HSE Researchers Train Neural Network to Predict Protein–Protein Interactions More Accurately

© iStock

Scientists at the AI and Digital Science Institute of the HSE Faculty of Computer Science have developed a model capable of predicting protein–protein interactions with 95% accuracy. GSMFormer-PPI integrates three types of protein data (including information about protein surface properties) to analyse relationships between proteins, rather than simply combining datasets as in previous models. The solution could accelerate the discovery of disease molecular mechanisms, biomarkers, and potential therapeutic targets. The paper has been published in Scientific Reports.

Almost all cellular processes depend on interactions between proteins. Cells use these interactions to transmit signals, initiate and regulate chemical reactions, and form molecular complexes essential for proper functioning. When such interactions are disrupted, cellular processes can malfunction, potentially leading to disease.

Therefore, to study disease mechanisms and identify therapeutic targets, it is important for scientists to understand which proteins can interact and which cannot. Determining this experimentally is difficult: when dozens or hundreds of proteins are considered, the number of possible pairs becomes too large to test individually. As a result, biologists use machine learning methods to predict these interactions based on the structure and properties of molecules.

HSE researchers have developed the GSMFormer-PPI system, which takes into account three types of data for each protein in a candidate pair: the amino acid sequence, the three-dimensional structure, and the properties of the molecular surface. To process this information, the authors used existing models that convert this data into numerical representations. A protein language model analyses the amino acid sequence—the order of amino acids that make up the protein. The three-dimensional structure of the protein is represented as a graph, in which amino acids are treated as nodes and their spatial contacts as edges; this representation is processed by a graph neural network. In addition, a separate algorithm captures protein surface properties—the shape and physicochemical characteristics of the regions through which proteins recognise one another.

These numerical representations of proteins were then fed into a transformer module developed by the authors—a neural network that jointly analyses different types of protein data. In contrast to many previous approaches, where features were often simply concatenated into a single vector, this model does not combine them mechanically but instead captures the relationships between them.

Maria Poptsova

'When proteins interact, their surface is particularly important: it is through the surface that molecules recognise one another, and it is where the physicochemical properties that determine binding are concentrated. In our model, we sought to incorporate this information alongside the protein’s sequence and three-dimensional structure and not merely concatenate these features but enable the algorithm to analyse the relationships between them. This is what allowed us to predict protein–protein interactions more accurately,' comments one of the authors, Maria Poptsova, Director of the Centre for Biomedical Research and Technology at the HSE FCS AI and Digital Science Institute.

General schema of the proposed GSMFormer-PPI model. Panel A illustrates the different types of protein representations used by the model: structural, sequential, and surface-based. Panel B shows how these representations are projected to a common dimensional space, processed by a transformer, and then used to generate the final prediction of the interaction.

© Arteaga, D., Chervov, N. & Poptsova, M. Multimodal graph, surface, and language-based model for protein protein interaction prediction. Sci Rep 16, 4772 (2026).

The researchers tested the new model’s performance on the PINDER dataset, a large database of known protein interactions. In these experiments, GSMFormer-PPI achieved an accuracy of 95.7%, outperforming popular graph-based models such as GCN and GAT. The researchers also tested a simpler version of GSMFormer-PPI—without the module that analyses relationships between different types of data. This version performed worse, demonstrating that it is not only the protein data itself but also how the model integrates and compares it that drives its accuracy.

Additional tests showed that all three types of data—sequence, spatial structure, and surface properties—are essential for accurate predictions. When the researchers removed any one component, prediction accuracy declined. In other words, the model performs better precisely because it considers the protein on multiple levels simultaneously. In the future, such systems could help identify protein pairs more efficiently when studying disease mechanisms and searching for drug targets.

The work was supported by a grant for research centres in AI provided by the Ministry of Economic Development of the Russian Federation and implemented at HSE University.