Abstract A022: Agentic AI–Driven Knowledge Graph Centrality with Reciprocal Rank Fusion for Interactive Discovery of Cancer Driver Genes and Pathways
Systematic identification of cancer driver genes and actionable pathways remains constrained by the complexity of multi-scale biological networks and the limitations of single-metric prioritization approaches. Knowledge graphs (KGs) integrating phenotype–disease–gene/protein–pathway relationships provide a comprehensive representation of disease biology, yet extracting high-confidence targets requires robust ranking and interpretability. We hypothesize that combining multi-centrality analysis with rank fusion and agentic AI can enable more stable, explainable, and interactive discovery of disease drivers. Methods: We developed an agentic AI framework that extends AI coding assistants into domain-specific “discovery agents” capable of executing graph analytics workflows interactively. The system operates on an integrated KG (>13K nodes, >600K edges) spanning disease, phenotype, gene/protein, and pathway relationships aggregated from public biomedical sources. For a given disease query, the framework: 1. Dynamically extracts disease-specific subgraphs (depth-controlled expansion: 1–3 hops) 2. Computes multiple centrality metrics (PageRank, betweenness, degree, eigenvector) 3. Applies weighted Reciprocal Rank Fusion (RRF) to aggregate rankings 4. Enables iterative, agent-guided exploration and explanation of results Evaluation was performed on HER2-positive breast cancer and PI3K/AKT signaling networks using known driver genes as reference signals. Results: Across evaluated cancer subgraphs (mean ∼10K–14K nodes), the proposed menthod showed: - Improved ranking stability: RRF reduced variance across centrality methods by ∼35–50% (measured via rank correlation dispersion) - High recall of known drivers: Top-10 ranked candidates included established oncogenes with >80% consistency across runs - Enhanced signal-to-noise ratio: Compared to single-metric rankings, RRF improved enrichment of known cancer genes in top-20 ranks by ∼1.6–2.1× - Novel candidate identification: The framework surfaced secondary genes with high consensus centrality that are underexplored in current literature - Interactive hypothesis generation: Agentic workflows enabled rapid iteration (<1–2 minutes per query cycle) for exploring pathway perturbations and assessing risks of targeting high-centrality hub proteins Importantly, hub genes with high betweenness/PageRank were associated with increased predicted off-target pathway disruption, highlighting the need for careful therapeutic prioritization. Conclusions: Integrating multi-centrality analysis with weighted RRF and agentic AI provides a robust and interpretable framework for cancer driver discovery. This approach mitigates bias from individual graph metrics while enabling interactive, explainable exploration of complex disease networks. Impact: This work introduces a scalable paradigm for AI-assisted target discovery in oncology, bridging knowledge graph analytics and agentic AI. The framework has the potential to accelerate identification of actionable targets and inform downstream drug discovery pipelines. Dony Ang. Agentic AI–Driven Knowledge Graph Centrality with Reciprocal Rank Fusion for Interactive Discovery of Cancer Driver Genes and Pathways [abstract]. In: Proceedings of AACR Drug Discovery and Development (AACR D3) Conference; 2026 Jul 21-24; Boston, MA. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(14_Suppl):Abstract nr A022.