Constructing Gene Activity Scores From ATAC-seq

From Accessible Chromatin to a Gene-Level Score

An ATAC-seq experiment records evidence of chromatin accessibility at genomic intervals. A gene activity score compresses accessibility near a gene into one number per cell and gene. It can provide a gene-oriented representation of ATAC-seq data, but it does not count RNA molecules. Its meaning depends on which intervals are included and how their contributions are combined.

This tutorial focuses on promoter and gene-body accessibility. It does not treat a score as evidence that a particular peak regulates a particular gene. Nearby peaks can be assigned to a gene by a rule for constructing the score, but that assignment alone does not establish a regulatory link.

Define the Regions Before Combining Counts

Let \(Y_{cp}\) be the ATAC-seq count for peak \(p\) in cell \(c\), using the counting convention of the data set, such as the number of fragments overlapping that peak. A gene annotation supplies the gene’s genomic coordinates, strand, and TSS. A promoter is defined relative to the TSS, often with a window extending upstream and downstream. For example, one analysis might use 2,000 bases upstream and 500 bases downstream. The gene body is the annotated gene interval, which may include introns as well as exons.

“Upstream” and “downstream” are relative to transcriptional direction, not always to increasing genomic coordinates. For a gene on the negative strand, upstream lies toward higher genomic coordinates. The annotation, genome assembly, promoter-window sizes, and gene-body definition therefore need to be specified. There is no single universal window that defines every gene activity score.

To combine counts cleanly, let \(P_g\) be the set of peaks assigned to the promoter of gene \(g\), and \(B_g\) the set assigned to its gene body. A peak overlapping both intervals should not be counted twice unless the method explicitly intends that. One practical convention is to make the sets disjoint by assigning promoter-overlapping peaks to the promoter set first and then assigning remaining gene-body peaks to the body set. A peak can still overlap regions assigned to different genes; independent gene scoring can count it for both genes.

What the Weighting Rule Changes

A simple rule is \(w_P=w_B=1\), so the score is the sum of promoter and body counts. This gives each observed fragment count the same contribution, regardless of region. A promoter-emphasizing rule chooses \(w_P>w_B\), reflecting the decision to give promoter accessibility greater influence. Neither rule is automatically correct for all genes or purposes: they define different summaries.

A different choice is to average within each region before combining them. If \(m_{Pg}\) and \(m_{Bg}\) are the numbers of eligible peaks in the promoter and body, respectively, one could use \(P_{cg}/m_{Pg}\) and \(B_{cg}/m_{Bg}\), then combine those region averages. This reduces the direct effect of having more peaks in one region, but changes the score’s meaning: it summarizes average count per eligible peak rather than total observed accessibility. It may also be unstable or undefined for a region with no eligible peaks. Empty regions require a prespecified handling rule, not an improvised denominator.

A useful justification for the sum is its additivity. If promoter and body counts are measured on disjoint peak sets, then the total count in their union is \(P_{cg}+B_{cg}\). More generally, for fixed weights, linearity of expectation gives \(E(S_{cg})=w_P E(P_{cg})+w_B E(B_{cg})\). This does not require the counts to be independent. It does require that the score be understood as a weighted sum of the selected counts; the equation does not establish that accessibility causes transcription or that the score is unbiased for RNA abundance.

Compare Rules on the Same Counts

Weight choices can change both score values and gene rankings. Thus, a score should not be reported without its region definitions and weighting rule. The following hypothetical comparison uses raw promoter and body totals in the same cell.

Region totals can also reflect how many peaks are available for counting. A gene body with many eligible peaks may accumulate more counts than a promoter with only a few, even when its average count per peak is lower. The next example contrasts totals with a simple per-peak summary.

Depth, Diagnostics, and Failure Modes

Raw scores can rise with sequencing depth because a cell with more measured fragments has more opportunities to contribute counts across peaks. Consequently, raw scores are not automatically comparable across cells with different library sizes. A practical workflow may scale counts by a cell-level depth factor before aggregation, or calculate raw gene scores and then apply an explicitly stated cell normalization. The choice should preserve clarity about what has been normalized: normalization changes the scale, but does not turn accessibility into RNA abundance. The earlier material on TF–IDF and LSI provides context for depth adjustment in ATAC-seq, but a gene activity score remains a distinct feature construction.

Before interpreting scores, check that the peak coordinates and gene annotations use the same genome assembly; that strand-aware TSS windows were constructed correctly; and that promoter-body overlaps were handled consistently. Inspect the number of eligible peaks per region and the fraction of genes with empty promoter or body sets. If one cell has unusually high scores across many genes, examine its total fragments and the normalization. If a handful of peaks dominate a score, inspect their counts and region assignments. For a sparse region, many zero counts can indicate limited coverage rather than biological absence of transcription.

Scores can also be sensitive to annotation choices. Alternative transcripts may have different TSSs, and overlapping genes can share accessible intervals. A score may change if the promoter window is widened, the gene body is defined differently, or peaks are assigned independently to multiple genes. These are reasons to document and, when relevant, compare reasonable constructions—not reasons to select whichever version produces a preferred result.

Why the Score Is a Proxy, Not a Transcript Measurement

ATAC-seq measures accessibility-related fragment counts, not transcription itself. Open chromatin can be present without a proportional amount of RNA, and transcription can depend on regulatory processes not represented by the selected promoter and gene-body peaks. A score can also miss relevant accessible regions outside the chosen intervals. Thus, a larger activity score means a larger value under the specified accessibility-summary rule; it does not, by itself, establish more RNA, active transcription, or a causal effect.

The score is useful as a compact gene-indexed representation of accessibility, especially when an analysis needs features associated with genes rather than only genomic peaks. Its limitations make interpretation conditional: state the scoring rule, check sensitivity to reasonable region and weight choices, and treat the result as a proxy. Direct assessment of how accessible peaks relate to gene expression requires additional analysis beyond constructing the score.

Common Mistakes

• Calling the score expression. The counts come from ATAC-seq fragments, not RNA molecules. Refer to the result as a gene activity score or accessibility-derived score.
• Leaving weights implicit. Equal weights, promoter emphasis, and per-peak averaging answer different summary questions. Report the rule used.
• Ignoring promoter orientation. Upstream must be defined relative to the gene’s strand; a fixed genomic-coordinate window can be wrong for negative-strand genes.
• Counting a shared interval without noticing. A peak can overlap the promoter and body or be assigned to more than one gene. State how such overlaps are handled.
• Comparing raw totals across unequal-depth cells as if directly comparable. Check library depth and describe any normalization before interpreting cell-to-cell differences.