01 — Method
Redundancy Analysis
RDA is a constrained multivariate ordination method that decomposes genomic variation into components explained by environmental predictors. It identifies which allele combinations are under selection by specific climate pressures — without requiring phenotypic observation.
The method projects both genotype space and environment space onto shared canonical axes. The distance between a query genotype's projection and the environmentally-predicted optimum is the genomic offset — a continuous, interpretable measure of maladaptation risk.
Published in Molecular Ecology and Nature Climate Change. Used operationally by UC Davis, CSIC, and IRTA. The scientific foundation is established and peer-reviewed.
Gradient Forest and BayPass are widely cited but are batch-only tools — designed for population-level analysis, not real-time query prediction. RDA's linear algebra structure allows the separation of model fitting from projection. This is the architectural insight that makes sub-millisecond inference possible.
Axis 1: 61.3% · Axis 2: 34.7% · Total: 99.98%
02 — Architecture
The Architectural Innovation
Standard RDA implementations run as batch scripts: load data, fit model, predict, repeat. Each query triggers a full model recomputation. For 10,000 varieties, that is 10,000 full RDA fits.
YOURTIMEONEARTH permanently separates model fitting from projection. Train once on your germplasm collection. The eigenvalue decomposition is stored — not the computation, the result.
Every subsequent prediction is pure matrix multiplication on pre-computed canonical axes. No refitting. No iteration. No overhead. The projection is deterministic and takes the same time regardless of training set size.
This is an architecture-level speed gain, not an optimisation-level one. The result is 0.015 milliseconds per prediction — seven orders of magnitude faster than a field trial confirmation.
Minutes
per variety, per environment. Full model refit each time. Not suitable for real-time query.
Seconds
per batch. Pre-fit, vectorised. Suitable for offline analysis. Not suitable for API deployment.
0.015 ms
per query. Eigendecomposition stored. Pure matrix projection. API-native. Scales to any volume.
03 — Validation
Validation Results
| Test | Genotype | Environment | Offset | Result |
|---|---|---|---|---|
| Galicia → Atlantic | Galicia ecotype | A Coruña · 14°C, 1000mm, pH 6.0 | 0.020 | Highly Suitable |
| Galicia → Semi-arid | Galicia ecotype | Sevilla · 25°C, 500mm, pH 7.8 | 0.809 | Avoid |
| Andalucía → Semi-arid | Andalucía ecotype | Almería · 22°C, 200mm, pH 8.0 | 0.031 | Highly Suitable |
| Andalucía → Atlantic | Andalucía ecotype | A Coruña · 14°C, 1000mm, pH 6.0 | 0.887 | Avoid |
| Gradient monotonicity | All ecotypes | All 12 locations | Confirmed | Pass |
| Confidence calibration | 3-factor scoring | Training size · Variance · Offset | Validated | Pass |
| Mediterranean ecotype | Mediterranean ecotype | Barcelona · 18°C, 600mm, pH 7.0 | 0.139 | Highly Suitable |
| Continental ecotype | Continental ecotype | Madrid · 15°C, 400mm, pH 7.2 | 0.050 | Highly Suitable |
Every prediction carries a confidence score derived from three independent factors: training set size (more samples → higher confidence), variance explained by the RDA model (higher R² → more reliable), and offset magnitude (extreme values reduce confidence as they are further from the training distribution). The Spain model achieves 99.98% variance explained, placing it in the highest reliability tier.
04 — Literature