Crops grown repeatedly in the same field often decline even when fertilizer and water are kept constant. This is called soil sickness, or replant disease. Part of the cause is autotoxicity: roots release chemicals that poison the next generation of the same crop. Researchers usually rank these chemicals by how strongly they harm seedlings in a dish. Living soil, however, is full of microbes that eat many of these chemicals. Catalpol, a toxin from Rehmannia glutinosa, lasts 43.39 h in pasteurized soil but only 12.04 h in live soil (Rehmannia glutinosa Replant Issues). So a compound that is potent in a dish can be harmless in the field. A weaker compound that microbes cannot break down may build up over several cropping cycles. This is the same lab-to-living-system translation gap that undermines drug discovery, and here it has a concrete, measurable cause: the soil's degradation filter.
We propose to predict that filter from the soil's DNA. We would sequence a soil's community genome (its metagenome), pull out the catabolic enzymes it encodes, and use molecular docking to test whether each autotoxin can sit in each enzyme's active site in a chemically reactive position. Molecular docking is a computational search that places a small molecule into a protein pocket and scores the fit. The prediction is a compound-by-soil persistence score. Combined with dish potency, that score should rank autotoxins by their actual harm in live soil better than potency alone. The key ingredient is reactive-pose docking against a soil-specific enzyme repertoire. We test explicitly whether it adds anything beyond cheaper gene-family annotation.
Core Claim
The ranking of autotoxins by in-soil harm differs from their ranking by in vitro potency, and most of that difference comes from compound-specific microbial degradation. That degradation can be predicted for a new soil from abundance-weighted, physically validated reactive docking of each compound into the soil's metagenome-encoded enzymes. The prediction should be better than enzyme-family annotation alone. Docking adds substrate specificity within families such as β-glucosidases or phenolic-acid decarboxylases, which determines whether catalpol, acteoside, ferulic acid or p-hydroxybenzoic acid is actually turned over.
We write the predicted live-soil degradation rate of compound $c$ in soil $s$ as
$$ k(c,s) = k_0(c) + \beta \sum_{g} a_g(s)\, r(c,g). $$
Here $k_0(c)$ is the abiotic rate measured in pasteurized soil, $a_g(s)$ is the metagenomic abundance of enzyme cluster $g$ in soil $s$, $r(c,g)\in[0,1]$ is the reactive-pose score of compound $c$ in enzyme $g$, and $\beta$ is a single fitted scale. The predicted in-soil harm is then potency times exposure, $H(c,s) = P(c)/k(c,s)$, with $P(c)$ the seedling-assay potency. The $1/k$ form is the time-integrated concentration under first-order decay. That is the kinetic form the half-life measurements in the file already assume.
Where This Sits in Prior Work
The soil sickness literature motivates the gap. Metabolomic surveys find hundreds of shifting compounds: over 1400 detected, with 416 down and 39 up between uncultivated and three-year Dictyophora rubrovalvata soils (Elucidation of Mechanism of Soil Degradation… Dictyophora rubrovalvata). These surveys cannot say which compounds matter causally. Tobacco work lists more than 19 synergistic autotoxins and shows that Pseudomonas putida removes 99.47% of p-hydroxybenzoic acid in 72 h in culture but fails to persist in the field (Autotoxins in continuous tobacco cropping soils). The Rehmannia half-life comparison (Rehmannia glutinosa Replant Issues) shows that microbes control exposure, but only for one compound at a time. We turn these isolated observations into a prediction across compounds and soils.
On the translation side, Cell-Level Virtual Screening argues that binding affinity alone misses the cellular context and models that context through gene-network rewiring. DeepSADR bridges cell lines and patients by aligning drug fragments with biological pathways. Both inspire our framing, but neither touches soil, and both model the context statistically. We model a single physical context variable, degradation, mechanistically, because soil datasets are far too small for domain-adaptation networks.
The docking literature constrains the method. PoseBusters shows that deep-learning docking frequently produces physically impossible poses on new proteins. Our enzymes are predicted structures from environmental sequences, which are about as far from training data as proteins get. So we take physically valid poses as a hard requirement. DiffDock motivates trying diffusion-based docking on predicted structures, where it retained 21.7% success against "10% or lower" for other methods. DOCKSTRING supplies a standardized Vina pipeline and warns that raw docking scores reward large, greasy molecules. That warning is one reason we score catalytic geometry rather than raw affinity.
Research Program
We would work with three soil systems from the file: Rehmannia glutinosa replant soil (catalpol, acteoside), continuous-tobacco soil (phenolic acids), and a Dictyophora cropping-age series (CC0 to CC3) whose microbiome shifts strongly with age. For each system we would collect fresh and replant soils from several fields, giving roughly a dozen soils in total. We would assemble a panel of about 40 reported autotoxins and structural analogues.
Three measurements are collected per compound and soil:
- Seedling potency on Lepidium sativum and Lolium perenne, as in the Rehmannia study.
- Half-life in live and pasteurized subsamples, measured by LC-MS. This gives an "oracle" ratio $\ln(t_{\text{past}}/t_{\text{live}})$.
- Shotgun metagenomes, from which we cluster candidate catabolic genes (glycoside hydrolases, decarboxylases, ring-cleaving dioxygenases, hydroxylases).
We would predict structures for the most abundant cluster representatives. Each compound is docked with AutoDock Vina through the DOCKSTRING pipeline and, separately, with DiffDock. Every pose passes through PoseBusters and a force-field minimization. The reactive-pose score $r(c,g)$ is the fraction of valid top poses in which the bond the enzyme would cut sits within a catalytic distance of the annotated catalytic residues.
The parameter $\beta$ is fit on held-out soils only. Evaluation uses leave-one-soil-out and leave-one-system-out splits, so that the model must predict persistence in a soil it has never seen.
Baselines
The naive reference is in vitro potency alone, $P(c)$, used to rank in-soil harm. This is what the field currently does implicitly, and our harm predictor must beat it by a clear margin in rank correlation with pot-assay damage.
The strongest non-docking competitor is enzyme-family annotation. It uses the same metagenomes and the same abundance weighting, but $r(c,g)$ is set to 1 whenever cluster $g$ belongs to a family with a known activity on the compound's chemical class. Because it receives identical data and enzyme clusters, any gain from docking reflects substrate specificity and not extra information.
A second competitor predicts biodegradability from molecular descriptors alone: glycoside status, aromatic substitution, logP and similar, fed to a small regression with the same held-out-soil protocol. It ignores the soil entirely and so tests whether the soil-specific term matters.
Two ablations isolate the key ingredient. The first replaces the reactive-pose score with the raw Vina docking score. The second drops the PoseBusters filter. This lets us separate catalytic geometry and physical validity from plain affinity.
For the choice of docking engine, the reported numbers favour classical tools on new proteins. On the PoseBusters Benchmark Set, physical validity for accurate poses was 58% for AutoDock Vina and 55% for Gold, against 12% for DiffDock (PoseBusters). We therefore treat Vina with validity filtering as the default and DiffDock as a challenger that must match it after filtering.
The oracle upper bound is the measured live/pasteurized half-life ratio. We report what fraction of the gap between potency-only and oracle-persistence ranking each predictor closes. There is no reported result on Lacuna for predicting soil autotoxin persistence from metagenomes, so this proposal establishes the first numbers.
Why not just…?
Why not just measure half-lives directly in live and pasteurized soil? For a fixed panel in a fixed soil this is cheaper and more accurate, which is why it serves as our oracle. It does not scale to the hundreds of candidate metabolites that metabolomics surfaces, or to soils whose microbiome is changing across cropping cycles. It also does not identify which genes and strains do the degrading, and that is what a targeted inoculant or rotation decision needs. If the reactive-docking predictor closes less than half the potency-to-oracle gap on held-out soils, direct measurement wins, and we would report the method as not useful.
Why not just annotate gene families and skip docking? This is the real threat. For well-characterized compounds such as phenolic acids, annotation may already be enough. The experiment settles it directly. Docking must beat annotation on within-family discrimination, for example between glycoside substrates that share hydrolase families, under leave-one-soil-out evaluation. If it does not, we drop docking and publish the annotation-based exposure model, which still addresses the translation gap.
Why not dock autotoxins into RgLRR29, the receptor shown to protect Rehmannia? This is the obvious intersection. Functional analysis of Rehmannia glutinosa key LRR-RLKs did no structural work, so it looks open. But there is no evidence that small-molecule exudates bind the receptor's leucine-rich-repeat domain at all. Docking into a predicted plant receptor sits exactly where PoseBusters reports the worst generalization, and the result would be a binding score with no in-soil readout. Enzymes have defined catalytic chemistry, and their predictions are directly testable as half-lives.
Why not just rotate crops or add biochar? Rotation restores suppression of the root-rot fungus Fusarium oxysporum by 45–56% (Crop rotation and native microbiome inoculation), but it is impractical for high-value perennials such as Rehmannia. Biochar is non-specific and also removes beneficial cues like the lysine that apple rootstocks use to recruit protective Bacillus (From metabolic fingerprints to field solutions). Our exposure model is complementary: it tells a grower or breeder which compounds actually need removing.
Decisive Checks
The first check is predictive. On held-out soils, predicted $H(c,s)$ must outrank potency alone in Spearman correlation with seedling damage in pot assays spiked into live soil. The reactive-docking persistence term must also beat annotation-only persistence.
The second check is interventional. We would isolate strains carrying the enzymes that receive the highest reactive scores for a specific compound, such as catalpol, building on the catalpol-degrading Pseudomonas aeruginosa Rh7 described in Rehmannia glutinosa Replant Issues. Adding these strains to pasteurized soil should shorten the half-life of the predicted substrates and leave non-predicted compounds unchanged. That compound-specific pattern is something neither annotation nor descriptors can fake.
The third check is longitudinal. Across the Dictyophora CC0–CC3 series, predicted persistence should shift with the measured community turnover. Compounds whose predicted degraders collapse alongside Sphingomonas (15.63% to 1.12%) should be among those that accumulate.
Failure on the first check falsifies the central claim. Failure on the second alone means docking finds correlates of degradation rather than causes.
References
- Rehmannia glutinosa Replant Issues: Root Exudate-Rhizobiome Interactions Clearly Influence Replant Success
- Elucidation of Mechanism of Soil Degradation Caused by Continuous Cropping of Dictyophora rubrovalvata Using Metagenomic and Metabolomic Technologies
- Autotoxins in continuous tobacco cropping soils and their management
- Cell-Level Virtual Screening
- DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
- PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences
- DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking
- DOCKSTRING: Easy Molecular Docking Yields Better Benchmarks for Ligand Design
- Functional analysis of Rehmannia glutinosa key LRR-RLKs during interaction of root exudates with Fusarium oxysporum reveals the roles of immune proteins in formation of replant disease
- Crop rotation and native microbiome inoculation restore soil capacity to suppress a root disease
- From metabolic fingerprints to field solutions: engineering the apple rhizosphere microbiome via host-directed Bacillus recruitment for sustainable apple replant disease control