JUMP-lite benchmark shrinks cell imaging data 1000x without losing signal

Image-based profiling has become a cornerstone of drug discovery and functional genomics, but the sheer scale of public resources like JUMP Cell Painting—115 TB of images—makes systematic benchmarking intractable for most labs. To address this, Alán F. Muñoz, Johan Fredin Haslum, Runxi Shen, Anne E. Carpenter, and Shantanu Singh (Broad Institute) present JUMP-lite, a curated 116 GB subset that is 1000x smaller than the original. The reduction is achieved through careful selection of perturbations with high-confidence annotations and lossy JPEG XL compression, which preserves phenotypic diversity while slashing storage requirements.

Alongside JUMP-lite, the team releases Nahual, an open-source framework for reproducible model deployment. Using these tools, they benchmarked five representation methods: classical features from CellProfiler and deep learning models MorphEM, OpenPhenom, SubCell, and DINOv2. Their standardized phenotypic activity and consistency metrics reveal meaningful performance differences, and critically, compression does not degrade downstream signal. This combination of a compact dataset and reproducible evaluation framework lowers the barrier to entry for researchers, making large-scale cell representation benchmarking accessible to any lab with a standard workstation.