Research map / Label-efficient and robust learning / Recognition and long-tail
InfoSEM: A Deep Generative Model with Informative Priors for 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.
Proposes InfoSEM, an unsupervised variational generative model for gene regulatory network inference that uses BioBERT gene-description embeddings (and optionally known ChIP-seq interactions) as priors, plus a new unseen-gene benchmark that exposes dataset-bias shortcuts in supervised GRN methods.
Problem
Inferring gene regulatory networks from scRNA-seq without costly ground-truth labels, and fairly evaluating GRN methods, since supervised methods may exploit gene-specific class-imbalance biases rather than regulatory mechanisms.
Approach
Extends DeepSEM's VAE-based linear SEM by replacing the Laplace prior on interaction effects with a linear prior driven by pretrained textual gene embeddings (BioBERT), and optionally adds known interactions as a prior on the logit of interaction probability; also introduces an 'interactions between unseen genes' test split.
Key results
- On the proposed unseen-gene benchmark with cell-type specific ChIP-seq, InfoSEM-B improves over DeepSEM on all datasets by 25% on average (abstract reports 38.5% overall over models without priors).
- InfoSEM-BC further boosts performance by 11.1% on average over InfoSEM-B, and achieves best AUPRC on all four datasets in the cell-type specific unseen-gene setting (e.g., hESC AUPRC 0.331 vs DeepSEM 0.265).
- Supervised methods (scGREAT, GENELink) drop on unseen genes, averaging 42% for cell-specific ChIP-seq and 79% for non-cell-specific ChIP-seq, even falling below random; one-hot LR and matrix completion match SOTA supervised models on the existing seen-gene benchmark.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add BioBERT gene-embedding priors to DeepSEM | InfoSEM-B · Unseen genes across four cell-type-specific and non-cell-type-specific ChIP-seq datasets | AUPRC and Hit@1%: 25% and 52% on average | Helped |
| Add known-interaction priors to InfoSEM-B | InfoSEM-BC · Unseen genes across datasets | AUPRC and Hit@1%: 31% and 66% | Helped |
| Replace BioBERT embeddings with GenePT embeddings (ablation) | InfoSEM-B · Cell-type-specific unseen genes | GRN inference performance | Helped |
| Replace BioBERT embeddings with gene ontology embeddings (ablation) | InfoSEM-B · Cell-type-specific unseen genes | GRN inference performance | Helped |
| Replace BioBERT embeddings with one-hot embeddings (ablation) | InfoSEM-B · Cell-type-specific unseen genes | GRN inference performance | No clear effect |
| Replace BioBERT embeddings with random embeddings (ablation) | InfoSEM-B · Cell-type-specific unseen genes | GRN inference performance | No clear effect |
| Replace BioBERT embeddings with scBERT embeddings (ablation) | InfoSEM-B · Cell-type-specific unseen genes | GRN inference performance | No clear effect |
| Train using only 20% of cells instead of all cells (ablation) | InfoSEM-B and InfoSEM-BC · hESC unseen genes | AUPRC: 20% | No clear effect |
Surprising findings
- Trivial baselines using only one-hot gene IDs or matrix completion (no scRNA-seq) perform on par with state-of-the-art supervised GRN methods on the existing benchmark, indicating exploitation of gene-specific class imbalance.
- InfoSEM-B with scBERT embeddings did not improve over DeepSEM, which the authors hypothesize is due to binning of scRNA-seq data in scBERT.
Limitations
- Gains from textual priors are mainly on the authors' new unseen-gene benchmark; InfoSEM-B is not uniformly best on all datasets (e.g., hHEP Hit@1% and mESC AUPRC are not top).
- Non-cell-type specific ChIP-seq ground truth is noisy (negatives include unknown positives), so absolute metrics are very small and recall analysis is used as a supplement.
- Hyperparameters for InfoSEM are cross-validated on the known GRN, and InfoSEM-BC requires ground-truth interactions that are not always available.
Experimental setup
Datasets: BEELINE hESC BEELINE hHEP BEELINE mDC BEELINE mESC cell-type specific ChIP-seq non-cell-type specific ChIP-seq (BEELINE)
Benchmarks: BEELINE hESC BEELINE hHEP BEELINE mDC BEELINE mESC
Metrics: AUPRC Hit@1% recall
Baselines: One-hot LR MatComp scGREAT GENELink SCENIC GRNBoost2 DeepSEM
Builds on: DeepSEM BEELINE GENIE3 GRNBoost2 TIGRESS
Closest papers by meaning
- KINDLE: Knowledge-Guided Distillation for Prior-Free Gene Regulatory Network Inference (NeurIPS 2025 · similarity 0.87)
- GeSubNet: Gene Interaction Inference for Disease Subtype Network Generation (ICLR 2025 · similarity 0.86)
- Pruning neural network models for gene regulatory dynamics using data and domain knowledge (NeurIPS 2024 · similarity 0.85)
- OgBench: A Framework for Evaluating Graph Neural Networks on Omics Data (NeurIPS 2026 · similarity 0.84)
- Gene Regulatory Network Inference in the Presence of Selection Bias and Latent Confounders (NeurIPS 2025 · similarity 0.84)
- Generative Intervention Models for Causal Perturbation Modeling (ICML 2025 · similarity 0.83)
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.