{
  "@context": {
    "@language": "en",
    "@vocab": "https://schema.org/",
    "citeAs": "cr:citeAs",
    "column": "cr:column",
    "conformsTo": "dct:conformsTo",
    "cr": "http://mlcommons.org/croissant/",
    "rai": "http://mlcommons.org/croissant/RAI/",
    "data": {
      "@id": "cr:data",
      "@type": "@json"
    },
    "dataType": {
      "@id": "cr:dataType",
      "@type": "@vocab"
    },
    "dct": "http://purl.org/dc/terms/",
    "equivalentProperty": "cr:equivalentProperty",
    "examples": {
      "@id": "cr:examples",
      "@type": "@json"
    },
    "extract": "cr:extract",
    "field": "cr:field",
    "fileProperty": "cr:fileProperty",
    "fileObject": "cr:fileObject",
    "fileSet": "cr:fileSet",
    "format": "cr:format",
    "includes": "cr:includes",
    "isLiveDataset": "cr:isLiveDataset",
    "jsonPath": "cr:jsonPath",
    "key": "cr:key",
    "md5": "cr:md5",
    "parentField": "cr:parentField",
    "path": "cr:path",
    "prov": "http://www.w3.org/ns/prov#",
    "recordSet": "cr:recordSet",
    "references": "cr:references",
    "regex": "cr:regex",
    "repeated": "cr:repeated",
    "replace": "cr:replace",
    "samplingRate": "cr:samplingRate",
    "sc": "https://schema.org/",
    "separator": "cr:separator",
    "source": "cr:source",
    "subField": "cr:subField",
    "transform": "cr:transform"
  },
  "@type": "sc:Dataset",
  "name": "OPS-Eval",
  "description": "OPS-Eval is a leakage-resistant evaluation protocol and benchmark suite for optical pooled screens, instantiated on the Vesuvius genome-wide CRISPR knockout screen with 5,072 genes, 20,537 guides, ~2.1M four-channel cell crops, and released phenotype features. It provides gene-disjoint splits, leakage and confound audits, a baseline ladder, and a one-command reproduction harness.",
  "url": "https://huggingface.co/datasets/cspeters119/ops-eval",
  "license": "https://spdx.org/licenses/MIT.html",
  "conformsTo": "http://mlcommons.org/croissant/1.1",
  "citeAs": "Anonymous. OPS-Eval: Leakage-Resistant Evaluation for Optical Pooled Screens. NeurIPS 2026 Evaluations and Datasets Track.",
  "version": "1.0.0",
  "datePublished": "2026-05-06",
  "keywords": [
    "optical pooled screens",
    "CRISPR",
    "evaluation",
    "data leakage",
    "gene-disjoint splits",
    "microscopy",
    "benchmark"
  ],
  "distribution": [
    {
      "@type": "cr:FileObject",
      "@id": "splits_v1",
      "name": "splits_v1.json",
      "description": "Gene-disjoint train/validation/test split assignments for 5,322 genes (5,072 targeted + 250 non-targeting controls). 3,628 train / 694 validation / 1,000 test genes.",
      "contentUrl": "splits_v1.json",
      "encodingFormat": "application/json",
      "sha256": "08dfbe378c94bfad670d3dff0836a90d322d3bb71e79a181e64628e9a18adb57"
    },
    {
      "@type": "cr:FileObject",
      "@id": "splits_v1_random",
      "name": "splits_v1_random.json",
      "description": "Random image-level comparison split (leaky). Used only for split-sensitivity analysis.",
      "contentUrl": "splits_v1_random.json",
      "encodingFormat": "application/json",
      "sha256": "3c1d90d0dd630d16cbd0cf8aa51e7aad0b92e40b14535a16a55393588ebbbaec"
    },
    {
      "@type": "cr:FileObject",
      "@id": "splits_v1_sgrna_disjoint",
      "name": "splits_v1_sgrna_disjoint.json",
      "description": "Guide-disjoint comparison split. Used only for split-sensitivity analysis.",
      "contentUrl": "splits_v1_sgrna_disjoint.json",
      "encodingFormat": "application/json",
      "sha256": "402d947816564b4e3e801a05ee2e44319caf2c1e0e6366f8efe197ebec0ec177"
    },
    {
      "@type": "cr:FileObject",
      "@id": "gene_metadata",
      "name": "gene_metadata.parquet",
      "description": "Per-gene metadata including 98-dimensional morphological feature vectors, cluster labels (222 interphase + 222 mitotic), guide counts, and cell counts.",
      "contentUrl": "gene_metadata.parquet",
      "encodingFormat": "application/x-parquet",
      "sha256": "815a6b4306b6cfec42caf962c3c75187027e4d14f6e266c9d1555b5b7e514040"
    },
    {
      "@type": "cr:FileObject",
      "@id": "manifest_v1",
      "name": "manifest_v1.json",
      "description": "Frozen manifest with SHA256 checksums for all data files, enabling exact replication.",
      "contentUrl": "manifest_v1.json",
      "encodingFormat": "application/json",
      "sha256": "a483d8c3df76cfbafadd838da92a77aa356547a8c676537aba74930f378bd126"
    },
    {
      "@type": "cr:FileSet",
      "@id": "montages",
      "name": "montages",
      "description": "Per-gene montage images (2700x2000 px) across 4 fluorescence channels (DNA, Tubulin, gH2AX, Actin) and 2 cell phases (interphase, mitotic). Each montage tiles ~100 single-cell crops per guide in horizontal bands.",
      "containedIn": "montages/",
      "includes": "montages/*/*.png",
      "encodingFormat": "image/png"
    },
    {
      "@type": "cr:FileSet",
      "@id": "cell_embeddings",
      "name": "cell_embeddings",
      "description": "Pre-extracted 512-dimensional cell-level embeddings (float16) from a frozen ResNet-18 encoder. One .npz file per gene with keys for each sgRNA.",
      "containedIn": "cell_embeddings/",
      "includes": "cell_embeddings/*.npz",
      "encodingFormat": "application/x-npz"
    },
    {
      "@type": "cr:FileSet",
      "@id": "tiny_sample",
      "name": "tiny_sample",
      "description": "A 50-gene reviewer-inspectable subset with sanitized filenames, sufficient for end-to-end smoke testing.",
      "containedIn": "tiny_sample/",
      "includes": "tiny_sample/**/*",
      "encodingFormat": "image/png"
    }
  ],
  "recordSet": [
    {
      "@type": "cr:RecordSet",
      "@id": "genes",
      "name": "genes",
      "description": "One record per gene with split assignment, cluster labels, and guide count.",
      "field": [
        {
          "@type": "cr:Field",
          "@id": "genes/gene_symbol",
          "name": "gene_symbol",
          "description": "HGNC gene symbol or non-targeting control identifier.",
          "dataType": "sc:Text",
          "source": {
            "fileObject": {"@id": "gene_metadata"},
            "extract": {"column": "gene"}
          }
        },
        {
          "@type": "cr:Field",
          "@id": "genes/interphase_cluster",
          "name": "interphase_cluster",
          "description": "Interphase phenotypic cluster label (0-221).",
          "dataType": "sc:Integer",
          "source": {
            "fileObject": {"@id": "gene_metadata"},
            "extract": {"column": "interphase_cluster"}
          }
        },
        {
          "@type": "cr:Field",
          "@id": "genes/guide_count",
          "name": "guide_count",
          "description": "Number of guide RNAs targeting this gene (median 4, range 1-7).",
          "dataType": "sc:Integer",
          "source": {
            "fileObject": {"@id": "gene_metadata"},
            "extract": {"column": "n_sgrnas"}
          }
        }
      ]
    }
  ],
  "rai:dataLimitations": "The benchmark is instantiated on a single cell line (HeLa) from a single screen (Vesuvius). Generalization to primary cells, other cell lines, or other perturbation modalities (CRISPRi, CRISPRa, Perturb-seq) is plausible based on structural arguments but has not been empirically validated beyond idr0071 (A549 cells). The 222 phenotypic cluster labels inherit any noise from the original unsupervised clustering pipeline of the source analysis. All image-based baselines use mean-pooling of frozen embeddings, which may understate the potential of fine-tuned or end-to-end approaches. The dataset is not intended for drawing biological conclusions about specific genes without independent experimental validation.",
  "rai:dataBiases": "HeLa is an immortalized cervical cancer cell line with well-documented chromosomal abnormalities and atypical biology. Perturbation effects observed in HeLa may not generalize to non-cancerous or primary cell lines. The screen covers 5,072 of approximately 20,000 human protein-coding genes; the sampled gene set reflects the Brunello CRISPR library design choices. Cluster label imbalance (2-66 genes per cluster, median 22) may affect per-class metric stability for small clusters.",
  "rai:personalSensitiveInformation": "None. All data are derived from an immortalized cell line (HeLa), not from identifiable human subjects. No personal, demographic, or sensitive information is present in the dataset.",
  "rai:dataUseCases": "OPS-Eval is designed for evaluating representation learning and classification methods on pooled CRISPR microscopy data under gene-disjoint splits. Validated use cases include: (1) benchmarking unseen-gene classification, (2) benchmarking guide-level retrieval, (3) quantifying metric inflation from leaky splits, and (4) auditing new OPS datasets for hierarchical leakage. The dataset should NOT be used for: clinical decision-making, drawing biological conclusions about specific genes without independent validation, or as evidence that any particular gene causes a specific phenotype.",
  "rai:dataSocialImpact": "Positive: promotes methodologically correct evaluation in computational biology, preventing false claims of algorithmic progress that could misdirect research effort and funding. Establishes a minimum reporting standard analogous to patient-level splitting in medical imaging. Negative: risk of over-reliance on a single benchmark; strong performance on OPS-Eval does not guarantee biological relevance or clinical utility.",
  "rai:hasSyntheticData": false,
  "prov:wasDerivedFrom": [
    "https://vesuvius.wi.mit.edu",
    "https://www.ebi.ac.uk/biostudies/bioimages/studies/S-BIAD394"
  ],
  "prov:wasGeneratedBy": "Large-scale pooled CRISPR knockout screen in HeLa cells (doxycycline-inducible Cas9) using the Brunello sgRNA library. Cells were imaged in four fluorescence channels (DNA/DAPI, Tubulin, gamma-H2AX, Actin). Cell segmentation, barcode calling, and phenotype scoring were performed by the original analysis pipeline (Funk et al., 2022). OPS-Eval adds gene-disjoint split generation, leakage audits, and baseline evaluation without modifying the underlying data."
}