Research map / Label-efficient and robust learning / Self-supervised and contrastive
scSSL-Bench: Benchmarking Self-Supervised Learning for Single-Cell Data
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.
Benchmarks nineteen SSL methods (generic, single-cell-specific, foundation models, baselines) and augmentation/architecture choices on nine single-cell datasets across batch correction, cell type annotation, and missing modality prediction.
Problem
There is no systematic comparison of SSL methods, hyperparameters, augmentations, and design choices for single-cell genomics data, and unclear whether specialized single-cell methods beat generic SSL.
Approach
A standardized benchmark that adapts generic image/text SSL methods to scRNA-seq and CITE-seq data, compares them with single-cell-specific contrastive and generative methods and foundation models, and systematically ablates augmentations (CLEAR, CLAIRE-style MNN, new BBKNN), embedding and projector size, temperature, DSBN, and multi-modal integration.
Key results
- scVI is the best scRNA-seq batch-correction method, with CLAIRE second; finetuned scGPT is strong on Immune Cell Atlas (total 0.781 vs VICReg 0.644).
- On multi-modal BMMC, SimCLR achieves the top total batch-correction score of 0.767, and totalVI scores 0.562.
- Random masking is the most effective augmentation alone and in combination, and embedding dimensions of 64 or 128 perform best; retaining the projector and DSBN do not improve results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use specialized single-cell methods rather than generic SSL for uni-modal batch correction | scRNA-seq batch correction · PBMC, Pancreas, and Immune Cell Atlas | Total | Mixed |
| Use generic methods such as SimCLR or VICReg rather than specialized methods | multi-modal batch correction · PBMC-M and BMMC | Total | Helped |
| Use generic SSL rather than specialized contrastive methods | cell type annotation · single-modal scRNA-seq datasets | Macro F1 | Helped |
| Finetune scGPT rather than use it zero-shot | scGPT · Immune Cell Atlas (10X 5' v2) | Macro F1: 0.358 -> 0.835 | Helped |
| Use random masking instead of other tested augmentations (ablation) | VICReg, SimCLR, and MoCo · batch correction datasets | Total | Helped |
| Use moderate embedding dimensionality (64 or 128) rather than lower dimensions (ablation) | generic SSL methods · HIC and MCA | Total | Helped |
| Increase embedding dimensionality (ablation) | generic SSL methods · HIC and MCA | biological conservation | Helped |
| Increase temperature (ablation) | SimCLR, MoCo, and NNCLR · HIC and MCA | batch correction | Hurt |
| Increase temperature (ablation) | SimCLR and NNCLR · PBMC-M | batch correction | Helped |
| Increase projector scaling factor (ablation) | Barlow Twins, BYOL, and VICReg · projector dimensionality ablation | performance | Helped |
| Retain the projector during inference (ablation) | SimCLR · MCA | Total: 0.625 -> 0.608 | Hurt |
| Apply domain-specific batch normalization instead of standard batch normalization (ablation) | generic SSL methods · HIC and MCA | Total | Hurt |
Surprising findings
- Generic SSL methods such as VICReg and SimCLR outperform specialized single-cell SSL frameworks on multi-modal integration and missing modality prediction.
- Retaining the projector during inference and domain-specific batch normalization, both recommended in prior work, do not improve and often reduce performance.
Limitations
- Multi-modal specialized methods (Concerto, scCLIP, scTEL) underperform, possibly due to limited tuning or implementation; results depend on specific preprocessing (4,000 HVGs)
- Hyperparameter tuning done on only two datasets (HIC and MCA)
- Concerto's DSBN not applied since public code lacks it
Experimental setup
Datasets: PBMC Pancreas Immune Cell Atlas Mouse Cell Atlas Human Immune Cells Lung Tabula Sapiens PBMC-M BMMC
Benchmarks: PBMC Pancreas Immune Cell Atlas PBMC-M BMMC MCA HIC Lung Tabula Sapiens
Metrics: scIB bio conservation scIB batch correction total score macro F1 accuracy Pearson correlation
Baselines: PCA SCDC scVI totalVI scGPT Geneformer scButterfly scTEL scCLIP
Builds on: SimCLR VICReg CLEAR CLAIRE Concerto scVI
Closest papers by meaning
- A Simple and Comprehensive Benchmark for Single-Cell Transcriptomics (AAAI 2025 · similarity 0.86)
- scPilot: Large Language Model Reasoning Toward Automated Single-Cell Analysis and Discovery (NeurIPS 2025 · similarity 0.85)
- Semi-supervised Knowledge Transfer Across Multi-omic Single-cell Data (NeurIPS 2024 · similarity 0.85)
- LangCell: Language-Cell Pre-training for Cell Identity Understanding (ICML 2024 · similarity 0.84)
- scMBERT: A Pre-Trained Deep Learning Model for Single-Cell Multiomic Data Representation and Prediction (Student Abstract) (AAAI 2025 · similarity 0.84)
- CellPLM: Pre-training of Cell Language Model Beyond Single Cells (ICLR 2024 · 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.