Genotype and structure above the probe
21. Genetic variation under the probe
Every section above treats a beta value as a measurement of methylation. It is also, in part, a measurement of genotype, and the two are not separable by looking at the number.
21.1 A probe sits at a genomic coordinate
Each probe has a chromosome and a base position, so the probe set can be intersected with any variant catalogue: ClinVar for clinically annotated variants, dbSNP for the general catalogue. The intersection is an ordinary range join, which GenomicRanges or bedtools will do, and it answers a question that matters before scoring: does this clock depend on probes whose signal a common variant can move.
21.2 Three ways a variant reaches the beta value
A SNP at the interrogated base. Single-base extension reads the variant rather than the methylation state, so the beta becomes a genotype readout. This is the population of probes the published masks remove. FALCONAge ships the Zhou masks (fa.preprocess.mask_report, apply_mask) but applies none automatically, because every clock was fitted on unmasked data; mask_report shows how much of each clock’s coefficient mass a mask would remove.
A SNP under the probe body. Hybridisation is weakened, intensity falls, and the beta drifts toward the middle of the range without any obvious signature of failure.
Cross-reactivity. A 50-mer against a three-base genome maps to more than one locus more often than the design intends. Chen and colleagues catalogued these for the 450K, and the affected probes carry a mixture of signals from several places in the genome.
21.3 meQTL: the variant is somewhere else entirely
A methylation quantitative trait locus is a variant associated with methylation at a CpG, usually nearby but sometimes far away. GoDMC and the EPIGEN meQTL resource map these genome-wide. The consequence for a clock is direct: if a clock CpG is under strong meQTL control, part of the between-person variance in that clock is genetic and fixed at conception rather than acquired, and no intervention will move it.
A clock’s between-person variance splits into an acquired component, which is what the clock is meant to measure, and a genetic component, which is not. The split is not visible in the score. Two people with the same clock age can differ because one has lived differently and the other inherited a different genotype at a handful of high-weight probes.
21.4 The arrays carry genotype probes on purpose
450K and EPIC include rsID-labelled SNP probes, 65 on the 450K, for sample identity. They confirm that two arrays came from the same person, which is the check that catches a swapped sample. MethylToSNP goes further and infers genotypes from the methylation pattern itself, which is useful and is also a reminder that array data is identifiable.
22. Structure above the single probe
22.1 A clock is a combination score
A clock sums hundreds of weighted features, which makes it the same kind of object as a polygenic risk score, and methylation risk scores for individual traits are built the same way: many published methylation biomarkers are proxies for proteins, metabolites and clinical values rather than clocks.
The shared property matters more than the shared arithmetic. A combination score is robust where a single probe is noisy, and it is opaque in the same measure: a change in the total does not say which component moved. That is the argument for clocks that report their components separately, as SystemsAge does for eleven systems.
22.2 Co-methylation modules
CpGs covary. Weighted correlation network analysis groups them into modules and summarises each by its first principal component, and the module can then be tested as one unit rather than as several thousand correlated tests. This is where the multiple-testing burden of an epigenome-wide scan becomes tractable, and it recovers structure that a per-probe analysis cannot see.
22.3 Allele-specific methylation
The two alleles at a locus can carry different methylation, and the pattern can follow a haplotype block. Array data cannot resolve this because a beta value averages both alleles into one number. Sequencing can, and tools such as MethHaplo combine allele-specific methylation with genotype to identify the regions involved. Imprinted loci are the familiar case and not the only one.
22.4 Interactions between variants are real and small
Genome-wide tests for SNP by SNP effects on methylation find them, and find that they explain little variance. The practical reading is that additive models are not leaving much on the table at this level.