Research map / Probabilistic and scientific modeling / Causal discovery
Generative Intervention Models for Causal Perturbation Modeling
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 generative intervention models (GIMs) that map perturbation features to distributions over atomic interventions in a jointly learned causal model, enabling prediction of unseen perturbation effects with mechanistic interpretability.
Problem
Predicting effects of perturbations with unseen features when the atomic causal interventions they induce are unknown, a gap for causal models that cannot handle novel non-atomic perturbations such as drugs.
Approach
Learn a generative model that maps perturbation features to Bernoulli intervention-target probabilities and intervention parameters in a jointly estimated SCM, marginalizing over interventions so that unseen perturbation features yield predictive distributions; MAP inference via gradient-based optimization with a continuous acyclicity-constrained graph representation.
Key results
- On nonlinear SCMs, GIMs significantly outperform baselines (including BaCaDI*) on structure learning metrics such as SID and Edge-F1.
- GIMs achieve Target-F1 comparable in partially OOD settings and outperform all causal baselines on intervention target prediction except UT-IGSP in linear settings.
- On SciPlex3 held-out highest dosage, GIMs and the MLP baseline give the most accurate predictions by Mean Distance and Pearson correlation, and GIMs are the only mechanistic model applicable to unseen perturbation features.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GIM instead of causal discovery baselines | Nonlinear SCMs · Synthetic nonlinear SCMs | causal structure accuracy | Helped |
| Use GIM instead of BaCaDI* (non-amortized point estimate) (ablation) | Nonlinear SCM causal discovery · Synthetic nonlinear SCMs | causal structure accuracy | Helped |
| Include perturbation features as context nodes without modeling their mechanisms | JCI-PC · Synthetic nonlinear SCMs | causal structure accuracy | No clear effect |
| Use GIM instead of causal baselines | Intervention target prediction · Synthetic perturbations, in-distribution and partially OOD | Target-F1 | Helped |
| Train GIM to map perturbation features to interventions and predict unseen feature settings | GIM · Synthetic partially OOD perturbations | Target-F1 | No clear effect |
| Train GIM to map perturbation features to interventions and predict unseen feature settings | GIM · Synthetic fully OOD perturbations | Target-F1 | Mixed |
| Use GIM instead of black-box and causal baselines | Perturbation distribution prediction · Synthetic in-distribution perturbations | distribution prediction accuracy | Mixed |
| Use GIM instead of black-box baselines | Perturbation distribution prediction · Synthetic partially OOD perturbations | distribution prediction accuracy | Mixed |
| Use GIM instead of black-box baselines | Perturbation distribution prediction · Synthetic fully OOD perturbations | distribution prediction accuracy | Mixed |
| Use GIM instead of CondOT | Perturbation distribution prediction · Synthetic fully OOD perturbations | distribution prediction accuracy | Helped |
| Use perturbation features with sufficient information about intervention effects | GIM · Synthetic fully OOD perturbations | predictive performance | Helped |
| Use GIM instead of the naive next-lower-dosage prediction | SciPlex3 drug perturbation prediction · SciPlex3 held-out highest drug dosages | Mean Distance | Helped |
Surprising findings
- Including perturbation features as context nodes in JCI-PC does not improve causal discovery over plain JCI-PC.
- On in-distribution perturbations, the black-box MLP consistently performs best at predicting seen perturbation effects, while GIMs tend to be better on fully OOD interventions.
Limitations
- Identifiability for unseen perturbations depends on how informative the perturbation features are.
- MAP approximation is a crude posterior approximation rather than full Bayesian inference.
- Linear settings: classical methods (GnIES, UT-IGSP) outperform GIMs on structure learning; MLP is best in-distribution.
Experimental setup
Datasets: synthetic linear and nonlinear SCMs (ER and SF graphs, d=20) SciPlex3
Benchmarks: synthetic linear and nonlinear SCMs SciPlex3
Metrics: SID Edge-F1 Target-F1 W2 Mean distance KDE-NLL Pearson correlation
Baselines: BaCaDI GnIES UT-IGSP JCI-PC JCI-PC Context CondOT MLP
Builds on: BaCaDI NOTEARS CondOT JCI-PC SID
Closest papers by meaning
- Identifying biological perturbation targets through causal differential networks (ICML 2025 · similarity 0.89)
- MapPFN: Learning Causal Perturbation Maps in Context (NeurIPS 2026 · similarity 0.89)
- Learning Identifiable Factorized Causal Representations of Cellular Responses (NeurIPS 2024 · similarity 0.88)
- Cradle-VAE: Enhancing Single-Cell Gene Perturbation Modeling with Counterfactual Reasoning-based Artifact Disentanglement (AAAI 2025 · similarity 0.87)
- Improving the Generation and Evaluation of Synthetic Data for Downstream Medical Causal Inference (NeurIPS 2025 · similarity 0.87)
- Coupling Generative Modeling and an Autoencoder with the Causal Bridge (NeurIPS 2025 · similarity 0.87)
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.