Decoding the ERS-CAF immunoregulatory axis via multimodal AI and its pan-cancer prognostic and therapeutic predictive value.
Authors
Affiliations (10)
Affiliations (10)
- Department of Spine Surgery, The First Affiliated Hospital, Hengyang Medical School, University of South China, Hengyang, Hunan, China.
- Department of Orthopedics, Wuhan Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, Hubei, China.
- Musculoskeletal Tumor Center, Peking University People's Hospital, Peking University, Beijing, Beijing, China.
- Department of Spine Surgery, The Second Xiangya Hospital, Central South University, Changsha, Hunan, China.
- The First Affiliated Hospital, Health Management Center, Hengyang Medical School, Hengyang, Hunan, China.
- Department of Orthopedics Surgery, General Hospital of the Central Theater Command, Wuhan, Hubei, China.
- Department of Ophthalmology, The Second Affiliated Hospital of Zhengzhou University, Zhengzhou, Henan, China.
- Department of Pharmacy, The First Affiliated Hospital, Hengyang Medical School, University of South China, Hengyang, Hunan, China. [email protected].
- Department of Radiology, The First Affiliated Hospital, Hengyang Medical School, University of South China, Hengyang, Hunan, China. [email protected].
- Department of Spine Surgery, The First Affiliated Hospital, Hengyang Medical School, University of South China, Hengyang, Hunan, China. [email protected].
Abstract
Endoplasmic reticulum stress-related cancer-associated fibroblasts (ERS-CAF) remodel the tumor microenvironment and drive immune exclusion and therapy resistance in chordoma, yet routine and non-invasive readouts of this biology are lacking. We hypothesized that standard pre-operative MRI and H&E whole-slide images (WSI) encode image-based surrogates of ERS-CAF-driven immunoregulation that can be learned and generalized across cancers. Three bulk-transcriptomic reference scores were defined for surrogate supervision, capturing ERS-program activity, ERS-CAF-immuneligand-receptor crosstalk and microenvironmental heterogeneity. In 126 chordoma cases, a stage-wise multimodal framework integrating calibrated WSI attention, gated radiopathomic fusion and domain alignment showed strong concordance with molecular profiles, independent prognostic value and biologically specific localization to fibrotic immune-excluded regions. These associations were generalized in zero-shot analyses to the TCGA pan-cancer. An MRI-only distilled model preserved most predictive performance with substantial gains in efficiency, supporting scalable non-invasive clinical application.