VirtualSkin integrates skin-cell identity with intervention-response prediction

Abstract

Human skin atlases resolve a diverse set of cellular states, yet connecting these states to their responses to intervention remains difficult across independently sampled tissues. Here we present VirtualSkin, a cell language model that aligns skin transcriptomes with biological descriptions and supports intervention conditioned response prediction. We evaluated the model across independent atlas, disease, anatomical and perturbation cohorts, with donors, studies and perturbations withheld according to the experimental design. VirtualSkin improved inflammatory skin annotation by 0.056 macro F1 points and cell text retrieval by 0.12 Recall@10 points relative to the strongest corresponding comparators. The representation retained cell identity during study integration and transferred across ageing, systemic sclerosis, lymphoma and anatomical sites. Across five intervention cohorts, response prediction improved by 0.03 0.09 response-score points, with agreement at the levels of gene expression change, gene recovery and transcriptional programs. Paired tissue trajectories provided a complementary assessment in two donors per intervention. Component ablations and pairing controls localized the predictive contributions, while uncertainty estimates supported selective prediction under distribution shift. These results connect skin-cell recognition to intervention response prediction within a common evaluation framework and define the conditions under which predictions can inform experimental prioritization.