anonymous-xyz96/MisLocus

MisLocus是一个用于变异错误定位基准测试的单细胞图像数据集，包含从Cell Painting Gallery中VarChAMP公共集合（cpg0020-varchamp）提取的U2OS细胞图像，这些细胞表达了GFP标记的人类参考和变异蛋白。每个图像裁剪尺寸为128x128像素，具有四个通道（DNA、GFP、AGP、Mito），数据类型为uint16。完整版本覆盖6个成像批次（B7、B8、B13、B14、B15、B16），每个批次约600个等位基因。目前上传了3,332,309个细胞，涉及1562个独特等位基因，分布在6个批次清单和6个批次的碎片压缩包中。数据集旨在评估学习图像表示在变异错误定位检测中的性能：给定来自同一基因的（参考、变异）细胞对，能否通过表示使每个等位基因分类器在批次内留一板交叉验证中区分参考与变异？数据集还包括预处理后的特征表示文件（如CellProfiler、Cytoself、SubCell等），并遵循CC-BY-4.0许可证。MisLocus — Single-Cell Crops for Variant Mislocalization Benchmarking. This dataset contains segmented single-cell crops of U2OS cells expressing GFP-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), ~600 alleles per batch. Currently uploaded: 3,332,309 cells across 1562 unique alleles in 6 batch manifest(s) and 6 batch(es) of shard tarballs. 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 enable per-allele classifiers to discriminate ref vs variant via within-batch leave-one-plate-out cross-validation? It includes preprocessed feature representations (e.g., CellProfiler, Cytoself, SubCell) and is licensed under CC-BY-4.0.