OncOmix AI: A Multimodal, Retrieval-Augmented Bioinformatics Framework for Personalized Cancer Care Support and Translational Research

The rapid growth of multimodal medical data has created substantial opportunities for biomedical and health informatics, particularly in personalized cancer care and translational oncology research, while simultaneously introducing major challenges related to heterogeneous data integration, scalable analysis, and interpretable reasoning across molecular and literature-based evidence. Existing approaches often analyze individual modalities in isolation or provide limited transparency when integrating multiomics data with biomedical literature. This paper presents OncOmix AI, a cloud-native, retrievalaugmented framework designed to address these challenges by enabling scalable, literature-grounded multi-omics interpretation at both patient and cohort levels. Using a BigQuery-native architecture, the system integrates clinical variables, RNA expression, somatic mutations, copy-number variation, protein expression, and DNA methylation from 500 TCGA Breast Cancer (BRCA) patients into unified patientlevel representations. Patient-specific summaries are generated using large language models (LLMs) based on observed molecular measurements and cohort-relative statistics. Relevant PubMed/PMC passages are retrieved and cited to ground the narrative explanation, while similaritybased indicators provide contextual trust signals; the Literature Validation Score (LVS) is applied exclusively for cohort-level event validation. At the cohort level, molecular events are ranked using prevalence, biological impact, and literature alignment quantified through LVS, supporting interpretable biomarker prioritization and hypothesis generation. Results demonstrate that OncOmix AI enables transparent patientlevel contextualization, interpretable cohort-level analysis, and fully scalable execution within BigQuery AI, addressing interpretability challenges commonly associated with blackbox language models. Although demonstrated in breast cancer, the proposed framework generalizes to additional diseases and multimodal biomedical datasets relevant to personalized cancer care and data-driven oncology research.