The genomics revolution has transformed plant breeding, using genomic prediction to accelerate variety development. Conventional breeding programs typically use a centralized approach, maximizing genetic diversity early and then selecting superior germplasm based on phenotypic observations in research stations. This approach may not fully capture genotype-by-environment-by-management (G × E × M) interactions, leading to suboptimal variety development for farmers in challenging environments. Smallholder farming systems, representing 80% of global farmers, require tailored solutions to ensure food security, particularly in the face of climate change. To address this, breeders need methods to accelerate variety development while directly addressing G × E × M interactions. Mobilizing farmers' knowledge of local adaptation can enhance adoption of improved varieties. A crowdsourced citizen science approach allows for cost-efficient on-farm variety testing, potentially outperforming breeder evaluations. This approach integrates E and M components into breeding by performing selection directly in target environments, generating large-scale data to match the genomic data. Combining genomic prediction with this citizen science approach – termed 'data-driven decentralized breeding' (3D-breeding) – allows simultaneous capture of G, E, and M. This study demonstrates 3D-breeding's potential benefits in the Ethiopian highlands, where durum wheat is grown. Ethiopian farmers often select local landraces which outperform modern varieties under local conditions. 3D-breeding aims to leverage this local diversity and knowledge, bringing breeding closer to target environments.

The paper reviews the existing literature on genomic selection in plant breeding, highlighting the limitations of centralized approaches in capturing G × E × M interactions, particularly in smallholder farming systems. It discusses the importance of integrating farmers' knowledge and participatory breeding methods to improve variety adoption and local adaptation. Existing studies on crowdsourced citizen science and participatory plant breeding are reviewed, showcasing the feasibility and potential benefits of decentralized approaches. The literature also supports the concept that farmers' assessments can be highly predictive of varietal performance under on-farm conditions, even surpassing those made by breeders. Prior research on durum wheat in Ethiopia is reviewed, highlighting the prevalence of locally adapted landraces and their superior performance compared to modern varieties.

The study compared 3D-breeding with a benchmark representing conventional genomic selection. First, 400 durum wheat genotypes were genotyped and phenotyped in centralized trials in two Ethiopian locations (Geregera and Hagreselam) across two years (2012 and 2013). Grain yield (GY) and overall appreciation (OA) by farmers were measured. Heritability was calculated for GY and OA in the centralized trials. A genomic prediction model was trained using the centralized data. Next, a subset of 41 genotypes were distributed to 1165 farmers across three regions (Amhara, Oromia, Tigray) for decentralized trials in three growing seasons (2013-2015). The tricot approach was used, allocating sets of three genotypes plus a commercial variety to each farmer. Farmers provided OA rankings, and GY was measured by technicians. The benchmark used the genomic prediction model trained on centralized data to predict GY and OA in the farmer fields. 3D-breeding was implemented using a Plackett-Luce Tree (PLT) model, incorporating farmers' OA rankings from the decentralized trials, along with genotype data and agroclimatic indices as covariates. Model accuracy was assessed using Kendall's tau correlation (r) between predicted and observed values. Additional analysis explored the effects of data availability, geographical location, and traits focused on (OA or GY) on model accuracy. Finally, the 3D-breeding model was used to predict the probability of selected genotypes outperforming currently recommended varieties, and the expected yield increase across simulated growing seasons. Statistical analysis was conducted using R software, and different methods were used to evaluate the heritability of traits and the performance of prediction models. Further statistical and computational details are given in the Supplementary Information.

The heritability of OA in centralized trials was higher than that of GY, suggesting farmers' evaluations are reliable. The benchmark (genomic prediction from centralized data) had low prediction accuracy for both GY and OA in farmer fields (average r = 0.046 for GY, 0.141 for OA). 3D-breeding significantly improved prediction accuracy. It provided higher accuracy for both GY and OA in farmer fields (r = 0.109 for GY, 0.251 for OA), consistently outperforming the benchmark across all seasons and even with reduced datasets. OA was a better predictor than GY in both centralized and decentralized trials. 3D-breeding identified genotypes with high OA that consistently outperformed currently recommended varieties, showing a substantial yield advantage (around 20%) across simulated seasons. These top-performing genotypes selected by 3D-breeding had a distinct genetic background, with a high proportion derived from landraces rather than improved varieties. The superior performance of 3D-breeding was consistent across various scenarios, including different data subsets and environmental conditions. The study also noted that the centralized approach provided more accurate GY predictions when restricted to cold-tolerant genotypes, suggesting that centralized stations may not adequately represent the full range of environmental variation in farmer fields.

The study's results demonstrate that 3D-breeding significantly improves prediction accuracy in challenging crop production environments compared to a conventional genomic selection benchmark. The superior performance of 3D-breeding likely stems from its integration of farmers' knowledge and on-farm data, capturing the complexity of G × E × M interactions better than centralized approaches. This highlights the importance of decentralizing breeding efforts to improve local adaptation and increase the effectiveness of breeding programs. Farmers' OA proved to be a valuable predictor of on-farm yield, further emphasizing the significance of incorporating farmer perspectives. The findings suggest that integrating farmers' knowledge can be key to improving breeding outcomes, especially in smallholder farming systems. The use of locally adapted landraces underscores the potential to tap into existing agrobiodiversity. The results have implications for the design of future breeding programs, suggesting a complementary approach where decentralized citizen science can enhance conventional breeding.

This study demonstrates the superior performance of 3D-breeding over conventional genomic selection for predicting crop performance in smallholder farming systems. The integration of farmers' knowledge and on-farm data significantly improves prediction accuracy and leads to the identification of locally adapted high-performing genotypes. This approach offers a promising strategy for accelerating varietal improvement, enhancing local adaptation, and improving food security in challenging environments. Future research should explore optimal farmer participation strategies, logistics of material distribution, and benefit-sharing mechanisms to maximize the impact of 3D-breeding. Combining 3D-breeding with conventional breeding, investigating larger germplasm collections, and incorporating multi-trait models could further enhance its effectiveness.

While the study used a large dataset, the generalizability of findings may be limited by the specific context of durum wheat in the Ethiopian highlands. Future studies should evaluate the 3D-breeding approach in different crops, agro-ecological zones, and farming systems. The study's reliance on farmer OA rankings as a primary trait may not fully capture all aspects of crop performance. Further research could explore the integration of more objective phenotypic traits in the model to ensure comprehensive assessment. The study's design did not explicitly investigate the influence of socioeconomic variables other than gender on farmer's evaluations, leaving room for further exploration.