HyperMap: An Efficient Framework for Transferring Perturbation Responses Across Diverse Biological Contexts

Abstract

Recent perturbation atlases profile transcriptional responses to thousands of targeted perturbations in a reference cell type. Generalising these datasets across lineages and individuals has been problematic, however, as similar baseline transcriptomes can yield highly divergent responses. To address this challenge, we present HyperMap, a meta-learning framework that translates existing atlases to predict perturbation responses in new biological contexts using a small number of perturbation "seeds." Applied to CRISPR gene knockdowns in induced pluripotent stem cells, HyperMap accurately captures responses of new iPSC donors. It generalises to additional cell lines, perturbations by small-molecule drugs, and knockdowns not yet performed in any context. HyperMap is highly efficient, obtaining best-in-class predictions with one-eighth the parameters of typical foundation models. Integrating across atlases yields HyperMapDB, a complete 18x19,036 (cell-line x perturbation) matrix expanding current data by 27-fold. HyperMap enables predictive maps spanning the combinatorial space of biological contexts, gene knockdowns and drugs.