Mapping the Genetics of Defense Memory in Rapeseed Cultivars

Plants can be "primed." After a harmless treatment, such as a silicon spray, they respond faster and more strongly when a real pathogen arrives days later, even though their resting gene expression has returned to normal in between. This is a form of innate immune memory. A recurring frustration is that priming treatments work in one cultivar and fail in another, and no one knows why. We propose to treat the memory itself as a heritable molecular trait. For each gene in each cultivar, we measure how much extra response to a pathogen challenge priming adds. We then ask which DNA variants control that extra response, and whether those variants explain which cultivars can be primed at all.

The core design has two parts. First, we use hybrids between one common rapeseed parent and fifteen diverse founder lines. Inside each hybrid, both parental copies of a gene sit in the same cell, so any difference in how the two copies respond to priming must come from the DNA next to the gene. Second, we use the recombinant inbred lines descended from the same crosses, a population whose genomes are random mixtures of two parents, to locate distant regulators of memory. The result would be the first map of "memory expression quantitative trait loci" (memory-eQTLs) in a crop. An expression quantitative trait locus is a DNA variant whose genotype predicts how strongly a gene is expressed. The map would come with a direct test of whether cultivar-specific priming failure is written in nearby gene regulation or in distant regulatory hubs.

Core Claim

The claim is that a subset of genetic variants changes a gene's response to a pathogen only in plants that were primed earlier. These variants are distinct from variants that control the ordinary pathogen response, and they account for much of the cultivar-to-cultivar variation in how well priming protects. The quantity being mapped is an interaction contrast.

Write $y_{a,c}$ for the log expression of one parental allele $a$ of a gene under condition $c$. The four conditions are mock-then-mock (MM), primed-then-mock (PM), mock-then-challenged (MC) and primed-then-challenged (PC). The memory effect of that allele is

$$m_a = (y_{a,PC} - y_{a,PM}) - (y_{a,MC} - y_{a,MM}),$$

the extra challenge response that priming adds. A cis memory effect is a difference $m_{\text{founder}} - m_{\text{ZS11}}$ between the two alleles inside one hybrid; cis means the variant sits next to the gene it controls.

We call an effect "memory" only if two conditions hold. The effect must be significant, and the primed-resting state must have returned to baseline, so that $y_{a,PM} \approx y_{a,MM}$. The second condition separates true memory from priming-induced expression that simply has not faded yet.

Why This Is Not Already Done

The literature in hand motivates the gap but does not fill it. The potato study Foliar Application of Silicon Enhances Resistance Against Phytophthora infestans Through the ET/JA- and NPR1- Dependent Signaling Pathways in Potato showed that silicon primes resistance through ethylene, jasmonic acid and NPR1 signalling. The work used one cultivar, 'Désirée', with targeted gene measurements, and the authors flag generalization across cultivars as their main open question. The review Advances in Plant Epigenetic Regulation of Abiotic Stress Response catalogs methylation, histone and small-RNA mechanisms of stress memory, largely from Arabidopsis, and calls for multi-cultivar crop studies; it contributes no genetic mapping. Mammalian trained-immunity work, such as miR–9-5p regulates immunometabolic and epigenetic pathways in β-glucan–trained immunity via IDH3α and the airway study DNA methylation changes during acute COVID-19 are associated with long-term transcriptional dysregulation in patients’ airway epithelial cells, dissects memory circuits in genetically unstructured samples. The COVID study explicitly lacks any test of how host genetics shapes the epigenetic scar.

On the genetics side, the eQTL methods we inherit were all developed on resting tissue. These are the multi-tissue design of The Genotype-Tissue Expression (GTEx) project, the outlier-resistant regression of Robust Linear Models for Cis-eQTL Analysis, and the count-based allelic model of A computational method for cell type-specific expression quantitative trait loci mapping using bulk RNA-seq data.

The closest generic idea is a "response eQTL," which compares expression before and after one stimulus. That idea is a known extension; the file lists context dependence as an open problem. Our difference is the four-arm design with a rest interval. It separates genetic control of the memory from genetic control of the basic pathogen response, which a single-stimulus response map conflates.

The population comes from Genetic Properties of a Nested Association Mapping Population Constructed With Semi-Winter and Spring Oilseed Rapes. It contains 2,425 recombinant inbred lines from ZS11 crossed with 15 founders, with 252 Gb of founder resequencing and over 4.4 million variants. That paper characterized only recombination and linkage. It measured no expression and no defense traits, and we add both.

Research Program

Choice of treatment and pathogen. The priming treatment is foliar sodium silicate, chosen because its ethylene and jasmonic acid route is mechanistically characterized in potato. The challenge pathogen is Botrytis cinerea, a fungus that kills host tissue and infects Brassica crops. Resistance to such fungi is governed by the same ethylene and jasmonic acid pathways that silicon engages. Botrytis stress expression has also been profiled across paralog pairs in the Brassicaceae relative Arabidopsis in The differential expression patterns of paralogs in response to stresses indicate expression and sequence divergences. We will use that resource to annotate whether memory genes fall in paralog classes under positive or purifying selection.

Pilot. A pilot on the 16 founders establishes whether silicon primes B. napus at all and how strongly protection varies across founders. Protection is measured as lesion area under PC relative to MC. If priming fails uniformly, the project switches to the abiotic memory route named in the review, heat priming in oilseed rape, keeping the same design.

Stage 1: allele-specific expression in hybrids. We grow F1 hybrids of ZS11 with each of the 15 founders under all four conditions. Each hybrid gets four biological replicates and two rest intervals, 3 and 10 days, which lets us measure how quickly memory decays. We then run RNA sequencing. Reads are aligned to a personalized diploid reference built from the founder resequencing data, rather than the single Darmor-bzh reference. This step addresses the reference mapping bias that would otherwise fake allelic imbalance in a polyploid with divergent A and C sub-genomes.

Allelic counts are modeled with a beta-binomial and total counts with a negative binomial, following CSeQTL's joint model, here with condition rather than cell type as the context. We fit robust MM-estimator regression as a sensitivity check because stressed tissue produces heavy-tailed outliers.

Stage 2: memory loci in recombinant inbred lines. We pick four families: the two founders whose priming efficacy is highest and the two whose efficacy is lowest relative to ZS11. From these we take about 100 lines each, sequenced at low coverage, following Powerful eQTL mapping through low-coverage RNA sequencing. All lines are sampled under MC and PC. A 25% subset is also sampled under PM to check that resting expression has returned to baseline. Lines are scored for lesion size.

Mapping is done within families and jointly on the existing linkage map. It yields local memory-eQTLs, which should replicate Stage 1, and distant memory hotspots.

Linkage artefacts. The A sub-genome loses linkage between markers quickly (average 280 kb) and the C sub-genome slowly (950 kb), and 11.75% of markers show segregation distortion. Together these could manufacture false distant hotspots in C. We therefore condition hotspot calls on local linkage blocks and report hotspot density per sub-genome before and after this correction.

Baselines

The naive reference point is the two-cultivar contrast a typical study would run. We compare bulk expression of ZS11 against one founder under MC and PC, call differentially expressed genes, and label genes with a cultivar-by-priming interaction as memory candidates. This baseline uses the same plants and reads as Stage 1. Our hybrid analysis beats it if hybrid allelic effects identify a set of local memory genes whose effects replicate in Stage 2 lines at a higher rate. Unlike the parental contrast, the hybrids can also separate local from distant control: where the parents differ but the alleles within the hybrid do not, the difference is distant.

The strongest existing method is count-based allelic and total-read modeling from CSeQTL. Its authors report greater than 90% directional concordance with single-cell data in human tissue. For plant or primed tissue we found no reported result on Lacuna for this method. Our contribution is not a new estimator here; it is the four-arm contrast and personalized reference. We will therefore also run CSeQTL-style modeling on the challenge-only arms, giving the same reads and the same model with the memory contrast removed.

The ordinary-regression baseline is genotype-by-condition least squares on log counts in the Stage 2 lines. The robust alternative is the MM-estimator. In human brain tissue the MM-estimator found 112 local eQTLs that least squares missed, with only 50% concordance between the two methods, per Robust Linear Models for Cis-eQTL Analysis. We report concordance between the two on primed tissue and expect it to fall toward the brain figure rather than the 91% seen in adipose tissue.

The key ablation drops the primed-only rest arm and the memory contrast. It maps ordinary challenge-response eQTLs (MC against MM) and asks how many memory-eQTLs are already explained by them. A second ablation aligns reads to Darmor-bzh alone, to measure how much apparent allelic memory is mapping artefact. All methods get the same reads, the same multiple-testing control (5% false discovery rate) and the same covariates.

Why not just…?

Why not just compare a primable and an unprimable cultivar by RNA-seq? This contrast finds thousands of differences but cannot say which are causes. Divergence in duplicated genes alone produces large stress-expression differences between lineages, as the Arabidopsis paralog study shows. In wheat, the Rht-B1b work needed a 245-line population to place the variance in two loci despite 1,993 differentially expressed genes (Impact of “Green Revolution” gene Rht-B1b on coleoptile length of wheat). Our first baseline runs exactly this contrast on the same plants, so the replication rate in the inbred lines settles the question directly.

Why not just map priming efficacy as a disease phenotype, without expression? This is a real threat if the goal is only breeding. Lesion-size QTL mapping in the same lines is cheaper and may find the loci that matter. We collect lesion size anyway and will map it. The expression layer earns its cost only if it answers what phenotype mapping cannot: whether a priming-failure locus acts by changing local regulation of a defense gene's memory, or through a distant hub. It also yields candidate genes inside the wide C sub-genome linkage blocks where phenotypic intervals span hundreds of genes. If memory-eQTLs do not colocalize with efficacy QTL, we will say the expression layer added nothing for prediction.

Why not just map ordinary pathogen-response eQTLs? Those loci mix basal resistance with memory. A cultivar may resist Botrytis well without being primable. The ablation tests this directly: if most memory-eQTLs are recovered by the challenge-only map, the four-arm design is unnecessary, and we will report that.

Why not just profile DNA methylation in the founders? Methylation is a plausible carrier of memory, as in the airway study, but describing it in 16 lines cannot assign causes to sequence variants. We will add methylation sequencing on the hybrids at the 3-day rest point as a follow-up layer only for genes with local memory effects.

Decisive Checks

The project succeeds if four results hold. First, the hybrids show genes with significant local memory effects at 5% false discovery rate that satisfy the return-to-baseline condition. Second, these effects replicate as local eQTLs in Stage 2 lines with high directional agreement, against which the count-model concordance figure above serves as a reference scale. Third, memory-eQTLs are largely not recovered by the challenge-only map. Fourth, across the 15 founders, the number or effect size of local memory alleles at ethylene and jasmonic acid pathway genes predicts each founder's measured priming efficacy better than the founders' challenge-only expression does.

For scale, the 'Désirée' potato study reports silicon lowering the disease index in whole plants from 55.21% to 35.42%. That is the size of protection difference we expect to see segregate among founders.

The idea fails if memory effects vanish once mapping bias is removed, or if memory-eQTLs collapse into ordinary response eQTLs. Either outcome would show that cultivar-specific priming is basal resistance in disguise.