Data-driven decentralized breeding increases prediction accuracy in a challenging crop production environment

The big data revolution in genomics has transformed plant breeding with inexpensive sequencing methods, enabling greatly accelerated variety development. At present, plant breeders use data-driven methods, including genomic prediction, to increase selection intensity while reducing the time of the breeding cycle and deriving greater genetic gain. Most conventional breeding programs still rely on a centralized scheme aimed at maximizing genetic diversity (G) in the early stages of selection and then identifying superior germplasm based on phenotypic observations made in a limited number of research stations with explicit environmental (E) and management (M) conditions. In this setting, genomic prediction may be used to predict the performance of untested new genotypes but is bound to the G × E × M interactions captured by the research stations that are used to train the selection models. This limitation of centralized breeding approaches may result in suboptimal development and deployment of crop varieties for use by farmers seeking local adaptation in challenging environments. This is especially relevant in smallholder farming systems, which involve about 80% of the world farmers and call for tailored solutions to support food security. To respond to local cropping needs impacted by climate change, breeders need to find new ways to accelerate variety development while directly addressing G × E × M interactions to the fullest. Mobilizing farmers’ traditional knowledge of crop varieties and local adaptation can address this challenge and enhance adoption of improved varieties in a coherent, decentralized breeding program relying on farmer-participatory selection. A crowdsourced citizen science approach has demonstrated the feasibility of a data-driven decentralized variety evaluation that enables on-farm variety testing in a digitally supported and cost-efficient way. Predictive accuracy of farmer selection criteria may outperform breeder evaluations even in a context of modern agriculture. Crowdsourced citizen science further integrates the E and M components into breeding by performing selection directly in target environments and using environmental data to analyze genotypic responses. Thus, the citizen science approach scales E and M data collection to generate a volume of data that matches the big data dimension of G. Combining genomic prediction with citizen science opens the possibility of simultaneously capturing the three dimensions of crop performance, G, E, and M, in a data-driven way. Here, we describe and demonstrate potential benefits of this approach that we call data-driven decentralized breeding, or 3D-breeding, for short. Potentially, 3D-breeding could benefit the ~500 million smallholder farmers around the world who often produce in challenging, low-input environments and work with diverse cropping and farming systems and respond to local consumption preferences. We applied the 3D-breeding approach in the Ethiopian highlands, where many smallholder farmers grow durum wheat (Triticum durum Desf.) and select landraces following criteria related to environmental adaptation, food culture, and market demand. Rich local wheat diversity has co-evolved with local cultures and landscapes over millennia. Consequently, Ethiopian farmers still often select and cultivate local landraces, which under local conditions tend to outperform modern varieties produced by centralized breeding. In this context, 3D-breeding can leverage local wheat diversity and knowledge and bring breeding closer to the target environments cutting through the complexity of G × E × M. Here, we collected data from the genotyping and phenotyping of 400 wheat varieties in centralized stations commonly used for varietal selection in Ethiopian highlands. We then selected and distributed a subset of 41 genotypes as packaged sets containing incomplete blocks of three genotypes, plus one commercial variety to each of 1,165 farmers located in the same breeding mega-environment. We tested 3D-breeding against a competitive benchmark that represents breeding based on a genomic prediction model trained on centralized stations to predict varietal performance in farmers’ decentralized fields. We focused on grain yield (GY) and farmers’ overall appreciation (OA) of wheat genotypes, which were both recorded in centralized and decentralized trials. To establish the benchmark, we used a genomic prediction model trained on data measured in stations to predict wheat GY and OA in farmer fields. We then developed 3D-breeding to move the selection to farmer fields, predicting wheat performance in farmers’ fields using a decentralized approach. Comparing side by side the accuracy of the two methods, we found that 3D-breeding could increase prediction accuracy in challenging environments and thus complement genomics assisted breeding.

Plant materials and DNA extraction: The study selected 400 durum wheat (Triticum durum Desf.) genotypes, including purified landrace accessions from the Ethiopian Biodiversity Institute (EBI) and improved lines cultivated in Ethiopia. Landraces were purified from single spikes as per Mengistu et al. (2016). Genomic DNA was extracted from pooled leaves (five seedlings per accession) using GenElute Plant Genomic DNA Miniprep Kit. DNA quality/quantity was checked by agarose gel electrophoresis and Nanodrop. Genotyping used the Infinium 90k wheat SNP array (TraitGenetics GmbH). SNPs were called with the tetraploid wheat pipeline in GenomeStudio v11 and filtered for quality (positions/samples with >80% failure rate and heterozygosity >50%). SNP calls and provenance metadata are available on Dataverse. Centralized trials: Conducted in 2012 and 2013 at Geregera (Amhara) and Hagreselam (Tigray), representative highland stations. Experimental design: replicated alpha lattice with 400 genotypes (800 plots per field). Agronomy: four rows per plot (2.5 m), seeding rate 100 kg ha^-1; fertilizer: 100 kg ha^-1 DAP + 50 kg ha^-1 urea at sowing and 50 kg ha^-1 urea at tillering; manual weeding. Farmer evaluation (2012): 60 experienced smallholder farmers (15 men and 15 women per location) provided overall appreciation (OA) using Likert 1–5 scales at flowering; farmers had no prior knowledge of genotypes, were trained, and evaluated individually while guided plot-by-plot. Grain yield (GY) measured post-harvest (converted to t ha^-1). Additional traits per Mengistu et al. (2016). Decentralized trials: 1,165 farmer-managed fields across Amhara (471), Oromia (399), and Tigray (295) from 2013–2015 (Season 1: 179 fields; Season 2: 651; Season 3: 335). A subset of 41 genotypes (38 purified landraces + 3 modern cultivars) was tested using the triadic comparison of technologies (tricot) approach. Each farmer received a randomized incomplete block of three local genotypes plus one improved variety (Asassa in Tigray/Amhara; Hitosa or Ude in Oromia), totaling four plots per farmer (0.4–1.6 m² per plot). Farmers planted, managed, and evaluated their plots. OA was captured as a rank (best to worst among the four entries). Field technicians collected GY. Farmers participated without financial incentives and kept harvested grain. Centralized trait data analysis: Best linear unbiased predictions (BLUPs) for GY_STATION and OA_STATION were computed with ASReml-R, treating location as fixed and other factors as random. Broad-sense (H^2) and narrow-sense (h^2) heritabilities were computed. Agreement between men and women in OA_STATION was assessed via linear modeling. Spearman correlations between station BLUPs and farm performance were computed. Decentralized trait data analysis: OA_FARM rankings analyzed using the Plackett–Luce model (R package PlackettLuce), estimating worth parameters (a) where probability of genotype i winning in a set is a_i/(a_i + … + a_n). Benchmark (centralized genomic prediction): A rrBLUP-based genomic prediction benchmark was trained on BLUPs from centralized trials, including farmer OA_STATION. Accuracy was assessed using Kendall’s tau (r) between predicted and observed values (Pearson correlations were also computed and were similar). Two primary scenarios: (1) train on 2012 station GY_STATION and OA_STATION for 400 genotypes; predict 2013 station GY_STATION for the subset of 41 genotypes; (2) train on combined station GY_STATION and OA_STATION; predict OA_FARM and GY_FARM for the 41 genotypes in decentralized trials. Cross-validation averaged Kendall r per season, weighted by sqrt(sample size). Additional scenarios tested non-overlapping training/test sets, restricting training to the 41-genotype subset, and stratifying farm predictions by environmental distance from stations. 3D-breeding model: Built from decentralized trial data using Plackett–Luce Trees (PLT), enabling recursive partitioning on environmental covariates, and incorporating genomic SNP data via an additive relationship matrix. Agroclimatic covariates (n=110) were derived from NASA POWER daily temperature and precipitation using nasapower and climatrends, summarized over vegetative, reproductive, grain filling, and full-season windows. Model selection used blocked cross-validation (seasons as blocks) with forward selection based on combined Akaike weights from validation season deviances. The final PLT used two covariates: (i) maximum night temperature during reproductive growth and (ii) minimum night temperature during vegetative growth. Model hyperparameters: cut-off a=0.01; minimal partition size = 20% of dataset. Accuracy was Kendall r between observed rankings and predicted coefficients. To assess performance under limited data, models were also trained on 75%, 50%, 25%, 15%, and 5% of decentralized plots. Generalization and reliability: Future seasonal variability was represented by extracting 15 years (2001–2015) of NASA POWER daily climate; three sowing-date windows per season were defined from observed planting dates. Genotype performance was predicted across 45 seasonal scenarios for 1,200 random points covering the trial region. Reliability (probability of outperforming check varieties Asassa, Hitosa, Ude) was computed using worth parameters as P(ij) = a_i/(a_i + a_j). Genetic diversity and selection: PCA of genotype diversity illustrated separation of top 3D-breeding selections (cold- vs warm-tolerant) from currently recommended varieties. Environmental characterization: Agroecological zonation from EIAR; climate variables (selected temperature indices) retrieved per field-season; PCA and MDS used to summarize site variation and compute climatic distance from stations to farms. Genotypes were classified as cold- or warm-adapted by altitude of origin (one-tailed unequal-variance t-test). Statistics and reproducibility: Centralized experiments: 2 locations × 2 seasons × replicated plots for 400 genotypes (3,200 plots). Decentralized: 1,165 farmer fields × 4 plots = 4,660 plots. Data management and analyses conducted in R with packages including data.table, caret, gosset, PlackettLuce, qvcalc, climatrends, nasapower, and visualization packages.

- Heritability: Across locations, H^2 for GY_STATION was 0.55 (2012) and 0.42 (2013); OA_STATION H^2 was 0.78. By gender, OA_STATION H^2 was 0.84 (men) and 0.67 (women), with consistent evaluations between genders. - Station-to-station prediction (subset of 41 genotypes) achieved up to r = 0.248 for predicting next-season GY_STATION. - Benchmark (centralized genomic prediction trained on station data) to predict decentralized farm performance had low accuracy: • OA_FARM: Kendall r per season = 0.134, 0.105, 0.183; combined r = 0.141 (±0.039). • GY_FARM: Kendall r per season = -0.012, 0.076, 0.073; combined r = 0.046 (±0.049). • OA_STATION showed more consistent positive accuracy than GY_STATION in predicting on-farm performance; station and farm measures were weakly correlated. - 3D-breeding (PLT on decentralized data with genomic prior and climate covariates) outperformed the benchmark: • OA_FARM: Kendall r per season = 0.270, 0.276, 0.203; combined r = 0.251 (±0.040). • GY_FARM: Kendall r per season = 0.160, 0.078, 0.119; combined r = 0.109 (±0.041). • Under reduced training data (5–75% of plots), OA_FARM accuracy ranged r = 0.162–0.230; GY_FARM r = 0.076–0.106, still exceeding the benchmark. - Environmental robustness: 3D-breeding accuracies were not biased toward specific environmental conditions and captured environmental diversity better than the benchmark. - OA as a predictor: Overall appreciation consistently provided higher prediction accuracy than GY across farmer fields; OA relates to multiple traits with generally higher heritability and aligns with adoption likelihood. - Selection outcomes: 3D-breeding selected genotypes (largely landrace-derived) genetically distinct from currently recommended varieties and with higher worth. Reliability of outperforming current checks in most fields was high (0.83–0.91). - Agronomic gain: Simulations across 15 growing seasons indicated expected on-farm yield increases of about 20% when selecting the top three genotypes per local conditions from 3D-breeding. - Environmental covariates: The most informative covariates were maximum night temperature (reproductive phase) and minimum night temperature (vegetative phase), enabling clear cold- vs warm-environment recommendations.

The study addresses a central challenge in crop improvement: the poor transferability of performance from centralized selection environments to heterogeneous production environments due to complex G × E × M interactions. By integrating farmer-generated phenotypes (overall appreciation and yield), genomics, and environment data directly from production fields, the 3D-breeding approach substantially improved prediction accuracy for untested environments relative to a robust centralized genomic prediction benchmark. The superiority was consistent across seasons, robust to reduced training data, and not restricted to specific environmental niches, demonstrating better capture of local environmental variability. Farmer overall appreciation emerged as a reliable, heritable, and predictive trait, likely reflecting a composite of multiple performance attributes (e.g., yield, seed size, biomass, phenology) relevant to local preferences and constraints. The strong performance of OA in predicting both OA and GY underscores the value of participatory phenotyping to guide selection toward varieties that are both productive and adoptable in smallholder contexts. 3D-breeding recommendations were not only more accurate but also translated into practical advantages: high reliability of outperforming existing recommended varieties and substantial expected yield gains (~20%) across variable seasons. The approach effectively identified landrace-derived genotypes with enhanced local adaptation, suggesting that incorporating diverse, locally adapted germplasm can outperform centralized, broadly adapted improved varieties under real-world conditions. These findings support rethinking breeding program design to complement centralized pipelines with decentralized, data-driven participatory evaluations. Analogous to dairy genomics where broad, farmer-generated phenotypic data accelerated genetic progress, crowdsourced on-farm data can expand the environmental scope of crop breeding. Incorporating environmental covariates into predictive models enhances the capacity to tailor recommendations to microclimates and management contexts, improving adoption and resilience in the face of climate variability. Overall, 3D-breeding provides a scalable, cost-effective way to align selection with target environments, potentially increasing genetic gain and adoption in smallholder systems while informing parent choice for the development of locally adapted recombinant lines.

The data-driven focus of 3D-breeding enables embracing the complexity of real-world G × E for the benefit of breeding. The approach leverages documented methods—from experimental design (tricot) to data curation and analysis—and can scale to larger field samples and denser genomic datasets for increased power. Expanding across seasons and management conditions can reveal additional drivers of local performance beyond temperature. Future 3D-breeding studies may stratify participants by socioeconomic features (e.g., gender, age, income) to characterize traditional knowledge more fully. Ideally, 3D-breeding should be combined with conventional, centralized breeding to improve training of prediction models for local adaptation and to accelerate the development of new varieties through the crowdsourced combination of breeder and farmer knowledge. Subsequent research should assess impacts on conservation and use of agrobiodiversity in situ and beyond. Open-source digital tools and citizen science make 3D-breeding applicable across contexts and crops, enabling the identification of untested production niches and fostering co-evolution between farming systems and data-driven breeding to complement traditional breeding.

- Transferability limits of centralized trials: Even within the same agroecological zone, stations did not capture the full pedoclimatic and management variability of farmer fields, leading to low centralized benchmark accuracy. - Open operational questions for decentralized breeding: How to sustain farmer participation, optimal seed quantities per farmer, logistics of distributing diverse genetic materials, and mechanisms for equitable benefit sharing when farmer knowledge informs new varieties. - Trait scope and genetic gain: Accuracy is only one component of genetic gain; future work should integrate multi-trait models (including highly heritable correlated traits) and optimize resource allocation across centralized and decentralized stages. - Data and model scope: This proof-of-concept used 1,165 fields over three seasons and two key environmental covariates; broader multi-season, multi-management datasets and additional covariates could further improve generalization and identify new environmental drivers. - Socioeconomic stratification: While gender differences were analyzed in station OA, broader socioeconomic stratification in decentralized trials (e.g., by age, income) was not fully explored and could affect preferences and adoption. - Implementation costs and logistics: While per-data-point costs are lower than conventional on-farm trials and genotyping costs are minimal, scaling requires investments in seed multiplication, distribution, and communication infrastructures.