Semantic-Guided Spatial Representation Learning for Spatial Domain Identification

Abstract Motivation Spatial domain identification is a fundamental task in spatial transcriptomics analysis, aiming to partition tissue sections into coherent regions that reflect underlying biological organization. Most existing methods rely on gene expression similarity and spatial proximity, which can be insufficient when expression measurements are sparse, noisy, or weakly discriminative. In such settings, representation learning faces ambiguity in determining how spatial neighborhood information and expression-based similarity should be jointly utilized. Results We present GreS, a spatial domain identification framework that incorporates gene-level semantic priors into spatial representation learning. GreS models spatial adjacency and expression-driven similarity using two complementary neighborhood graphs and leverages aggregated gene semantic information to guide how these views are weighted at the spot level. Rather than redefining neighborhood relationships, semantic priors act as contextual signals that modulate the relative contribution of spatial and expression-based cues during domain inference. We evaluate GreS on diverse spatial transcriptomics datasets, including layered brain tissue, embryonic development, and heterogeneous tumor microenvironments. Across these settings, GreS consistently identifies spatial domains that are structurally coherent and biologically interpretable, outperforming existing methods in quantitative accuracy and qualitative spatial organization. Our ablation analyses further demonstrate that performance gains arise from biologically meaningful and properly aligned semantic information, rather than from increased model complexity or auxiliary features. Availability and Implementation https://github.com/ai4nucleome/GreS Contact yanlinzhang@hkust-gz.edu.cn Supplementary Information A Supplementary file is submitted together with this manuscript.