SEVA: Grounding Gene-set Interpretations in Knowledge and Spatial Evidence

Many existing methods can retrieve biological hypotheses from curated databases. In spatial transcriptomics and spatial proteomics, each hypothesis is also tied to a tissue location, and this positional information enlarges the search space for supporting evidence. We introduce SEVA (Spatial Evidence Verification and Auditing) to carry out evidence-based interpretation in spatial omics. SEVA connects database-grounded gene-set interpretation to claim-directed execution of local expression and spatial checks. It links evidence to individual statements and filters unsupported spatial assertions while retaining supported functional interpretations. On a public human lymph-node section with 4,035 spots, we compare SEVA with a database-only GeneAgent workflow adaptation across three predeclared cases with three repetitions, using the same language model and knowledge snapshot. Both methods support all nine functional-statement instances; by incorporating local data, SEVA additionally supports nine sample-expression and six spatial-clustering instances.