Resolving the ambiguity between genomic rearrangements and gene fusions: An AI-augmented structural framework to define therapeutic eligibility in kinase-driven cancers.

1626 Background: A critical disconnect exists in precision oncology: clinical guidelines frequently conflate genomic "rearrangements" with functional gene "fusions", utilizing inconsistent terminology that obscures biological reality. While confirmatory RNA sequencing is recommended, it is often clinically infeasible due to tissue exhaustion or poor sample quality. Consequently, clinicians are forced to prescribe targeted therapies based on DNA-level proxies without knowing if a druggable protein actually exists. We addressed this unmet need with VeraFusionDx, an AI-augmented decision support system that moves beyond simple classification to determine therapeutic eligibility. Methods: We developed a universal framework integrating AI-driven report parsing with generative molecular modeling, trained on a massive real-world dataset (>10,000 gene fusions and >50,000 rearrangements) from a CAP/CLIA-certified lab. An AI normalization module standardizes heterogeneous inputs (unstructured NGS reports, gene+exon number pairs, or coordinates) into a unified format. Distinct from static database lookups, the core engine performs de novo characterization of each patient variant. It computationally reconstructs chimeric sequences and models 3D protein architecture to verify critical druggability criteria, including reading frame alignment and kinase domain integrity. Finally, an AI-literature agent cross-references validated targets with clinical evidence to define definitive therapeutic eligibility (https://verafusiondx.origimed.com). Results: Analysis using this generative engine revealed a heterogeneous landscape where genomic rearrangement proved a poor proxy for druggability, highlighting the necessity of structural validation. For NTRK genes, approximately two-thirds of detected variants were rearrangements rather than actionable fusions, with striking discordance by gene and tumor type. Notably, NTRK2 rearrangements were 4-fold more common than fusions, and FGFR1 were 90% rearrangements. Even among canonical partners ( EML4-ALK, KIF5B-RET, CD74-ROS1 ), the analysis identified a persistent 1-2% rate of non-functional mimics. To resolve this complexity, the system generated definitive outputs for each case, including de novo chimeric DNA and protein sequences, predicted 3D structural models, therapeutic actionability assessments, and an AI-curated summary of comparable literature evidence. Conclusions: Functioning as a comprehensive computational firewall, this framework integrates de novo sequence generation, structural modeling, and therapeutic intelligence to definitively filter out inert genomic noise, ensuring that life-altering TKI prescriptions are grounded in verified functional reality rather than ambiguous terminology.