Research map / Probabilistic and scientific modeling / Molecules and proteins
KINDLE: Knowledge-Guided Distillation for Prior-Free Gene Regulatory Network Inference
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.
KINDLE trains a prior-informed spatio-temporal attention teacher for gene expression forecasting and distills it into a prior-free student that infers GRNs from expression data alone.
Problem
Gene regulatory network inference from single-cell expression depends on prior knowledge (e.g., ATAC/Hi-C-derived candidate edges), which limits accuracy to the prior's overlap with ground truth and prevents discovery of interactions absent from the prior.
Approach
Use learning with privileged information: a teacher with a prior-derived spatial attention mask and temporal causal attention is distilled (via hard, soft, bilinear, or Gaussian RBF correlation losses) into a student that takes only expression time series and reads the GRN from its averaged spatial attention matrix.
Key results
- KINDLE-Gaussian RBF reaches AUROC 0.757, AUPRC 0.646, F1 0.529 on mESC, versus best baseline CEFCON AUPRC 0.253 and F1 0.429.
- KINDLE-Gaussian achieves best performance in 11 of 12 dataset-metric combinations across the four BEELINE datasets.
- On mHSC-GM, KINDLE-Gaussian AUPRC is 0.799 versus 0.444 for CEFCON; the soft distillation variant beats CellOracle by 0.546 AUPRC there.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use prior-free KINDLE with Gaussian RBF distillation instead of the strongest baseline | KINDLE · mESC | AUPRC: 0.253 -> 0.646 | Helped |
| use prior-free KINDLE with Gaussian RBF distillation instead of the strongest baseline | KINDLE · mHSC-E | AUPRC: 0.405 -> 0.601 | Helped |
| use prior-free KINDLE with Gaussian RBF distillation instead of the strongest baseline | KINDLE · mHSC-L | AUPRC: 0.659 -> 0.672 | Helped |
| use prior-free KINDLE with Gaussian RBF distillation instead of the strongest baseline | KINDLE · mHSC-GM | AUPRC: 0.444 -> 0.799 | Helped |
| replace Gaussian RBF with soft distillation (ablation) | KINDLE · mESC | AUPRC: 0.646 -> 0.636 | Hurt |
| replace Gaussian RBF with soft distillation (ablation) | KINDLE · mHSC-E | AUPRC: 0.601 -> 0.559 | Hurt |
| replace Gaussian RBF with soft distillation (ablation) | KINDLE · mHSC-L | AUPRC: 0.672 -> 0.670 | No clear effect |
| replace Gaussian RBF with soft distillation (ablation) | KINDLE · mHSC-GM | AUPRC: 0.799 -> 0.789 | Hurt |
| replace Gaussian RBF with hard distillation (ablation) | KINDLE · mESC | AUPRC: 0.646 -> 0.643 | Hurt |
| replace Gaussian RBF with hard distillation (ablation) | KINDLE · mHSC-E | AUPRC: 0.601 -> 0.578 | Hurt |
| replace Gaussian RBF with hard distillation (ablation) | KINDLE · mHSC-L | AUPRC: 0.672 -> 0.669 | No clear effect |
| replace Gaussian RBF with hard distillation (ablation) | KINDLE · mHSC-GM | AUPRC: 0.799 -> 0.793 | Hurt |
Surprising findings
- The prior-free student outperforms prior-guided baselines, including Prior_Random, on AUPRC and F1 in all four datasets.
- CEFCON achieves the best AUROC on mHSC-L (0.653), where KINDLE variants (about 0.60) are lower.
Limitations
- Requires temporal (longitudinal) expression data, limiting applicability to datasets without time series.
- Distillation may inherit biases and errors from the prior-dependent teacher.
- Only models transcriptional regulation, omitting post-transcriptional and epigenetic layers.
Experimental setup
Datasets: BEELINE mESC BEELINE mHSC-E BEELINE mHSC-GM BEELINE mHSC-L ChIP-seq-derived cell-type-specific ground truth networks (ENCODE, ChIP-Atlas, ESCAPE)
Benchmarks: BEELINE mESC BEELINE mHSC-E BEELINE mHSC-GM BEELINE mHSC-L
Metrics: AUROC AUPRC F1 adjusted p-value (ANOVA on AUCell scores) perturbation score
Baselines: GENIE3 GRNBoost2 CEFCON CellOracle NetREX Random Prior_Random
Builds on: learning using privileged information CellOracle BEELINE SCENIC AUCell TRIGON
Closest papers by meaning
- Gene Regulatory Network Inference in the Presence of Selection Bias and Latent Confounders (NeurIPS 2025 · similarity 0.88)
- InfoSEM: A Deep Generative Model with Informative Priors for Gene Regulatory Network Inference (ICML 2025 · similarity 0.87)
- Pruning neural network models for gene regulatory dynamics using data and domain knowledge (NeurIPS 2024 · similarity 0.87)
- Gene Regulatory Network Inference in the Presence of Dropouts: a Causal View (ICLR 2024 · similarity 0.86)
- GeSubNet: Gene Interaction Inference for Disease Subtype Network Generation (ICLR 2025 · similarity 0.86)
- A Generic Family of Graphical Models: Diversity, Efficiency, and Heterogeneity (ICML 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.