When a plant tissue piece is cut and placed on culture medium, it often releases phenolic compounds that oxidize, turn the medium brown, and kill the tissue before it can regenerate. Breeders treat this "lethal browning" as a recipe problem and add chemicals that soak up or neutralize the phenolics. Geneticists who map regeneration ability across many plant varieties treat the resulting failures as one more trait value. We argue both views miss something. A dead explant is a censored measurement: it tells you the tissue died before its regenerative capacity could be scored, not that the capacity was zero. And the anti-browning chemical does not only clean the medium. It changes which genes are switched on. So a genetic study of regeneration run on a single medium mixes two different heritable traits, browning susceptibility and regenerative capacity, and records each under one arbitrary chemical environment. Some of the "missing heritability", the gap between how heritable a trait looks in families and how much variance mapped DNA variants explain, may simply come from measuring it this way.
The proposal is to treat anti-browning additives as controlled environmental perturbations and to measure, variant by variant, how natural DNA differences in the regulatory regions of browning-pathway genes change gene activity under each additive. Gene activity is read out with reporter assays, in which a DNA fragment drives a light-emitting enzyme and the light measures how strongly that fragment switches a gene on. These medium-specific variant effects then become prior annotations in a censoring-aware heritability analysis. The core claim is testable and could fail. If the hypothesis is right, variants whose reporter effect appears under a particular additive should carry more of that medium's regeneration heritability than variants found by a reporter run on standard medium, and their allele effects should reappear in living, wounded explants.
What the literature gives us and what it leaves open
The plant side shows that additives reprogram metabolism, but only correlatively and in one genotype. In tea callus, Integrative Analysis of Transcriptomic and Metabolomic Profiles Uncovers the Mechanism of Color Variation in the Tea Plant Callus found that ascorbic acid pushes callus red by raising flavonoid genes (CHS, F3H, C4H, CYP75B1, ANR), that activated carbon pushes it green by raising photosynthesis genes (psa, psb, pet), and that polyvinylpyrrolidone (PVP) leaves it white. All of this was measured in the single cultivar Mengku and with only 14 phenolic acids quantified. Multi-omics joint analysis reveals the mechanism of flower color and fragrance variation in Lilium cernuum traced pigment loss to reduced UFGT expression, the enzyme that attaches a sugar to anthocyanins so they can be stored in the vacuole, and to upstream PAL/C4H/4CL repression. This too is expression correlation without functional tests. The cork oak and Illicium henryi culture papers (Tissue Culture and Rapid Micropropagation for Quercus suber L., Tissue culture requirements of the evergreen broad-leaved plant Illicium henryi D) establish that lethal oxidation is the bottleneck for woody species and that PVP helps. They are empirical recipe work. These papers motivate the question: they show the additive is an active environment. None of them asks whether genotypes differ in their response, or why.
The reporter side supplies the method and also warns about it. Dual-luciferase assays read the target fragment's firefly light against a second, constant Renilla luciferase to correct for how many cells took up the DNA. Massively parallel reporter assays (MPRAs) scale this to tens of thousands of barcoded fragments. Both are standard for testing single variants: A Functional Variant at miR-520a Binding Site in PIK3CA Alters Susceptibility to Colorectal Cancer in a Chinese Han Population tested one 3′ UTR variant this way, and Genome-wide regulatory model from MPRA data predicts functional regions, eQTLs, and GWAS hits trained models on MPRA data to score 14.3 million variants. Two constraints from these papers shape our design. First, in the PIK3CA study, endogenous blood mRNA went in the opposite direction to what the plasmid reporter predicted. Transient plasmids lack native chromatin and the cell's own trans environment, so reporter calls need an in vivo check. Second, the MPRA-trained model predicts activating elements well but repressive elements poorly. Our browning genes may well be controlled partly by repression, so we must report activators and repressors separately.
On the heritability side, SUMMIT: An integrative approach for better transcriptomic data imputation improves causal gene identification and Inference of nonlinear causal effects with application to TWAS with GWAS summary data recover hidden signal by improving the model that links genotype to expression. They gain power through larger reference panels or nonlinear causal estimation. A Module-Level Polygenic Risk Score-Based NetWAS Framework for Identifying AD Genetic Modules Mediated by Amygdala: An ADNI Study gains power by grouping variants into network modules. All three treat the phenotype as correctly measured and the environment as fixed. Our contribution is on the measurement side: we ask how much heritability returns once the phenotype is modeled as censored by death, and once each variant's regulatory effect is annotated for the chemical environment the phenotype was actually scored in.
The in vitro–in vivo translation work (DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction, Cell-Level Virtual Screening) shows that a perturbation's mechanism can change between dish and organism, and that naive alignment across that shift fails. We borrow the lesson but not the machinery. Here the "in vivo" target is the wounded explant itself, and the bridge we test is a direct one: allele effects in protoplast reporters should match allele effects inside heterozygous explants.
Core Claim
Browning-driven culture failure hides genetic variance in two ways: through censoring, and through gene-by-medium interaction at cis-regulatory variants of phenylpropanoid and oxidase genes. If both are modeled, by a survival-style phenotype plus medium-matched reporter annotations, measurably more regeneration heritability is captured than by a standard association study on a pooled or single medium. The gain should come specifically from matching the reporter's medium to the phenotype's medium, not merely from having reporter data.
Research Program
The system is tea (Camellia sinensis). It has a documented additive-specific transcriptional response, and its woody habit makes browning a real bottleneck. We will assemble and resequence a diversity panel of roughly 250 accessions; no such panel or benchmark exists on Lacuna, so this is new data collection. Each accession's stem explants are cultured on four media: no additive, PVP, activated carbon, and ascorbic acid. Every medium has replicate explants per accession. Imaging records medium darkening over time as a continuous browning curve, along with time to explant death and, for survivors, callus and shoot induction. Regeneration is then modeled as a two-part outcome: a survival component for death by browning, and a regenerative capacity component estimated only among explants alive at scoring. Death is treated as right-censoring of capacity, not as zero.
In parallel, we build an MPRA library in tea leaf protoplasts, which are plant cells with their walls removed so they take up DNA. The library contains promoter and 3′ UTR fragments, carrying each natural allele found in the panel, for the PAL, C4H, 4CL, CHS, F3H, ANR, UFGT, polyphenol oxidase, peroxidase, and psa/psb/pet gene families. These are the genes the tea and lily studies implicate. The pooled library is assayed under each of the four additives, plus a wounding-mimic condition with added phenolic extract, so that each variant gets an effect estimate per condition. The top 40 medium-specific variants are confirmed by dual-luciferase assays.
To close the translation gap, we take accessions heterozygous at reporter-active variants and measure allele-specific expression in RNA from explants harvested 24 and 72 hours after cutting on each medium. Because both alleles share one nucleus, chromatin context aside, this comparison controls for the trans environment that confused the PIK3CA reporter-versus-blood comparison. The medium-matched reporter effect should predict the sign and size of allelic imbalance in the explant on the same medium.
Finally, we partition heritability per medium. We use a linear mixed model that gives each class of variant its own variance component: reporter-active under the matched medium, reporter-active under a different medium only, reporter-inactive, and the rest of the genome. The model is fit to both the censored-survival phenotype and the conditional capacity phenotype.
Baselines
The do-nothing reference is a standard additive association study. Regeneration is scored as one number per accession on the no-additive medium, with dead explants coded zero, and heritability is estimated from all genotyped variants with no annotation. This is how a breeding program would run the study today, and it defines the "missing" variance we claim to recover. A second naive reference runs the same analysis pooled across media with medium as a fixed covariate. It tests whether simply averaging over media already does the work.
The strongest existing approach to turning reporter data into variant priorities is the model of Genome-wide regulatory model from MPRA data predicts functional regions, eQTLs, and GWAS hits. It trains elastic-net and gapped k-mer support vector models on MPRA outcomes plus genomic features, and reports an AUROC of 90–93% for activating regulatory hits in its HepG2/K562 tiling data. That number is in human cell lines, so it is a ceiling for comparison, not a matched result. No reported result was found on Lacuna for any plant regeneration or browning association study, and none for plant protoplast MPRAs. We will retrain this model class on our pooled, condition-agnostic protoplast MPRA data, with the same variants, features, and tuning budget, and use its scores as the annotation in the partitioning model. It is the right competitor because it asks whether a good regulatory annotation, without medium matching, already captures the gain. For the gene-level view, a linear expression-imputation association study using explant RNA-seq as the reference panel, in the spirit of SUMMIT, gets the same samples and is compared on the number of browning-pathway genes recovered at matched false discovery rate.
The key ablation swaps the annotation. Matched annotations use reporter effects measured under the same additive as the phenotype; the ablation uses effects from a mismatched additive. A second ablation removes censoring by coding death as zero while keeping the annotations. We will count the proposal as beating the baselines if three things hold together. First, matched annotations capture significantly more per-variant heritability than both mismatched and condition-agnostic annotations, judged by enrichment confidence intervals that do not overlap, in at least two of the three additive media. Second, censoring-aware phenotypes raise total captured heritability over zero-coding. Third, medium-matched reporter effects predict explant allelic imbalance better than condition-agnostic effects, measured by rank correlation on held-out variants.
Why not just…?
Why not just run the association study with a censoring-aware phenotype and skip the reporters? This is the most serious threat. Censoring alone could explain most of the recovered heritability. The zero-coding ablation, run with and without annotations, estimates censoring's contribution directly. If annotations add nothing beyond it, we will report that cleanly: missing regeneration heritability in browning species would then be largely a phenotyping artifact, a useful and cheap result. Censoring cannot say which genes or which additive interactions are involved, though, and that is what a breeder needs to pick a medium per genotype.
Why not just map response-eQTLs from explant RNA-seq across media instead of building an MPRA? Expression mapping in native tissue avoids the plasmid artifacts that undermined the PIK3CA reporter call. With about 250 accessions, though, it has the small-panel power problem that SUMMIT documents, where genes with low expression heritability are effectively untrainable. It also cannot separate which of several linked variants is causal. We keep it in the study as the allele-specific expression validation. If native allelic imbalance alone predicts heritability enrichment as well as the MPRA does, the reporter step is unnecessary, and the comparison of the two will show this.
Why not just screen media empirically for each genotype? For a handful of elite cultivars this is the practical answer, and the cork oak and Illicium work shows it can succeed. It does not generalize to new germplasm. It also ignores that additives redirect metabolism, as the tea callus study showed with ascorbic acid driving flavonoid accumulation. A held-out test addresses this: medium-specific variant genotypes should predict which additive rescues an unseen accession better than the panel-best single medium does.
Why not just apply an existing sequence-based regulatory model zero-shot? Existing models were trained on human cell lines, carry no information about chemical environment, and by their authors' account are weak on silencers. The condition-agnostic baseline above is the generous version of this alternative, retrained on our own plant data. A zero-shot human model is not expected to compete, and we will report it only as a lower reference.
Decisive Checks
Three outcomes would falsify the proposal. The first is that heritability enrichment does not depend on whether the reporter medium matches the phenotype medium. The second is that medium-specific reporter effects show no relationship with allelic imbalance in explants grown on the same medium; that would show the protoplast assay does not carry over to wounded tissue. The third is that reporter-active variants are almost entirely activators while heritability concentrates in regions the assay calls inactive, which would point to the known silencer blind spot rather than to biology. The ascorbic acid medium is the sharpest single test. The tea callus study predicts that flavonoid-gene variants, especially at CHS, F3H, and ANR, should show reporter effects specific to ascorbic acid and should carry ascorbic-acid-specific regeneration heritability. If that pattern does not appear in the one medium where prior work makes a concrete prediction, the central claim is wrong.
References
- Integrative Analysis of Transcriptomic and Metabolomic Profiles Uncovers the Mechanism of Color Variation in the Tea Plant Callus
- Multi-omics joint analysis reveals the mechanism of flower color and fragrance variation in Lilium cernuum
- Tissue Culture and Rapid Micropropagation for Quercus suber L.
- Tissue culture requirements of the evergreen broad-leaved plant Illicium henryi D
- A Functional Variant at miR-520a Binding Site in PIK3CA Alters Susceptibility to Colorectal Cancer in a Chinese Han Population
- Genome-wide regulatory model from MPRA data predicts functional regions, eQTLs, and GWAS hits
- SUMMIT: An integrative approach for better transcriptomic data imputation improves causal gene identification
- Inference of nonlinear causal effects with application to TWAS with GWAS summary data
- A Module-Level Polygenic Risk Score-Based NetWAS Framework for Identifying AD Genetic Modules Mediated by Amygdala: An ADNI Study
- DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
- Cell-Level Virtual Screening