Genomics-Assisted Breeding for Climate-Resilient Wheat: A Systematic Review

3.3.2. Integration of SB with GS for Disease Resistance

SB can facilitate accelerated breeding for disease resistance by producing more generations per year and advancing segregating material before field testing. However, disease presence and expression may be influenced by adult-plant development, pathogen population, humidity, temperature, and inoculation method [25]. Controlled-environment screening should therefore be followed by field validation, as demonstrated in wheat Fusarium head blight screening [26].

3.4. Theme 4: Integration of Genomic Selection, GWAS, and Speed Breeding

3.4.1. Synergies Between Genomic Tools

Integrating genomic, phenomic, environmental, transcriptomic, and other omics data improves predictions; however, more data does not always lead to better decisions [27]. High dimensionality can increase overfitting, opacity, and computational costs when individual modalities are not biologically relevant, sufficiently sampled, or harmonised; targeted integration can be more effective when the included modalities are biologically relevant, adequately sampled, and harmonised [28]. Studies should therefore use parsimonious baseline models for comparison, quantify each modality’s contribution in ablation analyses, and perform external/forward-cycle validation [29–31].

3.4.2. Overcoming Environmental Variability with Genomic Data

Environmental variables include stress timing, temperature, precipitation, VPD, soil variables, and management. Computational approaches for G×E include reaction-norm models, marker × environment kernels, multi-trait/multi-environment Bayesian models, factor-analytic mixed models, enviromic similarity matrices, crop-growth-model covariates, random forests, gradient boosting, and deep neural networks [32–33]. Models can also cluster and weight environments in the target population and stress-test models under future climate scenarios [34]. Their value depends on temporal resolution, biological relevance to the crop’s agronomic stage, representation of the target environment, and validation in new, real-world environments and future cycles. Leave-one-environment-out and forward prediction are thus more informative than random cross-validation alone [35].

3.4.3. Stepwise Integration Model for Genomic Selection, GWAS, and Speed Breeding

Table 4 presents a sequence of planned implementation steps derived from the reviewed evidence, not a validated pipeline. It links discovery, prediction, rapid progress, and field confirmation, but acknowledges that each stage can introduce error, cost, bias, and loss of diversity [36]. The framework should be prospectively evaluated against a conventional breeding comparator.

No included study implemented this complete model across repeated breeding cycles with a conventional breeding comparator (Table 5). Therefore, the framework is a testable research agenda, not evidence of operational effectiveness.

In general, the prediction performance of GS was moderate but was context-dependent; the discovery evidence from GWAS was evident; the cycle time under controlled conditions was decreased by the use of SB; and, in some cases, the combination of GS with environmental information gave better robustness [37]. There was a Wheat simulation that indicated that gains were possible using GS-SB integration, but no study was found that actually simulated a full pipeline. The prediction accuracy, realised response to selection, multi-environment yield stability, generation interval, genetic diversity, energy use, and cost per line selected should be reported jointly by prospective trials. Simulations must be used to plan, and should be clearly distinguished from empirical testing [38].

Quantitative reporting was uneven: GS studies emphasised accuracy, GWAS studies emphasised significance and marker discovery, and SB studies emphasised calendar acceleration. This outcome heterogeneity is the principal reason why a pooled meta-analysis was not performed and why Table 6 uses a structured comparative format.

4. DISCUSSION

4.1. Strengths and Limits of the Empirical Evidence

There is empirical evidence for useful but limited roles for them, according to the tool they are used with. Wheat GS results have shown that candidate rankings (in drought or heat or across water regimes) may differ depending on the validation set, the quality of the phenomes, the environmental representativeness, and the training population [5, 9, 12]. In the absence of causal function and deployment value, wheat GWAS results yield marker-trait associations and candidate regions [16, 17, 22, 25]. Controlled-environment SB may also decrease generation intervals, but it does not always increase realised genetic gain [36]. The prediction accuracy of GS (~0.45-0.51) reported here might be enough for some operations, but if used as a breeding tool, it relies on the heritability of the traits, selection pressure, cost, relatedness to the population of interest, and environmental stability [39–40].

4.2. Simulation Evidence and Inference Boundaries

Simulation can make comparisons between breeding designs which would take years to test empirically. However, the expected gain is dependent on several assumptions relating to genetic architecture, recombination, effective population size, prediction decay, cost, and selection intensity. The wheat evidence set used in this review had potential for use in together using GS and SB as a way to improve resistance against Fusarium head blight; however, it provided no long-term field validation [17]. Other aspects of the breeding design literature also emphasise the need to avoid loss of accuracy, selection intensity, genetic variation, and the cost of operations, to achieve more rapid cycles leading to larger annual gains [41–43]. It would be beneficial in future studies for pre-registration of assumptions, comparing models with field trials, and model updating based on the results of field trials to be performed.

4.3. Genotype × Environment Interaction and Computational Approaches

GxE is not just a statistical “noise,” but a change in performance and ranking of genotypes and environments, which depends on the developmental stage, soil, management, and intensity and timing of stresses. There are several computer-based methods that are relevant, such as reaction-norm models, marker × environment kernels, factor-analytic mixed models, multi-trait/multi-environment Bayesian models, enviromic similarity matrices, crop-growth-model covariates, and machine-learning ensembles [44–46]. Deep learning may represent interactions that are not linear and not just one type of data, but may be successful depending on the trait and data. Leave-one-environment-out, leave-one-year-out, and forward-cycle validation should be used to evaluate the models, and uncertainty should be reported for the calibrated point predictions. There also needs to be biological synchrony with environmental variables. For instance, a short heat period during anthesis or drought during grain filling could be covered by the seasonal means. Data in these areas in particular, such as meteorology data, soil data, management data and crop state data, should therefore be specifically harmonised if there are missing data and data imprecision [47–48]. Future climate scenarios are not observations but extrapolations and should be viewed as sensitivity analysis and/or stress testing.

4.4. Speed Breeding Trade-offs and Diversity Safeguards

The major advantage of SB is reduced calendar time. However, controlled-environment infrastructure, electricity and cooling requirements, genotype-specific protocol optimisation, space constraints, altered plant morphology, and microclimate-dependent disease expression create important trade-offs [49–51]. Rapid cycling can also cause genetic bottlenecking through repeated use of a small group of parents or reuse of genomic rankings without updated field information. Diversity safeguards should include maintaining a minimum number of families, applying optimum-contribution selection, limiting coancestry, monitoring effective population size, and establishing mechanisms for transferring tested foreign material when appropriate [52–53]. Recent contextual literature similarly concludes that speed breeding is most effective when combined with stress-aware protocols, genomics-assisted selection, and multi-location field testing, rather than used as an isolated substitute for conventional evaluation [54–56].

4.5. Artificial Intelligence, Digital Phenotyping, and Prediction Algorithms

Machine learning (ML) and deep learning may leverage markers, imagery, spectral measurements, weather, management, and omics data. They may also capture nonlinear interactions that simpler linear models cannot [57]. Candidate algorithms include random forests; gradient boosting models for structured covariate time series; models for images and spectra, including CNNs; models for longitudinal weather or phenology time series (recurrent or transformer); and models for genomic, pathway, or environmental relationships (graph models) [58–59]. These should be evaluated against approaches of well-known transparency, such as GBLUP, Bayesian genomic regression, and factor-analytic approaches. Evaluation of models should be done with nested cross-validation for hyperparameter tuning and leave-environment-out and forward-cycle tests to determine the transferability of operation. Important aspects are calibration, uncertainty quantification, ablation analysis, interpretability, data leakage checks, and reporting of computational and energy costs. Digital phenotyping may decrease some of the measurement costs yet increase a few measurement biases, such as from sensor drift and inconsistencies in platforms, missing data, and unequal access [60]. AI should be used within a model that is under the supervision of a human or a breeder and for which the changes in the model are documented, and there are clear acceptance/rejection criteria for breeding lines.

4.6. Climate Resilience as a Dynamic Multi-trait Target

Climate resilience comprises yield stability, recovery, phenology, and eventually root and canopy traits, disease resistance, quality, and adaptation to compound or sequential stresses. Trade-offs may occur between improvements in different components; for example, early maturity may reduce yield potential or grain quality under favourable conditions but help avoid stress. Selection indices and multi-trait models should therefore account for the target population of environments, farmer/market preferences, and/or the probability of compound events [61]. The identification of a marker or demonstration of stress tolerance under controlled conditions should not be used alone as evidence of the climate resilience of a cultivar.

4.7. Economic Feasibility, Equity, and Resource-Limited Programmes

The implementation costs are related to skilled personnel, electricity and cooling, genotyped and phenotyped plants, locations, equipment and data engineering, and logistic seed. The best design in terms of economic impact is not always the most technologically sophisticated, but one that maximises the realised benefit over time and investment, and maintains genetic diversity and relevance of the product [62–63]. Other challenges that occur in developing country programmes are unreliable electricity and internet connection, few cold chains, less laboratory capacity, high import prices for consumables, small and scattered networks of tests, weak quantitative genetics and data-engineering skills, limited access to cloud computing and proprietary platforms. A phased approach may start with smaller phenotyping sets, low-cost phenotyping panels, and types of genotyping platforms that are broadly available. Reducing recurrent costs might then be achieved via the establishment of regional phenotyping networks, open-source pipelines, interoperable data standards, local training and coordinated procurement and maintenance. Programmes should include capital and operating costs and cost per cycle, cost per candidate advanced, energy wastage, man-hours, opportunity cost, scale sensitivity and uncertainty in economic evaluation. Governance arrangements should ensure that germplasm and data are openly available at the local level and do not place too many restrictions on climate-adapted germplasm, and should be freely available to vulnerable farming systems [64–65].

4.8. Genome Editing, Pangenomics, Multi-Omics, and Genomic Foundation Models

The Wheat Genome Initiative is transitioning towards a multi-linear reference genome. Structural and presence-absence variation is captured by multiple assemblies and/or pangenomes, which may be missed by SNP-only analyses [62, 63, 64, 66]. Multi-omics integration can link sequence variation to gene expression, proteins, metabolites, epigenetic states, and physiological responses, and genome editing can test candidate-gene function and introduce defined alleles [39, 65]. Transformation capacity, off-target assessment, trait complexity, regulation, intellectual property, and access remain constraints for these tools; however, they could shorten the pathway from discovery to validation. More recently, large sequence, graph, or multimodal architectures have been pretrained on large quantities of genomic and functional information (called foundation models) and customised to specific tasks. Wheat applications include prediction and transfer learning for traits with limited labelled data, prediction of wheat variants, annotation of regulatory regions, imputation, and prioritisation of causal alleles. However, polyploid homoeology, biased reference panels, lack of phenotype labels, limited population representativeness, high computational demands, and lack of interpretability may generate high-confidence, non-transferable predictions. Evaluation should include wheat-specific benchmarks, held-out populations and chromosomes, forward-cycle testing, calibrated uncertainty, comparison to simpler models, and wheat germplasm and farmer-linked data-governance safeguards. Foundation models should currently be used as research tools to prioritise hypotheses and should not be treated as validated substitutes for breeding trials.

4.9. Transferable Lessons from Maize, Rice, and Barley

Data from other cereals provided context but were not added to the wheat data set. The allogamous-crop SpeedGS simulation by Jighly et al. [67] was not included in the analytical evidence set and is discussed here only as modelling evidence from a different context. Maize studies show the usefulness of integrating genetic data with environmental information from multi-environment (ME) trials, but performance can also be unstable due to environmental effects. In barley, clear evidence shows that explicit G×E genomic prediction and the pangenome can uncover genomic variation that a single reference genome can mask [50, 68]. Interconnected programmes for variety testing, rapid variety replacement, and development of product profiles in rice and maize also highlight the importance of these priorities [60–61]. Wheat transfer must consider polyploidy, self-pollination, disease ecology, grain-quality goals, transformation effectiveness, and programme size. These comparisons support methodological transfer only; they are not directly comparable evidence of wheat performance.

4.10. Evidence Gaps and Benchmarking Priorities

The primary knowledge issue is whether an integrated system can provide continuous, low-cost realised gains compared with a powerful conventional comparator; it is not about whether the parts can work. Such a credible benchmark would include multiple cycles of iteration, similar population sizes and budgets, evaluation in target environments, documentation of failures as well as successes, and explicit management of genetic diversity. Current pipeline-integrated and component-level evidence therefore does not support categorical claims about pipeline superiority.

FUTURE DIRECTIONS

Future programmes will need to be staged and forward-looking. First, specify the target population of environments and create a multi-trait product profile that lists farmer/market and climate-risk priorities. Second, standardise phenotyping and envirotyping across linked field sites. Third, compare transparent GS baselines with G×E, ML, deep-learning, and multimodel approaches via leave-environment-out and forward-cycle validation, uncertainty calibration, and breeder review. Fourth, prioritise alleles using GWAS, pangenomes, multi-omics, genome editing, and outputs from foundation models; however, use alleles only after independent functional validation and field testing. Fifth, introduce SB only with diversity thresholds, energy accounting, disease-screening equivalence, and contingency plans for facility failure. Operational systems should specify data standards, model-version control, retraining triggers, decision-making thresholds, staff responsibilities, cybersecurity, and audit trails. Finally, benchmark the full life cycle of this workflow against conventional breeding across multiple breeding cycles, incorporating multiple co-primary outcomes: realised response, stability, quality, diversity, time, cost, and energy. These tests would be more reproducible and applicable to programmes with limited resources if they use common protocols, open metadata, regional infrastructure, and tiered steps toward adoption.

STRENGTHS OF THE REVIEW

This review is wheat-specific, distinguishes empirical results from simulation projections, and compares the results of these methods: GS, GWAS, SB, environmental integration, and GS plus SB. It also flags the lack of end-to-end benchmarks, elaborates on the reasons for the unsuitability of meta-analysis, and links methodological evidence to computational, diversity, economic, equity, and implementation constraints.

LIMITATIONS

Limitations of the review include the possibility of publication bias, the restriction to English-language studies, restriction to studies analysed in the period between 2018 and 2025, and restriction to studies of wheat where a technique or tool used to address climate change was tested, and a measurable impact on the outcomes was reported. For no outcome was it possible to perform a formal meta-analysis because of heterogeneity. A five-domain tool developed by the author and not yet validated was used for the methodological appraisal. One simulation study performed on wheat was included along with 19 empirical/real data studies. To account for more recent developments, contextual sources between 2024 and 2026 were added to the source list to provide a more recent update, but not to alter the pre-specified 20-study analytical data set. The recommendations on cost, equity, AI governance, and operational aspects are based on evidence, given that limited economic evidence was available.