Determining a gene's functional significance within a specific cellular context has long been a challenge. We introduce a framework for quantifying gene importance by leveraging attributions learned by foundation models (FMs) trained on large corpora of single-cell RNA-sequencing (scRNA-seq) datasets. Attribution scores robustly quantify gene importance across datasets, emphasizing key genes in relevant cell populations, while minimizing technical artifacts. Therefore, we developed SIGnature, a tool that enables rapid search of gene signatures across multiple scRNA-seq atlases. We demonstrated its utility by querying the MS1 gene program across 400 diverse studies, identifying activation of this signature in myeloid cells from multiple hyperinflammatory conditions, including Kawasaki disease. In-house experimental validation confirmed that serum from Kawasaki disease patients induces the MS1 phenotype in monocytes. These findings highlight that SIGnature can help uncover shared mechanisms across conditions, demonstrating its power for large-scale signature scoring and cross-disease analysis.