Pancancer outcome prediction via a unified weakly supervised deep learning model

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Translational bioinformatics

Background:

• Accurate cancer prognosis is crucial for treatment guidance and patient outcomes.
• Existing histopathology-based models are often cancer-specific, lack external validation, and require non-routine molecular data.
• A unified, externally validated model integrating routinely available data is needed for pancancer prognosis.

Purpose of the Study:

• To develop and validate PROGPATH, a unified model for pancancer prognosis prediction.
• To integrate histopathological image features with routinely collected clinical variables.
• To overcome limitations of current cancer-specific and molecular data-dependent models.

Main Methods:

• PROGPATH utilizes a weakly supervised deep learning architecture with a foundation model for image encoding.
• Morphological features are aggregated via attention-guided multiple instance learning and fused with clinical data using a cross-attention transformer.
• A router-based classification strategy enhances prediction performance; trained on 7999 whole-slide images (WSIs) across 15 cancer types and validated on 17 external cohorts.

Main Results:

• PROGPATH achieved superior performance compared to state-of-the-art multimodal prognosis prediction models.
• Demonstrated strong generalizability across 12 cancer types and robustness in various patient subgroups (stage, treatment, biomarkers).
• Identified key pathological patterns (cell differentiation, necrosis) contributing to risk predictions, offering model interpretability.

Conclusions:

• PROGPATH effectively integrates histopathological and clinical data for pancancer prognosis prediction.
• The model shows significant potential for supporting personalized cancer management strategies.
• Highlights the value of foundation models and multimodal data integration in computational pathology.