Research map / Graph learning / Graph self-supervised learning
Gene-Gene Relationship Modeling Based on Genetic Evidence for Single-Cell RNA-Seq Data Imputation
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 scCR, a training-free propagation-based scRNA-seq imputation method that models both associating and dissociating gene-gene relationships by concatenating the data with its negation and standardizing each gene before building k-NN graphs.
Problem
scRNA-seq data contain many false zeros (dropouts) and noise; existing propagation-based imputation methods use k-NN graphs that capture only associating (co-occurring) gene relationships and ignore dissociating ones, and mix values across genes with different scales.
Approach
Concatenate the pre-imputed cell-gene matrix with its negation, standardize each column (gene), build a gene-gene k-NN graph on this matrix so both co-occurring and anti-correlated genes are connected, then run feature propagation across genes and cells, followed by a soft-FP denoising stage.
Key results
- On cell clustering, scCR improved ARI by 25.3% on Baron Mouse, 1.7% on Mouse Bladder, 6.4% on Zeisel, and 21.6% on Baron Human over the best prior baseline.
- On dropout recovery, scCR reduced RMSE from the most competitive baseline at most dropout rates across datasets (except Pancreas at 40% dropout).
- Under MNAR dropout simulation, scCR achieved RMSE of 0.304 on Baron Mouse vs 0.465 for scBFP and 0.517 for scFP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add concatenation and standardization to model complete gene-gene relationships (ablation) | scCR · Baron Mouse, Zeisel, Baron Human | ARI: 0.625±0.001 -> 0.827±0.093; 0.789±0.000 -> 0.902±0.000; 0.805±0.000 -> 0.823±0.000 | Helped |
| Add column-wise standardization (ablation) | scCR · Baron Mouse | ARI: 0.625±0.001 -> 0.623±0.000 | No clear effect |
| Add column-wise standardization (ablation) | scCR · Zeisel | ARI: 0.789±0.000 -> 0.898±0.000 | Helped |
| Add concatenation with the negated matrix (ablation) | scCR · Baron Mouse | ARI: 0.625±0.001 -> 0.629±0.001 | No clear effect |
| Add concatenation with the negated matrix (ablation) | scCR · Zeisel | ARI: 0.789±0.000 -> 0.821±0.000 | Helped |
| Add the complete relation stage (ablation) | scCR · Baron Mouse | ARI: 0.437 ± 0.061 -> 0.409 ± 0.008 | Hurt |
| Add the complete relation stage (ablation) | scCR · Zeisel | ARI: 0.682 ± 0.024 -> 0.732 ± 0.001 | Helped |
| Add the denoising stage (ablation) | scCR · Baron Mouse | ARI: 0.437 ± 0.061 -> 0.584 ± 0.000 | Helped |
| Add the complete relation and denoising stages (ablation) | scCR · Baron Human | ARI: 0.580 ± 0.036 -> 0.823 ± 0.000 | Helped |
| Increase α from 0.05 to 0.9 (ablation) | scCR · Baron Mouse | ARI: 0.827 ± 0.093 -> 0.448 ± 0.001 | Hurt |
| Increase β from 0.99 to 0.999 (ablation) | scCR · Baron Mouse | ARI: 0.827 ± 0.093 -> 0.925 ± 0.000 | Helped |
| Increase γ from 0.01 to 0.001 (ablation) | scCR · Baron Mouse | ARI: 0.827 ± 0.093 -> 0.927 ± 0.001 | Helped |
Surprising findings
- On Baron Mouse, the scCR ARI has high variance (0.827 ± 0.093), and some alternative hyperparameter settings (e.g., beta=0.999) gave higher ARI on Baron Mouse (0.925) than the default setting.
Limitations
- Method is specifically designed for scRNA-seq data and not validated on other domains.
- Performance on some datasets (e.g., Baron Mouse) shows high variance across runs and sensitivity to hyperparameters such as k and gamma.
- Pancreas dataset at 40% dropout is an exception where scCR does not reduce RMSE.
Experimental setup
Datasets: Baron Mouse Pancreas Mouse Bladder Zeisel Worm Neuron Baron Human
Benchmarks: Baron Mouse Pancreas Mouse Bladder Zeisel Worm Neuron Baron Human
Metrics: ARI NMI CA RMSE Median L1 Distance running time
Baselines: DCA AutoClass scTAG scGNN 2.0 scGCL MAGIC scFP scBFP
Builds on: Feature Propagation scFP scBFP MAGIC
Closest papers by meaning
- Gene Regulatory Network Inference in the Presence of Dropouts: a Causal View (ICLR 2024 · similarity 0.88)
- SpotDiff: Spatial Gene Expression Imputation Diffusion with Single-Cell RNA Sequencing Data Integration (AAAI 2025 · similarity 0.85)
- A Simple and Comprehensive Benchmark for Single-Cell Transcriptomics (AAAI 2025 · similarity 0.84)
- On the Recoverability of Causal Relations from Bulk Gene Expression Data (NeurIPS 2026 · similarity 0.84)
- A Generic Family of Graphical Models: Diversity, Efficiency, and Heterogeneity (ICML 2025 · similarity 0.84)
- Semi-supervised Knowledge Transfer Across Multi-omic Single-cell Data (NeurIPS 2024 · 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.