Ask questions across claims linked to experiments, exact results, limitations, and sources.Know what falsifiable science actually supports before you trust the answer.
Press Enter ↵ to search evidence
Explore by Goal
"For me, it is far better to grasp the Universe as it really is than to persist in delusion, however satisfying and reassuring."
- Carl Sagan
Quick Analysis Plan
Copied
Best-practice summary (concise)
Current best bioinformatics approaches for single-cell ATAC-seq (scATAC‑seq) center on: (1) careful preprocessing (adapter/trimming, alignment to genome, deduplication, barcode correction); (2) per-cell QC using fragment counts, TSS enrichment and mitochondrial fraction; (3) constructing a binary or TF-IDF normalized cell-by-peak matrix; (4) dimensionality reduction using LSI or topic models (cisTopic / latent semantic indexing) followed by UMAP/PHATE; (5) graph-based clustering; (6) peak-to-gene linkage and regulatory activity inference (gene score, co-accessibility / Cicero / ArchR methods); (7) integration with scRNA‑seq (label transfer, multi‑omic anchors, weighted-nearest neighbors); and (8) differential accessibility and TF motif/regulatory network analysis using chromVAR / footprinting and aggregate peak analyses. For robust, reproducible pipelines use Snakemake/nf-core or established packages (ArchR, Signac + Seurat, SnapATAC2, cisTopic). For a complete, runnable, evidence-backed analysis plan click the AI Scientist button below to run an automated, iterative bioinformatics agent that will fetch references, run code on your data, and produce figures.
Note: I can produce a full, fully-cited plan and runnable code; please run the AI Scientist agent or tell me your data format (10x fragments.tsv.gz, CellRanger-ATAC output, BAM, etc.) and goals (cell-type ID, trajectory, regulatory network) to proceed.
Below is a compact, evidence-oriented pipeline and tool-map for scATAC‑seq bioinformatics (preprocessing → QC → feature matrix → dimensionality reduction/clustering → regulatory inference → integration → DA/TF analysis → reproducible workflow). To run this end-to-end on your data, click the automated AI Scientist button at the end.
1) Input types & recommended first steps
Common inputs: 10x Genomics Cell Ranger-ATAC outputs (fragments.tsv.gz, peaks, barcodes), BAM/FASTQ, or plate-based scATAC fragment files.
Initial processing: adapter/quality trimming (if starting FASTQ), align to reference (BWA-MEM or Bowtie2 recommended for short ATAC fragments), mark duplicates, generate fragments file (chrom, start, end, barcode, read count), and correct/whitelist cell barcodes.
2) Per-cell quality control (QC) — critical metrics
TSS enrichment score: measures enrichment at transcription start sites — useful for calling high-quality nuclei.
Fraction of fragments in peaks (FRiP) and fraction in blacklist regions.
Mitochondrial or chrY fraction (if relevant) and nucleosome periodicity (mononucleosomal vs nucleosome-free fragments).
3) Feature construction: peaks, windows, or binning
Approaches trade off resolution vs sparsity:
Aggregate peak calling across cells (pseudo-bulk) to define a consensus peak set (commonly used).
Fixed-size genomic bins (e.g., 5kb, 10kb) — reduces bias from peak-calling but may dilute signal.
Gene-centric windows for gene score calculations (gene body + promoter) for integration with RNA data.
4) Normalization & dimensionality reduction
Leading practice: TF-IDF (term frequency–inverse document frequency) normalization of the binary cell-by-peak matrix followed by singular value decomposition (LSI) to reduce dimensionality, then UMAP for visualization. Alternative: topic modeling (cisTopic) or latent variable methods (scVI-style) adapted for accessibility.
5) Clustering & batch correction
Build a k-nearest neighbor graph on LSI/topic embeddings and apply Leiden or Louvain clustering.
Batch correction / integration: mutual nearest neighbors, Harmony, or combined multimodal anchors (e.g., Signac + Seurat anchors; WNN for multiome) — choose method based on whether multiomic data exist.
Peak linkage/co-accessibility: Cicero or co-accessibility modules (also accessible in ArchR) to predict enhancer–promoter links.
Gene activity scores: aggregate accessibility in promoter + gene body windows to estimate regulatory activity and enable label transfer from scRNA‑seq.
Use logistic regression or negative binomial generalized linear models adapted to sparse binary counts; aggregate pseudo-bulk per-cluster for robust peak calling and differential tests with tools that model overdispersion and batch effects. Correct for multiple testing (Benjamini–Hochberg).
8) Integration with scRNA‑seq or multiome
When paired multiome exists (same-cell scATAC + scRNA): use joint latent space methods (WNN, MOFA) or signac/Seurat multiomic anchors to transfer cell labels and improve peak-to-gene inference.
When separate datasets: compute gene activity scores from scATAC and use label transfer/classifiers to map scRNA labels onto scATAC cells.
Implement as a Snakemake or nf-core pipeline with containerization (Docker/Singularity) and explicit versions. Use GPU-accelerated steps only where supported (some latent models). Store intermediate fragments, peak matrices, and AnnData / Arrow (ArchR) objects for reproducibility.
10) Recommended software ecosystem (practical)
ArchR — end-to-end scATAC analysis with Arrow disk-backed objects, co-accessibility and integration features.
Signac + Seurat — R-based integration with scRNA, good for label transfer and visualization.
SnapATAC / SnapATAC2 — scalable single-cell ATAC toolkit focused on speed and clustering.
cisTopic — topic modeling for regulatory topic discovery.
chromVAR — TF motif deviation analysis for per-cell TF activity.
Cicero — co-accessibility to infer enhancer–promoter interactions from scATAC.
General stack: Python (Scanpy, anndata), R (Seurat, Signac, ArchR), Snakemake/nf-core for pipelines, and Plotly/ggplot2 for visualization.
Caveats, biases & blindspots (important)
scATAC is extremely sparse — many downstream analyses rely on aggregates (pseudo-bulk, cluster-level) and will inherit biases from peak calling and library complexity differences.
Cross-study comparisons require careful handling of batch effects and differences in peak definitions (consensus peaks recommended).
TF footprinting from scATAC is noisy; motif activity (chromVAR) is typically more robust than per-site footprinting at single-cell level.
Peak-to-gene assignments are probabilistic; independent validation (CRISPRi, reporter assays) is recommended when claiming regulatory links.
Next step — automated analysis:
To convert this plan into a fully-cited, runnable analysis (with dataset-specific cutoffs, code, and publication-grade figures), run the AI Scientist agent which will fetch precise references, produce plots, and generate reproducible pipelines for your input files.
If you want: specify your raw input type (FASTQ, fragments.tsv.gz, BAM), species/genome build, whether you have paired scRNA data, desired outputs (cell labels, regulatory network, differential peaks), and I will produce a custom, fully-cited Snakemake pipeline and runnable Python/R notebooks.
Generating an end-to-end, reproducible Snakemake pipeline that is creating processed fragments, computing QC metrics, building TF-IDF+LSI embeddings, clustering, and producing peak-to-gene links for your scATAC dataset.
Get emailed when your analysis is done!
We'll email you the results when your analysis is finished.
Hypothesis Graveyard
TF footprinting at single-cell resolution reliably identifies binding events — evidence shows this is usually too noisy at single-cell depth, so motif-deviation approaches are preferred.
Using raw read counts without TF-IDF/LSI for dimensionality reduction — performs poorly due to sparsity and sequencing depth confounding.