Tn5 Insertion Bias and ATAC-seq Fragmentation

Why Tn5 Bias and Fragment Length Matter

ATAC-seq uses the Tn5 transposase to insert sequencing adapters into DNA. In general, more insertions can indicate greater access of the enzyme to a genomic region, but insertion counts are not a pure measurement of accessibility: Tn5 has sequence preferences, and the distribution of fragment lengths also depends on chromatin structure and library preparation. Interpreting these signals requires separating what they can support from what they cannot establish on their own.

The introductory material on ATAC-seq establishes that transposase insertions are used as evidence about accessible chromatin. Here the focus is on two qualifications. First, Tn5 may insert more readily at some local DNA sequences than at others. Second, paired-end fragment lengths often show patterns associated with nucleosome-scale DNA protection, but those patterns can be altered by transposition conditions, sample quality, and size selection.

A Simple Model for Sequence Preference

Let \(Y_j\) be the number of observed Tn5 insertions at genomic position \(j\), \(A_j\) the local accessibility, \(B_j\) the relative insertion propensity associated with the local sequence, and \(L\) a library-depth factor. A useful working model is

This equation is a conceptual model, not a complete model of every source of sequencing variation. It assumes that, apart from depth and local sequence preference, expected insertions increase with accessibility. To use it for correction, one must have a credible estimate of \(B_j\), and that estimate must apply to the assay and analysis being performed. Other factors, including mapping and library properties, can also affect observed counts.

The model shows why raw insertion counts alone cannot separate accessibility from sequence preference. A large \(Y_j\) could arise from large \(A_j\), large \(B_j\), or both. If the sequence propensity is known, one can compare counts adjusted by \(B_j\), but this does not make the estimate exact: \(B_j\) is itself estimated, and correction can increase noise, especially where the estimated propensity is small.

Estimating and Using Sequence Bias

A sequence-bias model estimates how likely Tn5 is to insert near different local DNA sequences. One approach uses a control in which DNA is exposed to the transposase outside its native chromatin context, so the observed preference is less affected by chromatin accessibility. Another approach estimates preference from sequence context in the experimental data under explicit assumptions about which insertions provide a suitable reference. A model may represent sequence context using short motifs or a longer local sequence window.

Whichever approach is used, the estimate should be treated as a nuisance adjustment rather than a direct measurement of chromatin state. A correction is most defensible when the bias model is estimated with adequate data, its sequence context matches the assay, and it is applied consistently. If a model assigns near-zero propensity to a sequence, dividing by that estimate can produce an extremely large corrected value. Such unstable weights are a warning to inspect the model and its use, not a reason to regard the inflated value as strong biological evidence.

The exact interpretation of an insertion coordinate also depends on the read-alignment and fragment-processing convention. Some ATAC-seq workflows shift aligned read positions to approximate the Tn5 insertion site. Comparisons require consistent conventions: mixing shifted and unshifted coordinates can misalign insertion counts with local sequence and make a sequence-bias analysis misleading.

Reading Fragment-Length Periodicity

For paired-end data, the aligned read pair gives an estimate of the length of the DNA fragment between the two sequenced ends. A common qualitative pattern includes a short-fragment class, a longer class around the length of DNA wrapped around one nucleosome, and further classes at roughly additional nucleosome-scale intervals. Nucleosome-associated periodicity is supported when successive longer-fragment modes are separated by approximately one nucleosome repeat length, often near 180–200 base pairs. The precise positions and strength of modes vary with assay conditions and sample.

This pattern is not a direct count of nucleosomes or a stand-alone measure of accessibility. A short fragment may arise from nucleosome-depleted DNA, but short fragments can also be enriched by library processing or size selection. Longer fragments can be consistent with DNA protected by nucleosomes, but a fragment-length histogram does not locate those nucleosomes or establish their occupancy at a particular locus. Interpret the profile alongside other library diagnostics, including the quality-control measures covered in the earlier material on chromatin sequencing libraries.

Diagnosing Fragmentation Problems

Fragmentation profiles should be compared using the same fragment definition, alignment conventions, filtering rules, and plotting scale. A profile dominated by very short fragments with few longer classes may reflect excessive transposition, degraded nuclei, or library size selection, among other possibilities. A profile with few short fragments and weak or absent nucleosome-associated modes may be consistent with insufficient transposition or poor recovery of informative fragments. These are diagnostic clues, not unique explanations: no single profile shape identifies its cause without considering the protocol and other quality measures.

A practical assessment combines the fragment-length distribution with sequence-bias checks and broader library quality metrics. For example, strong sequence-dependent variation at insertion sites may indicate that raw insertion counts should not be interpreted as accessibility without qualification. A severe loss of longer fragments or an unusually narrow profile may prompt investigation of transposition conditions and size selection. A discrepancy between the fragment profile and other quality indicators should be investigated rather than resolved by selecting whichever metric gives the preferred conclusion.

Why Periodicity and Bias Are Not Simple Quality Scores

A common mistake is to reason, “The sample has many short fragments, so it must have high accessibility,” or “The longer-fragment periodicity is weak, so there are no nucleosomes.” Both statements overinterpret the profile. Fragment proportions depend on chromatin, transposition, DNA recovery, and size selection. Weak periodicity may reflect technical limitations or biological heterogeneity; a clear periodic pattern supports nucleosome-associated organization but does not quantify accessibility by itself.

A second pitfall is to treat sequence-bias correction as a universal repair. The incorrect reasoning is: “The raw counts are sequence-biased, so dividing by any estimated sequence preference gives unbiased accessibility.” The correction is only as credible as the bias estimate, the assumptions behind it, and the stability of the adjustment. In particular, very small propensity estimates can amplify noise. Inspect the distribution of estimated propensities and corrected values, and report the adjustment rather than presenting corrected counts as error-free measurements.

The main alternatives are to report raw insertion evidence with a clear caveat, to use a calibrated sequence-bias model, or to compare results across raw and adjusted representations as a sensitivity analysis. Raw counts are transparent but retain sequence preference; model-based adjustment targets that preference but introduces dependence on the model; a sensitivity analysis shows whether interpretation changes under the adjustment. None eliminates all technical or biological uncertainty. The later topic of differential chromatin accessibility testing addresses statistical comparisons between conditions; the fragment and bias checks here help assess the evidence those analyses may use.