RELAY: Relational In-Context Learning for Cellular Perturbation Prediction

RELAY, a relational in-context model, predicts unmeasured cellular perturbation coordinates from related measured responses and outperforms prior in-context and supervised methods on unseen variable values and unseen combinations.

• Why it matters: Perturbation responses vary across perturbation identity, cell type, donor, and dose, so predicting new conditions depends on how well models generalize beyond explicit variable representations.
• What they did: RELAY conditions on which experimental variables are shared (not their values) and uses control–perturbed demonstration pairs from related coordinates to predict target responses across chemical, genetic, and cytokine datasets.
• The result: In evaluation, RELAY beats existing in-context methods on unseen variable values and beats supervised perturbation-prediction methods on unseen combinations, and will open source code and data.

AI summary of the abstract · may contain errors · not medical advice · read the paper.