This paper introduces THREADS, a new molecular-driven foundation model for oncologic pathology designed to address the challenge of data scarcity in specialized oncology tasks. THREADS is a general-purpose encoder model that generates Whole-Slide Image (WSI) embeddings, pre-trained using multimodal contrastive learning guided by corresponding molecular profiles, which the authors posit offers an unbiased view of tissue morphology. The model was trained on the extensive MBTG-47K dataset, comprising over 47,000 paired WSI and molecular profiles from diverse sources like TCGA and various hospitals. Evaluation across a benchmark of 54 tasks demonstrates that THREADS frequently achieves superior predictive performance compared to existing baselines like PRISM and GIGAPATH for tasks including grading, mutation prediction, and patient survival analysis, highlighting its robustness and generalizability. The authors suggest that THREADS's innovative pretraining strategy and parameter efficiency make it a significant advancement for computational pathology.

References:

• Vaidya, A., et al. "Molecular-Driven Foundation Model for Oncologic Pathology. arXiv 2025." arXiv preprint arXiv:2501.16652.