A pipeline towards missing IS-A relationship discovery in the Gene Ontology

As the Gene Ontology (GO) serves as a foundational resource for biological knowledge, ensuring the completeness of its semantic structure is critically important. In this study, we introduce a novel automated pipeline to identify missing IS-A relationships in GO. The innovative core of our approach lies in employing a CNN-based model with attention, GO-FocusNet, to predict the existence of IS-A relationships between currently non-IS-A concepts. However, the large volume of non-IS-A concept pairs remains an obstacle to prediction efficiency. To address this, we retrieve concept pairs with a high likelihood of being missing IS-A relationships as candidates, by leveraging subtree structures and Box²EL-based concept embeddings. Furthermore, to minimize manual effort, we implement a two-step validation process: positive predictions are first verified using Deepseek and subsequently reviewed by domain experts. Experiments conducted on GO versions from 2022 to 2025 confirmed the effectiveness of our approach, demonstrating its value as a scalable and practical solution for ontology quality assurance.