A Cell Painting image of stained cells is turned into a vector so that similar treatments get similar vectors. But the same cells imaged on two different plates get different vectors, because focus, lighting, and dye brightness drift from plate to plate. That drift is a technical artifact the vector should ignore, and often does not. We measure the problem, fix it two ways (a control-anchored normalization and an encoder trained to ignore the plate), and then scale that encoder to the full eleven-laboratory JUMP dataset. Everything runs from a reproducible pipeline.
Cell Painting photographs stained cells, and a model turns each image into an embedding: a list of numbers meant to place similar treatments near each other. The problem is that the same biology, imaged on two different plates, produces two different embeddings. Focus, illumination, and dye brightness drift from plate to plate, and the embedding records that drift. So two identical wells can look different for reasons that have nothing to do with the cells. This is the batch effect.