# Norman 2019: Exploring Genetic Interaction Manifolds Constructed from Rich Single-Cell Phenotypes

**Authors:** Thomas M. Norman, Max A. Horlbeck, Joseph M. Replogle, Alex Y. Ge, Albert Xu, Marco Jost, Luke A. Gilbert, Jonathan S. Weissman
**Year:** 2019
**Venue:** Science, vol. 365, pp. 786–793
**DOI:** 10.1126/science.aax4438
**GEO accession:** GSE133344

## One-sentence summary

Systematic CRISPRa screen of 131 transcription factor perturbations (single + combinatorial) in K562 cells using Perturb-seq, revealing genetic interaction structure and enabling ML-based prediction of unseen combinations.

## Key contribution

The **go-to benchmark dataset** for perturbation response prediction. Norman 2019 is to perturbation biology what MNIST was to image recognition: small enough to run fast, complex enough to be non-trivial, biologically interpretable. Nearly every perturbation ML paper benchmarks against it.

The dataset maps the "genetic interaction manifold" — showing how combinations of TF perturbations produce cellular states that can be additive, synergistic, suppressive, or emergent.

## Experimental design

- **Cell line:** K562 (human myeloid leukemia)
- **Perturbation type:** CRISPRa (gene activation via dCas9-VP64)
- **Target genes:** 112 "hit genes" whose activation strongly affected K562 growth
- **Design:** Single perturbations (112) + all pairwise combinations (subset); ~57k sgRNA cassettes
- **Readout:** Single-cell RNA-seq (Perturb-seq)
- **Scale:** ~100k cells total

## Key findings

- Most double-gene perturbations are approximately **additive** (effects sum linearly)
- ~10–15% of pairs show significant non-additive interactions (epistasis)
- Genetic interactions cluster by functional pathway — pairs from the same pathway often suppress each other
- Single-cell resolution reveals subpopulation effects invisible in bulk measurements
- ML models can predict unseen combinations reasonably well from single perturbation data (enables GEARS, CPA evaluation)

## Why this dataset matters

1. **Gold standard benchmark** — GEARS, CPA, scGPT, Lingshu-Cell, Cell-JEPA all test on it
2. **Epistasis ground truth** — rare dataset where you have both single and combinatorial perturbations with readout
3. **Biologically interpretable** — TF overexpression is mechanistically cleaner than KO; effects are often large and directional
4. **Public** — fully available on GEO (GSE133344)

## Limitations as a benchmark

- Single cell line (K562) — myeloid leukemia; not primary, not primary
- CRISPRa only — gain-of-function; doesn't capture loss-of-function biology
- 112 target genes — not genome-scale; focused on growth-relevant TFs
- Well-characterized genes — models can cheat using prior biological knowledge
- Many papers report only "mean correlation" on held-out pairs — a forgiving metric

## Connections

- [../concepts/perturbation-biology.md](../concepts/perturbation-biology.md) — this dataset defines the standard benchmark
- [../papers/gears.md](gears.md) — trained/evaluated on Norman 2019
- [../papers/cpa.md](cpa.md) — evaluated on Norman 2019
- [../papers/replogle-2022.md](replogle-2022.md) — successor dataset; genome-scale but single-gene only
- [../entities/weissman-lab.md](../entities/weissman-lab.md) — lab that generated Norman + Replogle datasets