Research map / Probabilistic and scientific modeling / Causal discovery
Gene Regulatory Network Inference in the Presence of Dropouts: a Causal View
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 a causal dropout model and a test-wise zero-deletion CI test that lets standard causal discovery (PC, GES) recover gene regulatory networks from scRNA-seq data with dropouts.
Problem
Gene regulatory network inference from scRNA-seq data is biased by dropouts (technical zeros), which distort the joint distribution and make imputation-based approaches unidentifiable and unreliable.
Approach
Model dropout with a causal graph (self-masking Z_i -> D_i) and show that CI relations in the data with dropouts, computed after deleting samples with zeros in the conditioned variables, are asymptotically identical to those in the dropout-free data; this test-wise deletion plugs into PC and GES.
Key results
- On BoolODE/BEELINE, zero-deletion beats full samples on 64 of 90 dataset-algorithm pairs (71%) and beats imputation on 61 of 90 (68%).
- Imputed and binarized strategies are worse than full samples on 50% and 64% of the 90 pairs, respectively.
- On linear SEM simulations with 30 nodes, the method consistently outperforms competitors and performs close to the oracle without dropouts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use test-wise deletion instead of full samples | GRNI algorithms on BoolODE and curated BEELINE datasets · 6 synthetic and 4 curated datasets | F1-score: better than 'full samples' on 64 (71%) of them | Helped |
| Use test-wise deletion instead of imputation | GRNI algorithms on BoolODE and curated BEELINE datasets · 6 synthetic and 4 curated datasets | F1-score: better than 'imputed' on 61 (68%) of them | Helped |
| Use full samples with dropouts instead of oracle data | GRNI algorithms on BoolODE and curated BEELINE datasets · 6 synthetic and 4 curated datasets | F1-score: on 65 of them (72%), 'full samples' ... performs worse than 'oracle' | Hurt |
| Impute dropout values instead of using full samples | GRNI algorithms on BoolODE and curated BEELINE datasets · 6 synthetic and 4 curated datasets | F1-score: 'imputed' ... is even worse than 'full samples' on 45 (50%) of them | Hurt |
| Binarize the expression data instead of using full samples | GRNI algorithms on BoolODE and curated BEELINE datasets · 6 synthetic and 4 curated datasets | F1-score: 'binary' is even worse than 'full samples' on 58 (64%) of them | Hurt |
| Use test-wise deletion instead of direct application to full samples | PC and GES on linear SEM simulated data · Linear SEM, 30 nodes | SHD | Helped |
| Use test-wise deletion instead of direct application to full samples | PC on Perturb-seq data · Perturb-seq BMDC, 21 genes | agreement with known causal relations | Helped |
Surprising findings
- Imputation and binarization often perform worse than simply using full samples with dropouts, i.e., they can be counterproductive.
- Under the self-masking dropout mechanism, the joint distribution of the true expressions is unidentifiable, even though CI relations remain recoverable.
Limitations
- Sample size shrinks after test-wise deletion, reducing statistical power.
- Some structural identifiability (dropout edges between adjacent genes) remains unidentifiable without assumption A3.
Experimental setup
Datasets: BoolODE synthetic data (6 datasets) BEELINE curated datasets (4) Perturb-seq bone-marrow dendritic cells (9843 cells, 21 TFs) linear SEM simulations
Benchmarks: BEELINE BoolODE simulated data Perturb-seq BMDC
Metrics: structural Hamming distance F1-score of skeleton edges
Baselines: MAGIC mixedCCA PC-full GES binarization oracle
Builds on: PC algorithm Greedy Equivalence Search Causal Dropout Model BEELINE BoolODE
Closest papers by meaning
- Gene Regulatory Network Inference in the Presence of Selection Bias and Latent Confounders (NeurIPS 2025 · similarity 0.90)
- Gene-Gene Relationship Modeling Based on Genetic Evidence for Single-Cell RNA-Seq Data Imputation (NeurIPS 2024 · similarity 0.88)
- A Generic Family of Graphical Models: Diversity, Efficiency, and Heterogeneity (ICML 2025 · similarity 0.87)
- KINDLE: Knowledge-Guided Distillation for Prior-Free Gene Regulatory Network Inference (NeurIPS 2025 · similarity 0.86)
- PAIR-CI: Calibrated Conditional Independence Testing for Causal Discovery with Incomplete Data (NeurIPS 2026 · similarity 0.85)
- Pruning neural network models for gene regulatory dynamics using data and domain knowledge (NeurIPS 2024 · similarity 0.84)
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.