Accepted_test

A hybrid text analysis approach was developed for the ANDSystem, aiming to enhance the extraction and interpretation of complex biomedical information from the ever-growing volume of biomedical literature. Traditional text-mining methods, although useful, often fail to capture subtle interconnections within biological data. To address these shortcomings, the revised ANDSystem integrates classical text-mining techniques with advanced technologies like graph neural networks (GNNs) and large language models (LLMs).

The approach consists of several key steps: identifying molecular-biological entities using ontology dictionaries, verifying these entities with the aid of context and a PubMedBERT model, and extracting relationships between entities through predefined rules and templates. Additionally, it computes co-occurrences to assess interaction likelihoods, trains a graph neural network to vectorially represent knowledge graph vertices, and develops a binary classification model for predicting interactions based on these representations.

The effectiveness of this hybrid method is evident in its ability to more accurately identify and predict biologically relevant interactions, with significant improvements in detecting both known and novel interactions through contextual analysis performed by LLMs and GNNs. This integrated system not only enhances the reliability of reconstructing biological networks but also supports advanced research by providing a robust platform for detailed biological interpretation.