Representing Chromatin Accessibility Counts

Why the Representation Matters

An ATAC-seq experiment produces fragment observations at genomic coordinates, but most analyses compare cells using a feature matrix. The choice of features and entries determines what “similar accessibility” means: two cells may be close when their counts are compared, far apart when their accessible peaks are treated as sets, or indistinguishable after nearby peaks are aggregated. These are not merely alternative file formats. They encode different information and induce different distances.

This tutorial compares three choices: fragment counts assigned to peaks, binary peak-by-cell indicators, and measurements aggregated over broader genomic regions. The phrase “region-by-cell” describes a matrix whose rows are genomic regions and columns are cells; a peak-by-cell matrix is also region-by-cell in that broad sense. The substantive distinction is usually the resolution and definition of the genomic features, and whether each entry is a count or a binary indicator.

A fragment-count matrix is not the list of individual fragments: it is a summary of them after assigning fragments to a feature catalog. The result depends on how fragments are defined and counted, including choices about paired-end reads, duplicate handling, cell-barcode filtering, and overlap rules. A binary matrix discards the distinction between one and many assigned fragments. A broader-region matrix changes the feature catalog, for example by summing counts across several nearby peaks or across fixed genomic intervals.

Comparisons require a consistent definition. If one cell’s counts use a different peak catalog, fragment-counting rule, or barcode set from another’s, a difference between matrix entries may reflect processing rather than accessibility. Similarly, a binary entry of zero means “no fragment assigned under this assay and processing rule,” not proof that the chromatin is biologically closed.

How the Choices Change Distances

For two cells represented by feature vectors \(x\) and \(y\), Euclidean distance is \(d_E(x,y)=\sqrt{\sum_i(x_i-y_i)^2}\). It responds to both the features that differ and the magnitudes of their entries. If the vectors are binary, the squared Euclidean distance is simply the number of features present in exactly one of the two cells. For binary vectors, Jaccard distance instead compares the sets of detected features: it is one minus the size of their intersection divided by the size of their union. Jaccard distance ignores joint absences, which can be numerous in sparse ATAC data.

These definitions do not make one distance universally preferable. Euclidean distance on counts is sensitive to count magnitude and can be strongly affected by differences in total fragments. Jaccard distance on binary features focuses on overlap among detected features, but it discards how many fragments support each detection. Distances on broader regions reflect accessibility at that coarser resolution: separate nearby peak patterns can become identical if they fall within the same region.

The relation between binary Euclidean distance and set difference follows directly. Let \(A\) and \(C\) be the sets of features with value one in two binary vectors. For each feature, \((B_{iA}-B_{iC})^2\) is one exactly when that feature belongs to one set but not both, and is zero otherwise. Therefore,

So binary Euclidean distance counts features in the symmetric difference. Jaccard distance uses the same disagreement features but divides by the union size: \(d_J(A,C)=|A\mathbin{\triangle}C|/|A\cup C|\), when the union is nonempty. This normalization can make overlap comparisons less dependent on the total number of detected features. If both cells have no detected features, the union is empty and the Jaccard distance is undefined unless an explicit convention is adopted; that case should be handled and reported rather than silently treated as an ordinary comparison.

Changing the Genomic Resolution

A region-by-cell representation changes which genomic distinctions can be seen. Suppose a region is formed by grouping a specified set of peaks. A count-valued regional entry can be defined by summing the assigned counts over those peaks; a binary regional entry can be defined as one if any grouped peak is detected. Those rules are not equivalent: summing counts preserves the number of assigned fragments, whereas the binary rule records only whether the region had any detection.

The grouping rule must be explicit. “Region” might mean a merged interval, a fixed-width genomic bin, or another predefined unit. Different choices can combine biologically distinct peaks or split a signal across bins. The representation also depends on whether fragments crossing region boundaries are assigned once, assigned to every overlapping interval, or handled by another declared rule.

Choosing a Representation for Joint Analysis

For joint analysis with RNA, protein, or another modality, the ATAC matrix must remain traceable to the correct cells. Rows should refer to well-defined genomic features, columns to cell identities, and any transformations or filtering should preserve that mapping. A paired assay does not make the modalities interchangeable: an RNA count and an ATAC fragment count arise from different measurement processes and have different feature meanings.

A sensible representation depends on the question. Count-valued features preserve more information about observed support, but raw counts can make distance comparisons sensitive to depth. Binary features emphasize whether a peak was detected, but two cells with very different counts can become identical. Broader regions may provide a more stable coarse summary, but only if the loss of peak-level detail is acceptable. In each case, document the feature catalog, counting or thresholding rule, and treatment of empty or missing features.

When computing a distance, check that its inputs match the intended interpretation. Euclidean distance on count vectors is meaningful as a geometric comparison of those entries, but high-total-count cells may dominate it. Binary Jaccard distance avoids counting shared absences as overlap, but sparse data can still produce unstable comparisons when very few features are detected. A distance on aggregated regions answers a coarser question than one on individual peaks. No distance calculation alone resolves these design choices.

Diagnostics, Alternatives, and Common Mistakes

Useful checks include the distribution of total fragment counts per cell, the number of detected features per cell, the fraction of zero entries, and whether a small number of features dominate the count scale. Compare distance patterns with library size and processing batch: if cells group mainly by those quantities, the representation may be emphasizing technical variation. Check for cells with no detected features, inconsistent feature catalogs, duplicated or overlapping intervals, and cell-order mismatches between modalities.

These checks diagnose potential problems; they do not prove that a particular distance is biologically appropriate. A count-based comparison, a binary overlap measure, and a broader-region comparison are alternatives because they preserve different information. They should be chosen and interpreted in terms of the question, rather than selected only because one yields a visually appealing grouping. Transformations designed specifically for sparse ATAC data and methods for balancing modalities require additional choices beyond the raw representations described here.

• Mistake: treating raw counts as depth-free. “The count distance is larger, so the cells must differ more biologically” ignores that greater sequencing depth can increase counts. Check library size and state what the distance measures.
• Mistake: treating binarization as harmless normalization. Two cells with counts \((1,0)\) and \((100,0)\) both become \((1,0)\). Binarization removes magnitude, not merely a nuisance scale.
• Mistake: assuming broader regions preserve peak-level information. As the aggregation example shows, distinct peak patterns can produce the same regional vector. Choose the resolution that matches the scientific question.
• Mistake: interpreting a zero as biological closure. A zero is no assigned observation under the stated assay and processing pipeline; finite sampling and technical factors can also produce it.