Learning reproducible cellular perturbation responses from limited independent measurements

Abstract Large-scale perturbation atlases are increasingly used to train models that predict cellular responses, yet responses to individual perturbations are often measured only a few times. The resulting averages are commonly used as training targets, although they contain both reproducible perturbation effects and measurement-specific variation. Here we present RePert, a framework that uses independent measurements to construct reproducible targets for cellular-response prediction. RePert learns from separate measurements of the same perturbation, combines complementary estimates using independent calibration data, and uses a second assay to refine an existing response estimate. In two Cell Painting datasets and one gene-expression dataset, RePert estimates agreed more closely with held-out measurements than averages of the same measurements. In Cell Painting, they also better distinguished the correct compound from matched alternatives. In BBBC047, these estimates outperformed larger neural networks given identical inputs. Matched measurements from a second assay provided additional compound-specific information, with smaller gains when more Cell Painting measurements were available. The recovered responses also better preserved independently supported morphological features, response directions and patterns across doses. As training targets, they improved prediction of unseen compounds in BBBC047, whereas in sci-Plex3, with two measurements per condition, the effect depended on the prediction setting. Target construction is therefore a design dimension of perturbation modelling, distinct from model architecture and capacity.