Abstract
<title>Abstract</title> <p>Accurate cell-type identification in single-cell and spatial transcriptomics is crucial for understanding tissue heterogeneity and disease mechanisms. In semi-supervised settings, however, performance is frequently limited by label scarcity, distribution mismatch between labeled and unlabeled cells, and noisy connections in pre-defined cell graphs. To overcome these challenges, we propose GraphTransformer‑CoGNN, a two‑stage dynamic graph learning framework. First, a Graph Transformer learns an adaptive graph structure, incorporating global context through resistance‑distance‑based relative positional encoding. Second, Cooperative Graph Neural Network (CoGNN) refines this structure by suppressing spurious edges and enhancing biologically relevant interactions. The model is trained end‑to‑end with a joint objective combining supervised classification loss and triplet metric loss. Experiments on NanoString‑based spatial transcriptomics and the human DLPFC benchmark demonstrate that with only 18\%labeled cells, our method achieves state‑of‑the‑art accuracy and Cohen’s kappa, while significantly improving the discrimination of rare cell types.</p>