Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

ChromFound: Towards A Universal Foundation Model for Single-Cell Chromatin Accessibiltiy Data

NeurIPS 2025 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Zero-shot robustness and transfer · 1 citations (OpenAlex)

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.

Introduces ChromFound, a hybrid Mamba + window-attention foundation model pretrained on 1.97M scATAC-seq cells with genome-coordinate OCR tokenization, evaluated zero-shot on clustering and transferred to batch correction, cell annotation, and cross-omics prediction.

Problem

Single-cell ATAC-seq data are high-dimensional, sparse, and lack a standardized feature space, so there is no foundation model that supports zero-shot cell identification and multi-omics analysis.

Approach

Treat each open chromatin region as a token encoded by chromosome embedding, sinusoidal start/end genomic coordinate embeddings, and continuous accessibility value, and process the long genome-wide sequence with a hybrid of window-partitioned self-attention and Mamba blocks, pretrained with masked MSE reconstruction.

Key results

• Zero-shot cell clustering on PBMC169K (VIB_10xv1_1) reaches ARI 0.6953 vs best baseline PoissonVI 0.6314; average ARI gain reported as 17.02% across 8 datasets.
• Cross-omics prediction on Zhu11K cell lines reaches PCC 0.9449 vs best baseline 0.9196 (BABEL).
• Enhancer-gene link prediction for K562 achieves AUC 0.77 for COPZ1 and 0.61 for HNRNPA1.

Measured effects reported in the paper

Surprising findings

• Dropping the Mamba block or the WPSA attention leaves the other intact, but using only WPSA drops ARI to 0.3075, and Geneformer/scGPT Transformer-only variants reach ARI only 0.05 and 0.39 at 4k-32k peaks.
• Baseline VAE methods trained on cPeaks-aligned fixed-vocabulary data lose substantial performance on unseen tissues, while ChromFound performs well even on reference-aligned data.

Limitations

• Some baselines (BABEL, CMAE) are evaluated after filtering to top 10,000 highly variable OCRs, which the authors note makes comparison on enhancer perturbation unfair.
• Cell-type annotation and cross-omics results depend on fine-tuning or supervised MLP heads, not purely zero-shot.
• Reported ARI variance is high on some datasets (e.g., ±0.0387 and ±0.0500), and the 17.02% average gain is relative to a mix of baselines.

Experimental setup

Datasets: HumanAtlas FetalAtlas PBMCBenchmark Cortex130K Zhu45K Kuppe139K Liang154K To326K Morabito130K Zhu11K Fulco4K

Benchmarks: Morabito130K Kuppe139K Liang154K PBMC169K To326K Zhu45K Zhu11K Fulco4K

Metrics: ARI FMI NMI AMI AvgBIO AvgBATCH accuracy macro F1 PCC CCC

Baselines: PeakVI PoissonVI SnapATAC2 Signac SCALE SCALEX CASTLE scBasset Cellcano EpiAnno SANGO BABEL

Builds on: Mamba xTrimoGene scBasset

Closest papers by meaning

• A versatile informative diffusion model for single-cell ATAC-seq data generation and analysis (NeurIPS 2024 · similarity 0.87)
• scPilot: Large Language Model Reasoning Toward Automated Single-Cell Analysis and Discovery (NeurIPS 2025 · similarity 0.85)
• Cell ontology guided transcriptome foundation model (NeurIPS 2024 · similarity 0.85)
• Semi-supervised Knowledge Transfer Across Multi-omic Single-cell Data (NeurIPS 2024 · similarity 0.84)
• PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction (ICML 2025 · similarity 0.83)
• A Simple and Comprehensive Benchmark for Single-Cell Transcriptomics (AAAI 2025 · similarity 0.82)