Research map / Probabilistic and scientific modeling / Causal discovery
Gene Regulatory Network Inference in the Presence of Selection Bias and Latent Confounders
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 GISL, a nonparametric causal discovery algorithm that uses observational and single-gene perturbation data to distinguish regulation, latent confounding, and selection bias in gene regulatory network inference.
Problem
Gene regulatory network inference from expression data is confounded by spurious dependencies from latent confounders and selection bias (survival criteria), which standard GRNI and causal discovery methods do not handle.
Approach
Selection-induced dependencies are symmetric under perturbation whereas regulation and confounding are not; GISL uses CI patterns between perturbation indicators and gene variables (with d-separation correction) to classify each gene pair as regulatory, confounded, selection, or unknown, outputting a PAG-style equivalence class.
Key results
- On synthetic data with 10-25 variables, GISL identifies selection pairs with roughly 60-80% accuracy under hard and soft interventions (e.g., ~75% at n=2000, |X|=25 hard).
- GISL achieves 94.7% accuracy and SHD 1.0 versus PIDC (51.6%) and PPCOR (43.4%) on synthetic selection-biased data.
- On real data, GISL's selected selection pairs reach 77.7% (Dixit), 80.2% (Adamson), and 74.8% (Norman) agreement with a Z-score threshold of 0.15.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GISL rather than PIDC for synthetic regulatory relationships | Synthetic regulatory relationships · Synthetic data | F1: 55.1±0.03 -> 94.9±0.01 | Helped |
| Use GISL rather than PPCOR for synthetic regulatory relationships | Synthetic regulatory relationships · Synthetic data | Accuracy: 43.4±10.32 -> 94.7±0.01 | Helped |
| Use GISL rather than CDIS | Synthetic regulatory relationships · |X|=15, n=1500, soft intervention | F1: 63.8 ± 4.5 -> 80.0 ± 1.7 | Helped |
| Use GISL rather than CDIS | Synthetic regulatory relationships · |X|=15, n=1500, soft intervention | Precision (selection bias): 34.1 ± 9.1 -> 80.8 ± 11.3 | Helped |
| Use GISL rather than PIDC | Synthetic regulatory relationships · Synthetic data | Recall: 95.2±0 -> 95.1±0.01 | No clear effect |
| Increase Gaussian-noise mean range to μ~U(0,4) (ablation) | GISL · Synthetic regulatory relationships, n=1500 | F1: 69.6 ± 1.9 -> 72.6 ± 1.8 | Helped |
| Increase Gaussian-noise mean range to μ~U(2,4) (ablation) | GISL · Synthetic regulatory relationships, n=1500 | F1: 69.6 ± 1.9 -> 77.4 ± 0.8 | Helped |
| Increase edge count to 20 (ablation) | GISL · Synthetic regulatory relationships, n=1500 | F1: 69.6 ± 1.9 -> 67.6 ± 0.6 | Hurt |
| Increase variable count to |X|=25 (ablation) | GISL · Synthetic regulatory relationships, n=1500 | F1: 69.6 ± 1.9 -> 67.3 ± 0.3 | Hurt |
| Increase variable count to |X|=50 (ablation) | GISL · Synthetic regulatory relationships, n=1500 | F1: 69.6 ± 1.9 -> 71.2 ± 1.6 | Mixed |
| Use normalized-Hill differential-equation functions instead of random nonlinear functions (ablation) | GISL · Synthetic regulatory relationships, n=1500 | F1: 69.6 ± 1.9 -> 67.6 & \pm{\color{gray}{1.7}} | Hurt |
Surprising findings
- Ignoring selection bias causes existing causal and computational GRNI methods to output near fully connected graphs, as selection induces dependencies among variables linked through causal pathways.
- CDIS, which assumes selection is inactive after intervention, performs poorly (precision 61.4%) when biological survival constraints persist after perturbation.
Limitations
- Restricted to acyclic graphs; cycles are left for future work.
- Relies on CI tests (KCI) with faithfulness and Markov assumptions; accuracy drops as number of selection processes grows.
- Real-data selection validation uses Z-scores as a proxy rather than ground truth.
Experimental setup
Datasets: Dixit perturb-seq (K562) Adamson perturb-seq (K562) Norman perturb-seq (A549) synthetic SCM data
Benchmarks: Dixit Adamson Norman synthetic SCM data
Metrics: DAG F1 DAG precision DAG recall SHD SID selection accuracy Z-score agreement accuracy
Baselines: GIES IGSP UT-IGSP JCI-GSP FCI ICD PIDC PPCOR CDIS
Builds on: FCI FGES augmented DAG IGSP JCI
Closest papers by meaning
- Gene Regulatory Network Inference in the Presence of Dropouts: a Causal View (ICLR 2024 · similarity 0.90)
- KINDLE: Knowledge-Guided Distillation for Prior-Free Gene Regulatory Network Inference (NeurIPS 2025 · similarity 0.88)
- Identifying biological perturbation targets through causal differential networks (ICML 2025 · similarity 0.85)
- A Generic Family of Graphical Models: Diversity, Efficiency, and Heterogeneity (ICML 2025 · similarity 0.85)
- MapPFN: Learning Causal Perturbation Maps in Context (NeurIPS 2026 · similarity 0.85)
- When Selection Meets Intervention: Additional Complexities in Causal Discovery (ICLR 2025 · similarity 0.85)
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.