Abstract 2460: Judith: An agentic AI system for biomedical image analysis and scientific discovery in precision oncology

Abstract The rapidly increasing volume of multimodal biomedical data presents unprecedented opportunities in precision oncology but poses significant challenges for researchers analyzing complex modalities such as gigapixel pathology slides. To address these challenges, we developed Judith, an AI-driven research assistant that leverages foundation models, multimodal generative AI, and other advanced computational tools for streamlining complex workflows in biomedical image analysis, predictive modeling, and biomarker discovery. Judith enables researchers to execute computational tasks through natural language commands, automating tasks such as response prediction, multimodal cancer prognosis, and histopathology image retrieval. We demonstrate Judith’s capabilities in several precision oncology use cases. (1) Predictive modeling and biomarker discovery: Using a retrospective dataset of pretreatment biopsies from invasive breast carcinoma patients treated with Trastuzumab (n=85 patients), Judith trained an AI model that predicts pathologic complete response with an AUROC of 0.855. Judith could also interpret the model, identifying key cellular features linked to tumor proliferation and finding histologic biomarkers that explain cohort-level differences between responders and non-responders. (2) Multimodal data integration: Judith orchestrated the development of a survival prediction model integrating histology, radiology, and multi-omics data for progression-free survival prediction in low-grade glioma patients (n=127), achieving a C-Index of 0.684. Using an integrative analysis, Judith further stratified patients into high- and low-risk groups and identified unfavorable prognostic features for cancer progression such as high tumor cellularity and PSPH overexpression. (3) Histology slide retrieval: Judith automated the development of a histology slide search system (n=138), enabling pathologists to retrieve slides with similar cancer diagnoses and morphologies. Judith was then able to validate the performance of histology slide retrieval, with successful retrieval of rare cancer types such as leiomyosarcoma. Overall, Judith represents a major advancement in biomedical research by automating AI workflows and enhancing data-driven discovery. Its ability to analyze diverse data modalities, identify meaningful biomarkers, and generate interpretable insights empowers researchers to maximize the scientific insights in their data. Future developments will focus on expanding multimodal capabilities, refining interpretability tools, and integrating real-time collaboration features. Citation Format: Ming Y. Lu, Faisal Mahmood, Richard J. Chen. Judith: An agentic AI system for biomedical image analysis and scientific discovery in precision oncology. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2460.