Sample-Level Quality Control for Methylation Arrays

Why sample-level quality control matters

A methylation array measures many CpG sites for each sample. A sample with widespread weak signal, failed conversion, or other technical problems can produce unreliable measurements across many sites. This is different from a problematic individual probe: the question here is whether an array, considered as a whole, provides data that are fit for the planned analysis. Sample-level quality control (QC) uses several diagnostics together; no single plot or cutoff reliably answers that question in every study.

The diagnostics considered here are detection p-values, control-probe signals, distributions of methylated and unmethylated signal, and sample clustering. They are useful before downstream analyses because they can reveal broad technical failure and possible sample mix-ups. QC is an evidence-based decision, not a procedure for deleting every unusual-looking sample. A sample can be unusual for biological reasons, and a technical concern can be real even when the sample does not look like an extreme outlier.

Detection p-values and sample-level summaries

A common approach estimates background from negative-control probes and compares the observed probe signal with that background. Exact calculations depend on the array platform and processing software. The interpretation is consistent: a detection p-value near zero indicates a signal that is unlikely under the background model, while a larger value warrants concern about that probe's measurement. A failed detection does not establish that the CpG is biologically absent or unmethylated.

For a sample with \(m\) probes, choose a detection-p-value cutoff \(c\) as part of a documented QC plan, then summarize the fraction of probes that fail:

where \(p_j\) is the detection p-value for probe \(j\), and \(I(\cdot)\) equals 1 when its condition is true and 0 otherwise. This fraction is a descriptive sample-level summary. Probes with missing or invalid measurements should be handled explicitly rather than silently counted as passing. Also examine whether failures are concentrated in particular probe types or regions if that information is relevant to the platform and analysis.

A cutoff such as \(c=0.01\) is an example, not a universal rule. The suitable cutoff and sample-level action threshold depend on the platform, software, assay documentation, and study plan. Detection p-values across probes need not be independent, so the failed-probe fraction should not automatically be treated as a binomial test with independent trials.

Control probes and intensity distributions

Array control probes assess technical steps using signals whose expected behavior is known or specified by the platform. Depending on the array, control categories can provide information about processes such as staining, extension, hybridization, target removal, bisulfite conversion, and assay specificity. They are technical checks, not measurements of methylation at biological CpG sites. Examine the relevant categories separately: a broadly weak control pattern suggests a different concern from one isolated control outside its expected range.

Compare control signals with platform documentation and, when appropriate, other arrays processed under comparable conditions. The direction and acceptable range depend on the control type and platform; one generic intensity threshold cannot be applied to all controls. Record which controls failed, how far they departed from the expected range, and whether multiple controls point to the same technical problem. Control failures can flag a sample for review even when the overall signal looks plausible.

Also inspect methylated and unmethylated signal distributions separately. A sample with low signal in both channels may have inadequate overall intensity. A markedly abnormal pattern in one channel, a compressed distribution, or an unusual shape relative to comparable samples can also signal trouble. These observations are prompts for investigation, not diagnoses by themselves: legitimate biological differences can shift methylation-related summaries, and array design affects the intensity pattern. At this stage, the aim is to assess whether a sample's measurements are credible, not to adjust background or dye bias.

Using sample clustering as a diagnostic

A sample-level clustering plot can show whether an array is isolated from the others. Clustering may use QC summaries, control-probe signals, or measurements at a shared set of probes that passed relevant probe-level checks. If variables have very different scales, scale them appropriately before calculating distances; otherwise a variable with large numerical values can dominate. State what measurements were used and how missing values and scaling were handled.

An isolated sample is a flag, not a verdict. Clustering based on methylation profiles can separate samples because of real biology, such as tissue or cell composition, as well as technical problems. Conversely, a technical failure shared by several arrays may not appear as a single isolated branch. Evaluate clustering alongside sample annotations, control results, and intensity diagnostics rather than interpreting a dendrogram as an exclusion rule.

Combining evidence and making an exclusion decision

A defensible QC decision specifies the evidence and its relevance to the analysis. A sample with a high failed-probe fraction, multiple abnormal controls consistent with the same technical problem, and weak or distorted channel distributions has converging reasons for concern. A sample flagged by only one mild diagnostic, with the other checks satisfactory, generally calls for review rather than immediate removal. In either case, check sample labels, processing records, and other available metadata; a possible labeling error has implications beyond a low signal summary.

There is no universal numerical cutoff that justifies exclusion for every methylation array, tissue, and study. Establish criteria using platform guidance and study needs, and apply them consistently. Excluding an array changes the analyzed sample set; if the excluded samples differ systematically from those retained, the resulting conclusions may apply to a narrower population. Preserve QC summaries and the reasoning behind each decision so that the procedure is auditable.

Common pitfalls and limitations

• Treating a detection p-value as a methylation test. It assesses signal relative to background, not whether methylation differs between groups.
• Using one cutoff as a universal law. A threshold is a QC rule whose suitability depends on platform, processing, and study context; state and justify it.
• Removing every cluster outlier. A distinct biological profile can be valid. Check other diagnostics and sample context before exclusion.
• Trusting good overall intensity alone. Adequate channel medians do not rule out problems indicated by controls or detection failures.
• Confusing sample failure with probe failure. An individual bad probe does not by itself make the whole sample unusable; widespread or coherent problems are more relevant to sample-level decisions.

These checks have limits. They identify evidence of technical concern but do not guarantee that every retained measurement is unbiased, nor can they always identify the cause of an abnormal pattern. Keep sample QC separate from downstream correction and biological modeling, and avoid interpreting technical diagnostics as biological results.