[edit]
Mapping Three-Dimensional Tumor Heterogeneity through Deep Learning Inference of Spatial Transcriptomics from Routine Histopathology: A Proof-of-Concept Comparative Study
Proceedings of the 4th Machine Learning for Health Symposium, PMLR 259:73-85, 2025.
Abstract
Spatial transcriptomics (ST) technologies enable the mapping of gene and protein abundance within specific tissue architectures, representing a significant advancement over conventional bulk analyses that can obscure critical prognostic markers tied to spatial contexts. Expanding these analyses to three dimensions (3D) can further uncover intricate biomolecular phenomena that may be truncated or missed in two-dimensional (2D) studies. However, the widespread application of 3D ST profiling is limited by high costs and logistical challenges. Deep learning-based inference of ST data from routine histopathological staining offers a cost-effective alternative, allowing for the exploration of histologically associated biological pathways in 3D and enhancing our ability to detect structures linked to tumor progression. In this proof-of-concept study, we employed deep learning models to infer ST data from routine histopathology for 10 colorectal cancer patients, with 10 serial sections analyzed per patient. Our downstream analyses revealed several key instances where 3D approaches provided enhanced insights into cellular phenomena compared to traditional 2D methods. These findings lay the groundwork for future research aimed at leveraging these methods to investigate subtle 3D biomarkers associated with tumor metastasis and recurrence.