Deep learning integrates histopathology and proteogenomics at a pan-cancer level
Problem
Predicting clinical outcomes and molecular features (e.g., TP53 status) from routine H&E pathology slides alone, and explaining which histology features are predictive, was difficult.
Why it matters
Cheaper, slide-based prediction could improve cancer diagnosis and treatment planning without costly molecular tests.
What the AI did
Deep learning models trained on 2,755 slides from 657 patients across 6 cancer types classified tumor vs. normal, tissue-of-origin, and predicted TP53 and stage; multi-omics integration identified pathways behind predictive morphologies.
Results
AUROC 0.995 (tumor vs. normal), AUROC 0.979 (tissue-of-origin); generalizability confirmed using TCGA.
Impact
Research-level validation on an external cohort (TCGA); clinical application only suggested, not demonstrated.
Record
| Venue | Cell Reports Medicine |
|---|---|
| Published | 2023 |
| Work type | article |
| Authors | Joshua M. Wang, Runyu Hong, Elizabeth G. Demicco, Jimin Tan, Rossana Lazcano, Andre Luis Moreira, Yize Li, Anna Pamela Calinawan et al. |
| Citations | 41 (OpenAlex, at import time) |
| References | 46 |
| Model or method | not stated in source |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1016/j.xcrm.2023.101173 |
| Open access | gold |
| Keywords | deep learning, histopathology, proteogenomics, pan-cancer, AUROC, TCGA |
Importance rating
significant — thousands to millions affected
Genuinely new capability: predicting TP53 status and stage from routine H&E slides, potentially avoiding costly molecular tests in cancer care; only research-level validation on CPTAC/TCGA, clinical use merely suggested.