Ambiguity-Aware Multi-Stage Cell-Type Annotation for Spatial Transcriptomics.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics offers insights into tissue cellular organization.
• Accurate cell-type annotation is hindered by expression heterogeneity, mixed cell populations, and transitional cell states.
• Current methods often oversimplify by assigning single labels, masking biological ambiguity and leading to overconfident cell-type assignments.

Purpose of the Study:

• To develop an ambiguity-aware, multi-stage framework for robust spatial cell-type annotation.
• To improve the reliability and biological coherence of cell-type assignments in spatial transcriptomics data.
• To address the limitations of existing methods in handling ambiguous cell populations.

Main Methods:

• A hybrid spatial-feature clustering approach combined with constrained language-model inference.
• Assigning confidence scores based on marker gene expression, candidate cell-type separation, and entropy.
• Implementing a refinement strategy for low-confidence clusters through local reclustering and preserving unresolved clusters as mixed populations.

Main Results:

• Reduced cluster-level ambiguity from 16.1% to 2.27% and cell-level ambiguity from 18.4% to 0.86% in cholangiocarcinoma data.
• Improved confidence calibration of cell-type assignments.
• Demonstrated the effectiveness of topological integration and language-model inference for scalable and biologically coherent annotations.

Conclusions:

• Explicitly handling ambiguity is crucial for reliable spatial cell-type annotation, especially in heterogeneous tumors.
• The proposed framework enhances the accuracy and interpretability of spatial transcriptomics data.
• This approach provides a more nuanced understanding of cellular composition in complex biological systems.