SNR-ST-Mix: sample-specific neighborhood regression mixup for augmented spatial transcriptomics imputation with deep...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics (ST) provides gene expression data within tissue context.
• ST data often suffers from noise, low resolution, and sparse sampling, limiting fine spatial structure recovery.
• Current deep learning augmentation methods for ST are often designed for classification, neglecting spatial and transcriptomic relationships, leading to biologically implausible results.

Purpose of the Study:

• To develop a novel data augmentation framework, SNR-ST-Mix, specifically for spatial transcriptomics regression tasks.
• To address limitations of existing augmentation strategies, including noise, low resolution, and biologically implausible interpolations.
• To improve the performance and stability of gene expression imputation in spatial transcriptomics.

Main Methods:

• Proposed SNR-ST-Mix, a geometry- and expression-aware data augmentation framework for ST data.
• Constrained data mixing to k-nearest spatial neighbors and adaptively weighted interpolation coefficients based on expression similarity.
• Generated augmented samples preserving local biological structure and ensuring spatial smoothness.

Main Results:

• SNR-ST-Mix consistently outperformed conventional augmentation methods across various tissue types.
• The framework improved prediction stability and expanded the effective training manifold.
• No architectural changes or additional computation were required for implementation.

Conclusions:

• SNR-ST-Mix offers an effective and biologically principled augmentation strategy for ST regression.
• The method leverages spatial geometry and transcriptomic similarity to enhance predictive performance.
• Improved generalization and prediction stability were achieved without increasing model complexity.