Deep Learning on Histologic Slides Accurately Predicts Consensus Molecular Subtypes and Spatial Heterogeneity in...

Area of Science:

• Oncology
• Computational Pathology
• Bioinformatics

Background:

• Colon cancer (CC) is a heterogeneous disease with molecular subtypes (CMS) impacting treatment efficacy.
• Current adjuvant CC treatment relies on T and N staging, potentially overlooking molecular information.
• Intra-tumor heterogeneity (ITH) within CMS subtypes is linked to poorer prognosis.

Purpose of the Study:

• To develop and validate a deep learning model for predicting CMS and ITH in CC using whole slide images (WSIs).
• To assess the model's performance in internal and external validation cohorts.
• To explore the model's ability to map CMS spatial distribution and associate histological features with CMS subtypes.

Main Methods:

• Utilized WSIs from 1,996 CC patients across PETACC-8, TCGA-COAD, and PRODIGE-13 cohorts.
• Developed a deep learning framework involving self-supervised patch embedding and weakly-supervised CMS prediction.
• Employed CMSclassifier for ground-truth CMS scoring and performed interpretability analyses.

Main Results:

• High macro-average AUC achieved: 93.0% (internal CV) and 94.4% (external validation) for homogeneous tumors (PETACC-8 model).
• The model accurately predicted CMS and characterized ITH, demonstrating robustness across cohorts.
• Spatial CMS distribution and associations between histology and CMS subtypes were successfully visualized.

Conclusions:

• Deep learning on routine histology WSIs offers an efficient method for predicting CC CMS and ITH.
• This approach can enhance patient stratification for adjuvant therapy, moving beyond traditional staging.
• The findings support the routine clinical integration of CMS and ITH assessment in CC management.