SIGEL: a context-aware genomic representation learning framework for spatial genomics analysis
Wenlin Li, Maocheng Zhu, Yucheng Xu, Mengqian Huang, Jin Chen, Hao Wu*, Xiaobo Sun*

• We develop a “context-aware, self-supervised Spatially Informed Gene Emedding Learning (SIGEL)” framework, addressing the gap in methods for generating gene manifolds and utilizing spatial genomic context.
• SIGEL generates gene representations (SGRs) that are context-aware, semantically rich, and robust against technical artifacts across different samples, enhancing the biological relevance and functional understanding of genomic contexts.
• SGRs are applied to important biomedical research tasks such as imputing missing genes in FISH-based ST, detecting genes with specific spatial expression patterns, identifying disease-associated genes and gene-gene interactions, and improving spatial clustering.
• Extensive experiments with diverse ST datasets across various platforms, species, and tissues demonstrate the superior performance of SGRs in facilitating various analytical tasks, affirming their potential in data-driven genomic research.

Dual Advancement of Representation Learning and Clustering for Sparse and Noisy Images
Wenlin Li#, Yucheng Xu#, Xiaoqing Zheng, Suoya Han, Jun Wang, Xiaobo Sun*

• DARLC is the first unified framework for joint representation learning and clustering of sparse and noisy images (SNIs).
• DARLC integrates contrastive learning, masked image modeling, and a graph attention-based augmentation to enhance representation learning.
• DARLC employs a Student’s t mixture model for robust SNI clustering, addressing class collision and improving adaptability.
• DARLC outperforms state-of-the-art methods on 12 SNI datasets in clustering quality and downstream task performance.

MEATRD: Multimodal Anomalous Tissue Region Detection Enhanced with Spatial Transcriptomics
Kaicheng Xu#, Qilong Wu#, Yan Lu, Yinan Zheng, Wenlin Li, Xingjie Tang, Jun Wang, Xiaobo Sun*

• MEATRD is the first method to integrate histology images and ST data for anomalous tissue region (ATR) detection across tissue regions.
• MEATRD leverages a masked graph dual-attention transformer to enable cross-modality information sharing and mitigate over-generalization.
• MEATRD combines reconstruction-based learning with one-class classification to detect ATRs with minimal visual differences from normal tissues.
• MEATRD outperforms state-of-the-art methods across eight ST datasets, demonstrating superior ATR detection performance.