Skip to content

Perturbation Datasets

Datasets for training perturbation prediction models (scPPDM, JEPA) on CRISPR screening and Perturb-seq data.

Overview

Perturbation datasets capture cellular responses to genetic interventions (CRISPR knockouts, knockdowns, overexpression). These are essential for:

  • scPPDM: Single-cell Perturbation Prediction via Diffusion Models
  • JEPA: Joint Embedding Predictive Architecture for perturbation response
  • Counterfactual generation: Predicting "what if" scenarios

Key Datasets

Dataset Cells Perturbations Cell Type Document
scPerturb Multiple Multiple Various scperturb.md
Replogle 2022 ~2.5M >5,000 K562 scperturb.md
Norman 2019 ~100k ~300 (combinatorial) K562 scperturb.md
Adamson 2016 ~10k ~100 K562 scperturb.md

Data Structure

Perturb-seq data typically includes:

AnnData object:
├── X: Expression matrix (cells × genes)
├── obs: Cell metadata
│   ├── perturbation: Gene targeted
│   ├── is_control: Boolean (NT control)
│   └── cell_type, batch, etc.
└── var: Gene metadata
    └── gene_name, highly_variable, etc.

Key Fields

  • Control cells: Non-targeting (NT) guide, baseline expression
  • Perturbed cells: Expression after intervention
  • Perturbation label: Which gene was targeted

Use Cases in genai-lab

1. scPPDM (Diffusion-based)

Train diffusion model to predict post-perturbation expression:

Control expression + Perturbation embedding → Diffusion → Perturbed expression

2. JEPA (Embedding-based)

Predict perturbation effects in latent space:

Control latent + Perturbation → Predictor → Perturbed latent

3. Counterfactual Generation

Generate "what if" scenarios:

  • What if gene X was knocked out in this cell?
  • What is the distribution of possible outcomes?

Documents