DyGFormer:Transformer With Resistance-Distance Bias for Semi-Supervised Cell-Type Identification

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell spatial transcriptomics is vital for cancer research, enabling analysis of cell types within tissue context.
• Accurate cell identification is challenged by technical noise, data variability, and annotation limitations.

Purpose of the Study:

• To develop DyGFormer, a semi-supervised model for robust cell-type classification in spatial transcriptomics.
• To enhance cell-type identification by jointly modeling transcriptional data and spatial organization.

Main Methods:

• DyGFormer utilizes a geometric spatial prior and self-attention to learn dynamic cell-cell connectivities.
• Incorporates resistance-distance positional bias and a Neighbor Interaction History encoder for stable cross-layer analysis.
• Employs triplet-guided metric supervision with a small labeled subset to improve class separation.

Main Results:

• DyGFormer demonstrates consistent improvements in Accuracy and Macro-F1 scores across diverse spatial datasets (NanoString/CosMx, MERFISH).
• The model shows superior performance compared to baseline methods under various label-sampling scenarios, including spatially disjoint settings.
• Validation across multiple tissues and platforms confirms the model's generalizability and robustness.

Conclusions:

• DyGFormer offers a powerful semi-supervised approach for accurate cell-type classification in spatial transcriptomics.
• The model effectively addresses key challenges, paving the way for more reliable insights in cancer studies.
• DyGFormer's ability to integrate spatial and expression data advances the field of computational spatial biology.