Yield stability — the ability of a variety to produce consistent output across varied growing conditions — matters more than peak yield in most commercial breeding contexts. A variety that averages 7.2 t/ha across six environments is often more valuable to a farmer than one that hits 9.0 t/ha in a favorable year but falls to 4.5 t/ha under stress. The problem is that stability can only be observed retrospectively, through multi-environment trials that take years to complete. Our interest is whether sequence-based models can offer a prospective estimate of stability before the trials happen.
The Climate Scenario Framework
We use three Representative Concentration Pathways as the structural frame for our scenario analysis: RCP 4.5 (moderate emissions, approximately 2°C warming above pre-industrial by 2100), RCP 6.0 (intermediate pathway, roughly 2.2–2.5°C), and RCP 8.5 (high-emissions scenario, 3.5–4.0°C). These are standard IPCC scenarios, and we use them not because we are climate scientists making probabilistic forecasts, but because they provide a concrete vocabulary for communicating range of outcomes to breeders planning a 10–15 year varietal development timeline.
The translation from climate scenario to agronomic conditions follows published crop modeling literature for European and North African wheat production zones. Under RCP 4.5, average French grain belt growing-season temperatures increase by approximately 0.8–1.1°C relative to the 1990–2020 baseline, with modest precipitation reductions in June–July. Under RCP 8.5, the same comparison is 2.2–2.8°C with substantially increased terminal drought frequency. These are not our estimates — they are the agro-climate projections we layer onto our genomic prediction framework.
How the Model Integrates Climate and Sequence Information
Our foundation model operates on sequence data alone. It does not natively ingest temperature or precipitation parameters. The integration happens at the fine-tuning stage, where we treat the multi-environment trial data from CIMMYT historical nurseries as a GxE-structured training set: the same variety tested in 8 environments gives 8 phenotype observations, and the environment index (soil type, latitude, growing season temperature and water availability) becomes a conditioning variable for the prediction head.
Technically, this is a Genomic Selection model with environment as a fixed effect. The innovation relative to standard genomic selection is that the G-matrix representation — which in classical gBLUP is a simple pairwise relatedness matrix — is replaced by our 512-dimensional embedding, which carries non-linear epistatic context that the relatedness matrix cannot represent. The environment conditioning allows the model to produce environment-specific breeding values: not just "this line has high yield potential" but "this line's yield potential is relatively insensitive to the temperature increase expected under RCP 6.0".
What We Observed Across 3,700 Variety-Environment Records
Training on 80% of the CIMMYT wheat dataset and evaluating on the held-out 20%, we asked three questions:
Q1: Can the model rank varieties by yield stability under current climate conditions? Spearman ρ between predicted and observed Finlay–Wilkinson stability coefficients: 0.74. This is the in-distribution case and represents what the model can do without any extrapolation.
Q2: Can the model predict variety-by-environment interactions for environments not in the training data? Using leave-one-environment-out cross-validation: Pearson r for environment-specific yield dropped to 0.61, compared to 0.79 on randomly held-out variety-environment combinations. The degradation is expected — the model cannot perfectly reconstruct a novel environment's interaction pattern — but the signal remains meaningful for rank-ordering varieties within an environment class.
Q3: Can we extrapolate predictions to the RCP scenarios using the climate-variety interaction patterns learned from historical nurseries? This is the hardest question because it requires extrapolation beyond the observed range of environmental conditions. Our answer is tentative: the model identifies a consistent subgroup of approximately 12–18% of the wheat accession space that shows reduced yield sensitivity across the spectrum from current to RCP 8.5-simulated conditions. These lines have embedding profiles consistent with known drought and heat tolerance QTL regions in the Triticum aestivum reference. But we cannot validate this directly — we would need trials conducted under physically controlled climate-analogue conditions to verify the RCP 8.5 prediction specifically.
The Stability-Yield Tradeoff in Embedding Space
A consistent finding across our internal analysis is that high stability and high peak yield tend to occupy different regions of the embedding space. The lines that we predict as most stable under RCP 6.0 conditions are, on average, 7–11% below the highest-yielding varieties in optimal conditions. This reflects a real biological tradeoff: defensive mechanisms (osmotic adjustment, conservative stomatal regulation, deeper root architecture) divert photosynthate away from harvestable grain under non-stressed conditions.
Breeders know this tradeoff exists. What is useful about the embedding representation is that it makes the tradeoff visible before field testing. You can identify lines that appear to lie on the efficient frontier — above-average stability without proportionally sacrificed yield potential — because the embedding clusters them separately from both extreme types. Whether this translates to real phenotypic advantage remains to be confirmed in multi-year trials, but it provides a principled basis for prioritizing which lines to advance to that testing.
A Worked Scenario: French Dryland Winter Wheat, 2035 Planning Horizon
A breeding program planning varieties for commercial release in 2035 is currently selecting from a population of F4–F5 lines generated in 2024–2025. The target environment is the Loire Valley dryland zone, where summer precipitation is projected to decrease by 10–20% under RCP 4.5 by 2035 and growing season temperatures are projected to increase by 0.7–1.0°C. The target agronomic zone is not one of the CIMMYT nursery locations, but its climate indices map reasonably well to several Mediterranean nursery environments in the historical dataset.
In this scenario, our model produces a ranked list of the 3,200 F4 lines with predicted stability scores under each climate scenario. The breeder advances the 300 highest-ranking lines to replicated field trials in 2025–2026. The ranking is imperfect — we expect perhaps 65–70% overlap between our predicted top-300 and what field trials would identify as the top-300 — but it is substantially better than advancing lines by pedigree intuition alone, and it prioritizes climate stability explicitly rather than only current-environment yield.
What This Does Not Replace
We are not saying that climate-scenario genomic prediction eliminates the need for multi-environment trials. It does not. The RCP 8.5 extrapolation in particular should be treated with caution: we are projecting into conditions that have no close historical analogue in the training data, and model extrapolation beyond the training distribution is inherently uncertain. No genomic selection model can predict how a variety will respond to conditions that no variety in the training set has experienced.
The appropriate use of this framework is to prioritize which lines enter field trials, not to replace field trials as the final selection step. For commercial variety release, real yield and quality data across multiple years and locations will always be required. Sequence-based stability prediction is an early-stage tool that narrows the funnel efficiently — it does not close it.
We are also limited by the current model's underperformance on hexaploid wheat relative to diploid species. The RCP scenario analysis above is honest about its uncertainty bands. For wheat, those bands are wider than for maize or sorghum, and breeders should weight wheat-specific predictions accordingly until we release the polyploid-aware version of the foundation model.
