ATAC-seq has a reputation for being simple and elegant. The protocol is compact, the input requirement is low, and the output claims to capture chromatin accessibility at a genome-wide scale. Many experimental biologists believe the data should be straightforward to analyze - like just running a few commands and getting nice peaks. But we know that is not the case.
In reality, ATAC-seq data analysis is full of traps. Even experienced labs sometimes fall into them. It’s not because people are careless - most of them are quite careful. But because this data behaves in non-obvious ways, and some pitfalls are almost invisible unless you’ve seen them before.
We have worked on dozens of ATAC-seq projects - some designed beautifully, others recovered after collapse. The problems we share here are not theoretical. They are what we have actually encountered in real-world projects - sometimes in peer-reviewed studies, sometimes in datasets submitted for rescue. Each one of them has taught us painful but important lessons.
This series is not for beginners. We will not explain how to install MACS2 or how to interpret a PCA plot. Instead, we will discuss nine key failure points - and how experienced bioinformaticians avoid them.
Feeling unsure about your ATAC-seq results? We help research teams troubleshoot confusing data and regain clarity in chromatin accessibility analysis. Request a free consultation →
Everyone talks about quality control, but few realize it can hurt as much as it helps. Many teams apply default thresholds for metrics like TSS enrichment, FRiP score, or fragment length distributions, assuming these cutoffs will clean the data and make it more “trustworthy.” But the truth is: biology does not follow our preferred QC rules.
We’ve seen strong immune activation signals eliminated because the short fragments didn’t pass the arbitrary size cutoff. We’ve seen stressed tissue samples entirely removed because their FRiP scores looked low - even though those samples carried the core biology of the study. Bioinformaticians with deeper experience always ask the key question first: what exactly are we filtering out, and why?
Instead of trusting numeric thresholds blindly, we always visualize the signal over known regions - active promoters, enhancers, marker loci. If filtering removes real signal from these places, then it is the filters, not the sample, that need adjustment. Sometimes a sample with lower “QC” actually gives you more biology than the one that looks perfect.
Not sure if your ATAC-seq conclusions will hold up? We provide in-depth review and second opinions to help avoid surprises during manuscript submission or peer review. Request a free consultation →
One of the most common mistakes is discarding all mitochondrial and duplicate reads as if they are technical garbage. In reality, these features are highly context-dependent. We’ve worked on multiple datasets where mtDNA content actually reflected a biological phenomenon - like metabolic stress or apoptosis. And in low-input samples, high duplicate rate is simply unavoidable - sometimes even informative.
The problem is that people often use Picard or SAMtools defaults without thinking. These tools were made for general-purpose sequencing, not ATAC-seq in complex or low-yield samples. When you remove reads based on default thresholds, you risk discarding your best signals - and later, you won’t be able to explain to the reviewer why those samples are gone.
What we do instead is more careful. We generate library complexity curves to assess duplication from a saturation perspective. We correlate mitochondrial content with sample conditions to see whether the pattern reflects biology. We only remove reads when we are confident they are noise - not just because the manual says so.
Fragment size distribution is widely used as a QC metric in ATAC-seq. People expect to see clear nucleosome laddering - mono-, di-, tri-nucleosomal peaks with sharp boundaries. And when the pattern is not there, they assume the sample failed. This is a big mistake.
In some tissues - especially those with highly open chromatin - the laddering is not visible, even if the signal is strong. In one study we analyzed, an embryonic tissue showed a broad peak of short fragments and no clear nucleosome structure. The team was ready to discard it until we showed that signal enrichment at key loci was excellent.
Fragment size pattern should be interpreted in context. Instead of relying on the shape of the curve, we look at signal enrichment at TSS, active enhancers, and known regulatory elements. If the data is informative and reproducible, then the absence of laddering is irrelevant. Blind reliance on these patterns has led many people to throw away their best samples.
Peak calling is a central step, but it’s also one of the most misunderstood. Many people run MACS2 or Genrich on each sample separately, thinking it’s just like RNA-seq - find the features first, then compare across conditions. But this is a trap. Without a shared coordinate space, statistical comparison becomes meaningless.
We’ve seen teams compare peaks that do not even overlap between samples, apply differential accessibility tests on mismatched regions, and produce volcano plots that look dramatic but reflect artifacts. In some projects, reviewers directly questioned the validity of the comparisons - and they were right to do so.
What we do is always define a consensus peak set. We merge peaks across replicates or conditions, and count reads against this unified set. This ensures that we are comparing apples to apples. If you don’t do this, you may still get p-values - but they are not tied to any biological truth.
Struggling with peak comparability across conditions? We help labs establish robust consensus peak sets for meaningful statistical comparisons in ATAC-seq studies. Request a free consultation →
Some peaks look beautiful - sharp, consistent, high signal - but are biologically meaningless. These include Tn5 hypersensitive sites that arise from sequence bias, unmappable regions that trap reads, or simply noisy loci that happen to look consistent by accident. If you don’t control for these, they will distort every downstream step.
In one rescue project, we found that over 20% of reported differential peaks were located in known ENCODE blacklist regions. Another dataset had dozens of high-signal peaks that occurred identically across all samples - yet the team used them to drive motif analysis and drew mechanistic conclusions. It was a disaster waiting to happen.
We always cross-check peak lists against artifact databases. We inspect reproducibility across replicates - real biological peaks should vary with condition, not be identical everywhere. And we treat too-good-to-be-true signal with skepticism. Beautiful peaks are not always true ones.
Continue Reading Part 2