Example: protein localization with MisLocus

Biological question: does a protein variant change where the protein is located in a cell?

MisLocus is a related example of using image-based representations to compare reference and variant proteins. It uses single-cell images from VarChAMP, not the JUMP Cell Painting dataset.

Start with precomputed features

The public dataset includes cleaned features from CellProfiler, Cytoself, MorphEm, and SubCell variants, plus cell metadata and optional four-channel image crops. You can reuse the features without running an image encoder.

1. Follow the repository’s feature-download instructions to select one representation and imaging batch rather than downloading the full dataset.
2. Follow the CPU scoring guide to check cell identities and comparison cohorts, then compare variant cells with matched reference cells using classification or phenotypic-activity scores.
3. Use the benchmark workflow to evaluate scores against variant annotations or retrieve reference-protein localizations using Human Protein Atlas labels.

Published feature files are already cleaned: do not normalize them again. Preserve the documented replicate splits and control calibration when comparing representations. Detectable image differences alone do not establish a variant’s clinical pathogenicity.

This example runs in the companion repository’s own environment; it is not one of the Hub’s Colab notebooks. Consult its maintained guides for installation, dataset revisions, and analysis settings.