HONeYBEE: Enabling Scalable Multimodal AI in Oncology Through Foundation Model-Driven Embeddings

The increasing availability of multimodal biomedical data has created new opportunities for AI-driven oncology. However, integrating these heterogeneous data sources into unified, high-quality representations remains a critical challenge. We introduce Harmonized ONcologY Biomedical Embedding Encoder (HONeYBEE), a modular, open-source framework that preprocesses diverse oncology datasets--including clinical notes, pathology slides, radiology scans, and molecular profiles--to generate consistent, high-dimensional data representations (also known as embeddings) using domain-specific foundation models. HONeYBEE simplifies downstream applications such as prognosis estimation, cancer subtype classification, and patient retrieval through a unified API. Using multimodal data from over 11,400 patients in The Cancer Genome Atlas (TCGA), we show that HONeYBEEs integrated representations improve cancer-type separability in t-SNE space and enhance survival prediction. Multimodal embeddings achieved superior performance across all survival models (C-index: CoxPH = 0.6587, RSF = 0.6252, DeepSurv = 0.6621), outperforming single-modality baselines. Fine-tuned clinical embeddings further improved classification accuracy by 14.8% compared to pretrained models. By releasing all representations, models, and pipelines, HONeYBEE enables reproducible, scalable, and clinically meaningful AI workflows for precision oncology.