A Hybrid AI Architecture for Transparent and Trustworthy Biomedical Hypothesis Evaluation

The exponential growth of biomedical publications and structured data resources has created new opportunities for data-driven discovery of novel hypotheses, such as potential interactions among proteins, diseases, and chemical compounds. However, current approaches to hypothesis generation often lack transparent explainability. In this paper, we propose a novel hybrid Artificial Intelligence evaluation framework to support experts in hypothesis generation analysis through the integration of Large Language Models (LLMs) and Dimensions Knowledge Graphs (DKG), a specialized Knowledge Graphs structure to represent large-scale structured representation of scientific knowledge that interlinks publications, entities, and metadata across the research lifecycle. The DKG to provide an explicit, machine-readable model of entities and relationships, enabling traceability, contextualization, and multi-hop exploration of connections between biomedical concepts. On the other hand, LLMs contribute by synthesizing graph-derived evidence from the DKG, thereby supporting interpretability of the analysis while remaining grounded in structured knowledge. The approach has been developed through the Design Science Research methodology. The evaluation phase demonstrates the usefulness and effectiveness of the approach in supporting hypothesis analysis and prioritisation tasks, with domain experts involved in the process.