University of California San Diego / Ideker Lab
Meta-learning framework that transfers perturbation responses across cell lines, donors, and drugs from a few seed perturbations, using one-eighth the parameters of typical single-cell foundation models.
HyperMap is a meta-learning framework for predicting how cells respond to genetic and chemical perturbations across biological contexts that were never directly profiled. Developed by Bhupinder Dhaka, Jiahao Gao, and Trey Ideker at the University of California San Diego and released as a preprint in April 2026, it tackles a persistent obstacle in functional genomics: recent perturbation atlases catalogue transcriptional responses to thousands of targeted perturbations, but these maps generalize poorly across lineages and individuals. Cells with nearly identical baseline transcriptomes can mount strikingly divergent responses to the same perturbation, so an atlas built in one reference cell type cannot simply be reused for another donor, cell line, or compound.
Rather than retraining a large model for each new context, HyperMap treats each context as a task in a meta-learning setting. At inference time it is given a small number of perturbation "seeds" — measured responses for a handful of perturbations in the target context — and uses them as support examples to adapt an existing atlas to that context. Crucially, these seeds are inference-time support data, not new training data: a single fixed checkpoint adapts on the fly without any gradient updates or re-training. This lets HyperMap translate a well-characterized reference atlas into accurate predictions for new iPSC donors, additional cell lines, small-molecule drug treatments, and even gene knockdowns that have not yet been performed in any context.
By integrating predictions across the atlases it has access to, HyperMap assembles HyperMapDB, a complete 18 x 19,036 matrix spanning 18 cell lines and 19,036 perturbations. A single fixed checkpoint generates this entire dense matrix without per-context re-training, expanding the available perturbation data roughly 27-fold relative to the source atlases.
HyperMap is a meta-learning model that conditions on a small support set of measured perturbation responses ("seeds") to predict transcriptional outcomes in a target context, framing each cell line, donor, or drug treatment as a distinct task. It was developed primarily on CRISPR gene-knockdown atlases in induced pluripotent stem cells (iPSCs), where it accurately reconstructs the responses of previously unseen iPSC donors. The authors report that it generalizes to additional cell lines, to perturbations induced by small-molecule drugs, and to knockdowns not yet performed in any profiled context. A central efficiency claim is that HyperMap reaches best-in-class predictive performance with roughly one-eighth the parameters of typical single-cell foundation models, and that a fixed checkpoint can populate the full HyperMapDB matrix — 18 cell lines by 19,036 perturbations — without re-training, yielding an approximately 27-fold expansion over the source atlases. As an April 2026 bioRxiv preprint in systems biology, these results have not yet undergone peer review.
HyperMap is aimed at researchers who need perturbation response predictions in contexts that are expensive or impractical to profile exhaustively. Drug-discovery and functional-genomics teams can use it to forecast CRISPR knockdown or small-molecule effects in new cell lines and donor backgrounds from only a handful of seed experiments, prioritizing which perturbations to run at the bench. Because it spans the combinatorial space of contexts, gene knockdowns, and drugs, it is well suited to building predictive maps for disease-relevant cell types where reference atlases exist but context-specific data is sparse, and to studying donor-to-donor variability in perturbation responses across iPSC panels.
HyperMap reframes cross-context perturbation prediction as a meta-learning problem solved with a few inference-time seeds rather than a large, separately trained model per context, and reports competitive accuracy at a fraction of the parameter count of mainstream single-cell foundation models. Its most tangible contribution is HyperMapDB, a dense 18 x 19,036 perturbation matrix produced from a single checkpoint that expands available data roughly 27-fold and could serve as a resource for downstream analysis and benchmarking. As a recent, not-yet-peer-reviewed preprint, its generalization claims await independent validation, and no public code or dataset release was identified at the time of writing; the parameter-efficiency and transfer results nonetheless point toward lighter-weight, adaptable alternatives to large perturbation foundation models.