Image2Gene: A Minimalist and Weakly-Supervised Framework for Morphology-Aligned Gene Expression Prediction From...

Area of Science:

• Computational Biology
• Genomics
• Histopathology

Background:

• Spatial transcriptome analysis is crucial for understanding tissue architecture and cellular function.
• Current methods often require expensive sequencing or complex computational modules.
• Predicting gene expression from histological images offers a cost-effective alternative.

Purpose of the Study:

• To develop a simple, effective framework for gene expression prediction directly from histological images.
• To enable scalable spatial transcriptome analysis without the need for expensive sequencing.
• To create a robust method for learning biologically meaningful similarities between image and gene expression data.

Main Methods:

• Image2Gene framework utilizing a weakly supervised contrastive learning approach.
• Employs an image encoder and fully connected layers for prediction.
• Encodes spatial coordinates via learnable embeddings.
• Novel contrastive loss function aligning image and gene expression similarity.
• Gene expression inference using k-Nearest Neighbor interpolation.

Main Results:

• Image2Gene achieves highly competitive performance on HER+ and cSCC spatial transcriptome datasets.
• Demonstrates the ability to predict gene expression profiles directly from tissue morphology.
• The framework successfully learns structured image representations reflecting gene expression variations.
• Achieves robust and biologically meaningful similarity learning by aligning samples in image space.

Conclusions:

• Image2Gene presents a scalable, annotation-free alternative for inferring transcriptome patterns from histological sections.
• The proposed method offers a promising approach for spatial transcriptome analysis.
• Highlights the potential of weakly supervised contrastive learning in integrating imaging and genomic data.