CTMAP: an adversarial cross-modal learning framework for accurate and robust cell-type annotation in single-cell resolution spatial transcriptomics.

Recent advances in single-cell-resolution spatial transcriptomics (scST) have enabled the measurement of gene expression profiles for individual cells while preserving the spatial organization of the tissue microenvironment. However, accurate cell-type annotation remains challenging due to sparse gene coverage, platform-specific technical biases, and the difficulty of identifying rare cell populations. Here, we propose CTMAP, a deep learning-based cross-modal integration framework for robust cell-type annotation of scST cells. CTMAP employs an adversarial learning strategy to align features between scRNA-seq reference data and scST data, and performs cell-type annotation of scST cells in a shared latent space using cell-type centroids derived from the reference data. We systematically evaluated CTMAP on six real scST datasets together with their corresponding scRNA-seq reference datasets. Compared with a wide range of state-of-the-art methods, CTMAP demonstrates marked advantages in overall annotation accuracy, robustness to cell-type composition mismatch, cross-platform generalization, and sensitivity to rare cell populations. Furthermore, across a series of robustness tests involving noise injection and cell-type imbalance, CTMAP consistently exhibits strong stability and resistance to perturbations, highlighting its potential as a general and reliable solution for cell-type annotation in scST.