Multi-Scale Mapping of Gene Expression from Whole-slide Images for Identifying Phenotype-Associated Subpopulations

Area of Science:

• Computational biology
• Genomics
• Pathology

Background:

• Discovering phenotype-associated subpopulations is crucial for targeted therapies and prognostic biomarker development.
• Current methods for inferring genetic alterations from whole-slide images (WSIs) using deep learning often operate at a single scale.
• Multi-scale gene expression analysis is essential for comprehensive understanding.

Purpose of the Study:

• To introduce BiSCALE, a deep learning framework for predicting gene expression from WSIs at both tissue (bulk) and near-cellular (spot) levels.
• To link these multi-scale gene expression predictions to clinical phenotypes.
• To establish a cost-effective method for multi-scale gene analysis and phenotype-associated feature discovery.

Main Methods:

• Developed BiSCALE, a deep learning framework integrating a WSI foundation encoder and a Vision-Mamba fusion module.
• Employed a two-stage training strategy to address scale and distribution differences between bulk and spot data.
• Trained the model on 2109 bulk tumor samples and 141,000 spatial transcriptomics spots across three cancer types.

Main Results:

• BiSCALE outperformed established bulk and spatial baselines and generalized well to independent cohorts.
• Demonstrated strong concordance between predicted bulk and spot gene expression profiles.
• Successfully recovered biologically relevant pathway activity and supported downstream applications like risk stratification and cell-identity annotation.

Conclusions:

• BiSCALE provides a cost-effective approach for multi-scale gene analysis from routine pathology images.
• The framework facilitates the discovery of phenotype-associated subpopulations, including those linked to recurrence and hypoxia.
• BiSCALE enables advanced applications such as patient-level risk stratification and spot-level cell annotation.