BiNVaST: niche boundary-preserving and synthetic data-guided variational autoencoders for concordant inference across high-resolution spatial transcriptomic platforms
High-resolution spatial transcriptomics (ST) enables subcellular spatially resolved transcript profiling but remains difficult to annotate in complex tissues because of gene coverage and super sparseness in expression. These challenges are amplified in cancer research, where neoplastic cell states and rare immune cells in tumor microenvironment are often poorly captured by generic references and/or cell typing tools exclusively built on molecular profile. Here, we present BiNVaST, a niche boundary-preserving and dual-branch variational autoencoder fused with statistical simulator for reliable cell typing across multiple spatial platforms. It generates platform-adjusted in silico training data by integrating query spatial profile and user-defined pre-annotated single-cell reference through Singular Value Decomposition and Optimal Transport, combining molecular similarity and spatial proximity to predict segmented cell labels. BiNVaST was benchmarked against existing methods across Xenium, CosMx, Stereo-seq, Visium HD platforms using six specimens from four tissue types. This tool is sensitive in detecting divergent tumor cell states and improves the cross-platform concordance, distinguishing disease-specific neoplastic and non-neoplastic cellular subpopulations and supporting robust cellular or niche inferences in heterogeneous tissues.