transhumanist-already-exists/human-crispr-fitness-agreements

Human CRISPR Fitness Agreements

Traceable, group-aware evidence for recurrent gene fitness effects in human cell cultures

8,633 gene-symbol records supported by at least three audited authorship/supervision groups in retained human-cell CRISPR fitness screens. This is a versioned derivative of BioGRID ORCS 2.0.18 and the project's publication/provenance audit.

One row per gene. Every counted agreement links to the publications and exact screens that support it. Choose the minimum number of supporting groups at loading time.

Scope: experimentally derived CRISPR fitness-screen hits in cultured human cells. Rescue experiments are not an inclusion requirement. The dataset does not certify clinical validity, universal gene essentiality, or fully independent replication.

Why this resource is useful

Genetic loss-of-function experiments directly perturb gene activity and measure a cellular outcome. Recurrent hits across research groups can help prioritize genes for which an experimental fitness association extends beyond a single publication. This is useful evidence for hypothesis generation, experimental follow-up, and evidence-aware model development.

Publication volume can overstate breadth of support when papers share investigators or reuse screens. This resource preserves an audited grouping of papers, counts each supporting group once per gene, and retains the source chain needed to inspect the result. A researcher can therefore change the breadth-of-support threshold without rebuilding the dataset or losing provenance.

The principal contribution is a filterable evidence index, with experimental contexts accessible through linked screen metadata. It contains no generated gene-mechanism summaries, model-predicted labels, expression matrices, or patient-level records.

At a glance

Property	Value
Unit of observation	One ORCS gene symbol
Included genes	8,633
Minimum positive supporting groups	3
Audited publication scope	44 papers
Publications contributing retained screens	42
Retained screen records	886
Biological setting	Human cell cultures; frequently cancer-derived models
Readout	Proliferation/fitness in negative-selection loss-of-function screens
Version	orcs-2.0.18-audit-2026-09-04-v1

Screen records are not a count of independent cell lines, biological replicates, or laboratories.

What an agreement means

An agreement records at least one source HIT=YES for a gene in a retained screen from a publication assigned to that group. Each group counts once per gene, irrespective of its number of publications or screens. agreement_count equals the length of agreements.

Groups are connected components of publications sharing verified corresponding/co-senior authors or explicitly identified supervisors. Transitive connections are included. This authorship-based proxy does not certify independent laboratories, samples, or experiments. Group IDs are frozen within this dataset version.

The selection covers human-cell proliferation/fitness, loss-of-function, negative-selection time-course screens with the source condition label -. Support is reconstructed from individual source hit tables. The audit retained 886 screen records after excluding or holding known/suspected reuse, unresolved provenance, and one permissive-threshold screen. This dataset expands gene coverage within the existing 44-publication audit; it does not cover all ORCS screens or all published studies.

Files and fields

• data.jsonl: one gene per line, with gene_symbol, source gene_identifiers, agreement_count, agreements, and total supporting publication_count and screen_count.
• Each agreement contains group_id, supporting publication/screen counts, and publications. Each publication contains pmid, source_url, and the exact supporting screen_ids for that gene.
• publications.jsonl: publication titles, verified leadership, and evidence URLs with source locators.
• screens.jsonl: retained screen metadata, cell line/type, library, significance criteria, and source links.
• groups.jsonl: full audited component membership. A component can include a paper whose screens were excluded; only retained positive support contributes to gene agreements.
• metadata.json: version, counts, and validation summary.
• example.json: one complete real record.
• sha256.json: file integrity hashes.

Gene identifiers/symbols are retained from ORCS; no additional HGNC canonicalization or protein-coding-only filter was applied.

Dynamic subsets

Load directly from Hugging Face:

from datasets import load_dataset

repo = "transhumanist-already-exists/human-crispr-fitness-agreements"
genes = load_dataset(repo, "default", split="train")

# Change this threshold without changing the source dataset.
minimum_groups = 4
subset = genes.filter(lambda row: row["agreement_count"] >= minimum_groups)
assert len(subset) == 5999

# Follow a gene's support to publications and individual screens.
record = next(row for row in genes if row["gene_symbol"] == "ABCA5")
for agreement in record["agreements"]:
    print(agreement["group_id"], agreement["publications"])

screens = load_dataset(repo, "screens", split="train")
publications = load_dataset(repo, "publications", split="train")

The train split is a storage convention for the complete table. No independently held-out validation or test split is provided. For reproducible experiments, pin a Hub commit using revision=.

Minimum agreement count	Genes
3	8,633
4	5,999
5	4,472
10	2,034
15	774
20	30

Python, without additional dependencies (run from this directory):

import json

with open("data.jsonl") as source:
    records = [json.loads(line) for line in source]

minimum_groups = 4
subset = [r for r in records if r["agreement_count"] >= minimum_groups]
assert len(subset) == 5999

Interpretation limits

These records support recurrent fitness-screen hits in cultured human cells, frequently cancer-derived models. They do not require knockout followed by gene rescue, establish an organism-level mechanism, or provide a harmonized effect size/direction. Source significance thresholds and experimental designs vary. Four or ten groups is a filtering choice, not a guarantee of biological truth.

This is a positive-support dataset. Missing support does not automatically indicate disagreement; the gene may be untested or fail a study-specific threshold. Non-hit observations and denominators must be evaluated separately before estimating concordance rates. No claim is made that every counted group reproduced the same phenotype in the same cellular context.

Additional considerations:

• Context dependence: a fitness effect in one cell line may not generalize to primary cells, other tissues, or a living person. Cross-context support and same-context replication answer different questions.
• Assay specificity: off-target effects, copy-number-related cutting toxicity, variable guide efficiency, and culture conditions can influence CRISPR measurements. Group counts alone cannot eliminate these concerns.
• Unequal opportunity for support: well-studied genes and commonly screened cell systems have more opportunities to accumulate hits. agreement_count is a count, not a calibrated probability, effect magnitude, or vote fraction.
• Grouping uncertainty: shared supervisory authors connect papers transitively. Conversely, separate components can still share samples, infrastructure, or historical data. Known/suspected reuse was screened, but exhaustive raw-sample independence has not been established.
• Audit limits: authorship and provenance fields are research curation outputs and remain subject to correction; this release is not an independently certified or peer-reviewed benchmark.
• Snapshot scope: the 44-paper audit is a selected corpus. Inclusion does not imply an exhaustive literature search for each gene.

Intended uses

• Prioritize genes with recurrent experimental fitness-screen support.
• Build evidence-aware retrieval or training datasets with explicit provenance.
• Compare conclusions across thresholds such as ≥3, ≥4, ≥5, or ≥10 groups.
• Join hits to cell-line, publication, and assay metadata before context-specific analysis.

For model evaluation, account for publication, authorship-group, cell-line, and source-data overlap when designing splits. A random gene split alone does not establish independence of the underlying evidence. This dataset should not be used as a clinical decision tool, a rescue-validated gene catalogue, or a negative-label catalogue for genes without recorded support.

Curation and validation

1. Restrict the previously audited human-cell screen corpus to the selected loss-of-function fitness design.
2. Apply the frozen screen-level provenance and eligibility decisions.
3. Assign publications to audited supervision components.
4. Re-read original ORCS hit tables and collect each gene's exact positive screen/publication support.
5. Deduplicate supporting components per gene and retain counts ≥3.
6. Validate source hit counts, publication counts, group uniqueness, nested references, and all published threshold totals.

The retained scope has 886 of 892 candidate screen records. Six were excluded or held for historical reuse, suspected shared observations, unresolved provenance/leadership, or a permissive significance threshold. These decisions reduce identifiable sources of inflated support; they do not prove complete independence.

Reproduction

The project builder is scripts/build_gene_agreement_dataset.py. It uses frozen local audit tables and /tmp/BIOGRID-ORCS-human.tar.gz. The Hub release contains the derived records and linked evidence metadata; the complete upstream local audit workspace and original archive are not bundled. Full end-to-end rebuilding requires those additional inputs. Counts and nested group/publication/screen references can be checked directly from the published files. sha256.json records release-file integrity hashes.

Sources, attribution, and license

Source screen data: BioGRID ORCS and the ORCS 2.0.18 release. Field definitions are documented in BioGRID's ORCS download format reference.

BioGRID distributes its downloadable datasets under the MIT License; see its download and attribution statement. This derived dataset is distributed under MIT. Linked articles and their contents retain their respective publishers' and authors' terms; their full text is not redistributed here.

Please cite the original contributing studies relevant to your selected genes, using the PMIDs in each record, and acknowledge BioGRID. BioGRID requests citation of: Stark C, Breitkreutz BJ, Reguly T, Boucher L, Breitkreutz A, Tyers M. BioGRID: a general repository for interaction datasets. Nucleic Acids Research. 2006;34:D535–D539. Also report this dataset's version, Hub revision, selected threshold, and any context filters used in your analysis.