Research map / Probabilistic and scientific modeling / Causal discovery

Identifying biological perturbation targets through causal differential networks

ICML 2025 · Probabilistic and scientific modeling › Causal discovery · approach: Interventional causal discovery · 1 citations (OpenAlex)

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 Causal Differential Networks (CDN), an amortized supervised model that infers noisy causal graphs from observational and interventional datasets and classifies intervention targets from their graph differences, outperforming perturbation-modeling and causal discovery baselines.

Problem

Identifying which observed variables (genes or compounds' targets) were intervened upon, given observational and interventional datasets with thousands of variables and only tens of interventional samples, where biological systems violate classical causal assumptions.

Approach

Jointly train an attention-based causal structure learner (built on SEA, with FCI local estimates and pairwise correlation features) and a differential network that maps the difference between observational and interventional graph representations, plus node-level statistics, to intervention targets, pretraining on simulated biological-style interventions and finetuning on real Perturb-seq data.

Key results

• On seen cell lines, CDN (Perturb-seq) achieves normalized rank 0.92-0.98 across five Perturb-seq datasets vs. 0.41-0.56 for the best baselines.
• On synthetic hard-intervention linear (1 target) N=20 data, CDN reaches mAP 0.82 and AUC 0.96, versus 0.58 mAP for the best baseline DCI.
• CDN finds the true target within the top 100 candidates in 3 of 6 chemical perturbation cases (Sci-Plex).

Measured effects reported in the paper

Surprising findings

• No other deep learning method exceeds a simple Wilcoxon differential-expression test (DGE) on the Perturb-seq benchmark.
• Most existing causal discovery baselines (DCDI, BaCaDI, UT-IGSP) perform poorly at recovering intervention targets on synthetic data.

Limitations

• Does not handle cyclic feedback or dynamic systems.
• Assumes causal sufficiency (no unseen confounders), which is unrealistic in biology.
• Domain shift between simulated pretraining data and real single-cell data.

Experimental setup

Datasets: Perturb-seq K562 genome-wide Perturb-seq K562 essential Perturb-seq RPE1 Perturb-seq HepG2 Perturb-seq Jurkat Sci-Plex A172 Sci-Plex T98G synthetic linear and neural-network SCMs

Benchmarks: Perturb-seq K562 (gw) Perturb-seq K562 (es) Perturb-seq RPE1 Perturb-seq HepG2 Perturb-seq Jurkat Sci-Plex A172 Sci-Plex T98G synthetic ER graphs N=10,20

Metrics: normalized rank Spearman correlation Pearson correlation recall at p mAP AUC

Baselines: GEARS GenePt PDGrapher DGE Linear MLP UT-IGSP DCDI BaCaDI DCI MB+CI

Builds on: SEA FCI DCI GEARS axial attention

Closest papers by meaning

• MapPFN: Learning Causal Perturbation Maps in Context (NeurIPS 2026 · similarity 0.90)
• Generative Intervention Models for Causal Perturbation Modeling (ICML 2025 · similarity 0.89)
• Interpretable Causal Representation Learning for Biological Data in the Pathway Space (ICLR 2025 · similarity 0.89)
• Differentiable Cyclic Causal Discovery Under Unmeasured Confounders (NeurIPS 2025 · similarity 0.88)
• Learning Identifiable Factorized Causal Representations of Cellular Responses (NeurIPS 2024 · similarity 0.88)
• Scalable and Flexible Causal Discovery with an Efficient Test for Adjacency (ICML 2024 · similarity 0.88)