MisLocus: A single-cell imaging benchmark for protein mislocalization

Dataset summary

This dataset contains segmented single-cell crops of U2OS cells expressing mNeonGreen-tagged human reference and variant proteins, drawn from the public VarChAMP collection in the Cell Painting Gallery (cpg0020-varchamp). Each crop is 128 x 128 pixels, four channels (DNA, GFP, AGP, Mito), uint16. The full release covers 6 imaging batches (B7, B8, B13, B14, B15, B16), 232–919 alleles per batch (1,562 unique alleles including controls across the release).

Currently uploaded (including controls): 3,332,309 cells and 1,562 unique alleles across six batch manifests and six batches of crop shards. Excluding control observations and recurring control identities, the biological cohort contains 3,002,837 cells, 1,541 alleles (250 references and 1,291 variants) across 282 genes; 1,055 variants in 125 genes have a matched reference.

Evaluative claim

The dataset enables evaluation of learned image representations on variant mislocalization detection: given pairs of (reference, variant) cells from the same gene, can a representation distinguish variant from matched reference cells on an independently imaged technical replicate? The paper's headline results use a single T4-held-out evaluation: when representation training is required, T1 and T2 are used for training, T3 for validation, and T4 is never used during representation learning. Downstream supervised classifiers are trained on T1--T3 and evaluated on T4; unsupervised mAP_norm is likewise computed using T4 profiles only. This is the primary held-out evaluation. Frozen representations and CellProfiler features are evaluated on the same T4 plates. Benchmark numbers and the complete protocol are reported in the companion paper.

Release status

The complete six-batch dataset is available. The fine-tuned SubCell feature files were replaced with the allele-supervised RYBG-v2 seed-42 results.

Component	Status
Inspection sample (sample/, ~1.3 GB)	uploaded
Per-batch manifests (manifest/)	6/6 batches uploaded
Full single-cell crops (single_cell_crops/)	6/6 batches uploaded

Manifest batches present: Batch_13, Batch_14, Batch_15, Batch_16, Batch_7, Batch_8 Shard batches present: 2024_01_23_Batch_7, 2024_02_06_Batch_8, 2025_01_27_Batch_13, 2025_01_28_Batch_14, 2025_03_17_Batch_15, 2025_03_17_Batch_16

Repository layout

metadata/
    allele_inventory.parquet                # 1,532 rows × 1,031 cols (1,282 distinct gene_variant values);
                                              ClinVar, sequence and plate/well metadata for 6 batches.
                                              Not the QC manifest's 1,291 biological variants.
manifest/                                   # cell-level catalog per batch (with s3:// tiffpaths)
sample/                                     # curated inspection subset (~1.3 GB)
single_cell_crops/                          # raw 128x128 crops, sharded per batch (~6 GB / shard)
    2025_03_17_Batch_15/shard-NN.tar.gz
    ...
representations/                            # processed feature parquets per (representation, batch)
    cellprofiler/{batch}/features.parquet           # ~1100 cols, post-pycytominer feature selection
    cytoself/{batch}/features.parquet               # ~500 cols, VQ-VAE-derived
    subcell_finetuned_mae/{batch}/features.parquet  # allele-supervised RYBG-v2 MAE, seed 42
    subcell_finetuned_vit/{batch}/features.parquet  # allele-supervised RYBG-v2 ViT, seed 42
    subcell_portable_rbg_mae/{batch}/features.parquet
    subcell_portable_rbg_vit/{batch}/features.parquet
    morphem/{batch}/features.parquet               # frozen MorphEm, 4 x 384 features
MisLocus_croissant.json                     # Croissant metadata (core + RAI fields)
README.md                                   # this file
LICENSE                                     # CC-BY-4.0

Representations

Each representations/{rep}/{batch}/features.parquet is the post-preprocessed, ready-to-use feature set for that (representation, batch) pair. Feature schema:

Representation	Backbone	Feature cols (approx.)	Input
cellprofiler	classic CellProfiler ImageMath features	903–1,091	TIFFs
cytoself	VQ-VAE (Cytoself), trained B7+B8+B13-B16	471–714 (combined: global + spectrum)	128x128 GFP+nuc-distance crops
subcell_finetuned_mae	Allele-supervised SubCell RYBG-v2 MAE, seed 42	1,533–1,536	128x128 4-channel crops
subcell_finetuned_vit	Allele-supervised SubCell RYBG-v2 ViT, seed 42	1,535–1,536	128x128 4-channel crops
subcell_portable_rbg_mae	Frozen SubCell-Portable MAE (RBG channel mapping)	1,513–1,533	128x128 4-channel crops
subcell_portable_rbg_vit	Frozen SubCell-Portable ViT (RBG channel mapping)	1,501–1,516	128x128 4-channel crops
morphem	Frozen CaicedoLab MorphEm ViT-S/16	1,536 (4 x 384)	each channel processed independently from 128x128 crops

morphem is the public release name for the pipeline's internal vit representation. The four independent CLS-token embeddings are ViT_gfp_0..383, ViT_dna_0..383, ViT_agp_0..383, and ViT_mito_0..383. Downstream groups: GFP (384), Morph (DNA + AGP + Mito, 1,152), and ALL (1,536). All four channels are retained even when a benchmark reports only GFP and ALL.

The updated fine-tuned SubCell encoders were trained on technical replicates T1+T2, selected on T3 by validation macro allele average precision at pass 100, and applied to T1–T4 (including the held-out T4). They use four-channel AGP/Mito/DNA/GFP input in RYBG order and allele-level supervision. Each released parquet is a processed, per-plate normalized and feature-selected matrix, not raw encoder output; model checkpoints are not included in this dataset.

All seven feature parquets per batch share metadata columns including Metadata_Plate, Metadata_Well, Metadata_Site, Metadata_ImageNumber, Metadata_ObjectNumber, Metadata_gene_allele, and Metadata_node_type. Feature rows are a subset of manifest cells after preprocessing; join on (Metadata_Plate, Metadata_Well, Metadata_ImageNumber, Metadata_ObjectNumber). Metadata_CellID is present in the manifest and in the updated fine-tuned SubCell files, but not in the other released feature parquets. The new SubCell files additionally include Metadata_Split, Metadata_Batch, and Metadata_BatchQualifiedCellID.

Quick start (smoke test)

To validate the pipeline end-to-end on real data without downloading the full dataset (~375 GB of crop tarballs and feature parquets), pull a single batch:

hf download anonymous-xyz96/MisLocus --repo-type dataset \
    --include "single_cell_crops/2025_03_17_Batch_15/*" \
    --include "manifest/manifest_Batch_15.parquet" \
    --local-dir ./MisLocus

cd ./MisLocus/single_cell_crops/2025_03_17_Batch_15
for f in shard-*.tar.gz; do tar -xzf "$f"; done

After extraction, each allele directory contains dna.npy, gfp.npy, agp.npy, mito.npy (each shape (N_cells, 128, 128) uint16) plus metadata.parquet with cell-level identifiers.

Splits

Train/val/test splits are not encoded in the directory structure. Per-batch manifests include Metadata_Plate and technical-replicate information (T1/T2/T3/T4 suffixes on the plate name), so users can derive the paper's primary split: T1/T2 for representation training, T3 for validation, and T4 held out for evaluation. Downstream supervised classifiers are trained on T1--T3 and scored on T4; unsupervised retrieval also uses T4 profiles.

Reproducibility numbers

See the companion paper for AUROC tables, ablations, and the primary single-T4-held-out benchmark protocol. Dataset versions are tracked via HF Hub.

License

CC-BY-4.0. The underlying VarChAMP imaging data is licensed under CC-BY-4.0 by the Cell Painting Gallery.