Generative Herbicide Search Guided by Hyperspectral Stress Fingerprints

Weeds cut yields in major growing regions by up to half, and our main weapon against them is failing: decades of uniform spraying have selected for resistant weed populations, while no genuinely new herbicide mode of action — the biochemical pathway a herbicide disrupts — has reached the field in roughly thirty years Niu et al.. Meanwhile, computational chemists have built generative methods that search astronomically large molecular spaces ($10^{23}$ to $10^{60}$ candidate compounds) for drugs and materials, but none of these tools has ever been pointed at a plant biochemistry problem. The two literatures sit next to each other without touching: generative molecular design optimizes against scalar, in-silico property scores, while agrochemical discovery still relies on manual chemistry and weeks of greenhouse waiting.

This proposal closes that loop with one change that matters: the reward signal for molecular generation comes not from a predicted potency number but from a measured, high-dimensional stress fingerprint. When a plant is hit by a herbicide, its light-reflectance spectrum shifts within a day in a way that is specific to which biochemical pathway was attacked — well before any visible wilting. An automated hyperspectral system has already shown it can read these fingerprints and classify the site of action of eight commercial herbicides at 81.5% accuracy one day after treatment Niu et al.. We propose to use exactly this kind of measured fingerprint as the oracle in a closed-loop, synthesis-constrained generative search for new herbicide leads — so that the search can be steered toward novel mechanisms and away from the known modes of action that weeds have already evolved resistance to, something no scalar potency score can express.

Core Claim

The claim is that a closed loop of generative molecular proposal, rapid hyperspectral phenotyping, and batch active learning will find more herbicidal hits per synthesized compound than either manual analogue chemistry or a generative loop driven by a scalar efficacy surrogate — and, more importantly, will find hits whose spectral signatures lie outside the signature hull of all known sites of action, i.e., candidate new mechanisms. Two design choices carry the weight. First, generation is constrained to a reaction grammar built from a proven two-step agrochemical synthesis, so every proposed molecule is synthetically reachable by construction — importing the lesson that template-constrained search eliminates the invalid "hallucinated chemistry" that plagues free-form generators (Zhao et al.; Podda et al.; SCENT). Second, the reward is a vector derived from the plant's measured response, not a predicted activity score, which both exposes mechanism-level information and partially immunizes the loop against the reward hacking that generative models routinely commit against learned oracles (BoMolLLM; Pandey et al.).

Concretely, for a candidate molecule $m$ whose measured early spectral signature is embedded as a vector $s(m)$, the per-compound reward is

$$r(m) \;=\; \alpha\, d(m) \;+\; \beta\, I_{\text{weed}}(m) \;-\; \gamma\, I_{\text{crop}}(m) \;-\; \delta\, C(m),$$

where $d(m)$ is the distance from $s(m)$ to the convex hull of the measured signatures of the eight known sites of action (the mechanism-novelty term), $I_{\text{weed}}$ and $I_{\text{crop}}$ are early injury scores in barnyard grass and in a maize safety control, $C(m)$ is the estimated synthesis cost of the two-step route, and the weights are fixed per search round. The novelty term is the ingredient that no prior system has: it rewards a molecule for stressing the weed in a way no commercial herbicide does.

Research Program

The program has four components, each reusing a technique that already works somewhere else in the file.

The generator is a template-constrained search — Monte Carlo tree search or a generative flow network — over a reaction graph whose edges are the two chemical steps validated by Ma et al.: solvent-free benzylation of 2-hydroxybenzyl alcohol with an N-heterocycle, followed by nucleophilic aromatic substitution onto the 4,6-dimethoxypyrimidine core. Their protocol ran at 39–98% yield across 34 compounds, so every molecule the generator proposes is plausibly synthesizable in one to two steps. The building-block sets are then expanded beyond their 9 heterocyclic skeletons to a commodity reagent catalog; this is where the search space grows from tens to thousands of reachable compounds. This mirrors the move Zhao et al. made from unguided to guided tree search over validated reaction templates, where guidance raised the on-target generation rate from roughly 0.27 to 0.73–0.80, and the cost- and diversity-aware design of SCENT, which cut average synthesis costs by 30–50% while finding nearly eight times more high-reward scaffolds than its baseline.

The oracle is a physical experiment, not a prediction. Each search round proposes a batch of roughly 40 compounds; each is sprayed onto replicated pots of Echinochloa crus-galli (barnyard grass, the weed used in the existing hyperspectral study) and of maize as the crop-safety control, then scanned at 1, 3, and 7 days after treatment on a line-scanning hyperspectral platform of the same design as Niu et al. (380–1030 nm). A classifier of the same family — support vector machines on spectral features — is trained on their published eight-herbicide dataset to embed new compounds relative to known sites of action; the leave-one-site-of-action-out experiment below tests whether a held-out mechanism lands outside the hull of the remaining seven, which is precisely the discrimination the novelty term requires.

The inner surrogate and acquisition close the gap between rounds. Because a greenhouse batch is inherently parallel, the acquisition step uses batch Thompson sampling with an ensemble surrogate over molecular fingerprints, following the parallel and distributed Thompson sampling design of Hernández-Lobato et al., which retrieved top-performing molecules at 20x the recall of random search on a large virtual screen. The surrogate predicts the reward vector — including the novelty component — for unmade molecules, and the generator is retrained or re-guided each round on the accumulated structure–fingerprint data.

The confirmation stage attacks the target-identification gap that Ma et al. left open: any hit whose signature falls outside the known-SOA hull is prioritized for dose–response characterization, chemoproteomic target pull-down, and — in a secondary robustness check drawn from the "resistance flip" of Guo et al., where the glyphosate resistance index of Eleusine indica fell from 8.9 to 3.1 under cooler temperatures — dose–response under both 30°C/20°C and 20°C/15°C regimes, so that final leads are selected for efficacy that does not collapse under field temperature variation.

Proposed Experiments

The first experiment is retrospective and cheap: on the published barnyard-grass hyperspectral dataset, retrain the site-of-action classifier and verify the reported 81.5% overall accuracy at 1 day after treatment, with per-class accuracies up to 96.9% for Indaziflam and Atrazine and above 90% for the untreated control Niu et al.. Then run the leave-one-site-of-action-out test: does the held-out herbicide's signature fall measurably outside the hull of the other seven? If not, the novelty term has no signal and the core mechanism fails here, before any synthesis. A second retrospective test uses the 34-compound library of Ma et al. as a simulated loop: with their growth-inhibition outcomes (11 of 34 compounds exceeding 60% inhibition on crabgrass, and lead compound 5q above 80% on velvetleaf and sicklepod) as ground truth, we ask whether spectral-informed batch acquisition recovers the 11 hits in fewer than 34 syntheses.

The prospective experiment runs three full loops at 40 compounds per round. Primary metrics are hits (>60% growth inhibition at 21 days, confirmed by biomass) per compound synthesized, the fraction of hits with signatures outside the known-SOA hull, a crop-safety index from the maize control scans, and the dose required relative to the 750 g/ha screening rate used by Ma et al.. For context on what guidance buys over unguided generation, Zhao et al. report that unguided template search reached an on-target rate of only ~0.27 versus 0.73–0.80 when guided; our unguided-template baseline plays the same role. No reported result found on Lacuna for any closed-loop generative pipeline evaluated against herbicidal efficacy or hyperspectral phenotyping, so the prospective numbers we produce will be the first of their kind.

Baselines

The naive reference point is the manual protocol of Ma et al.: an unguided, intuition-driven 34-compound library around the pyrimidine core, screened post-emergence in the greenhouse and scored visually at the 750 g/ha rate. It is the right "do nothing" comparator because it is what the field actually does, and it demonstrably works — 11 hits and a lead compound. It will be run under our identical weed species, greenhouse conditions, and scoring endpoint, so the comparison is compounds-in versus hits-out. We beat it if the loop delivers a higher confirmed hit count per compound synthesized at a lower application rate, and — decisively — if at least one hit has a spectral signature outside the known-SOA hull, since their library is structurally confined to the acetohydroxyacid-synthase-inhibitor pharmacophore and reported no confirmed target.

The strongest existing competitor is a scalar-oracle generative loop: the multi-objective graph-convolutional reinforcement learning design of Khemchandani et al., which generates 100%-valid molecules and recovers known high-affinity binders at Tanimoto similarity 1.0, coupled to a learned herbicidal-efficacy surrogate trained on our accumulated data, with crop selectivity handled as a second predicted objective in the style of their transporter-selectivity term. It gets the same building blocks, the same greenhouse batch budget, the same compute and tuning effort, and the same number of loops; only the reward source differs (predicted scalar efficacy versus measured spectral fingerprint). We beat it if the fingerprint-driven loop produces a strictly better set of hits on the primary metrics, and specifically if it finds novel-signature hits that the scalar loop cannot even express as an objective. No reported result found on Lacuna for this baseline run on any plant target, so we will establish its agrochemical performance ourselves.

The key ablation is the same loop with the same generator, acquisition, and batch schedule, but with the reward collapsed to a scalar: late visual biomass score at 21 days, no spectra. This isolates the spectral oracle itself as the proposal's active ingredient. If the scalar-reward ablation matches the full loop's hit rate and novel-signature yield, then hyperspectral phenotyping is merely an accelerator, not a steering signal, and the core claim reduces to "batch active learning helps screening" — a much weaker and already-demonstrated proposition Hernández-Lobato et al.. An unguided-template baseline (uniformly random legal reaction combinations) completes the set and mirrors the ~0.27 unguided rate that Zhao et al. report against their 0.73–0.80 guided rate.

Why not just…?

Why not just screen existing chemical catalogs with high-throughput assays? This is the cheapest objection, and the file answers it quantitatively: chemical space is too large for brute force by an incomprehensible margin, and intelligent sampling beats random selection by large factors wherever it has been measured — 20x higher recall of top molecules for parallel Thompson sampling versus random search, at 1,500 CPU years instead of 30,000 Hernández-Lobato et al., a 17x measurement acceleration on a 1,011-composition catalyst space Jenewein et al., and 4 experiments versus 7 for random search even in a tiny formulation space Khongkomolsakul et al.. Catalog screening also inherits pharmaceutical-biased libraries, which is exactly the bias that a plant-pathway-driven generator is designed to escape. The random-screening arm is nevertheless retained as the unguided baseline so the objection is settled empirically rather than by citation.

Why not just let synthetic chemists elaborate analogues around an active core? This is the real threat, because manual diversity-oriented synthesis already succeeded: 34 compounds, 11 hits, a lead Ma et al.. We do not argue that chemists are slow; we argue that this particular strategy optimizes the wrong thing for a resistance crisis. Their library circles a pharmacophore shared with commercial acetohydroxyacid-synthase inhibitors — the known mode of action that resistance mutations already defeat — and manual elaboration stayed within 9 heterocyclic skeletons tested at a high 750 g/ha rate with no confirmed biochemical target. The comparison that settles it is in the experimental design: per-compound hit yield against the manual baseline run in the same greenhouse, plus the fraction of hits with signatures outside the known-SOA hull and Tanimoto distance of hits from existing AHAS inhibitors. If manual elaboration matches the loop on all three, the proposal is not worth its machinery, and we will say so.

Why not just use standard Bayesian optimization with a Gaussian process and expected improvement on a scalar efficacy score? This collapses to the ablation above plus a known scalability defect: expected improvement is inherently sequential, and batching it through "fantasization" loops does not parallelize cleanly, which is precisely why Hernández-Lobato et al. replaced it with parallel Thompson sampling — and greenhouse rounds are batches by nature. The deeper problem is representational: a scalar score compresses away the mechanism information in the spectral signature, so the search cannot be asked to find a new mode of action, and a purely predicted surrogate invites the reward hacking documented when generative models exploit oracle inaccuracies BoMolLLM; Pandey et al.. Our measured-fingerprint reward keeps the outer oracle physical. If, however, the scalar ablation matches the full loop on hit rate, this objection wins in practice and the conclusion of the study becomes that acceleration, not mechanism steering, is the real value — a finding we would report honestly.

Decisive Checks

Three outcomes would falsify or sharply limit the claim. First, the leave-one-site-of-action-out test: if held-out mechanisms are not separable from the remaining signatures at 1–3 days after treatment, the novelty reward is built on sand and no amount of generation machinery fixes it. Second, if after three loops the fingerprint-guided arm does not exceed the manual baseline's hits-per-compound figure (11 of 34, or roughly one hit per three syntheses Ma et al.), the loop has not earned its complexity. Third, if none of the confirmed hits carries a signature outside the known-SOA hull, then the search found potency but not novelty, and the resistance-motivation collapses to ordinary lead optimization. Conversely, one confirmed novel-signature hit with a deconvoluted target, dose–response below the 750 g/ha screening rate, crop safety in the maize control, and stable efficacy across both temperature regimes of Guo et al. would be the first empirically grounded demonstration that chemical space exploration can be steered by what a plant, rather than a predictor, says about a molecule.