Beta Values and M Values for CpG Sites

Two Scales for a CpG Methylation Signal

A methylation array reports signal associated with methylated and unmethylated versions of a CpG site. A compact summary of those signals is useful, but the choice of scale changes how the result looks and how it should be interpreted. Beta values express relative methylated signal on a bounded scale, while M values express a log-ratio on an unbounded scale. Neither scale turns array intensity into a direct count of methylated molecules.

The earlier material on methylation array probe chemistry and signal channels establishes the source of these two signal components. Here, write \(X\) for the measured methylated intensity and \(Y\) for the measured unmethylated intensity for one CpG site in one sample. Initially assume both are positive and measured on a comparable scale. These are signal intensities, not necessarily integer counts.

The formulas are summaries of signal. A beta value near 0.8 means that the methylated channel contributes about 80% of the sum of these two intensities. It does not, by itself, establish that exactly 80% of cells, DNA molecules, or alleles are methylated. Probe behavior, signal processing, cell mixture, and measurement noise can all affect the observed intensities.

Deriving the Conversion Between Scales

The two summaries are linked exactly when they are calculated from the same positive pair \(X,Y\). Starting with the M value, exponentiation gives the methylated-to-unmethylated intensity ratio. Substituting that ratio into the beta formula gives the conversion:

Conversely, because \(1-\beta=Y/(X+Y)\), the ratio \(\beta/(1-\beta)\) equals \(X/Y\). Taking its base-2 logarithm recovers M:

Thus \(M\) is the base-2 log-odds of the beta value. The word “odds” describes the mathematical transformation of the relative signal; it does not make beta a calibrated probability of methylation for an individual molecule. If \(X=Y\), then \(\beta=0.5\) and \(M=0\). If \(X>Y\), then \(\beta>0.5\) and \(M>0\); if \(X<Y\), then \(\beta<0.5\) and \(M<0\).

The conversion also explains why equal changes on the two scales do not have a constant correspondence. Differentiating \(\beta=2^M/(1+2^M)\) with respect to \(M\) gives:

This slope is largest at \(\beta=0.5\) and shrinks as beta approaches 0 or 1. Therefore, a fixed M-scale change corresponds to a smaller beta-scale change near the extremes than near the middle. The transformation is monotone, so it preserves the ordering of samples at a site when all are calculated from the same positive intensities, but it does not preserve differences or distances.

Worked Comparisons Across Samples

Beta differences are often easy to communicate as differences in relative signal, while M differences describe ratios of ratios. To see the distinction, use two hypothetical samples measured at the same site with comparable processing. Assume their intensities are positive and use the unadjusted formulas for both samples.

Offsets, Zeros, and Implementation Choices

Real measurements can be near zero, and the logarithm in an M value is undefined when either intensity is zero. Implementations therefore may use an offset or other preprocessing. There is no single offset convention that can be assumed without checking how the values were produced. If a common positive offset \(a\) is added to both channels, define adjusted values using \(X+a\) and \(Y+a\). The exact beta–M conversion still holds if both summaries use that same adjusted pair:

An offset can have a substantial effect when the signals are small, because it contributes a larger fraction of the adjusted values. For large signals relative to \(a\), its proportional influence is smaller. Offsets also change the numerical estimand: the adjusted beta and M summarize the adjusted signals, not the original unadjusted pair.

Some definitions use an offset in the beta denominator only, for example \(X/(X+Y+c)\), while the M value may be calculated as \(\log_2(X/Y)\). Those two definitions are not exact inverses of one another. For the example above with \(c=1\), the denominator-only beta is \(8/11\approx0.7273\), whereas the unadjusted M value remains 2. When comparing results across software or studies, check the actual formulas, offset choices, and signal preprocessing rather than assuming the conversion applies unchanged.

Choosing a Scale for Interpretation and Modeling

Beta values are often useful for plots and summaries intended to communicate relative methylated signal in an intuitive way. M values are often considered for statistical modeling because a log-ratio can be more symmetric than a bounded proportion-like measure and avoids piling values against the beta boundaries. These are tendencies, not guarantees: the distribution and variance depend on the measurements and preprocessing. A transformation alone does not ensure normal residuals, constant variance, independence, or valid inferential calibration.

The scale should match the question and the downstream model. A beta difference is in units of relative signal share; an M difference is in log-base-2 ratio units. If results are reported on one scale after fitting or summarizing on another, state the conversion and its assumptions. A nonlinear transformation means that a group difference on the M scale cannot generally be converted into one constant beta difference that applies at every baseline beta value.

Diagnostics and Common Pitfalls

Before interpreting either scale, check that intensities are finite and that the chosen formula is defined. Inspect the distributions of both channels, the resulting values, and the frequency of values at or near the boundaries. Unexpected beta values outside the intended range, non-finite M values, or substantial numbers of zeros can indicate a mismatch in signal handling or a need to examine preprocessing. Extreme values may also reflect saturation or weak signals; the transformation cannot resolve the underlying measurement problem.

A common mistake is to call beta a literal methylation percentage and then interpret a difference of 0.10 as exactly ten percentage points of molecules methylated. The correction is to describe beta as a relative signal share under the specified formula; stronger biological interpretations require appropriate measurement assumptions and evidence. Another mistake is to treat beta and M differences as numerically equivalent. The conversion is nonlinear, so report the scale and units, and do not translate an M difference into a fixed beta difference without a baseline value.

A further pitfall is comparing values produced with different offsets or definitions as though they were the same quantity. Record whether an offset was used, whether it was added to both channels or only the denominator, and whether values were transformed before calculation. The exact conversion derived above requires both summaries to use the same positive pair of intensities.