Ministry of Education and Research, the European Regional Development Fund (ERDF),  Saxony-Anhalt, and the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation).

Study: Skalidis I., Kyrilis F.L., Tüting C., Hamdi F., Chojnowski G., Kastritis P.L. Cryo-EM and artificial intelligence visualize endogenous protein community members. Structure (2022). doi: 10.1016/j.str.2022.01.001