S387 Systematic Review of Deep Learning Methods for Molecular Pathway Prediction in Colorectal Cancer Histology

Introduction: Histopathological evaluation remains central to colorectal cancer (CRC) diagnosis and staging, yet conventional methods lack the ability to infer molecular pathway activation—critical for prognosis and therapy selection. Recent advances in deep learning (DL) have enabled histology-based prediction of key molecular subtypes and signaling pathways, promising to bridge the gap between digital pathology and precision oncology. To systematically review the application of DL algorithms in predicting molecular pathways directly from hematoxylin and eosin (H&E) stained slides of CRC, and to evaluate their diagnostic performance and translational potential. Methods: A comprehensive literature search was conducted across PubMed, Scopus, and IEEE Xplore (2015–2024) using PRISMA guidelines. Studies were included if they employed convolutional neural networks (CNNs) or other DL methods to predict molecular features (e.g., MSI status, CMS classification, Wnt/PI3K/TGF-β signaling) from H&E histology in CRC. Using R (metafor) and Python (scikit-learn, PyTorch), we extracted and meta-analyzed pooled area under the receiver operating characteristic curve (AUROC), accuracy, and sensitivity across studies. Study quality was assessed using PROBAST and TRIPOD-AI criteria. Results: Out of 364 screened articles, 18 met inclusion criteria, encompassing over 9,200 patient samples. DL models achieved high accuracy in predicting MSI (mean AUROC = 0.87), consensus molecular subtypes (AUROC = 0.81), and pathway activation signatures (Wnt, TGF-β, PI3K/AKT; AUROC range = 0.75–0.89). Models leveraging attention mechanisms and multi-resolution tiles outperformed conventional CNNs. External validation was conducted in only 39% of studies. Several studies demonstrated prognostic utility, linking DL-predicted pathway activity with progression-free survival (hazard ratio = 0.53; P < 0.01). However, lack of model interpretability and standardization across cohorts were key limitations. Conclusion: DL models show strong promise in predicting molecular pathways from standard histology, supporting their role as scalable tools for personalized CRC care. This systematic review highlights their diagnostic accuracy and potential for clinical integration, while underscoring the need for validation on multi-institutional datasets and explainable artificial intelligence (AI) frameworks. As AI-driven pathology matures, these tools may enhance treatment stratification without additional testing burden.