Toward next-generation machine learning and deep learning for spatial omics.

Area of Science:

• Computational biology
• Bioinformatics
• Data science

Background:

• Spatial omics technologies generate complex, high-dimensional data requiring advanced computational models.
• Existing machine learning (ML) and deep learning (DL) methods lack clear guidance for specific spatial omics challenges.
• There is a need for methodological clarity in applying ML/DL to spatial omics data.

Purpose of the Study:

• To critically synthesize ML/DL approaches for core spatial omics tasks.
• To provide methodological guidance for selecting appropriate ML/DL models.
• To propose a decision framework for the deployment of ML/DL in spatial omics.

Main Methods:

• Comparative synthesis of classical ML and modern DL architectures (CNNs, GNNs, transformers, generative models).
• Review of emerging strategies like optimal transport and foundation models.
• Discussion of practical solutions including self-supervised pretraining and federated learning.

Main Results:

• Classical ML offers interpretable baselines but struggles with spatial dependencies.
• DL models capture complex patterns and multi-omics integration but face data scarcity and computational challenges.
• Emerging strategies improve cross-modality alignment but need rigorous evaluation.

Conclusions:

• A decision framework is proposed to guide ML/DL model selection based on data characteristics and application.
• Recommendations are provided for enhancing scalability, reproducibility, and clinical translation.
• The review aims to facilitate reproducible, interpretable, and clinically translatable ML/DL deployment in spatial omics.