Benchmarking cell-type-specific spatially variable gene detection methods

Cell-type-specific spatially variable genes (ctSVGs) integrate cell-type composition with spatially varying gene expression, revealing novel insights into molecular mechanisms through spatial transcriptomics. However, the emerging methods devoted to ctSVG detection have not been systematically evaluated, limiting their broader application and future algorithmic improvements. Here, we present a comprehensive evaluation of six state-of-the-art ctSVG detection methods, in terms of consistency, predictive performance, rotational robustness, scalability, and biological interpretability. We benchmark these methods on 46 real datasets, and a wide range of simulated datasets encompassing diverse biological and technical scenarios. Several key observations are yielded: (i) current algorithms complement each other in predictive performance and computational efficiency; (ii) STANCE and Celina achieve a better predictive performance on multiple spatial patterns, while C-SIDE, spVC, ctSVG, and CTSV exhibit tighter control of false positives; (iii) STANCE, ctSVG, CTSV, and Celina have better overall performance on single-cell resolution data; (iv) rotation invariance still warrants further investigation; (v) Celina appears to have a relative superiority in most metrics, though it tends to generate spurious signals affected by nontargeted cell types; and (vi) algorithmic choice strongly influences downstream biological interpretation. Together, our study offers valuable guidance for tool selection and algorithmic development.