Reproduce “Must: Maximizing Latent Capacity of Spatial Transcriptomics Data” on OpenResearch

About this paper

Zelin Zang, Liangyu Li, Yongjie Xu, Chenrui Duan, Kai Wang, Yang You, Yi Sun, Stan Z. Li

A framework called Multiple-modality Structure Transformation (MuST) integrates multi-modal spatial transcriptomics data into a universal latent space, effectively mitigating modality bias. This approach enhances the precision of identifying complex and fine-grained tissue structures across various platforms and improves downstream analyses like spatial clustering, deconvolution, and trajectory inference.

Key insights

• Modality bias is a fundamental challenge in multi-modal spatial transcriptomics, where inconsistent contributions from different data types can lead to significant information loss if not properly addressed.
• Dynamically estimating and balancing the contributions of individual modalities through an adaptive topology fusion mechanism, rather than fixed weights, is crucial for effectively integrating multi-modal data and preserving both local and global tissue structures.
• Combining data augmentation with graph neural networks within a deep learning framework allows for the creation of a robust, unified latent space that supports a wide array of downstream analytical tasks while mitigating modality bias.

Summary by alphaXiv.