Research map / Probabilistic and scientific modeling / Molecules and proteins
Cradle-VAE: Enhancing Single-Cell Gene Perturbation Modeling with Counterfactual Reasoning-based Artifact Disentanglement
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Cradle-VAE, a VAE for single-cell gene perturbation response prediction that models technical QC artifacts as a latent variable and uses a counterfactual KL alignment loss to disentangle them, improving treatment effect prediction and QC pass rate.
Problem
Predicting single-cell gene perturbation responses from Perturb-seq data, where technical artifacts (low-quality or doublet cells) in training and generated data degrade model predictability and output quality.
Approach
A causal VAE that separately encodes perturbation effects and a latent artifact embedding, and adds an auxiliary KL loss that aligns the artifact-present counterfactual basal state with a reference basal state built from QC-failed cells under the same perturbation, so artifacts are disentangled from the basal state and generated data passes QC more often.
Key results
- On Norman at 3σ, Cradle-VAE reaches ATE-ρ 0.7119 and QCPR 93.53%, versus best baseline Conditional VAE 0.5314 and 74.05% QCPR.
- On Replogle at 5σ, Cradle-VAE reaches ATE-ρ 0.7638 and QCPR 98.41%, versus SAMS-VAE 0.7167 ATE-ρ.
- Ablations show removing the counterfactual KL loss (w/o CF) or modeling artifacts as fixed embedding (w/o Causal) lowers ATE-ρ and QCPR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Cradle-VAE instead of the strongest baseline | Norman · Norman | ATE-ρ: 0.5814 ± 0.08 -> 0.7518 ± 0.03 | Helped |
| use Cradle-VAE instead of the strongest baseline | Dixit · Dixit | ATE-ρ: 0.3990 ± 0.04 -> 0.6572 ± 0.03 | Helped |
| use Cradle-VAE instead of the strongest baseline | Replogle · Replogle | ATE-ρ: 0.7296 ± 0.01 -> 0.7638 ± 0.01 | Helped |
| use Cradle-VAE instead of the strongest baseline | Adamson · Adamson | ATE-ρ: 0.6484 ± 0.01 -> 0.7636 ± 0.01 | Helped |
| use Cradle-VAE instead of the strongest baseline | Norman · Norman | QCPR (%): 87.19 ± 0.71 -> 96.62 ± 0.38 | Helped |
| use Cradle-VAE instead of the strongest baseline | Dixit · Dixit | QCPR (%): 83.36 ± 0.50 -> 91.40 ± 1.80 | Helped |
| use Cradle-VAE instead of the strongest baseline | Replogle · Replogle | QCPR (%): 88.45 ± 0.27 -> 98.41 ± 0.13 | Helped |
| use Cradle-VAE instead of the strongest baseline | Adamson · Adamson | QCPR (%): 89.05 ± 0.49 -> 94.34 ± 0.42 | Helped |
| add the counterfactual KL auxiliary loss (ablation) | Cradle-VAE · Norman, 3σ | ATE-ρ: 0.6505 ± 0.02 -> 0.7119 ± 0.03 | Helped |
| model artifacts with a parameterized prior distribution instead of a fixed learnable embedding (ablation) | Cradle-VAE · Norman, 3σ | ATE-ρ: 0.7018 ± 0.02 -> 0.7119 ± 0.03 | Helped |
| add the counterfactual KL auxiliary loss (ablation) | Cradle-VAE · Norman, 4σ | ATE-ρ: 0.7111 ± 0.03 -> 0.7477 ± 0.03 | Helped |
| model artifacts with a parameterized prior distribution instead of a fixed learnable embedding (ablation) | Cradle-VAE · Norman, 4σ | ATE-ρ: 0.7058 ± 0.03 -> 0.7477 ± 0.03 | Helped |
Surprising findings
- Removing the counterfactual loss hurts most at the strictest 3σ threshold, while the embedding-based artifact ablation hurts most at 5σ.
Limitations
- Gains are reported at multiple QC thresholds but QC pass rate is a self-defined metric tied to the same QC criteria used for labeling artifacts.
- Multi-gene perturbation test sets are constructed by the authors, and the method is evaluated only on K562 cell line datasets.
- Baselines are retrained with five seeds, but the paper does not report compute resources.
Experimental setup
Datasets: Norman Dixit Replogle Adamson
Benchmarks: Norman Dixit Replogle Adamson
Metrics: ATE-rho ATE-R2 Jaccard top-50 DEGs QC Pass Rate
Baselines: Conditional VAE CPA-VAE sVAE+ SAMS-VAE
Builds on: SAMS-VAE CPA sVAE+
Closest papers by meaning
- PerturbReason: A Knowledge-Grounded Benchmark and Framework for Cell-State–Conditioned Mechanistic Reasoning of Perturbation Effects (NeurIPS 2026 · similarity 0.87)
- Generative Intervention Models for Causal Perturbation Modeling (ICML 2025 · similarity 0.87)
- PerturBench: Benchmarking Machine Learning Models for Cellular Perturbation Analysis (NeurIPS 2025 · similarity 0.87)
- Learning Identifiable Factorized Causal Representations of Cellular Responses (NeurIPS 2024 · similarity 0.86)
- PRESCRIBE: Predicting Single-Cell Responses with Bayesian Estimation (NeurIPS 2025 · similarity 0.86)
- Interpretable Causal Representation Learning for Biological Data in the Pathway Space (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.