README.md

metadata

license: cc-by-4.0
extra_gated_heading: Access C2DBench
extra_gated_prompt: >-
  C2DBench is derived from public sources (Norman et al. 2019, GEO GSE133344,
  via scPerturb; DGIdb 5.0). It contains no raw sequencing data and no
  human-subject data — K562 is an immortalised cell line. Please confirm the
  following before access.
extra_gated_fields:
  Full name: text
  Email: text
  Affiliation: text
  Intended use: text
  I will cite Norman et al 2019 and DGIdb alongside this dataset: checkbox
  I understand drug_reference lists are computational nominations, not evidence of efficacy, and will not be used to inform clinical or treatment decisions: checkbox
extra_gated_button_content: Request access
task_categories:
  - feature-extraction
  - question-answering
tags:
  - single-cell
  - perturb-seq
  - crispra
  - drug-discovery
  - agents
  - biology
configs:
  - config_name: default
    data_files:
      - split: eval
        path: data/eval.parquet
      - split: reference_library
        path: data/reference_library.parquet

C2DBench — Cell Shift to Drug (Norman v1)

Given two cell populations — untouched controls and cells where genes were switched on — infer what was changed, what else is functionally involved, and what drugs might act on it.

Built from Norman et al. 2019, CRISPR activation in K562. Self-contained: no external files needed.

Task

Stage	Target	Graded
List 1	the activated gene(s)	✅ list1_gold
List 2	functionally related genes	✅ list2_gold
Drug list	compounds acting on those pathways	⚠️ drug_reference provided; no efficacy key

Splits

There is no train/val/test. This is an agentic evaluation set — every row is scored. The perception model (soft-prompt adapter over cell embeddings) is trained elsewhere; nothing here is training data.

Split	Rows
eval	690	every task item
reference_library	105	one profile per single-gene condition — the agent may consult this

Within eval, stratify on n_target_genes:

n_target_genes	Rows	Difficulty
2	340	the real test — the pair's combined profile was never observed, only its two constituents, so it must be composed
1	310	near-lookup — the gene's own profile sits in reference_library
0	40	sham (control-vs-control), empty list1_gold — naming genes here is a hallucination

Every row is the same full trace: cell pair → list 1 → list 2 → drug list.

The difficulty gap is the point. Nearest-centroid solves single-gene items at 0.905; the best trivial method on two-gene items reaches 0.335. Report the two strata separately — a headline number pooled across them mostly measures the easy half.

reference_library is study material, not leakage: composing those profiles into an unseen pair is the task.

Columns

Column
control_profile, perturbed_profile, delta	1,983-dim log1p CP10K pseudobulk
top_up_genes, top_down_genes	top 25 movers with logFC (JSON) — for text/LLM arms
list1_gold, list2_gold	answer keys
drug_reference	specificity-ranked drug nominations (JSON, median 20)
drug_gold	null — Norman has no drug treatments
n_target_genes	2 = real test, 1 = near-lookup, 0 = sham
effect_tier, snr	strong / moderate / weak / marginal
qc_energy_distance, qc_permutation_p	per-item separation statistics
leakage_control_target_logfc	held-out target genes — do not add to features
control_cell_idx, perturbed_cell_idx	row offsets into the source h5ad

Baselines (two-gene items, n=340)

Method	top-1	top-5	top-20
Additive composition — the bar	0.335	0.603	0.732
Chance (5,460 candidate pairs)	0.00018	—	—

run_baseline.py reproduces this from the parquet files alone.

Three things to know

Targets are masked. CRISPRa raises the activated gene's own transcript, so reading the top mover recovers a true gene 89.6% of the time. All 102 target genes are removed from every visible field; that baseline then scores 0.000. Unmasked values are kept in leakage_control_target_logfc for measuring the leak, not for features.

List 2 is measured, not looked up. Two genes are related if activating them produces similar shifts (top-10, cross-half reproducibility 0.847). A GO/Reactome key would be circular — solvers can query those databases. This one recovers known families 9× above chance.

Pairs are lane- and depth-matched. Perturbed cells run ~18% shallower than controls, and some encoders read depth almost perfectly, so unmatched a model can detect "perturbed" from read depth alone. After matching, 98.5% of real pairs separate at p<0.05 with a 0% sham false-positive rate.

Limits

One cell line, one lab, CRISPR activation only. 340 two-gene items ranks methods but won't resolve small differences. Scoring drug efficacy needs measured drug responses (sciPlex, LINCS, Tahoe) — not available in Norman.

Sources

Derived data only; no raw sequencing redistributed.

• Norman TM et al. Exploring genetic interaction manifolds constructed from rich single-cell phenotypes. Science 365:786–793 (2019). 10.1126/science.aax4438 · GEO GSE133344
• Peidli S et al. scPerturb: harmonized single-cell perturbation data. Nature Methods (2024) — source of the processed object.
• DGIdb 5.0, Nucleic Acids Research (2024) — drug_reference, queried once and frozen. DGIdb aggregates upstream sources with varying licences; see dgidb.org before redistributing that column.

K562 is an immortalised cell line, not human-subject material. No embeddings included. Please cite Norman et al. and DGIdb alongside this dataset.