SNR-ST-Mix: Sample-specific Neighborhood Regression Mixup for Augmented Spatial Transcriptomics Imputation with Deep Neural Network

2026-06-07 · Source: Machine Learning · Field: Science & Research — Life Sciences & Biology, Artificial Intelligence & Machine Learning, Research Methodology & Innovation · Depth: Expert, quick

Summary

SNR-ST-Mix is a new data augmentation framework for spatial transcriptomics (ST) gene expression imputation. ST data often presents challenges like noise, low resolution, and sparse sampling. Deep neural networks for imputation are constrained by limited sample sizes. Existing augmentation methods, designed for classification, neglect spatial and transcriptomic relationships. This leads to biologically implausible interpolations. SNR-ST-Mix addresses these issues by constraining data mixing to a spot's k-nearest spatial neighbors. It adaptively weights interpolation coefficients based on expression similarity. This geometry- and expression-aware approach generates synthetic samples. These samples preserve local biological structure and ensure spatial smoothness. This expands the effective training manifold. Experiments across various tissue types show SNR-ST-Mix consistently outperforms conventional augmentation. It achieves this without architectural changes or additional computation, improving prediction stability and generalization.

Key takeaway

For Research Scientists developing deep learning models for spatial transcriptomics imputation, you should consider integrating SNR-ST-Mix. This framework offers a biologically principled data augmentation strategy. It significantly improves predictive performance and generalization, especially with limited sample sizes. By leveraging spatial geometry and expression similarity, SNR-ST-Mix generates more plausible synthetic data. This approach enhances model stability without requiring architectural changes or additional computational overhead, streamlining your development process.

Key insights

SNR-ST-Mix augments spatial transcriptomics data by leveraging spatial geometry and expression similarity for biologically plausible interpolations.

Principles

• Augmentation should be geometry-aware.
• Incorporate expression similarity for weighting.
• Preserve local biological structure.

Method

SNR-ST-Mix constrains data mixing to k-nearest spatial neighbors. It adaptively weights interpolation coefficients based on expression similarity, generating augmented samples that preserve local biological structure and ensure spatial smoothness.

In practice

• Apply SNR-ST-Mix for ST regression tasks.
• Improve deep learning generalization on ST.
• Enhance prediction stability without new architecture.

Topics

• Spatial Transcriptomics
• Gene Expression Imputation
• Data Augmentation
• Deep Learning Models
• k-Nearest Neighbors
• SNR-ST-Mix

Best for: AI Scientist, Research Scientist

Related on AIssential

• Scalable Mean-Field Variational Inference via Preconditioned Primal-Dual Optimization — Primal-dual optimization with preconditioning offers scalable, robust, and faster convergence for large-scale mean-field variational inference.
• Estimating genotype-tissue specific gene expression using hybrid deep learning — A hybrid deep learning model accurately estimates genotype-tissue specific gene expression, outperforming distance-based methods by ~30%.
• AugMask: Training Diffusion Models on Incomplete Tabular Data via Stochastic Augmentation and Masking — AugMask adapts diffusion models for incomplete tabular data by separating conditioning from supervision during training.
• SFT-GRPO Data Overlap as a Post-Training Hyperparameter for Autoformalization — Minimizing SFT-GRPO data overlap significantly improves autoformalization model performance and semantic accuracy.
• Small Data, Big Maps: Training Geospatial ML Models When Samples Are Scarce — Geospatial ML with scarce data demands pragmatic strategies focusing on data quality, appropriate models, and transparent uncertainty.
• On the Limits of Latent Reuse in Diffusion Models — Latent reuse in diffusion models is reliable only when source and target data subspaces are closely aligned.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.