Beyond Gene Accuracy? Auditing Whether Ligand-Receptor Evaluation Adds Information in Virtual Spatial Transcriptomics

Virtual spatial transcriptomics (vST) is increasingly used for biological analyses beyond gene reconstruction. Downstream scores are expected to depend on gene accuracy, since they are computed from predicted expression. The key question is whether added relational and spatial structure provides enough independent information to warrant a new evaluation axis. We study this question for ligand--receptor (LR) communication, applying a fixed, training-free LR operator identically to measured and H\&E-predicted expression on 61 HEST-Bench Visium slides across six tissues and six predictors. LR concordance is strongly above the spatial-shuffle null for every model (mean $z=21.6$), demonstrating recoverability. However, LR concordance is tightly coupled to per-gene accuracy: across 366 slide--model observations, PCC and LR concordance have Spearman $\rho=0.941$, while PCC alone accounts for $R^2=0.890$ of LR concordance. This coupling persists across neighborhood scales, rank scoring, individual predictors, and 202,230 LR-pair observations. When magnitude is discarded and only spatial overlap of high-interaction hotspots is scored, coupling falls to $\rho=0.746$ and $R^2=0.471$ while remaining above chance. Thus, magnitude-based LR concordance preserves spatial structure but provides little additional information for distinguishing model accuracy beyond gene accuracy; spatial localization is a less redundant signal.