Integrating multiomics and whole slide imaging for predicting the malignant transformation of precancerous rectal lesions

Background and objectives

Predicting the malignant transformation of rectal precancerous lesions remains challenging because conventional Whole Slide Images (WSIs) capture morphological information but lack molecular insight. Multiomics data provide complementary biological signals that often precede visible morphological changes. This study aimed to develop an artificial intelligence (AI)-based multimodal framework integrating WSI and multiomics data for accurate early prediction of malignant transformation.

Methods

WSI patches (512×512 px at 20× magnification) and matched multiomics profiles were used for 450 rectal tissue samples from the publicly available The Cancer Genome Atlas dataset. A multimodal architecture was designed, employing a Vision Transformer (ViT-B/16) for WSI feature extraction and a Variational Autoencoder for multiomics representation learning. Features were fused via a cross-attention mechanism to capture inter-modality dependencies. Baseline models, including a convolutional neural network-only image model and an omics-only multilayer perceptron, were trained for comparison. Five-fold cross-validation was applied, with binary cross-entropy loss, the AdamW optimizer, early stopping, and hyperparameter tuning to ensure reproducibility.

Results

The multimodal Vision Transformer–Variational Autoencoder fusion model outperformed unimodal baselines, achieving an accuracy of 0.892 ± 0.012 and an area under the receiver operating characteristic curve of 0.927 ± 0.009, corresponding to a 7–10% improvement over WSI-only and omics-only models. Cross-attention–based fusion improved prediction stability and classification performance, while interpretability analyses (Grad-CAM and SHAP) highlighted biologically meaningful histopathological regions and molecular feature contributions.

Conclusions

This study presents a robust and scalable AI-based framework for integrating WSI and multiomics data in rectal precancerous lesions. The model improves predictive precision compared with unimodal baselines and offers preliminary interpretability insights through attention mechanisms. These findings support the potential of multimodal AI for early cancer risk assessment and precision pathology.

Full text

https://www.xiahepublishing.com/2835-3315/CSP-2025-00026

The study was recently published in the Cancer Screening and Prevention.

Cancer Screening and Prevention (CSP) publishes high-quality research and review articles related to cancer screening and prevention. It aims to provide a platform for studies that develop innovative and creative strategies and precise models for screening, early detection, and prevention of various cancers. Studies on the integration of precision cancer prevention multiomics where cancer screening, early detection and prevention regimens can precisely reflect the risk of cancer from dissected genomic and environmental parameters are particularly welcome.

Follow us on X: @xiahepublishing

Follow us on LinkedIn:  Xia & He Publishing Inc.