Abstract P24: Multi-Scale Modeling for Prediction of Spatial Single-Cell and Bulk Transcriptomic Profiles from Whole-Slide Images of Colorectal Cancer

Recent advances in self-supervised foundation models for computational cancer histopathology have shown great promise for predicting clinical phenotypes from routine hematoxylin and eosin (H&E)-stained images. However, application of these models in the context of discovery, for example for identification of novel biomarkers or drug targets, remains challenging as this typically requires linking whole slide imaging (WSI) data to molecular data, which has historically been scarce. Recently, spatial omics (SO) has emerged as a powerful technology for high-throughput profiling of RNAs and proteins directly within cancer tissues. This presents new and exciting opportunities to improve histopathology models for making molecular predictions about the tumor microenvironment. In this work, we develop machine learning models that can generate transcriptomic profiles at two different spatial scales: (1) cellular-level resolution, i.e. SO, and (2) slide-level resolution, i.e. bulk RNA-sequencing. We present a preprocessing and benchmarking framework to enable rigorous comparison of models on an in-house SO dataset, as well as a public bulk RNA-seq dataset (TCGA-COAD/READ), of colorectal cancer. We fine-tune recently published digital pathology foundation models for these tasks, and find that they outperform both general pathology-naïve image foundation models, as well as specialized digital pathology deep learning models. We also find that existing models cannot be simultaneously trained using supervision from multiple spatial resolutions. To address this, we propose a graph neural network (GNN) modeling approach that can model both local interactions between cells within tumor niches, as well as global interactions across tumor regions. We show that our GNN model outperforms existing models for both cell-level and slide-level tasks. This work presents a novel multi-scale framework for predicting transcriptome profiles from WSIs. Citation Format: Tianyi Zhang, Malay Singh, Dwivedi Naman, Timothy Antoni, Ju Can, Firas Elbayoumi, Iain Bee Huat Tan, Tiantian He, Hwee Kuan Lee, Shyam Prabhakar, Grace Hui Ting Yeo. Multi-Scale Modeling for Prediction of Spatial Single-Cell and Bulk Transcriptomic Profiles from Whole-Slide Images of Colorectal Cancer [abstract]. In: Proceedings of Frontiers in Cancer Science 2024; 2024 Nov 13-15; Singapore. Philadelphia (PA): AACR; Cancer Res 2025;85(15_Suppl):Abstract nr P24.