Yan, X.; Zheng, R.; Chen, J.; Li, M.; Lan, W.

By resolving cell-type compositions from multi-cellular spatial measurements, deconvolution is central to resolving the cellular landscape of complex tissues. Existing deconvolution methods fit continuous expression values and are therefore sensitive to batch effects between single-cell references and spatial data, requiring retraining for each new context. Here we present SpaRank, a context-aware framework that performs spatial deconvolution by representing spots as ranked feature sequences. Adapting the rank-based encodings of single-cell foundation models, this formulation is inherently robust to technical variation, enabling a pretrain-transfer paradigm. On simulated benchmarks, SpaRank achieves strong deconvolution accuracy, robustness to expression perturbations, and substantial computational efficiency. On experimental datasets, pretrained models generalize across diverse biological contexts: a model pretrained on a multi-organ lymphoid atlas accurately resolved cell-type distributions across distinct tissues and sequencing platforms; likewise, a model pretrained on an integrated breast atlas delineated cell-type compositions across normal and malignant disease states. Furthermore, the framework naturally extends to multimodal spatial deconvolution by employing gated fusion to adaptively integrate diverse omics signals, improving accuracy over single-modality approaches. Overall, SpaRank establishes a transferable deconvolution paradigm, enabling unified cellular atlases to support direct, context-aware inference across diverse biological states and profiling modalities.