Breeders increasingly design ideal crop genotypes ("ideotypes") in simulation. The approach is to wire a virtual genome into a mechanistic plant growth model and let a genetic algorithm search allele combinations for the highest simulated yield. These growth models contain a feedback loop: the plant's ratio of biomass supply to biomass demand decides whether new fruits set or abort. That loop can turn simple additive genetics into a rugged fitness landscape with several distinct ways of being productive, such as many small fruits, few large ones, or cyclic bearing. We propose that the standard search collapses onto whichever developmental regime it finds first. Measuring and preserving diversity among regimes, rather than among genotypes, should find better and more environment-robust ideotypes for the same number of simulations.
The idea in two sentences: describe every candidate genotype by the shape of its simulated supply/demand trajectory, and keep the search population spread across those trajectory shapes. Then test causally whether the benefit disappears when the model's feedback loop is switched off. If it does not disappear, or if a well-tuned restart genetic algorithm already covers the regimes, the method is unnecessary, and the experiments are designed to show that.
Core Claim
Letort et al. built a virtual maize genome whose alleles additively set internal parameters of the GreenLab growth model, such as organ sink strengths. A plain genetic algorithm found an ideotype with roughly 60% higher cob weight than the reference model. That paper supplies our testbed and our strongest reported number. It never asked whether the search had converged prematurely.
Mathieu et al. supply the reason to suspect it did. In the same model family, the supply-to-demand ratio $Q/D$ alone produces threshold-driven organ abortion and self-sustained oscillations: high $Q/D$ triggers many buds, which raises demand and causes later abortions. The same model predicted fruit set versus abortion on greenhouse cucumber stems with 89% accuracy.
Our claim has three parts.
- Feedback creates epistasis. Even when allele-to-parameter effects are strictly additive, the $Q/D$ feedback creates measurable non-additive interaction between alleles (epistasis) at the yield level.
- The landscape splits into regimes. The resulting landscape breaks into a small number of basins. Each basin corresponds to a qualitatively distinct developmental regime.
- Regime diversity is what matters. Search that keeps diversity across regimes beats search that keeps diversity across genotypes, because many genotypes map to the same regime.
This differs from the closest work in where diversity is measured. Biological Sequence Design with GFlowNets and Diversity By Design prevent collapse in sequence space against static learned scoring models. Here the score comes from a dynamical simulator, and diversity is defined on the simulator's behavior. The closest methodological analogue is Iterative Neural Similarity Deflation. That work pushes recurrent networks away from previously found fixed-point solutions by penalizing similarity of internal dynamics rather than of weights, and it found oscillatory alternatives. We borrow the principle of diversifying on dynamics, applied to plant development, but use an explicit archive instead of a learned similarity penalty. The archive also avoids that paper's reliance on a linear similarity proxy.
Research Program
Testbed. We first reimplement a GreenLab-style source–sink model from the descriptions in Letort et al. and Mathieu et al.. The virtual genome follows Letort's design: biallelic loci with additive effects on sink strengths, expansion rates and the bud-initiation threshold. Validation means reproducing Letort's qualitative result, a genetic algorithm reaching about a 60% cob-weight gain over the reference. Reimplementation is a real risk, and we treat failure to reproduce as a stop-and-diagnose point.
Diagnosing the landscape. For each genotype we record the simulated $Q/D$ time series. From it we compute a low-dimensional regime descriptor with three components: the fraction of fruit positions aborted, the dominant period of fruit set from the spectrum of the $Q/D$ series, and the mean $Q/D$ ratio.
We then sample genotypes at scale, run many independent optimizer restarts, and fit an additive model of yield on allele dosages. The variance this additive model leaves unexplained is our measure of feedback-induced epistasis. Basins are identified by clustering local-search endpoints by their descriptor.
Search method. The method keeps the crossover-plus-elevated-mutation genetic algorithm that Dang et al. proved escapes plateaus via an emergent diversity burst, at mutation rate $p_m=(1+\delta)/n$ with $n$ the number of loci. It changes only selection. The descriptor space is divided into cells, the best genotype in each cell is retained, and parents are drawn across cells.
This mirrors the strategy-space idea of Gupta et al., who spread batch Bayesian optimization across exploration–exploitation trade-offs rather than across input locations. Here the axis being spread is developmental behavior. The method adds no learned critic, avoiding the roughly 180% runtime overhead reported for DynAMO. Because the simulator itself scores candidates, a learned scoring model cannot be exploited.
Robustness. Letort optimized in one stable environment. We additionally re-simulate every archived ideotype under a set of perturbed light and temperature trajectories. This tests whether regime diversity yields candidates that hold up under genotype-by-environment shifts, the replication failure documented by Santure et al..
Baselines
Do-nothing reference. The naive reference is Letort's standard genetic algorithm, re-run on our reimplementation with its original settings. The only reported figure for this setting is the approximately 60% cob-weight gain in Letort et al.. Matching it validates the testbed, and exceeding it is the minimum bar. Uniform random search over genotypes at equal simulation budget is the second floor. Multiple-Frequencies PBT shows greedy population methods can fall below random search: on Brax Humanoid, standard population-based training reached an interquartile-mean reward of 16,171 against 17,713 for random search.
Strongest competitors. The first is the same crossover genetic algorithm with Dang's elevated mutation and independent restarts, tuned per budget. Its tuning budget equals ours, and each restart counts against the total simulation count. It is the strongest competitor because it has a theoretical escape guarantee and needs no descriptor.
The second is CMA-ES, an evolution strategy that adapts a Gaussian search distribution, run on a continuous relaxation of allele dosages with rounding and restarts. It is included because Design-Bench found classical CMA-ES and whitened gradient ascent matched or beat newer methods on 4 of 8 tasks.
The third is divergence-guided particle swarm optimization from Da Costa and Modenesi, a swarm optimizer that repels particles from the global best. That paper reports 2x to 8x gains over standard particle swarm on multimodal functions. No reported result was found on Lacuna for any of these competitors on GreenLab.
Ablations isolating the key ingredient. The first ablation replaces the regime descriptor with Hamming-distance fitness sharing, which keeps genotype diversity only. The second replaces it with a random projection of the genotype of the same dimension. Beating these competitors means a higher best cob weight at matched simulation budgets, with a run-to-run spread that does not overlap. It also means a higher probability across seeds of reaching the best-known basin.
Why not just…?
Why not just run a crossover genetic algorithm with higher mutation and restarts? This is the most serious threat. Dang et al. prove that crossover generates enough diversity to cross a fitness gap on stylized jump functions, landscapes where a flat plateau sits a fixed number of bit-flips from the optimum. If plant regimes are separated by such gaps, restarts may suffice. The landscape diagnostic settles it directly. We count how often restarts land in each basin. If restarts cover all basins at a cost comparable to the archive method, we report that the simple method wins and drop the claim.
Why not just use CMA-ES? The Design-Bench result makes CMA-ES a real contender. However, its single adapting Gaussian is built to commit to one basin. Its regime coverage per run, and its best-of-budget yield, are measured directly against ours.
Why not just map alleles to intermediate parameters, where effects are additive, and optimize there? Letort showed that mapping genes to internal parameters recovers loci that mapping to final yield misses. But the map from those parameters to yield passes through the $Q/D$ threshold, which Mathieu et al. show is non-monotonic. Our additive-fit residual quantifies exactly how much this matters. A near-zero residual would mean the landscape is effectively smooth and our premise fails.
Why not just sample in proportion to reward with a GFlowNet? A GFlowNet is a generative policy trained to sample candidates with probability proportional to their reward. It yields genotype diversity, but a reward-proportional sampler still concentrates on whichever regime holds the most high-scoring genotypes. We run the BioSeq-GFN-AL codebase with the simulator as the reward and compare regime coverage at equal budget.
Decisive Checks
The central check is causal. We repeat every comparison on a control version of the model in which bud initiation no longer depends on $Q/D$, which removes abortion thresholds and oscillations. Our prediction has three parts:
- In the control model, the epistasis residual collapses toward zero.
- In the control model, basins merge into one, and the archive method's advantage vanishes or turns into a small overhead. This mirrors the precision loss Da Costa and Modenesi report for repulsive search on the single-basin Sphere function.
- In the full model, the advantage appears and grows with the strength of the feedback.
An advantage that persists without feedback would mean the descriptor is helping for some reason unrelated to developmental regimes. That outcome would falsify the mechanism we propose even if yield improves.
The secondary check is whether ideotypes from non-dominant regimes lose less yield under perturbed environments than the single best ideotype from the greedy search. This would test the practical claim that preserving regimes, not only peaks, is what breeding pipelines should keep.
References
- Quantitative Genetics and Functional-Structural Plant Growth Models: Simulation of Quantitative Trait Loci Detection for Model Parameters and Application to Potential Yield Optimization
- Oscillations in Functional Structural Plant Growth Models
- Biological Sequence Design with GFlowNets
- Diversity By Design: Leveraging Distribution Matching for Offline Model-Based Optimization
- Discovering alternative solutions beyond the simplicity bias in recurrent neural networks
- Escaping Local Optima Using Crossover With Emergent Diversity
- Exploiting Strategy-Space Diversity for Batch Bayesian Optimization
- Replicated analysis of the genetic architecture of quantitative traits in two wild great tit populations
- Multiple-Frequencies Population-Based Training
- Design-Bench: Benchmarks for Data-Driven Offline Model-Based Optimization
- Divergence-Guided Particle Swarm Optimization