Research map / Probabilistic and scientific modeling / Causal discovery
Learning Identifiable Factorized Causal Representations of Cellular Responses
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 Factorized Causal Representation (FCR), an identifiable VAE-based model that splits cell latents into covariate-, treatment-, and interaction-specific blocks, with identifiability proofs and adversarial regularizers, evaluated on four single-cell perturbation datasets.
Problem
Modeling single-cell perturbation responses that depend on cellular context, where existing identifiable or latent additive models either handle only a single context or cannot capture treatment-covariate interactions.
Approach
Factorizes the latent space into covariate-specific, treatment-specific and interaction-specific blocks with a product-based interaction prior, proves component-wise identifiability of the interaction block and block-wise identifiability of the other two, and enforces conditional independence via permutation discriminators and a causal structure regularizer.
Key results
- On conditional cellular response prediction, FCR achieves R^2 0.87 on sciPlex, 0.90 on multiPlex-Tram and 0.83 on multiPlex-7, outperforming VCI, CPA, scGEN, sVAE and CINEMA-OT.
- On multiPlex-9, CPA (0.79) slightly outperforms FCR (0.78), which the authors attribute to fewer covariate-specific responses in that dataset.
- Baseline latent representations fail the conditional independence tests (p<0.05) for z_x independent of t given x, while FCR's permutation discriminators yield p-values generally above 0.05.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FCR instead of VCI for conditional cellular response prediction | Conditional cellular response prediction · sciPlex | R²: 0.86 ± 0.03 -> 0.87 ± 0.02 | Helped |
| Use FCR instead of VCI for conditional cellular response prediction | Conditional cellular response prediction · multiPlex-Tram | R²: 0.89 ± 0.04 -> 0.90 ± 0.03 | Helped |
| Use FCR instead of VCI for conditional cellular response prediction | Conditional cellular response prediction · multiPlex-7 | R²: 0.81 ± 0.04 -> 0.83 ± 0.03 | Helped |
| Use FCR instead of CPA for conditional cellular response prediction | Conditional cellular response prediction · multiPlex-9 | R²: 0.79 ± 0.02 -> 0.78 ± 0.03 | Hurt |
Surprising findings
- CPA outperforms FCR on multiPlex-9, attributed to the dataset having fewer covariate-specific responses.
- On sciPlex, clustering on covariates alone yields higher NMI than clustering on treatment-covariate pairs or treatments alone, attributed to HDAC inhibitors having minimal effect on cell lines below 10 uM.
Limitations
- Current implementation does not use general embeddings for treatments or covariates, so it cannot predict responses to unseen drugs or cell types.
- Generative model assumes deterministic mixing with no observation noise in the main method.
- Identifiability theory requires strong experimental sufficiency (at least 2n_tx+1 covariate-treatment settings with switched controls).
Experimental setup
Datasets: sciPlex multiPlex-Tram multiPlex-7 multiPlex-9
Benchmarks: sciPlex multiPlex-Tram multiPlex-7 multiPlex-9
Metrics: NMI p-value of KCI conditional independence test HSIC R^2
Baselines: scVI iVAE betaVAE factorVAE VCI CPA scGEN sVAE CINEMA-OT
Builds on: iVAE CPA VCI nonlinear ICA identifiability
Closest papers by meaning
- Interpretable Causal Representation Learning for Biological Data in the Pathway Space (ICLR 2025 · similarity 0.89)
- Identifying biological perturbation targets through causal differential networks (ICML 2025 · similarity 0.88)
- Generative Intervention Models for Causal Perturbation Modeling (ICML 2025 · similarity 0.88)
- Learning Cross-Domain Representations for Transferable Drug Perturbations on Single-Cell Transcriptional Responses (AAAI 2025 · similarity 0.87)
- MapPFN: Learning Causal Perturbation Maps in Context (NeurIPS 2026 · similarity 0.86)
- Causal Representation Learning from Multimodal Biomedical Observations (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.