Research map / Probabilistic and scientific modeling / Molecules and proteins
Pruning neural network models for gene regulatory dynamics using data and domain knowledge
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 DASH, a domain-aware iterative pruning method that blends learned weight magnitudes with prior gene-regulatory knowledge to yield sparse, biologically meaningful neural ODE models of gene regulatory dynamics.
Problem
Pruned neural networks for gene regulatory dynamics are sparse but fail to recover biologically meaningful regulatory structure; existing pruning ignores available domain knowledge.
Approach
Score weights for iterative magnitude pruning as a convex combination of learned absolute weights and domain prior matrices (input-output priors P and input-input priors C), with pseudo-inverse projections propagating priors through layers so pruning aligns with known biology while staying data-driven.
Key results
- On SIM350 with 5% noise, DASH reaches 94.6% sparsity and 90.7% balanced accuracy versus 81.9% sparsity and 61.7% balanced accuracy for IMP.
- On breast cancer data, DASH achieves 92.71% sparsity and 95.69% balanced accuracy, with test MSE 7.76e-5, comparable to BioPrune but better than other pruning baselines that score near 50% balanced accuracy.
- On yeast cell cycle data, DASH attains 97.18% sparsity and 88.43% balanced accuracy, versus 79.23% for BioPrune.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use domain-aware pruning rather than fully prior-based pruning | PHOENIX with DASH · SIM350, 5% noise | Bal. Acc.(%): 87.3 \pm 0.8 -> 90.7 \pm 0.4 | Helped |
| Use domain-aware pruning rather than fully prior-based pruning | PHOENIX with DASH · Breast cancer in pseudotime | Bal. Acc.: 95.67% -> 95.69% | No clear effect |
| Use domain-aware pruning rather than fully prior-based pruning | PHOENIX with DASH · Yeast cell cycle | Bal. Acc.: 79.23% -> 88.43% | Helped |
| Use domain-aware pruning rather than fully prior-based pruning | PHOENIX with DASH · Erythroid lineage, human bone marrow | Sparsity: 87.09% -> 95.95% | Helped |
| Corrupt prior information by 20% rather than use an uncorrupted prior (ablation) | DASH · SIM350, 0% expression noise | Bal. Acc.: 91.1 -> 86.2 | Hurt |
| Corrupt prior information by 40% rather than use an uncorrupted prior (ablation) | DASH · SIM350, 0% expression noise | Bal. Acc.: 91.1 -> 79.5 | Hurt |
| Corrupt prior information by 20% rather than use an uncorrupted prior (ablation) | BioPrune · SIM350, 0% expression noise | Bal. Acc.: 88.0 -> 81.5 | Hurt |
| Corrupt prior information by 40% rather than use an uncorrupted prior (ablation) | BioPrune · SIM350, 0% expression noise | Bal. Acc.: 88.0 -> 80.1 | Hurt |
| Apply DASH instead of no pruning | PHOENIX · SIM350, 5% noise | Bal Acc: 51.8% -> 90.7% | Helped |
| Apply DASH instead of no pruning | PHOENIX · Erythroid lineage, human bone marrow | Test MSE ($10^{-4}$): 2.12 -> 2.12 | No clear effect |
Surprising findings
- Prior-only pruning (BioPrune) performs nearly as well as DASH on breast cancer data, suggesting DASH relies mainly on the prior there.
- Implicit penalty methods and data-only pruning baselines recover gene regulatory relationships no better than random chance (around 50% balanced accuracy) on real data.
Limitations
- Performance depends on prior quality; wrong or noisy priors can steer the solution to incorrect models (sensitivity analysis shows degradation at 20-40% corruption).
- Propagation of priors is derived for L=1 and L=2 layers; extension to deeper networks and to convolution or attention architectures is not established.
- No gold-standard ground truth for the bone marrow data, so evaluation relies on out-degree and pathway analysis.
Experimental setup
Datasets: SIM350 SIM690 breast cancer pseudotime single-cell data yeast synchronized cell cycle data human bone marrow hematopoiesis data STRING database GRAND breast motif map ChIP-seq TF-gene binding
Benchmarks: SIM350 SIM690 breast cancer pseudotime yeast cell cycle human bone marrow Erythroid lineage
Metrics: sparsity balanced accuracy test MSE out-degree pathway enrichment z-score
Baselines: Magnitude pruning (PINN+MP) BioPrune IMP Iter. SynFlow SparseFlow L0 regularization C-NODE PathReg DST PINN
Builds on: PHOENIX Iterative magnitude pruning SparseFlow PathReg C-NODE
Closest papers by meaning
- KINDLE: Knowledge-Guided Distillation for Prior-Free Gene Regulatory Network Inference (NeurIPS 2025 · similarity 0.87)
- InfoSEM: A Deep Generative Model with Informative Priors for Gene Regulatory Network Inference (ICML 2025 · similarity 0.85)
- Gene Regulatory Network Inference in the Presence of Dropouts: a Causal View (ICLR 2024 · similarity 0.84)
- A Generic Family of Graphical Models: Diversity, Efficiency, and Heterogeneity (ICML 2025 · similarity 0.84)
- MapPFN: Learning Causal Perturbation Maps in Context (NeurIPS 2026 · similarity 0.84)
- Gene Regulatory Network Inference in the Presence of Selection Bias and Latent Confounders (NeurIPS 2025 · 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.