Why Correct Intensities Before Interpreting Methylation
A methylation array measures fluorescent intensities, not methylation percentages directly. The recorded signal for a probe can include both probe-specific foreground and nonspecific background. In addition, the red and green channels can respond differently because of dye, scanning, or other technical properties. If these effects are left unaddressed, the resulting methylated and unmethylated signals may be distorted before they are converted to beta values or M-values.
Background correction and dye-bias adjustment target different problems. Background correction estimates signal unrelated to specific probe binding and reduces its contribution. Dye-bias adjustment addresses systematic differences between the two color channels. Both are usually applied to the intensity data, before calculating methylation summaries. Neither operation repairs a failed sample or probe, and neither guarantees that technical artifacts have been eliminated.
The probe design matters. As established in the earlier material on Infinium probe designs, Type I and Type II probes use the color channels differently. A correction must respect those design features and the controls available on the array; a single rule applied indiscriminately to every intensity can mistake expected probe behavior for an artifact. The aim is to correct a technical pattern supported by control information, not to make all probes or samples look alike.
Background Correction: Estimate Noise, Then Reduce It
Arrays include control measurements that can inform background estimation. Negative controls are intended to measure signal in the absence of specific target binding. Some preprocessing approaches also use out-of-band measurements: signal recorded in a channel not expected to carry the specific probe signal under that probe design. These measurements provide information about nonspecific signal, although their suitability depends on the array and the implementation.
A simple offset correction estimates a background level \(b\) from controls and computes a corrected intensity \(\max(Y-b,0)\). This is useful for understanding the basic idea, but it treats background as a fixed amount. More model-based methods, including normal-exponential approaches such as Noob, represent observed intensity as a combination of a background distribution and a nonnegative foreground signal. They estimate a corrected foreground intensity using the fitted background and signal distributions. The correction is therefore not necessarily identical to subtracting one control average from every probe.
Worked Example: A Simple Background Offset
Suppose a hypothetical array has negative-control intensities whose estimated background level is \(b=180\) intensity units. A probe has observed intensity \(Y=1{,}200\) units. Under the simple offset rule, the corrected intensity is:
The correction reduces this probe's intensity by 180 units, leaving 1,020 units. For a second hypothetical probe with \(Y=130\), the same rule gives \(\max(130-180,0)=0\), not \(-50\), because intensity is constrained to be nonnegative under this rule. A zero after correction is a warning that the probe signal is at or below the estimated background, not evidence that the underlying biological methylation is exactly zero.
This arithmetic illustrates offset subtraction only. It does not claim that every array pipeline estimates background with one constant or uses this exact correction. A distribution-based method can produce a different corrected value because it accounts for uncertainty in the signal and background components.
A short justification shows why low intensities are especially sensitive to background estimation. If the observed signal is \(Y=S+B\) and the estimated background is \(b\), then an untruncated offset estimate is \(Y-b=S+(B-b)\). Its error relative to the foreground \(S\) is \(B-b\). Thus, if \(b\) is too small, residual background remains; if it is too large, genuine foreground is removed. When \(S\) is large, a modest error may be a small fraction of the corrected value. When \(S\) is near zero, the same error can dominate it. This is why correction quality should be assessed across the intensity range, not just by checking a single average.
Dye-Bias Adjustment: Correct Channel-Dependent Scale
A dye-related artifact occurs when otherwise comparable signal is systematically stronger in one channel than the other. It can shift methylated and unmethylated intensities differently, depending on probe design and channel assignment. A simple model for illustration is that a channel's recorded intensity equals its underlying signal multiplied by a channel-specific scale factor. If the red channel is consistently amplified relative to green, raw red-to-green comparisons can therefore be biased.
Array control probes can help estimate channel differences. A simple adjustment might rescale each channel using factors estimated from suitable controls. More flexible procedures may estimate a systematic relationship across control intensities rather than assume one constant multiplier. The defining principle is that the correction should be estimated from information intended to reflect technical channel behavior, and then applied in a way that respects the array's probe design. The precise control set and fitting method vary by preprocessing procedure.
Worked Example: Rescaling Two Channels
Consider a hypothetical array with suitable red- and green-channel controls. Their median intensities are 1,200 and 800 units, respectively. For a simple illustration, choose the geometric mean of these medians as a common target:
The corresponding scale factors are \(f_R=979.8/1{,}200\approx0.8165\) for red and \(f_G=979.8/800\approx1.225\) for green. Suppose a hypothetical probe has methylated signal 1,600 units in red and unmethylated signal 800 units in green. After these illustrative channel adjustments:
The corrected intensity ratio is approximately \(1{,}306/980=1.333\), compared with the unadjusted ratio \(1{,}600/800=2.000\). The example shows how channel scaling can change a ratio used in a methylation summary. It does not establish that the raw ratio was entirely technical: the adjustment is justified only if the control-based scale difference represents dye bias relevant to these signals. Actual procedures may use different controls and may estimate intensity-dependent corrections rather than these two factors.
For context, beta values summarize the relative methylated and unmethylated intensities, often with an offset to stabilize low-intensity values; M-values use a log ratio of those intensities. Because both summaries depend on corrected intensities, background and dye adjustments can alter their values. A correction should be evaluated at the intensity level as well as in the resulting methylation summaries, and the same preprocessing strategy should be applied consistently to samples being compared.
Diagnostic Plots and Failure Modes
Diagnostics ask whether the correction removed a plausible technical pattern without creating a new one. No single plot proves that the adjustment is correct. Examine plots by sample and, where useful, by probe design or channel. Control measurements are particularly informative because they help separate correction behavior from real differences among methylation probes.
| Diagnostic | Pattern to inspect | Possible concern |
|---|---|---|
| Control intensity distributions by channel | Red and green controls before and after adjustment | A persistent channel separation can indicate inadequate correction; a forced match despite incompatible control behavior can indicate overcorrection. |
| Negative-control or background distribution | Low-end intensities and residual control signal | A long residual tail may suggest background remains; excessive collapse of low intensities can indicate signal has been removed too aggressively. |
| Intensity distributions by sample | Density or quantile plots before and after correction | One sample remaining markedly different may need further investigation; identical-looking distributions are not automatically desirable if genuine sample differences are expected. |
| MA-style channel diagnostic | Difference between channel log intensities versus their average, particularly among controls | A systematic trend in controls can indicate intensity-dependent dye bias; a curve removed from all probes indiscriminately may also remove biological signal. |
An MA-style plot can be defined for paired channel measurements as \(M=\log_2(R/G)\) against \(A=\tfrac12\log_2(RG)\), where \(R\) and \(G\) are positive red- and green-channel intensities. A trend in \(M\) across \(A\) among suitable control measurements suggests that channel imbalance depends on intensity. For methylation probes, however, a nonzero or structured \(M\) can reflect genuine methylation differences or probe design. It is not generally appropriate to force every methylation probe's \(M\) value to zero.
Worked Example: Reading a Control MA Pattern
Suppose a hypothetical diagnostic plot shows that, among the selected control pairs, mean \(M\) is 0.40 at low \(A\) and 0.10 at high \(A\). Since \(M=\log_2(R/G)\), these values correspond to channel ratios \(R/G=2^{0.40}\approx1.320\) and \(R/G=2^{0.10}\approx1.072\), respectively. The controls therefore show a larger red-versus-green imbalance at lower average intensity.
If an intensity-dependent correction is fitted from appropriate controls, a useful follow-up is to make the same diagnostic plot after adjustment. If the control trend is greatly reduced and the low-end control distributions remain plausible, that supports the intended correction. If the plot instead shows a reversed trend, the adjustment may have gone too far or the fitting procedure may be unsuitable. This plot alone cannot determine whether biological probe signals were preserved; that requires checking the correction's scope and comparing probe behavior by design.
Common warning signs include residual channel trends among controls, markedly different background distributions across samples after correction, or a large shift in many corrected intensities that is driven by a small number of atypical controls. Investigate these patterns rather than automatically increasing correction strength. Check control quality, sample-level quality-control results, channel assignments, and whether the chosen procedure matches the array type. A correction cannot compensate for failed hybridization, saturation, or poor probe performance.
Choosing and Reporting a Correction
A simple offset is transparent and can be useful for illustrating the role of background, but it assumes a common additive background estimate. A model-based background method uses a distribution for signal and noise and can handle uncertainty more explicitly, but depends on its model and controls. A simple channel rescaling is easy to interpret, while an intensity-dependent adjustment can address trends that a constant factor misses. More flexibility is not automatically better: a fitted curve can also track biological structure if it is estimated from inappropriate measurements.
The correction should be chosen with the array's control structure and probe design in mind. Inspect diagnostics before and after adjustment, retain a record of the method and settings, and avoid choosing a procedure solely because it makes all sample distributions look similar. These steps support reproducibility and help distinguish an appropriate technical adjustment from a transformation that changes the biological quantity of interest.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A probe has observed intensity 75 units and the estimated offset is 100 units. What does the simple offset rule return, and what does that result mean?
- Why can a control-based channel correction change a methylation beta value?
- A control MA plot has a trend before correction but no trend afterward. Is this alone enough to establish that the correction was valid? Explain.
- What is the key difference between a constant channel rescaling and an intensity-dependent dye correction?
- Why is it a mistake to force all methylation-probe MA values to zero?
Solutions. (1) The result is \(\max(75-100,0)=0\) units. The observed signal is at or below the estimated background under this rule; it does not prove that the true biological signal is zero. (2) A beta value depends on the relative methylated and unmethylated intensities. If the correction rescales or otherwise changes either intensity, their ratio and therefore the beta value can change. (3) No. The control trend's reduction is supportive, but one must also check that the controls are suitable, the adjustment has not reversed the trend, and the procedure has not imposed an inappropriate change on biological probes. (4) A constant rescaling uses one factor per channel across intensities; an intensity-dependent correction allows the estimated channel difference to vary with signal level. (5) Probe-level differences can reflect genuine methylation and probe-design behavior, not just dye bias. Forcing every value to zero can erase biological signal; correction should target justified technical structure, especially as assessed in suitable controls.