We propose CellCLIP, a multimodal contrastive learning framework for Cell Painting data. It aligns image profiles and perturbation descriptions using pretrained encoders and a novel channel-wise scheme.