Why Probe-Level Checks Matter
A methylation array does not observe methylation directly. It measures fluorescence generated when probe oligonucleotides bind and are extended, then uses channel intensities to estimate methylation at a CpG site. The earlier material on methylation array probe chemistry and beta values establishes how those signals produce a CpG-level estimate. Here the focus is whether the probe itself can be trusted to represent that CpG.
A probe can be problematic for different reasons. Its sequence may also bind elsewhere in the genome, a common variant may change how well it binds or extends, or the probe may produce weak or inconsistent signal. These problems are distinct: a probe can have a unique sequence but poor signal, or strong signal that is partly generated by an unintended genomic target. In either case, the resulting estimate can be misleading even if the calculation from intensity to beta value is performed correctly.
Three Sources of Probe Problems
Cross-reactivity occurs when a probe sequence can hybridize to more than one genomic location with sufficient similarity to contribute signal. The array then measures a combination of signal from the intended target and one or more off-target sequences. An off-target sequence can contribute signal whether or not it contains a methylated CpG relevant to the intended measurement.
Probe-overlapping variants are genetic variants that fall within the probe-binding sequence or at a base used in probe extension. A mismatch between the probe and a sample’s allele can change hybridization or extension efficiency. If the variant is near the assayed CpG, it may also change the CpG itself or its immediate sequence context. Either way, measured intensity can depend on genotype as well as methylation.
Poor performance refers to weak, unstable, or otherwise unreliable signal. Potential signs include low total intensity, frequent failure under the platform’s detection metric, large variation between technical replicates, or extreme values that are not reproducible. A weak probe may yield a numerical beta value, but that does not make the value informative.
The probe manifest supplies a starting point: the intended target, probe sequence, genomic coordinates, and, where available, annotations for known variants or non-unique alignments. Annotations depend on a reference genome build and on the variant databases and populations used. Before acting on an annotation, confirm that the coordinates and sequence correspond to the array version and reference build in use.
How a Probe Problem Distorts a CpG Estimate
A simple intensity model helps explain the direction of some biases. Suppose the target contributes methylated-channel intensity \(M_T\) and unmethylated-channel intensity \(U_T\). If an off-target sequence contributes \(M_O\) and \(U_O\), the measured channels are approximately the sums of the target and off-target contributions. Ignoring any offset for clarity, the measured beta value is then:
This is a conceptual additive model, not a claim that all array chemistry is perfectly additive. It shows that the measured estimate is influenced by the relative signal contributions and methylation states of both targets. If the off-target signal is substantial, the result can be pulled away from the intended CpG’s methylation level.
A probe-overlapping variant can create a different distortion. If allele-dependent binding reduces intensity in one channel, the beta value can shift even when the true methylation fraction is the same across genotypes. The artifact may therefore resemble a biological association between genotype and methylation. Conversely, reduced signal in both channels can make a probe noisy or cause it to fail detection rather than creating a consistent directional shift.
Worked Example: Off-Target Signal Pulls a Beta Value
Suppose a hypothetical probe targets a CpG with target intensities \(M_T=80\) and \(U_T=20\), corresponding to a target beta value of \(80/(80+20)=0.80\). A similar sequence elsewhere contributes off-target intensities \(M_O=10\) and \(U_O=90\). Under the simple additive model, the measured intensities are:
The measured beta value is \(90/(90+110)=0.45\). Thus, an off-target contribution equal in total intensity to the target contribution changes the estimate from 0.80 to 0.45. The result does not imply that the intended CpG is 45% methylated; it is the estimate from a mixture of the two hypothetical sources.
Combining Annotation and Signal Evidence
A defensible review uses both the sequence-based evidence and the observed behavior of the probe. Annotation can predict a plausible source of bias even when the observed signals look orderly. Signal summaries can reveal practical failure even when the available annotations do not identify an obvious sequence problem. Neither type of evidence alone is a complete validation.
Check the array manifest, probe sequence, target CpG, and reference genome build. A coordinate mismatch can produce a false overlap or hide a real one.
Determine whether the sequence has plausible additional genomic matches and whether known variants overlap the binding or extension region. Note the variant source and population context.
Summarize intensity and detection behavior across samples, and examine technical replicates where available. Look for persistent low signal, frequent detection failure, or poor reproducibility.
Use the study’s prespecified criteria and the strength of combined evidence. Record the reason and assess whether conclusions depend on the decision.
Detection metrics and recommended thresholds vary by platform, software, and preprocessing stage. A detection p-value, for example, is a platform-defined measure comparing observed signal with a background reference; it is not a universal probability that the probe is biologically correct. Report the metric and rule actually used rather than presenting one cutoff as a law of methylation measurement.
Worked Example: A Variant Near the Probe Extension Site
Consider a hypothetical probe whose annotation places a common variant at the base used for extension. In samples with one genotype, the probe yields methylated and unmethylated intensities of 70 and 30, giving \(\hat{p}=70/(70+30)=0.70\). In samples with another genotype, a mismatch reduces the measured methylated-channel intensity, producing intensities of 35 and 65, so \(\hat{p}=35/(35+65)=0.35\).
The numerical difference is \(0.70-0.35=0.35\), or 35 percentage points. If the true methylation were identical in the two genotype groups, this difference would be a measurement artifact in this hypothetical example, not a methylation difference. The annotation gives a mechanism that could explain the pattern; the two sets of signal values alone do not prove that the variant caused it.
A careful conclusion is therefore that the probe has a plausible genotype-dependent measurement risk. The analyst should flag or exclude it under the prespecified probe criteria and avoid interpreting its apparent genotype pattern as methylation biology without independent evidence.
Signal Evidence and Poor Performance
Probe performance should be assessed across the samples in the data set, but the question here is probe-specific: does this measurement work reliably enough to represent its intended CpG? Useful summaries include the fraction of samples failing a platform-defined detection rule, the distribution of total signal, and variation among technical replicates. A single weak observation may reflect a sample-specific issue; repeated weakness for one probe across many samples is stronger evidence of a probe-level limitation.
A beta-value distribution can also prompt investigation. For example, a probe with values concentrated at implausible extremes or with unusually erratic replicate values may warrant review. However, a bimodal distribution is not itself proof of failure: it could reflect biology, genotype, cell composition, or another feature of the samples. Distributional patterns are diagnostic clues, not unique explanations.
Worked Example: Flagging a Weak, Inconsistent Probe
For a hypothetical probe measured in 20 samples, suppose the platform-defined detection rule is failed in 7 samples. The failure fraction is:
Suppose also that its total intensity is low relative to most other probes on the same array, and that technical replicates give beta values 0.20, 0.80, 0.30, and 0.70. Their range is \(0.80-0.20=0.60\). These facts provide mutually consistent evidence of poor performance: frequent detection failure, weak signal, and substantial replicate disagreement.
The conclusion is not that 35% is a universal exclusion threshold or that every measured beta value is wrong. Rather, under a prespecified rule that treats repeated detection failure and poor replicate agreement as probe-level concerns, this probe should be flagged for exclusion or sensitivity analysis. The stated evidence supports that decision for this hypothetical data set.
Choosing a Response and Recognizing Limitations
Common responses include excluding probes with strong, well-documented sequence risks; flagging uncertain probes; or retaining probes while reporting sensitivity analyses with and without them. The choice depends on the purpose of the analysis, the quality of annotations, and the prespecified rules. Filtering solely by annotation may miss poor performers, while filtering solely by signal may retain a consistently strong but cross-reactive probe. A combined review is generally more informative.
No annotation set is complete. Known variant frequencies depend on sampled populations, and a variant described as uncommon in one reference population may be more frequent in the study population. Sequence alignment depends on the reference build and alignment choices. Signal patterns, in turn, do not uniquely identify their cause. These limitations argue for documenting the evidence and uncertainty rather than treating a flag as a definitive diagnosis.
Probe-level review is distinct from sample-level quality control. A probe can be unreliable across many samples, whereas a sample can have poor measurements at many otherwise acceptable probes. The follow-on sample-level assessment considers that complementary question; probe review does not substitute for it. Likewise, identifying a problematic probe is not the same as correcting its signal or testing for differential methylation.
Common Mistake
- Wrong reasoning: “The probe has a high total intensity, so its measurement must be trustworthy.”
- Correction: High intensity does not establish specificity. A cross-reactive probe can produce strong signal from both the intended and unintended targets. Evaluate sequence specificity and signal behavior together.
- Wrong reasoning: “A common variant is annotated nearby, so this probe must be discarded.”
- Correction: Confirm that the variant overlaps a relevant part of the probe or target, and consider the annotation’s population and build. A nearby variant alone is not evidence of altered hybridization.
- Wrong reasoning: “A bimodal beta distribution proves cross-reactivity.”
- Correction: Bimodality is a clue, not a mechanism. Genetic effects, biological subgroups, or other sample features can also produce two modes; seek supporting annotation or replicate evidence.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A probe’s target contributes methylated and unmethylated intensities of 60 and 40. An off-target sequence contributes 20 and 80. Under the additive model, what beta value is measured?
- Why can a variant at a probe-extension base create an apparent methylation difference between genotype groups even if their true methylation is the same?
- A probe fails a detection rule in 3 of 15 samples. Calculate the failure fraction and state one reason not to treat that fraction as a universal decision threshold.
- Why does strong total intensity not rule out cross-reactivity?
- A probe’s beta values are bimodal, but there is no variant or cross-reactivity annotation. What is a cautious interpretation?
Solutions. (1) The measured methylated intensity is \(60+20=80\), and total intensity is \(60+40+20+80=200\), so the measured beta value is \(80/200=0.40\). (2) The variant can change binding or extension efficiency, altering the measured channel intensities and therefore the beta value independently of methylation. (3) \(3/15=0.20=20\%\). Detection metrics and appropriate thresholds vary by platform and analysis criteria, so the fraction is evidence to review, not a universal cutoff. (4) Off-target binding can add substantial fluorescence; high signal does not show that all signal came from the intended CpG. (5) Bimodality is a diagnostic clue, not proof of a faulty probe. Consider other biological or sample-related explanations and seek corroborating evidence before flagging or excluding it.