Single-Cell RNA and ATAC Integration for Cell-State Programs

Why Pair RNA With Chromatin Accessibility?

Single-cell RNA sequencing measures transcripts, while single-cell ATAC sequencing measures accessible regions of chromatin. The measurements answer related but different questions: which genes are expressed, and which regulatory regions are accessible? When both assays are performed on the same cells, their measurements can be aligned cell by cell. This pairing helps characterize cell states and investigate candidate regulatory programs, but it does not by itself show that an accessible region caused a change in gene expression.

The data are naturally represented by two matrices with the same cells as rows. Let \(R_{ig}\) denote the RNA measurement for gene \(g\) in cell \(i\), and let \(A_{ip}\) denote the ATAC measurement for peak \(p\) in cell \(i\). An ATAC peak is a genomic interval with detected accessibility. RNA features are genes; ATAC features are genomic intervals, so matching the cells does not automatically identify which peak regulates which gene.

Reliable integration requires more than a shared cell barcode. Cell identities must be matched correctly, both assays must have usable quality, and technical differences in depth or batch must not be mistaken for biology. Cell-state comparisons also need biological replication: many cells from one donor provide detailed information about that donor, but they do not substitute for independent donors.

From Paired Measurements to Cell-State Evidence

A typical analysis first checks cell and assay quality, then represents each modality in a form appropriate to its data. RNA counts reflect transcript capture; ATAC counts reflect fragments overlapping peaks. Their raw values have different scales and sparsity, so adding raw counts across modalities is generally not meaningful. Analysts commonly normalize within each assay and construct modality-specific representations before examining how the same cells relate across them.

A joint representation can help identify cells with similar profiles in both layers. It should be checked for domination by one modality, separation by library size or batch, and loss of meaningful variation. A cluster is a statistical grouping, not automatically a cell type or a biological state. Labels should be supported by interpretable features, such as expression of characteristic genes and accessibility at relevant promoters or candidate regulatory regions. A marker used to define a state is not independent validation of that same state.

Peak-to-gene links require additional evidence. A peak near a gene’s promoter is a plausible candidate, but proximity alone does not establish a functional link. A distal peak may be considered when accessibility and expression covary across suitable observations, or when prior biological evidence supports the connection. Such associations can arise because both features respond to cell state, donor, or another factor. When feasible, stronger regulatory claims require perturbation or other evidence beyond observational co-variation.

Make Replicates, Not Cells, the Basis for Inference

For a state comparison, a defensible strategy is often to aggregate measurements within each biological replicate and state, then compare replicate-level summaries. For example, aggregate cells from each donor assigned to state A and state B separately, and calculate a normalized expression or accessibility summary for each donor-state combination. The donor, not each cell, is then the independent unit for a population-level comparison.

When each donor contributes both states, define a within-donor difference \(d_j\), such as the state-A minus state-B expression value for donor \(j\). With \(n\) independent donors, the estimated mean difference is \(\bar{d}\), its standard error is \(s_d/\sqrt{n}\), and a paired \(t\) interval is \(\bar{d}\mathbin{\pm}t_{1-\alpha/2,n-1}s_d/\sqrt{n}\). This approach assumes independent donor-level differences and, for the small-sample \(t\) interval, an approximately normal distribution of those differences. Pairing can account for donor-wide baseline variation; it does not remove confounding if state membership is systematically tied to batch or another factor.

This identity explains why paired comparisons can be more precise: if measurements from the two states tend to move together within a donor, their positive covariance reduces the variability of the difference. It also clarifies why treating thousands of cells from one donor as thousands of independent replicates gives misleading precision. Models for counts or repeated samples may be appropriate in other designs, but the inferential unit and dependence structure still need to be represented correctly.

Evaluating Candidate Peak–Gene Links

A candidate link can be assessed by asking whether accessibility at a peak and expression of a gene vary together across comparable observations. For example, one may calculate a correlation after aggregating cells into donor-by-state summaries. Such a correlation is descriptive evidence, not a causal estimate. A pooled cell-level correlation can be especially misleading if both measurements differ between cell states: the association may reflect state separation rather than a within-state regulatory relationship.

Diagnostics, Alternatives, and Limits

Before reporting a cross-layer program, inspect whether the integrated result is biologically interpretable and technically credible. Warning signs include clusters that mainly separate by donor or assay batch, state labels supported by only one modality, candidate links that disappear when one donor is omitted, and apparent precision based on cell counts rather than biological replicates. Also examine whether conclusions change under reasonable normalization or integration choices. Stability is useful evidence, but it does not replace independent validation.

A common mistake is to treat a nearby accessible peak and higher gene expression as proof that the peak activates that gene. The correction is to describe the observations as a candidate regulatory association, report how the link was proposed and tested, and distinguish cross-layer consistency from causal evidence. A second frequent error is to use cells as independent replicates in a state comparison. The correction is to respect donor-level dependence and base population inference on biological replicates or a model that explicitly represents the dependence.

These methods are valuable because they connect cellular identity to both expressed genes and candidate regulatory context. Their conclusions remain limited by sparse measurements, uncertainty in cell labels and peak-to-gene assignments, donor and batch structure, and the observational nature of the data. Report which cells and donors were analyzed, how states and candidate links were defined, how uncertainty and multiple testing were handled, and whether the claim is descriptive, associational, or supported by perturbation.