Research map / Probabilistic and scientific modeling / Causal discovery
Interpretable Causal Representation Learning for Biological Data in the Pathway Space
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 SENA-discrepancy-VAE, which adds a pathway-masked sparse encoder so causal latent factors in a discrepancy-VAE map to gene-set biological processes, keeping predictive performance on Perturb-seq double perturbations.
Problem
Causal representation learning models for Perturb-seq data produce latent factors that cannot be reconciled with known biological processes, so they are hard to interpret, even though they can predict unseen perturbation combinations.
Approach
Replace the discrepancy-VAE's fully connected encoder with a two-layer masked encoder (SENA-delta) in which gene-to-biological-process weights are masked by GO gene sets (with a leak parameter lambda), so each causal latent factor is a linear combination of pathway activities while retaining the discrepancy-VAE identifiability assumptions.
Key results
- On Norman2019, SENA-delta (lambda=0.1) achieved MMD 1.58489 vs 1.59811 for the original MLP encoder at latent dimension 105, i.e. comparable or slightly better.
- GEARS had much higher MMD (12.6 to 15.3) than both discrepancy-VAE variants on double-perturbation prediction.
- On Wessels2023, SENA-discrepancy-VAE is comparable to the original only at higher lambda (e.g. lambda=0.5 with MMD 0.1753 at latent dim 14 vs 0.1886 for MLP).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the original MLP encoder with SENA-δ (λ=0.1) | discrepancy-VAE · Norman2019 double perturbations, 105 latent dimensions | MMD: 1.59811 -> 1.58489 | Helped |
| Replace the original MLP encoder with SENA-δ (λ=0.1) | discrepancy-VAE · Norman2019 control samples | MSE: 0.02152 -> 0.02134 | Helped |
| Set the SENA-δ mask parameter λ to 0 (ablation) | SENA-discrepancy-VAE · Norman2019 double perturbations, 105 latent dimensions | KLD: 0.00022 -> 0.00018 | Helped |
| Set the SENA-δ mask parameter λ to 0 (ablation) | SENA-discrepancy-VAE · Norman2019 double perturbations, 105 latent dimensions | MMD: 1.59811 -> 1.74588 | Hurt |
| Replace the original MLP encoder with SENA-δ (λ=0.5) | discrepancy-VAE · Wessels2023 double perturbations, 50 latent dimensions | MMD: 0.1810 -> 0.2285 | Hurt |
| Set the SENA-δ mask parameter λ to 0 (ablation) | SENA-discrepancy-VAE · Wessels2023 double perturbations, 50 latent dimensions | MMD: 0.1810 -> 0.5675 | Hurt |
| Use SENA-δ encoder with λ=0.1 in an AE instead of a fully connected MLP encoder (ablation) | AE · Norman2019 test set | MSE: 1.350 ± 0.029 -> $\mathbf{1.012 \pm 0.000}$ | Helped |
| Use SENA-δ encoder with λ=0.001 in a VAE instead of a fully connected MLP encoder (ablation) | VAE · Norman2019 test set | KLD: 3.816 ± 1.331 -> $\mathbf{0.001 \pm 0.001}$ | Helped |
| Use GEARS for double-perturbation prediction instead of discrepancy-VAE | Norman2019 · Norman2019 double perturbations, 105 latent dimensions | MMD: 1.59811 -> 14.9420 | Hurt |
Surprising findings
- Both models assign most interventions to a small number of latent factors, which the authors attribute to the single-target assumption of the CRL theory.
- On Wessels2023, all single-gene perturbations are assigned to a single latent factor by both models.
Limitations
- Assuming each intervention targets a single latent factor forces overlapping biological processes into the same factor, limiting disentanglement.
- Predictive performance on the Wessels2023 dataset is only comparable at higher lambda values, showing an interpretability-accuracy trade-off.
- Encoder is deliberately simple (two linear-type layers) to prioritize interpretability over representational capacity.
Experimental setup
Datasets: Norman2019 (K562 Perturb-seq) Wessels2023 (THP1 CaRPool-seq) Gene Ontology biological processes
Benchmarks: Norman2019 double-perturbation prediction Wessels2023 double-perturbation prediction
Metrics: MMD MSE KLD L1 sparsity of causal graph Hits@100 differential activation ratio (DAR)
Baselines: discrepancy-VAE (MLP encoder) GEARS l1-regularized MLP encoder
Builds on: discrepancy-VAE GEARS
Closest papers by meaning
- Learning Identifiable Factorized Causal Representations of Cellular Responses (NeurIPS 2024 · similarity 0.89)
- Identifying biological perturbation targets through causal differential networks (ICML 2025 · similarity 0.89)
- Causal Representation Learning from Multimodal Biomedical Observations (ICLR 2025 · similarity 0.86)
- Causal Representation Learning Made Identifiable by Grouping of Observational Variables (ICML 2024 · similarity 0.86)
- MapPFN: Learning Causal Perturbation Maps in Context (NeurIPS 2026 · similarity 0.86)
- Generative Intervention Models for Causal Perturbation Modeling (ICML 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.