The explosion of biomedical data in recent decades has created a critical challenge: validating automatically generated biological annotations. While computational algorithms can produce thousands of candidates in hours, determining which are biologically valid still requires costly and slow expert review. This bottleneck limits the speed of biomedical curation and, by extension, progress in areas such as genomics, pharmacology, and personalized medicine. To address this problem, researchers have proposed using biomedical knowledge graphs (bioKGs), which capture entities and their functional associations, combined with machine learning techniques that estimate the plausibility of each candidate annotation. This approach not only accelerates prioritization but also preserves human control in the final decision.
The core idea consists of training relation-specific binary classifiers using knowledge graph embeddings. However, the quality of these classifiers largely depends on the negative samples used during training. An innovative strategy is community-based negative sampling, which selects pairs of entities that likely have no true relation but share semantic context. This improves classifier robustness, increasing balanced accuracy by an average of 5.8% according to experiments on five large bioKGs. Furthermore, a family of plausibility measures is introduced that combines classifier confidence, classifier reliability, and the semantic context provided by alternative relationships between the same entities. These measures significantly outperform classifier confidence alone, enabling more effective prioritization of annotations for expert review.
From a technical and business perspective, this workflow is perfectly transferable to other sectors where validation of automatically generated data is critical. At Q2BSTUDIO, as a company specialized in software and technology development, we understand that combining artificial intelligence and knowledge graphs can be applied to domains such as logistics, finance, or cybersecurity. For example, for a client needing to curate large volumes of security incident reports, we can develop a custom AI solution that prioritizes the most plausible events using a threat knowledge graph. Cloud infrastructure, whether AWS or Azure, provides the scalability needed to process these graphs and deploy models efficiently and securely.
Moreover, cybersecurity plays a fundamental role when handling sensitive data, both in biomedicine and other sectors. At Q2BSTUDIO we integrate advanced protection measures in all our developments, ensuring that training data and predictions are safeguarded. On the other hand, monitoring and visualization of results greatly benefit from Business Intelligence tools such as Power BI, allowing experts to explore plausibility scores, filter by relationships, and make informed decisions. The combination of these technologies —custom applications, AI, cloud, cybersecurity, and BI— forms a complete ecosystem that accelerates curation without sacrificing quality.
A key aspect of the plausibility measures proposed in the study is their ability to consider alternative relationships between the same entities. For instance, a gene and a disease may be linked by multiple types of association (mutation, expression, protein interaction). Instead of treating each relation in isolation, the model evaluates the weight of each within the global context. This holistic view is similar to how modern AI agents integrate multiple information sources to make autonomous decisions. At Q2BSTUDIO we develop cloud infrastructure for intelligent agents that operate in real time, merging heterogeneous data and applying plausibility rules to prioritize tasks.
Implementing a prioritization system like the one described requires not only robust AI models but also a custom software platform that manages data flow, classifier training, and the expert review interface. At Q2BSTUDIO we have developed multiplatform applications for healthcare and R&D environments, integrating machine learning pipelines in the cloud. Our team of AWS and Azure experts ensures that the infrastructure is elastic, fault-tolerant, and cost-optimized. Additionally, Power BI dashboards allow curators to visualize the evolution of annotation quality and adjust plausibility thresholds in real time.
In short, prioritizing biomedical annotations using plausibility measures based on knowledge graphs and machine learning represents a significant advance for AI-assisted curation. This approach, although born in the biomedical field, is perfectly applicable to any sector where automatic data validation is a bottleneck. At Q2BSTUDIO we offer the necessary capabilities to design and implement similar solutions: from custom application development to integration with cloud services, including cybersecurity, Business Intelligence, and the creation of intelligent agents. If your organization needs to speed up the review of AI-generated candidates, contact us to explore how we can collaborate.
