February 03, 2022

"Scientists could develop new medicines more quickly with the help of the machine-learning model."

The researchers at MIT developed a machine-learning model that can accurately predict the complex formed when two proteins interact. Their technique frequently predicts protein structures similar to experimentally observed structures.

This technique may aid scientists in gaining a better understanding of specific biological processes; it may also expedite the development of new medicines.

"Deep learning is extremely effective at capturing interactions between different proteins that would be difficult for chemists or biologists to write experimentally," explains Octavian-Eugen Ganea, a postdoctoral researcher at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL).

Protein Attachment

According to Equidock, a new docking model developed by scientists, proteins attach via rotation or translation in three-dimensional space. Still, the proteins' shapes don't squeeze or bend during the process.

The neural network can process the 3D graphs created by the model from the 3D structures of two proteins. A node in the chart represents each amino acid in a protein chain, and each node represents a chain of amino acids.

Geometric knowledge was incorporated into the model by the researchers to understand how objects in 3D space can change when rotated or translated. Even in a 3D environment, proteins always attach similarly, thanks to mathematical knowledge built into the model. It's how proteins interact with each other in the human body.

Binding-pocket points are atoms in the two proteins that are most likely to interact and form chemical reactions, and the machine-learning system uses this information to identify them. Finally, this method combines the two proteins into one complex.

Moreover, this model was overcoming the model's lack of training data. Due to the scarcity of experimental 3D data for proteins, Ganea emphasises the importance of incorporating geometric knowledge into Equidock. Without these geometric constraints, the model may identify spurious correlations in the data.

Seconds Vs Hours

Following training, scientists put through the model its paces against a set of four software approaches. Additionally, this model measures how well the predicted protein complex matches the actual protein complex, and Equidock performed well in most cases, but it occasionally fell short of the baseline.

Along with applying this method to traditional models, the team hopes to incorporate specific atomic interactions into Equidock to improve prediction accuracy. For instance, atoms in proteins can occasionally bond via hydrophobic interactions involving water molecules.

Ganea notes that their technique could develop small, drug-like molecules. Because these molecules bind to protein surfaces in unique ways, rapidly determining how they attach could significantly shorten the time required for drug development.

They intend to enhance Equidock in the future to make predictions for flexible protein docking. The primary impediment is a lack of training data, which is why Ganea and his colleagues are working to generate synthetic data to improve the model.

Image source: Unsplash

Article reference: https://news.mit.edu/2022/ai-predicts-protein-docking-0201

Dr Nivash Jeevanandam PhD,
Researcher | Senior Technology Journalist