Russian scientists have significantly improved the accuracy of neural networks in one of the most complex tasks in biology — predicting interactions between proteins. The new approach allows predicting such processes with an accuracy of up to 95.7%, which opens up additional opportunities for drug development and disease study.
The work was published in the journal Scientific Reports. The research was conducted by a team led by Maria Poptsova, Director of the Center for Biomedical Research at the Institute of Artificial Intelligence and Digital Sciences of HSE University.
The main idea is to consider not only the structure and sequence of proteins, as most existing models do, but also the properties of their surface. This factor is often overlooked, although it is critical for understanding how molecules "interact" with each other.
The method is based on a combination of several types of neural networks that simultaneously analyze the amino acid sequence, three-dimensional structure, and physical characteristics of the protein surface. This approach allows significantly increasing accuracy compared to popular algorithms like GCN and GAT.
The technology was tested on one of the largest datasets on protein interactions. As a result, the new model showed an accuracy of 95.7% — higher than existing solutions.
This is an important step for science: disruptions in protein interactions underlie many diseases. The more accurately such processes can be predicted, the faster it is possible to find potential "targets" for drugs and better understand the mechanisms of diseases — without expensive and lengthy experiments.