What ATAC-seq measures

ATAC-seq (Assay for Transposase-Accessible Chromatin with sequencing) tells you which parts of the genome are physically open.

The method is simple. A hyperactive Tn5 transposase is loaded with sequencing adapters and added to nuclei. Tn5 can only cut where DNA isn’t tightly wrapped around nucleosomes, and it inserts the adapters as it cuts. Open regions get fragmented and tagged, closed chromatin doesn’t, and after sequencing the read density is a map of accessibility.

This is useful because a lot of regulation happens in accessible chromatin. Promoters, enhancers and active transcription-factor binding sites are in open regions, so an accessibility map narrows down where to look for regulatory elements a lot, and you don’t need to know in advance which factor you’re after.

How it differs from ChIP-seq

The two get mentioned together but answer different questions:

In practice ATAC-seq often comes first to show where to look, and ChIP-seq follows to show which protein is bound there. The low input requirement is also why ATAC-seq became the standard for single-cell chromatin work.

Quality control

Three checks tell you whether a library is usable:

• Fragment size distribution. A good library shows clear periodicity: a sub-nucleosomal peak below ~100 bp, then peaks near 200 and 400 bp for mono- and di-nucleosomes. If that ladder is missing, the digestion was over- or under-done.
• TSS enrichment. Reads should pile up sharply at transcription start sites. A flat profile means the signal is mostly background.
• Mitochondrial contamination. Mitochondrial DNA has no nucleosomes, so Tn5 cuts it freely and it can take up a large share of reads. Omni-ATAC and similar protocol revisions were largely designed to reduce this.

Peaks are then usually called with MACS2. Its peak model was designed for ChIP-seq, so the assumptions don’t carry over perfectly; shifting reads to centre on the Tn5 insertion site is a standard correction.

Deep learning

Accessibility data suits sequence models well. The input is a one-dimensional string over a four-letter alphabet, the output is a per-position signal, and there’s a lot of data. Convolutional models more or less automatically learn motif-like filters in their first layer, which is why this combination took off early.

The main caution is the one in the Whalen et al. paper below. Genomic data is highly structured, and a random train/test split puts adjacent or homologous regions on both sides, so the reported accuracy is partly measuring memorisation. Splitting by chromosome is the usual minimum precaution. It’s the same kind of leakage you get from any careless split.

Reading list

ATAC-seq

1. Buenrostro JD, et al. Transposition of native chromatin for fast and sensitive epigenomic profiling of open chromatin, DNA-binding proteins and nucleosome position. Nat Methods. 2013. The original method paper.
2. Yan F, et al. From reads to insight: a hitchhiker’s guide to ATAC-seq data analysis. Genome Biol. 2020. The most practical walkthrough of the full analysis.
3. Ou J, et al. ATACseqQC: a Bioconductor package for post-alignment quality assessment of ATAC-seq data. BMC Genomics. 2018.
4. Landt SG, et al. ChIP-seq guidelines and practices of the ENCODE and modENCODE consortia. Genome Res. 2012. Still the reference for what good quality control looks like.
5. Zhang Y, et al. Model-based Analysis of ChIP-Seq (MACS). Genome Biol. 2008.

Deep learning in genomics

1. Eraslan G, et al. Deep learning: new computational modelling techniques for genomics. Nat Rev Genet. 2019.
2. Zou J, et al. A primer on deep learning in genomics. Nat Genet. 2019.
3. Whalen S, et al. Navigating the pitfalls of applying machine learning in genomics. Nat Rev Genet. 2022. Worth reading alongside the others rather than after them.