Autonomous Phenotypic Biomarker Discovery with Self-Guided Prognostic Foundation Model

Phenotypic biomarkers in histopathology hold strong prognostic value but are difficult to discover systematically due to reliance on subjective annotation, poor scalability, and limited reproducibility. We introduce ProgSeer, an end-to-end platform for autonomous phenotypic biomarker discovery. ProgSeer integrates a self-guided foundation model (ProgFM) that analyzes whole-slide images at both tissue and cellular scales, a frequency-differential filtering strategy to prioritize outcome-associated morphological features, and a clinician-in-the-loop validation interface. Applied to 8 multicenter cohorts across 7 cancer types, ProgFM outperforms current state-of-the-art models in prognostic prediction by about 5% average margin. Using top three frequency-differential filtering strategy, the framework identifies 21 robust biomarkers, including 4 previously underexplored phenotypic biomarkers. All candidates are validated through multimodal evidence—survival stratification, Immunohistochemical visualization, bulk transcriptomics, and spatial transcriptomic colocalization. ProgSeer provides a scalable, reproducible framework for accelerating biomarker-driven precision oncology.