Decoding the ERS-CAF immunoregulatory axis via multimodal AI and its pan-cancer prognostic and therapeutic predictive...

Area of Science:

• Oncology
• Radiology
• Computational Pathology

Background:

• Endoplasmic reticulum stress-related cancer-associated fibroblasts (ERS-CAF) influence the tumor microenvironment, promoting immune exclusion and therapy resistance in chordoma.
• Current methods for assessing ERS-CAF biology are invasive, lacking routine non-invasive readouts for clinical application.

Purpose of the Study:

• To investigate if standard pre-operative MRI and H&E whole-slide images (WSI) can serve as non-invasive surrogates for ERS-CAF-driven immunoregulation.
• To develop and generalize image-based models for detecting ERS-CAF biology across different cancer types.

Main Methods:

• Defined three bulk-transcriptomic reference scores for surrogate supervision: ERS-program activity, ERS-CAF-immuneligand-receptor crosstalk, and microenvironmental heterogeneity.
• Developed a stage-wise multimodal framework integrating WSI attention, radiopathomic fusion, and domain alignment in 126 chordoma cases.
• Performed zero-shot generalization analysis on TCGA pan-cancer data and created an MRI-only distilled model.

Main Results:

• The multimodal framework demonstrated strong concordance with molecular profiles and independent prognostic value in chordoma.
• Image-based surrogates localized to fibrotic, immune-excluded regions, confirming biological specificity.
• The MRI-only model retained significant predictive performance while enhancing computational efficiency.

Conclusions:

• Standard MRI and WSI can provide non-invasive readouts of ERS-CAF biology and its impact on the tumor microenvironment.
• The developed image-based models show promise for generalization across cancers, supporting scalable clinical application.
• This approach offers a pathway for non-invasive assessment of tumor immunoregulation and resistance mechanisms.