Reliable feature extraction from multichannel microscopy images is crucial for biological discovery, but existing models typically require fixed channel architectures or artificial RGB compositing. We introduce CellRep, a channel-invariant representation learning model that generates consistent feature representations across varying experimental conditions. By employing content-aware patch embedding and channel-mixing transformer encoding, CellRep learns to identify and represent biological structures independent of channel position or type. Our evaluations demonstrate CellRep’s strong performance as a microscopy image featurizer for perturbation prediction, particularly when generalizing to novel cell types, imaging techniques, and channel configurations not seen during training.