MIT researchers build neural network to speed up drug development
Antibodies (small proteins) produced by the immune system can attach to specific parts of a virus to neutralize it. As scientists continue to fight SARS-CoV-2, the virus that causes Covid-19, one possible weapon is a synthetic antibody that binds to the virus's spike proteins to prevent it from entering human cells.
To develop a successful synthetic antibody, researchers must understand exactly how this happens. Proteins with lumpy three-dimensional structures containing many folds can stick together in millions of combinations, so finding the right protein complex among almost countless candidates takes an extremely long time.
To simplify the process, researchers at the Massachusetts Institute of Technology created a machine learning model that can directly predict the complex formed when two proteins join. Their method is 80–500 times faster than modern software methods and often predicts protein structures that are closer to the structures observed experimentally.
This method could help scientists better understand some biological processes involving protein interactions, such as DNA replication and repair. It could also speed up the development of new drugs.
“Deep learning is very good at capturing interactions between different proteins that are difficult for chemists or biologists to record experimentally. Some of these interactions are very complex, and people have not found good ways to express them. This deep learning model can learn these types of interactions from data”
The model developed by the researchers, called Equidock, focuses on rigid-body docking, which occurs when two proteins join by rotating or moving in three-dimensional space, but their shapes do not compress or bend.
The model takes the three-dimensional structures of two proteins and converts them into three-dimensional graphs that can be processed by a neural network. Proteins consist of chains of amino acids, each represented by a node in the graph.
The researchers built geometric knowledge into the model, so it understands how objects change when rotated or moved in three-dimensional space. The model also has built-in mathematical knowledge that ensures proteins always attach the same way regardless of where they are in three-dimensional space. This is how proteins dock in the human body.
Using this information, the machine learning system identifies atoms on the two proteins that are most likely to interact and form chemical bonds, known as binding points. It then uses these points to join the two proteins into a complex.
“If we can tell from the proteins which specific parts are likely to be these binding points, that provides all the information needed to join two proteins together. Assuming we can find these two sets of points, we can simply work out how to rotate and move the proteins so that one set matches the other”
One of the biggest challenges in building this model was overcoming the shortage of training data. According to Ganea, because there is so little experimental three-dimensional data for proteins, it was especially important to include geometric knowledge in Equidock. Without these geometric constraints, the model could pick up false correlations in the dataset.
After training the model, the researchers compared it with four software methods. Equidock can predict the final protein complex in just one to five seconds. All baselines took much longer, from 10 minutes to an hour or more.
On quality metrics that measure how closely the predicted protein complex matches the actual one, Equidock was often comparable to the baselines but sometimes fell short of them.
“We still lag behind one of the baselines. Our method can still be improved, and it can still be useful. It could be used in very large virtual screening, where we want to understand how thousands of proteins may interact and form complexes. Our method could be used to very quickly generate an initial set of candidates, which could then be fine-tuned with some of the more accurate but slower traditional methods”
In addition to using this method with traditional models, the team wants to incorporate specific atomic interactions into Equidock so it can make more accurate predictions. For example, sometimes atoms in proteins join through hydrophobic interactions, which involve water molecules.
According to Ganea, their method could also be applied to designing small, drug-like molecules. These molecules bind to the protein surface in a specific way, so quickly determining how this attachment occurs could shorten drug development timelines.
In the future, they plan to improve Equidock so it can predict flexible protein docking. The biggest obstacle is the lack of training data, so Ganea and his colleagues are working on creating synthetic data they could use to improve the model.
