# GEARS: Predicting Transcriptional Outcomes of Novel Multigene Perturbations

**Authors:** Yusuf Roohani, Kexin Huang, Jure Leskovec
**Year:** 2023 (online Aug 17); 2024 (print, vol. 42 issue 6 pp. 927–935)
**Venue:** Nature Biotechnology
**DOI:** 10.1038/s41587-023-01905-6
**PDF:** https://cs.stanford.edu/people/jure/pubs/gears-natbio23.pdf

## One-sentence summary

GEARS uses a gene–gene knowledge graph (GO) + deep learning to predict whole-transcriptome responses for genetic perturbations, including unseen 2-gene combinations never observed during training.

## Key contribution

First model to generalize perturbation response prediction to **novel multi-gene combinations**. Prior models (scGen, CPA) could transfer across cell types or dosages but couldn't predict what happens when you knock out two genes simultaneously if that combination wasn't in training data. GEARS does this by grounding the model in biological prior knowledge (Gene Ontology network).

## Methods

- **Architecture:** Graph Neural Network over a GO-derived gene–gene relationship graph + transformer for expression
- **Key idea:** Instead of learning gene interactions from scratch, use GO as structural prior; genes with known regulatory/pathway relationships get informed message passing
- **Training data:** Primarily Norman 2019 (CRISPRa, K562), Replogle Perturb-seq datasets
- **Prediction target:** Mean post-perturbation expression profile (differential expression from control)
- **Evaluation:** Held-out 2-gene combinations; measures correlation of predicted vs. actual fold-change

## Key findings

- Significantly outperforms CPA and scGen on unseen 2-gene perturbation prediction
- The GO graph prior is essential — ablating it degrades performance substantially
- Captures both additive interactions (most pairs) and non-additive genetic interactions
- Works across multiple datasets (Norman K562 CRISPRa, RPE1 CRISPRi)

## Limitations

- Predicts **mean** response; doesn't model cell-to-cell heterogeneity or distribution
- Limited to perturbations with known GO annotations; rare/uncharacterized genes underperform
- Trained on cell line data; cross-cell-type generalization not demonstrated
- 2-gene combinations only; 3+ gene interactions untested
- Still interpolates within the space spanned by training perturbations; true OOD generalization unclear

## Architecture insight

GEARS is the best argument for **biological prior knowledge in ML models**. The graph is not learned end-to-end — it's injected as structure. This is the opposite of the "learn everything from data" paradigm in foundation models. The fact that it works so well at 2-gene generalization suggests the field under-uses structured biological priors.

## Connections

- [../concepts/perturbation-biology.md](../concepts/perturbation-biology.md) — the primary application
- [../concepts/virtual-cell.md](../concepts/virtual-cell.md) — GEARS as a narrow but rigorous perturbation predictor
- [../papers/cpa.md](cpa.md) — main baseline model; CPA focuses on chemical perturbations
- [../papers/norman-2019.md](norman-2019.md) — primary training/eval dataset
- [../papers/replogle-2022.md](replogle-2022.md) — larger training dataset used
- [../papers/scgpt.md](scgpt.md) — scGPT also does perturbation prediction; different approach (pure data-driven)

## Bo's notes

GEARS is the field's clearest demonstration that structured biological priors improve generalization. The tension with foundation model thinking is productive: pure scaling (scGPT) vs. graph-grounded prediction (GEARS). Neither is clearly dominant. The real question is whether a foundation model trained on more diverse perturbation data can eventually match GEARS's combinatorial generalization without the graph prior.