Plant epimutations are not random drift: REED reads one stretch of DNA and predicts where methylation will go

A stretch of DNA itself specifies which way methylation at each CG site will drift, so the direction of epimutation can be read from sequence without waiting for generations of pedigree data. Previously a single cytosine's steady state could only be estimated from the epigenome itself, such as methylation variability across hundreds of accessions or chromatin states; such estimates cannot tell whether chromatin or DNA carries the steady state, and the data they need exist for almost no plant other than Arabidopsis. On held-out Arabidopsis chromosome 5, REED reached Pearson correlations of 0.92, 0.92 and 0.83 for CG, CHG and CHH, was more accurate than INTERACT in 89–94% of 12-kb windows, and had 38–70% lower error than three published methods.

Gains and losses of spontaneous methylation are not an unbiased random walk but are pulled toward the sequence-predicted set point: high-set-point sites gain methylation more readily and low-set-point sites lose it more readily. In two independent pedigrees, the Shaw population of nine lines over 31 generations and MA3 of two lineages over 11 generations, only the events themselves had been observed, with no per-site directional expectation to compare against. Among unmethylated sites, those in the top third of the set point gained methylation about 2,500 times more often than those in the bottom third, and among methylated sites the top third lost it 21 times less often; across 20 bins of the set point the rate-implied equilibrium matched it in value, with slopes of 0.95 and 0.99 in the two laboratories' data and 1.01 in MA3.

The set point belongs to no single genotype or environment: it predicts the direction of spontaneous change in wild populations, a crop germline and a long-lived tree alike. Earlier pedigree experiments were confined to a few Arabidopsis accessions, with no cross-species or cross-tissue test and no separation of whether stress changes direction or only speed. Salicylic acid and soil salinity accelerated gains and losses by 21–40% while the set-point effect and equilibrium slope were unchanged, moving the equilibrium by at most 0.06 within a preregistered equivalence margin of 0.1; in rice Huhan3, none of 330 gains fell among unmethylated sites in the bottom third of the set point and none of 180 losses among methylated sites in the top third.

The set point is determined by two kinds of rule written in DNA: cooperation between nearby CG sites, and ACGT-core transcription-factor motifs that lower methylation. The cooperative-maintenance model had taken methylable regions from chromatin data, and motif effects lacked a quantitative test against natural variation. Across 294,969 variant–site pairs in 586 accessions, natural variants that create or destroy a nearby CG changed methylation as REED predicted, with r of 0.54 and slope 0.82; among 805 plant motifs, the ACGT-core families made up 84% of the 80 strongest lowering motifs but only 16% of all motifs.

The demonstrated scope is Arabidopsis pedigrees and wild populations, the germline of two rice cultivars, somatic branches of a poplar tree about 330 years old, and Arabidopsis lines expressing a bacterial SssI fusion; set points come from each species' own REED model over 10 kb of sequence, and tests were run on chromosomes held out from training.

Borrowed set points gave the expected direction in 35 of 42 model–target combinations, with accuracy declining as divergence time grows, and own-species models remained the most accurate; stress sped gains and losses but barely shifted direction, with the equilibrium shift bounded within a preregistered equivalence margin of 0.1.