Claudio Mirabello is one of the researchers that Björn Wallner collaborates with. They combine the computer models’ predictions with experiments, to better understand what different proteins look like and how they interact. Photo credit Thor Balkhed Almost everything in our bodies is dependent on proteins. They are the building blocks that make up the core of our cells, and make the cells stick together in tissue. But also, and perhaps more importantly, proteins are the machinery of cells. Everything from sending signals within and between cells, repairing damage to DNA and regulating blood sugar is done by proteins. This often requires a complex machinery of many different proteins that come together at the right time, do their thing, and then dissolve.
“To understand life, both in its healthy, normal state and when affected by disease, we need to understand how proteins interact. It’s that interaction, how proteins talk to each other, that we need to get an insight into through our research,” says Björn Wallner, professor of bioinformatics at the Department of Physics, Chemistry and Biology, IFM, at LiU.The LiU researchers develop computer models to understand what proteins look like, and how they interact. They then apply their methods in projects, together with research colleagues who carryout experiments.
One of the major questions
The human body has some 20,000 different proteins. A protein is basically a long chain of 20 different amino acids, bound together like the pearls of a necklace. But the function of the protein depends on how this chain is folded into a three-dimensional shape.Björn Wallner leads a research group working on structural bioinformatics, which is part of the large research field of data-driven life science. Photo credit Thor Balkhed
“Finding out the order of the amino acids, what we call the protein sequence, is quite easy. But figuring out the three-dimensional structure of the protein is much harder,” says Björn Wallner.
The problem of calculating the shape of a protein based on its amino acid sequence is something that researchers have wrestled with for about 50 years now. Researchers nowadays use computers to create models and carry out the extensive calculations necessary. Every two years, researchers can compare and assess methods, to find the best ones, in the international CASP (Critical Assessment of protein Structure Prediction) experiment. Björn Wallner is one of many researchers who have taken part in this assessment. All participants had a number of protein sequences sent to them, and then used their own computer models to try to predict the structure of the proteins.
But in 2020, something happened. The Google company DeepMind presented AlphaFold – its latest version of an artificial intelligence, AI, developed to predict protein structures. Some years earlier, DeepMind had developed an AI that can play Go (a strategic board game that has been played in Asian countries for more than 2,500 years) better than a professional Go player. At the 2020 CASP, AlphaFold outclassed the other models used to calculate protein structure.
“This is a revolution in structural biology. AlphaFold is one of the first AI-based methods to solve a major problem in biology. In principle, they solved the problem of predicting structures of individual proteins,” says Björn Wallner.
LiU’s model in the top tier
In the summer of 2021, AlphaFold was released as open-source software, and the method behind it was published in Nature. It was thereby made available to the public. At the following CASP assessment, all participating research groups had developed AlphaFold further, as it was so good. The December 2022 comparison showed that Björn Wallner’s model was in the top tier in the category of prediction of interaction between several different proteins.Björn Wallner. Photo credit Thor Balkhed
“It’s a big deal that we can compete with large groups worldwide that have much larger resources.”
Björn Wallner’s further development of AlphaFold is based on a simple principle.
“I realised that AlphaFold got stuck on its first guess, even if you repeated the run several times. You’d want it to make different guesses, as the first one might not be correct. So, by introducing random perturbations, I forced it to come up with more guesses.”
According to Björn Wallner, one big advantage of AlphaFold is that it is very good at assessing itself and saying how good its guess is. That is why he wanted the model to generate more possible structures and choose the very best ones.
“It’s a simple principle, which makes it possible for everyone else to do it in the same way now. My code is available for everyone to run. A research group only interested in a few proteins can do the calculations on an ordinary computer using one or more good graphics cards.”
A strategy that could boost research
He believes that many people in his field will use his strategy in their models, mainly because it is simple. The model has already proven useful in some of his research collaborations, and he thinks that it may turn out to be an effective tool, especially for researchers interested in proteins that take on several different shapes and that integrate with several different proteins.
The AI model has already led to the discovery of new interaction surfaces between two proteins.
“We first thought that there was something wrong with the run. But since the model is so good at assessing its guesses, we double-checked the result using several experimental methods. It turned out that the model was right, and that there was an alternative binding site that no one had found before. So, I think that the model has many research applications.”