Combined Approach to Associative Network Reconstruction: Integrating GraphSAGE and Co-Occurrence Statistics

This study focuses on developing a hybrid approach for predicting molecular-genetic interactions, combining graph neural networks (GNNs) and co-occurrence analysis of entities in scientific literature. The method’s effectiveness is demonstrated using the associative network of Escherichia coli, reconstructed using the ANDSystem and its ANDDigest module. Results showed a significant improvement in the accuracy of interaction predictions, in terms of conformity to the original graph topology, compared to using GNNs alone. The combination of approaches improved the Fl-score from 0.815 to 0.97 and reduced the loss function value from 0.405 to 0.08. Evaluation on experimentally confirmed protein-protein interactions also demonstrated high model efficiency (Fl-score 0.9799, Matthews correlation coefficient 0.9597). The proposed method can be applied in analyzing complex biological systems, planning experiments, and optimizing biotechnological processes.