Proteins are among the most important building blocks of nature and play a central role in biological processes in all organisms. Accordingly, scientists are keen to understand them as precisely as possible. As polymers of different amino acids, proteins can have different 3-dimensional structures and various functions. However, it is often difficult to determine which amino acids influence protein function and which influence structural stability. Using the so-called Function-Structure-Adaptability (FSA) approach, a team led by Andreas Winkler and Oliver Eder from the Institute of Biochemistry at Graz University of Technology (TU Graz) have achieved a breakthrough, which they have now published in the journal “Structure”. FSA compares machine-learning-generated, idealised protein sequences with natural sequences that have developed over millions of years of evolution. This allows the amino acids that are crucial for function and stability to be identified with unprecedented accuracy. This knowledge provides an important basis for the production and modification of proteins and thus for the development of new drugs, for the targeted improvement of proteins in industrial applications and for a better understanding of protein changes, for example in connection with antibiotic resistance.

Understanding the building blocks of life better

“As biochemists, we want to understand how proteins have evolved in nature and thus find out which amino acids are relevant for specific functions,” says Andreas Winkler. “To do this, we combined what nature has conserved during evolution with what an AI model considers relevant for the stability and structure of a protein. This combination of millions of years of evolutionary history and the latest technology greatly simplifies the analysis and understanding of proteins.”

For its method, the team used the deep learning model ProteinMPNN, which generates new protein sequences with the aim of ensuring that they adopt a predetermined stable, three-dimensional structure. The researchers compared these sequences with those in natural proteins. As a test system, the bacteriophytochrome protein family was utilized, which in nature serves as a photoreceptor for some bacteria and plays a central role in the perception of environmental influences such as light. The new analysis method revealed that if an amino acid is repeatedly represented in the natural sequences, but does not appear to be significant for ProteinMPNN, this indicates a functional role. However, if it is strongly present in both sequence collections, this is an indication of structural significance.

Validation in the laboratory

For their approach, the researchers had to group the amino acids based on chemical properties in order to then statistically compare natural and AI-generated proteins. This made it possible to classify amino acids into three categories: “functional” (important for the specific role of the protein), “structural” (relevant for stability and folding) and “adaptable” (a third category that still requires further research). The team validated the results by means of extensive laboratory experiments in which they were able to influence the functional properties of proteins by making specific changes to correspondingly classified amino acids. This made it possible, for example, to significantly influence the light perception of the photoreceptor test system. The comparison with functional residues already known from the literature also confirmed the high hit rate of the new analysis method.

“In the past, it often took several months or even years of preparatory work and laboratory work to carry out an analysis like this,” says Oliver Eder. “The preliminary work to identify potentially interesting natural protein sequences is now possible for a new protein within a week. And because our method allows us to pre-filter the functional amino acids much more specifically, we don’t have to spend so much time in the laboratory on testing and characterisation. As the method can in principle be applied to all protein classes, we can now appreciate the intricate details of how proteins work in a more targeted way.”