Multimodal deep learning improves cross-environment prediction of durum wheat yield components

Abstract

Durum wheat breeding increasingly requires tools that can anticipate how genotypes respond to the shifting mix of heat and drought typical of Mediterranean drylands. In this study, we explored whether a multimodal deep learning (MM-DL) framework—one that brings together genomic markers, environmental covariates (ECs), near-infrared spectroscopy (NIRS), and phenology—can improve predictions for five key yield-related traits: grain number per spike (GN), grain weight per spike (GW), number of spikelets per spike (NS), spike length (SL), and spike weight (SW). The evaluation relied on multi-environment data from three contrasting seasons, using two scenarios with direct relevance for breeding: predicting an entirely unseen season (Prediction of Year, PoY) and predicting new sowing environments within the same season (Prediction of Sowing Year, PoSY). Across all traits, integrating multiple data sources improved prediction accuracy relative to genomics alone, with gains varying according to the physiological basis of each trait. SL was consistently the easiest trait to predict (PoY ≈ 0.56–0.71; PoSY ≈ 0.74–0.80), followed by NS (PoY ≈ 0.23–0.47; PoSY ≈ 0.56–0.67). GW showed moderate and dependable accuracy (PoY ≈ 0.24–0.30; PoSY ≈ 0.34–0.46), while SW yielded intermediate values (PoY ≈ 0.27–0.38; PoSY ≈ 0.34–0.41). GN remained the most environmentally sensitive trait, with modest accuracy under PoY (0.28–0.32) and clearer gains within PoSY (0.43–0.53). ECs contributed most to improving cross-season transferability, whereas NIRS and phenology added smaller but still useful trait-specific signals. Overall, the results show that combining genomics with enviromics and phenomics produces more stable and biologically informed predictions. For durum wheat breeding under increasing climate variability, MM-DL offers a practical path toward more reliable selection decisions across seasons.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12870-026-08537-3.

Introduction

Durum wheat [Triticum turgidum L. ssp. durum (Desf.) Husn.] is central to the agri-food economies of many regions and remains the essential raw material for pasta, couscous, freekeh, and a wide range of semolina-based foods. Globally, the crop covers close to 17 million hectares, with the Mediterranean Basin, North Africa, and parts of North America (i.e., Canada) serving as its main production hubs. These areas rely predominantly on rainfed agriculture, where seasonal rainfall is uncertain, and farmers must contend with increasingly unstable weather patterns. Over the last decade, prolonged droughts, erratic precipitation, and frequent episodes of heat stress have become more common. These conditions often coincide with sensitive growth stages—stem elongation, heading, flowering, and grain filling—making durum wheat particularly vulnerable during the physiological windows that most determine final yield [1–4].

This growing climatic volatility has slowed the progress of breeding programs that rely on predictable environmental conditions to identify stable, high-performing lines. Genomic selection (GS) helped accelerate breeding by enabling early selection, shortening generation intervals, and improving the rate of genetic gain [5–7]. However, GS-based predictions often decline when tested on seasons with climatic profiles that differ from those used for training. Differences in stress timing, temperature patterns, and moisture availability can alter genotype performance and amplify genotype × environment (G×E) interactions. As a result, predictions that appear reliable within a given season can fail when applied to a future year, limiting their usefulness for making year-ahead decisions [8–10].

In recent years, breeders and quantitative geneticists have increasingly turned to enviromics to address this challenge. Enviromics aims to represent environmental conditions in a biologically meaningful way by transforming raw weather variables—such as radiation, vapor pressure deficit, temperature, and water availability—into covariates aligned with key developmental phases. Instead of treating the environment as a nuisance factor, enviromic approaches characterize the drivers of phenotypic plasticity and allow models to understand how genotypes respond across diverse conditions [11–15]. This shift has improved the physiological interpretability of prediction models and enhanced their ability to generalize to new seasons.

Parallel to these developments, high-throughput phenotyping technologies have become more accessible. Near-infrared spectroscopy (NIRS), in particular, provides a rapid and cost-effective way to assess grain biochemical composition, which often reflects how plants respond metabolically to heat, drought, and nutrient limitations. When combined with genomic markers, NIRS adds an additional biological layer that helps distinguish genotypes based on their functional responses, not just their genetic background [16–19].

These advances naturally converge toward multimodal modeling frameworks—approaches capable of integrating multiple complementary data streams. Deep learning (DL), especially multimodal architectures, has shown promise for capturing nonlinear relationships and complex interactions among genetics, physiology, and environment. Several studies have already demonstrated the value of neural networks for genomic prediction (GP) and multi-environment modeling in wheat and maize [20–24]. Yet, there remains a gap in understanding how these multimodal systems perform when multiple biological and environmental layers are fused in the context of durum wheat, a crop particularly sensitive to environmental variability.

This gap is especially critical in the Mediterranean drylands, where climate uncertainty is no longer an exception but the norm. For breeding programs operating under these conditions, prediction frameworks must not only exploit genomic information but also incorporate environmental and physiological descriptors that capture the dynamic nature of genotype performance across seasons and sowing environments.

With this motivation, the present study evaluates a multimodal deep learning (MM-DL) framework designed to integrate four major sources of information—genomic markers, environmental covariates (ECs), NIRS, and phenological traits—to predict five important yield-related traits in durum wheat: grain number per spike (GN), grain weight per spike (GW), number of spikelets per spike (NS), spike length (SL), and spike weight (SW).

More specifically, we aim to: (1) Assess the predictive performance of MM-DL models under two deployment-relevant scenarios: (i) Prediction of Year (PoY): forecasting performance in a future season not present in the training data; (ii) Prediction of Sowing Year (PoSY): predicting new sowing environments within the same season, where climatic context is shared. (2) Quantify the individual and combined contribution of each data modality—genomics, ECs, NIRS, and phenology—to predictive ability. (3) Examine why traits differ in their cross-environment predictability, linking model behavior to known physiological mechanisms. (4) Identify which traits and scenarios offer the greatest potential for reliable selection in climate-variable durum wheat breeding programs.

By bringing together multiple complementary layers of information, this study provides a comprehensive evaluation of how MM-DL can improve cross-season prediction and support more resilient breeding strategies in durum wheat.

Although grain yield is the ultimate target of selection in durum wheat breeding, yield integrates multiple developmental and physiological processes that occur at different growth stages and respond differently to environmental stress. In climate-variable environments, direct yield prediction can obscure the mechanisms underlying genotype × environment interaction. For this reason, yield components are routinely analyzed as biologically meaningful intermediates that help diagnose stress sensitivity and interpret yield stability. Focusing on these components allows us to disentangle genetic, environmental, and physiological contributions to yield formation and to evaluate how multimodal prediction models capture these processes across seasons and sowing environments.

Materials and methods

Plant material and experimental design

The study utilized a diversity panel of 186 durum wheat genotypes, including 43 landraces (LR), 45 old cultivars (OC), and 98 modern cultivars (MC) (Supplementary Table S1). A detailed genomic characterization of this panel, including genotyping procedures, along with the experimental design details, was previously reported by Puglisi et al. [25]. The panel was evaluated across three cropping seasons in Foggia, Italy (41°28′N, 15°32′ E; 75 m a.s.l.), assessing multiple sowing dates (early indicated as 1; optimal as 2; delayed as 3) to maximize environmental variability.

Phenotypic data and near-infrared spectroscopy

Phenotypic data were collected for five yield-related traits: GN, GW, NS, SL, and SW (Supplementary Tables S2 and S3). Traits were measured at the ripening stage, Zadoks’ growth stage 90 [26] on manually harvested spikes. In addition to yield components, morpho-phenological traits, including plant height (PH) and heading date (HD), were recorded and utilized as trait covariates (TCs) in the prediction models.

Whole-grain flour samples were scanned using a FOSS Infratec 1241 analyzer to obtain NIRS. Spectral data were preprocessed using Standard Normal Variate (SNV) transformation, followed by Savitzky-Golay smoothing (second-order polynomial, 11-point window) to correct for baseline shifts and scattering effects [27].

Environmental covariates (Enviromics)

Environmental characterization was performed using the EnvRtype R package [28]. Daily weather records (minimum and maximum temperature, precipitation, solar radiation) were obtained from an on-site meteorological station and integrated with NASA POWER reanalysis data. A total of 32 ECs were computed to capture thermal, radiation, and water-deficit stresses. To account for phenological variation, ECs were aggregated into three developmental windows: sowing time-heading date (ST–HD), heading date-flowering time (HD–FT), flowering time-harvest time (FT–HT), and across all developmental timing (All DT). These enviromic attributes summarize heat and drought exposure, timing, and intensity.

Deep learning models and multimodal data integration

In this section, we describe the MM-DL framework used to integrate genomic, enviromic, and NIRS data for cross-environment prediction in durum wheat. Because our objective is to emphasize biological interpretation and practical deployment rather than algorithmic formalism, the models are presented functionally rather than algebraically. Traditional neural network equations, while rigorous, often obscure the rationale for multimodal fusion and its relevance to G×E modeling. Instead, we adopt a conceptual description that clarifies how modality-specific encoders extract complementary signals, how skip connections preserve additive genetic and environmental effects, and how fusion layers capture genotype-specific adaptive responses.

Instead of focusing on complex equations, we emphasized clear descriptions of the model structure, how each data source contributes, and why the architecture is biologically consistent. This follows current best practices in GP and MM-DL, where transparency, interpretability, and the integration of complementary information are essential for real progress in crop improvement.

A total of 20 models (Table 1) were evaluated for each target trait. These included both unimodal models (M01)—using information from a single data type—and multimodal models (M02-M20) that combined multiple biological and environmental signals. The modalities considered were: (i) genomic markers (G) capturing additive genetic potential; (ii) ECs summarizing weather conditions at specific growth phases; (iii) NIRS features reflecting biochemical and metabolic status; and (iv) trait covariates (TCs), including PH and HD, which reflect phenological adaptation. Together, these modalities represent the four major components that determine yield formation: genetic potential (G), environmental stress exposure (ECs), physiological and metabolic responses (NIRS), and developmental timing, enabling models to learn the biological pathways leading to GN, GW, NS, SL, and SW expression across environments. These 20 models were designed to:

Table 1.

Summary of the unimodal and multimodal prediction models evaluated in this study (M01–M20), indicating the data sources included in each model: genomic markers (G), environmental covariates (ECs) aggregated for different developmental timing [sowing time-heading date (ST–HD), heading date-flowering time (HD–FT), flowering time-harvest time (FT–HT)] and across all developmental timing (All DT), Near-infrared spectroscopy (NIRS), and trait covariates [TCs: plant height (PH) and heading date (HD)]

ID	Genomic (G)	Environmental covariates (ECs): developmental timing			Near-infrared spectroscopy (NIRS)	Trait covariates (TCs: PH and HD)
		ST–HD	HD–FT	FT–HT
M01	✓
M02	✓					✓
M03	✓				✓
M04	✓				✓	✓
M05	✓	✓
M06	✓	✓				✓
M07	✓	✓			✓
M08	✓	✓			✓	✓
M09	✓		✓
M10	✓		✓			✓
M11	✓		✓		✓
M12	✓		✓		✓	✓
M13	✓			✓
M14	✓			✓		✓
M15	✓			✓	✓
M16	✓			✓	✓	✓
M17	✓	✓*	✓*	✓*
M18	✓	✓*	✓*	✓*		✓
M19	✓	✓*	✓*	✓*	✓
M20	✓	✓*	✓*	✓*	✓	✓

^*All DT (across all developmental timing)

1. quantify the individual contribution of each data type to predictive ability,

2. evaluate the added value of multimodal integration under G×E, and

3. Identify the biologically complementary signals most helpful in predicting yield components in durum wheat.

Because EC, NIRS, and TC are not known in real breeding deployments for future growing seasons, test-environment values for these modalities were replaced with the mean of the training environments, ensuring realistic and operational predictions. This strategy simulates a practical breeder’s situation in which only G is known before the crop season starts, while environmental and phenotyping information becomes available only after environmental exposure. This choice avoids optimistic bias and ensures that reported accuracies reflect realistic breeder deployment conditions rather than idealized access to future environmental information.

A complete summary of the unimodal (M01) and multimodal (M02-M20) model configurations evaluated in this study is provided in Table 1.

Cross-environment evaluation

Predictive performance was evaluated using a Leave-One-Environment-Out (LOEO) strategy:

1. Training: all environments except one sowing date × year environment, and

2. Testing: the held-out environment.

This validation strategy forces the model to extrapolate into environmental conditions not observed during training, which is the key deployment challenge in climate-variable breeding pipelines. This enabled comparison of two relevant breeding deployment scenarios:

1. Prediction of Year (PoY): Training on earlier years and predicting performance in a future growing season with unseen environmental conditions.

   • → Strongest G×E challenge → Most important for increasing genetic gain per year.
2. Prediction of Sowing Year (PoSY):Training and testing within the same season, but predicting a new sowing date in a different environment.

   • → Captures intra-season heat/drought timing and photoperiod effects

This validation structure ensures that conclusions are directly relevant for breeding decisions where candidate genotypes must be selected before environmental outcomes are known.

Prediction accuracy was measured using Pearson’s correlation (r) between predicted and observed phenotypes.

Model architecture: multimodal residual neural networks

All prediction models were implemented as multimodal residual neural networks (ResNets). ResNets are effective for plant-breeding prediction because they can represent multiple layers of biological organization. Early layers capture simple additive signals, such as SNP effects or local spectral features. As information flows through the network, deeper layers combine these simple features into more complex patterns, including genotype-specific responses to environments and trait inter-dependencies. The skip connections ensure that stable genetic effects are preserved, while the nonlinear transformations learn how different data sources (genomic, environmental, and phenomic) interact to drive final yield performance.

Architecture components:

• Modality-specific encoders

  • Dense → BatchNorm → ReLU
  • → capture the structure of each biological layer independently
• Multimodal fusion layer

  • → concatenates encoded representations to learn joint effects.
• Two residual blocks

  • Dense → BatchNorm → ReLU → Dense → BatchNorm + skip connection
  • → enable flexible learning of G×E patterns
• Single linear output neuron

  • → predicts continuous phenotypes (GN, GW, NS, SL, SW)

Residual transformation:

1

This structure prevents vanishing gradients, stabilizes convergence, and is effective in GP problems where preserving genetic signal is crucial while learning complex interactions.

Training used the Adam optimizer, minimizing mean squared error (MSE), with early stopping and L2 regularization to reduce overfitting. Each model was trained independently per trait and modality combination. Dropout regularization was applied within each residual block to reduce co-adaptation of neurons and mitigate overfitting.

MM-DL architecture integrating G, ECs, and NIRS, for predicting GN, GW, NS, SL, and SW in durum wheat (Fig. 1). Modality-specific encoders feed a fusion layer followed by two ResNet blocks with skip connections, preserving stable genetic signals while capturing nonlinear environment-responsive effects. Arrows indicate sequential information flow from modality encoders to fusion and residual blocks.

Fig. 1.

Conceptual overview of the multimodal deep learning (MM-DL) framework. Genomic (G), environmental covariates (ECs) across different developmental timing [sowing time-heading date (ST–HD), heading date-flowering time (HD–FT), flowering time-harvest time (FT–HT), and across all developmental timing (All DT)], near-infrared spectroscopy (NIRS), and Trait covariates [TCs: plant height (PH), and heading date (HD)] data are processed through modality-specific encoders. Skip connections preserve additive information, fusion layers integrate nonlinear cross-modal interactions, and the final layer predicts yield component traits [grain number per spike (GN), grain weight per spike (GW), number of spikelets per spike (NS), spike length (SL), and spike weight (SW)] across environments

Unlike black-box models, the multimodal ResNet architecture enables mapping of specific trait-environment interactions, improving physiological interpretability and model trustworthiness for breeders.

Hyperparameter optimization

A Bayesian Optimization (BO) framework [29] was used to jointly fine-tune learning rate, number of layers and hidden units, dropout rate, L2 regularization, and weight decay. The search space was defined on log-scaled parameters, and the Expected Improvement (EI) acquisition function guided 250 optimization iterations. Five random initial evaluations improved exploration of the hyperparameter space.

This systematic procedure ensures that differences in predictive ability reflect the value of modality combinations, rather than arbitrary hyperparameter choices.

Reproducibility and deployment readiness

All models were developed using Python v3.9.19 and PyTorch v2.4.0 with GPU acceleration. Full training logs, hyperparameters, and random seeds were recorded to guarantee repeatability. All model code, preprocessed data, and training logs are available upon reasonable request from the corresponding authors.

Overall, this framework provides a scalable and operational strategy to support informed selection and resource allocation in climate-variable durum wheat breeding pipelines.

Results

The results are organized by trait, and within each trait, two prediction model analyses were performed: (1) prediction of a year, and (2) prediction of a combination of sowing time – year (environment). Figures 2, 3, 4, 5 and 6 show only the top five prediction accuracies for the PoY and PoSY breeding deployment scenarios for each yield component. The top 20 prediction accuracies, representing a subset of the evaluated unimodal and multimodal model configurations (Table 1), are reported in Supplementary Tables S4 and S5.

Fig. 2.

Top five prediction correlations for grain number per spike (GN) across two scenarios: A Prediction of Year (PoY), colored by training years; and B Prediction of Sowing Year (PoSY), colored by environment and training-year scenarios. The dashed line, together with the numeric values, indicates the mean prediction accuracy of the top-20 models reported in Supplementary Tables S4 and S5

Fig. 3.

Top five prediction correlations for grain weight per spike (GW) across two scenarios: A Prediction of Year (PoY), colored by training years; and B Prediction of Sowing Year (PoSY), colored by environment and training-year scenarios. The dashed line, together with the numeric values, indicates the mean prediction accuracy of the top-20 models reported in Supplementary Tables S4 and S5

Fig. 4.

Top five prediction correlations for number of spikelets per spike (NS) across two scenarios: A Prediction of Year (PoY), colored by training years; and B Prediction of Sowing Year (PoSY), colored by environment and training-year scenarios. The dashed line, together with the numeric values, indicates the mean prediction accuracy of the top-20 models reported in Supplementary Tables S4 and S5

Fig. 5.

Top five prediction correlations for spike length (SL) across two scenarios: A Prediction of Year (PoY), colored by training years; and B Prediction of Sowing Year (PoSY), colored by environment and training-year scenarios. The dashed line, together with the numeric values, indicates the mean prediction accuracy of the top-20 models reported in Supplementary Tables S4 and S5

Fig. 6.

Top five prediction correlations for spike weight (SW) across two scenarios: A Prediction of Year (PoY), colored by training years; and B Prediction of Sowing Year (PoSY), colored by environment and training-year scenarios. The dashed line, together with the numeric values, indicates the mean prediction accuracy of the top-20 models reported in Supplementary Tables S4 and S5

Grain number per spike

Prediction of year

PoY results for GN showed low-to-moderate accuracy (Fig. 2A; Supplementary Table S4), reflecting the strong environmental sensitivity of this trait and the difficulty of extrapolating across seasons with contrasting stress patterns. The highest correlation obtained under PoY was r = 0.32 for the model M14 [G + ECs (FT–HT) + TCs], followed by other multimodal combinations that integrated ECs and developmental covariables. Models using ECs aligned with reproductive phases (FT–HT and ST–HD) consistently outperformed M01 models, indicating that heat, moisture, and vapor pressure deficit around anthesis must be explicitly modeled for GN. NIRS contributed small but positive gains. M01 models showed limited performance (r ≈ 0.29). Overall, Supplementary Table S4 shows that GN has limited cross-year transferability, even under multimodal modeling, due to its dependence on short, stress-prone physiological windows governing floret survival and grain set.

Prediction of sowing year

Within-season prediction (PoSY) produced substantially higher accuracies for GN (Fig. 2B; Supplementary Table S5). The best-performing model, M04, reached r = 0.53, with M14 (r = 0.52) and M05 (r = 0.50) also ranking among the top five models. Because sowing environments within the same year share the same overall climatic trajectory, the models were able to capture genotype-specific physiological responses with much greater stability. NIRS features were particularly informative, reflecting their ability to encode biochemical signals related to reproductive metabolism and spike fertility. TCs (PH and HD) improved performance by capturing sowing-dependent developmental timing. M01 models performed notably better in PoSY (r = 0.45) than in PoY (r = 0.29), though still below multimodal architectures. Supplementary Table S5 thus demonstrates that GN can be predicted with operational accuracy within a season, provided that environmental and phenomic information are included.

Comparison between PoY and PoSY for GN

A direct comparison of Fig. 2A and B highlights a pronounced scenario effect. The maximum PoY accuracy (r ≈ 0.32) was far lower than the maximum PoSY accuracy (r ≈ 0.53), corresponding to an absolute gain of ~ 0.21 and a relative improvement of ~ 65%. This contrast indicates that GN is highly dependent on climatological synchrony. Year-to-year variation in heat and drought timing, especially during booting and anthesis, leads to poor cross-year extrapolation (Fig. 2A). In contrast, the shared seasonal context among sowing dates allows models to capture GN plasticity much more effectively (Fig.2B). The most influential modality under PoY was ECs (i.e., M14), while under PoSY the strongest gains came from NIRS + TC (i.e., M04). In both scenarios, genomic information alone was insufficient.

Summary for GN

Figure 2 shows that GN is the most environmentally plastic yield component. Under PoY (Fig. 2A), accuracies were modest (top r ≈ 0.32), limited by the unpredictable nature of annual stress patterns during the critical reproductive window. In contrast, PoSY results (Fig. 2B) showed substantially higher accuracies (top r ≈ 0.53), as within-season climatic consistency allows multimodal models—especially those integrating NIRS and developmental covariables—to more effectively represent genotype-specific reproductive responses. GN can therefore be predicted reliably within the same season, but cross-year prediction remains challenging, even when environmental information is available.

Grain weight per spike

Prediction of year

PoY results for GW showed moderate accuracy (Fig. 3A; Supplementary Table S4), with prediction performance comparable to GN under M01 models. However, GW displayed more consistent gains across models incorporating environmental covariates, whereas GN remained highly sensitive to environmental timing and stress effects even under multimodal configurations. As shown in Fig. 3A, the best-performing model under PoY was M13 [G + ECs (FT–HT)] with r = 0.30, followed closely by M18 (r = 0.29) and M05 (r = 0.29). These results indicate that enviromic covariates, particularly those aligned with the grain-filling period (FT–HT), contribute meaningful improvements by capturing environmental influences on assimilate availability and kernel filling duration. Models incorporating NIRS and TCs showed positive but smaller gains. M01 models reached accuracies around r ≈ 0.29, confirming that GW is genetically more stable across years than GN, but still benefits from the inclusion of environmental information when predicting unobserved seasons. Overall, Fig. 3A demonstrates that GW maintains moderate cross-year predictability, with enviromic descriptors providing consistent improvements over genomic baselines.

Prediction of sowing year

PoSY results for GW were significantly higher than PoY, with accuracies reaching operationally useful levels (Fig. 3B; Supplementary Table S5). The top model was M05 [G + ECs (ST–HD)] with r = 0.46, followed by M03 (r = 0.44) and M14 (r = 0.43). Within-season climatic consistency allows models to more accurately capture genotype-by-environment interactions affecting grain filling, including temperature-driven variation in kernel weight and stress effects on carbohydrate deposition. NIRS was particularly valuable in PoSY for GW, reflecting its ability to encode biochemical and compositional signals related to kernel density and endosperm structure. M01 models also performed well (r ≈ 0.42), outperforming their PoY counterparts and confirming the inherent stability of GW under shared seasonal context. Altogether, Fig. 3B shows that GW can be reliably predicted within a season, with the highest accuracies arising from combinations of genomics, ECs, and NIRS.

Comparison between PoY and PoSY for GW

Comparing Fig. 3A and B reveals a clear scenario-dependent improvement. The maximum PoY accuracy (r ≈ 0.30) is substantially lower than the maximum PoSY accuracy (r ≈ 0.46), representing an absolute gain of ~ 0.16 and a relative improvement of ~ 53%. This pattern reflects the physiology of GW: although kernel weight is influenced by post-anthesis stress, the trait is less sensitive to short-term stress timing than GN, making it more predictable across sowing environments within the same year. Enviromic covariates—especially those capturing heat and moisture conditions during grain filling—were essential for improving both PoY and PoSY, though their influence was more pronounced in cross-year predictions. NIRS contributed more strongly under PoSY, capturing quality-related and biochemical attributes that vary with in-season environmental differences. Genomics performed reasonably well in both scenarios, consistent with the high heritability and stable genetic control of GW.

Summary for GW

Figure 3 shows that GW is a moderately stable and genetically driven trait with clear improvements from multimodal integration. Under PoY (Fig. 3A), accuracies reached r ≈ 0.30–0.29, with ECs providing the largest boosts by accounting for stress during the grain-filling period. Under PoSY (Fig. 3B), accuracies increased sharply, reaching r ≈ 0.46, as within-season climatic synchrony allowed models to better capture genotype-specific responses. NIRS further improved prediction in PoSY, reflecting its ability to encode kernel biochemical signatures. Genomic-only models performed better for GW than for GN, confirming the trait’s stronger heritable component. Overall, GW demonstrates robust predictability both within and across seasons, with the highest accuracy obtained under PoSY and consistent benefits from multimodal data integration.

Number of spikelets per spike

Prediction of year

Across seasons, NS was one of the most stable and predictable traits in the study. As shown in Fig. 4A, M01 models already produced moderate correlations (r ≈ 0.47), reflecting the strong genetic control of spikelet formation (Supplementary Table S4). The best-performing PoY models were those that incorporated enviromic covariates. These additions improved prediction by ~ 0.04–0.10 r-units, suggesting that pre-anthesis temperature, photoperiod, and early-season stress patterns play a measurable role in shaping spikelet number across years. TCs (PH and HD) and NIRS contributed small but positive gains.

Overall, Fig. 4A highlights that NS maintains good predictability even when forecasting an entirely new season, thanks to its early determination and relatively low exposure to late-season heat or drought.

Prediction of sowing year

Prediction within the same season led to even higher accuracies for NS (Fig. 4B; Supplementary Table S5). The top multimodal models reached r ≈ 0.65–0.67, making NS one of the most reliably predicted traits in the dataset. Because sowing environments within a year share similar temperature and stress trajectories during spike development, the models were better able to capture genotype-specific responses. ECs again played an important role, while the integration of NIRS and TCs added useful information about metabolic patterns and phenological differences among genotypes. M01 models also performed strongly under PoSY and, in fact, achieved the highest accuracy in Fig. 4B (r = 0.67). Multimodal configurations (e.g., G + NIRS and G + ECs) frequently ranked among the top-five models and showed comparable performance, but did not uniformly outperform M01 prediction for NS under PoSY. Taken together, Fig. 4B shows that NS can be predicted with high confidence within season, and that combining complementary data sources leads to clear improvements.

Comparison between PoY and PoSY for NS

Across both prediction scenarios (Fig. 4), NS exhibited consistently more stable prediction accuracy than most other yield components. M01 models already achieved moderate correlations, and the inclusion of environmental covariates and NIRS produced incremental but consistent improvements. Comparing Fig. 4A and B reveals a consistent increase in accuracy when predictions are made within the same season. While PoY correlations topped around r ≈ 0.47, PoSY values reached r ≈ 0.67, representing an improvement of roughly 0.20 r-units. This scenario gap is moderately smaller than the one observed for GN, which reflects the biology of NS: spikelet number is determined early, before the most intense and unpredictable heat or drought events occur. As a result, NS shows much less environmental plasticity and maintains strong genetic determinism across seasons. ECs improve prediction in both scenarios, though their effect is more moderate than for highly stress-sensitive traits. NIRS and developmental covariables contribute more under PoSY, where subtle differences in phenology and carbon allocation are more consistently expressed.

Summary for NS

Figure 4 clearly shows that NS is one of the most robust and predictable traits in durum wheat. Its early developmental timing and strong genetic basis make it far less sensitive to the erratic stress patterns that limit cross-year predictions for other traits like GN. Under PoY, NS reaches moderate accuracy (r ≈ 0.35), and under PoSY, prediction improves further (r ≈ 0.62). Under PoSY, both genomic-only and multimodal models achieved high prediction accuracy (up to r ≈ 0.67), indicating that NS can be predicted with high confidence within season. MM-DL models provided competitive performance and ranked consistently among the top configurations, but genomic prediction alone achieved the single highest accuracy for this trait. Overall, NS stands out as an excellent candidate for both year-advance predictions and within-season selection, offering breeders a stable and dependable component of spike architecture.

Spike length

Prediction of year

SL was the most predictable trait in the study, and its performance under PoY highlights just how stable it is across years. As reported in Fig. 5A (Supplementary Table S4), even the M01 models achieved high accuracy (r ≈ 0.71), far above what was observed for GN, GW, NS, or SW. This reflects the biology of SL: this trait is determined early during stem elongation, a period that typically occurs before the most severe heat or drought events in Mediterranean environments. Because its developmental window is protected from large environmental shocks, SL carries a strong and consistent genetic signal across seasons. Adding ECs resulted in small but meaningful gains. For example, the best model, M17 [G + ECs (All DT)], reached r ≈ 0.71, while other ECs-based models (e.g. M13, M09, and M05) improved performance by + 0.01 to + 0.02 r-units. These increases likely capture subtle differences in temperature and photoperiod during early spike growth. NIRS contributed positively but modestly. Overall, Fig. 5A shows that SL is exceptionally stable across years, making it a highly dependable trait for year-advance selection.

Prediction of sowing year

SL also performed extremely well under the PoSY scenario. According to Fig. 5B (Supplementary Table S5), the top multimodal and genomic models reached r ≈ 0.78–0.80, the highest accuracies observed for any trait in the study. The shared climatic context within a season allows early spike development to unfold under similar temperature and photoperiod conditions, helping the models capture genotype-specific differences with remarkable consistency. Enviromic covariates again provided small, steady benefits, while NIRS offered slight improvements in environments where carbon allocation and metabolic signatures were more influential. The small difference between multimodal and genomic-only results under PoSY reinforces the idea that SL is largely governed by genetic factors, with environmental influences playing a secondary role. Together, Fig. 5B confirms that SL predictability is extremely high within season, with very little noise introduced by sowing-date differences.

Comparison between PoY and PoSY for SL

Comparing Fig. 5A and B makes it clear that SL is the most stable and reliably predicted trait across all scenarios. The PoY and PoSY accuracies—r ≈ 0.71 and r ≈ 0.80, respectively—are both significantly higher than for the other four traits. The gap between scenarios is small (~ 0.07–0.09), which aligns with the biological understanding that SL is formed before major environmental stresses strike. Because SL depends less on stress timing and more on inherent genetic differences in spike architecture, it maintains strong predictability whether forecasting a new season or a new sowing environment. While enviromic covariates help refine predictions, the consistent performance of M01 models highlights that SL is largely driven by strong additive genetic effects. Multimodal models provide incremental gains, but not to the extent seen for traits like GN or SW.

Summary for SL

Figure 5 shows that SL is the most predictable and genetically stable trait among those evaluated. Its early developmental determination shields it from the unpredictable stresses that complicate cross-year predictions for other traits. Under PoY, SL already reaches very high accuracy (r ≈ 0.71), and under PoSY it improves even further (r ≈ 0.78–0.80). ECs and NIRS contribute small but consistent gains, yet the core of SL predictability comes from its strong genetic signal. These results make SL an excellent candidate for both year-advance forecasting and within-season selection, offering breeders a reliable and stable architectural trait that is minimally affected by environmental variation.

Spike weight

Prediction of year

SW showed moderate predictability across years, reflecting its intermediate biological nature—part architectural, part physiological. As shown in Fig. 6A, M01 models reached accuracies around r ≈ 0.32, indicating that SW has a meaningful genetic component but is also influenced by environmental conditions during later developmental stages. Adding environmental covariates improved prediction slightly, with models such as M13 [G + ECs (FT–HT)] reaching r ≈ 0.38 (Fig. 6A; Supplementary Table S4), suggesting that heat and moisture conditions during the grain-filling period help explain differences in spike biomass across years. Some multimodal models integrating ECs, NIRS, and TCs offered small gains (e.g. M16, and M03), though these improvements were generally modest. This mild response reflects the complexity of SW: it is shaped early by spike architecture (which is genetically stable) and later by assimilate partitioning and stress during grain filling (which is environmentally variable). Overall, Fig. 6A highlights that SW can be predicted moderately well across years, but its cross-season stability is lower than that of NS or SL.

Prediction of sowing year

Prediction of SW within the same season (Fig. 6B; Supplementary Table S5) resulted in noticeably higher accuracies, with the best multimodal models reaching r ≈ 0.40–0.41 (e.g. M09, and M20). This improvement reflects the more consistent climatic context shared across sowing dates within a year, which allows the models to better capture environmental effects on spike biomass accumulation. Enviromic covariates again played a key role, particularly those describing temperature and water availability during spike elongation and grain filling. The contribution of NIRS was positive but moderate, likely capturing variation in grain composition and carbon allocation linked to SW. Genomic-only models also performed better under PoSY than PoY, confirming that within-season predictions reduce the environmental noise that often masks genetic signals in cross-year forecasts. Figure 6B demonstrates that SW is more predictable within a season, making PoSY a more reliable scenario for operational decision-making.

Comparison between PoY and PoSY for SW

Comparing Fig. 6A and B reveals a clear scenario effect: PoSY correlates (up to r ≈ 0.41) consistently exceed those under PoY (r ≈ 0.32–0.38). This improvement of ~ 0.10 r-units reflects the fact that SW is partly controlled by late-season processes, such as carbohydrate deposition and spike maturation, which are highly sensitive to environmental variation across years. Within a single season, however, these processes unfold under much more similar climatic trajectories, allowing the models to capture genotype-specific differences more effectively. Unlike SL or NS, which benefit from strong genetic determination and early formation, SW incorporates both early spike development and later biomass accumulation. This dual nature explains why SW is less stable than SL or NS but far more predictable than GN. MM-DL integration—especially through ECs—helps bridge these environmental effects, though the gains remain moderate compared with more stress-responsive traits.

Summary for SW

Figure 6 shows that SW is a moderately predictable trait with clear improvements when predictions are made within the same season. Under PoY (Fig. 6A), accuracies reached r ≈ 0.30–0.38, reflecting the combined influence of genetics and late-season environmental conditions. Under PoSY (Fig. 6B), the best models achieved r ≈ 0.40–0.44, as within-season consistency allowed the model to better capture the environmental and physiological drivers of spike biomass. ECs contributed the most to improving prediction (e.g., M13, and M09, respectively), while genomic, NIRS, and TCs each added moderate gains. Overall, SW represents an intermediate trait: less stable than early architectural traits like NS and SL, but significantly more predictable than highly sensitive traits like GN. Its mixed genetic–environmental basis makes multimodal modeling especially useful for improving prediction accuracy.

Comparative performance across traits and scenarios

Across all traits and prediction settings, MM-DL generally improved predictive accuracy compared with genomic-only baselines. However, apparent trait-dependent differences emerged, reflecting the underlying physiological determinants, timing of trait formation, and environmental sensitivity controlling each yield component.

A coherent hierarchy of predictability was observed:

• SL > NS > GW > SW > GN

This ranking reflects the developmental and physiological context in which each trait is formed:

1. SL exhibited the highest predictive accuracy across all environments and scenarios (PoY ≈ 0.56–0.71; PoSY ≈ 0.74–0.80).

Its early developmental determination during stem elongation—before central heat and drought stress—confers high genetic control and exceptional stability across seasons.

• 2.NS showed high and stable predictability (PoY ≈ 0.23–0.47; PoSY ≈ 0.56–0.67).

Determination during early reproductive development makes NS less sensitive to unpredictable stress spikes, leading to strong transferability in both scenarios.

• 3.GW displayed moderate-to-high predictability (PoY ≈ 0.24–0.30; PoSY ≈ 0.34–0.46).

While mainly genetically driven, GW responds to stress during grain filling, making it more environmentally influenced than SL and NS, but still reliable for cross-year selection.

• 4.SW achieved moderate predictability (PoY ≈ 0.27–0.38; PoSY ≈ 0.34–0.41).

As a biomass-related trait integrating early spike development and assimilate partitioning, SW is more sensitive to intra-season climatic variation and therefore less stable across years.

• 5.GN remained the most environmentally plastic trait (PoY ≈ 0.28–0.32; PoSY ≈ 0.43–0.53).

GN depends heavily on short-window reproductive processes around booting and anthesis, making it highly sensitive to heat and drought timing and the most challenging trait to predict in unobserved seasons.

In summary, across all five yield-related traits, prediction accuracy differed markedly by trait and prediction scenario. SL consistently exhibited the highest accuracy (PoY ≈ 0.56–0.71; PoSY ≈ 0.74–0.80), followed by the NS (PoY ≈ 0.23–0.47; PoSY ≈ 0.56–0.67). GW showed moderate predictability (PoY ≈ 0.24–0.30; PoSY ≈ 0.34–0.46), while SW achieved intermediate values (PoY ≈ 0.27–0.38; PoSY ≈ 0.34–0.41). GN displayed the lowest cross-year accuracy (PoY ≈ 0.28–0.32) but improved substantially within season (PoSY ≈ 0.43–0.53). Across all traits, PoSY consistently outperformed PoY, reflecting higher predictability when training and testing environments shared the same seasonal context.

Scenario-level differences: PoY vs. PoSY

A consistent pattern emerged across all traits:

• PoSY > PoY (Fig. 7A).

Fig. 7.

Prediction accuracy of unimodal and multimodal models implemented for each yield component traits (grain number per spike, GN; grain weight per spike, GW; number of spikelets per spike, NS; spike length, SL; spike weight, SW) under two breeding-relevant scenarios: Prediction of Year (PoY), where models are trained on past seasons and used to forecast performance in a future year with unobserved climatic conditions, and Prediction of Sowing Year (PoSY), where predictions are made for new sowing-date × environment combinations within the same season. Bar height represents Pearson’s correlation (r) between observed and predicted values for each trait (prediction accuracy). A Bars show the mean correlation of the top-20 models reported in Supplementary Tables S4 (PoY) and S5 (PoSY). Higher prediction accuracy under PoSY reflects reduced temporal climatic uncertainty and shared seasonal context, rather than differences in model structure or data availability. For each pair of bars at the top, the numerical value indicates the percentage advantage of the PoSY scenario over the PoY scenario. B Bars show, for each trait and scenario, the best-performing unimodal model (M01) and the best-performing multimodal model (M02-M20) reported in Supplementary Tables S4 (PoY) and S5 (PoSY). For each pair of bars, the numerical label indicates the percentage advantage of the multimodal model relative to the unimodal model

Figure 7A shows that across traits, PoSY consistently outperforms PoY due to reduced temporal uncertainty, while multimodal integration (G + ECs, G+NIRS, TC) yields meaningful accuracy gains, particularly for GN, SW, and GW, with improvements ranging from 12.7 to 18.6%, 19.7–8.8%, and 4.3–9.4%, respectively, across the two breeding scenarios (Fig. 7B). These findings demonstrate the value of MM-DL models for supporting operational in-season selection and accelerating genetic gain in climate-variable durum wheat breeding.

This reflects the fundamental limitation imposed by temporal climatic uncertainty when forecasting future seasons. When the climatic context between training and target environments is shared (PoSY), MM-DL models can more effectively reproduce genotype-specific physiological responses, substantially improving predictive accuracy.

Traits SL and SW showed minimal decline from PoSY to PoY, 22% and 22.8% respectively, demonstrating robust year-to-year stability. By contrast, GW, GN and especially NS exhibited the largest performance gap, with increases of 49.5%, 55.8% and 78.4%, respectively (Fig. 7A).

These results demonstrate that environmental predictability is the dominant factor influencing the transferability of GP across growing seasons.

Contribution of multimodal information

Across all five traits, multimodal fusion provided consistent accuracy gains (Fig. 7B), with a clear pattern:

• ECs > NIRS > TC > G alone.

ECs provided the most significant improvement, especially under PoY, by capturing the severity and timing of heat and drought stress. ECs were indispensable for environmentally responsive traits (GN, SW).

NIRS phenomics provided trait-specific biochemical signals, particularly valuable for GN and SW, where stress responses and assimilate composition play roles, and TCs (HD and PH) contributed modest but meaningful improvements in traits influenced by developmental timing (GN, NS, GW). Genomic effects dominated prediction for highly heritable traits (SL, NS), but ECs was required to explain G×E for GN, GW, and SW.

The combination of complementary biological layers allowed MM-DL to learn the pathway from genotype → environment → physiology → phenotype, enhancing robustness across variable conditions.

Overall synthesis and breeding implications

The comparative analysis highlights fundamental principles for durum wheat breeding under climate variability: SL and NS are ideal targets for both year-advance selection (PoY) and in-season decisions (PoSY) due to their genetic stability and high predictability. GW is a strong candidate for cross-year prediction, supporting reliable selection for sink stability. SW benefits from within-season prediction and multimodal integration but shows moderate cross-year transferability. GN, though essential for yield potential, requires enviromic integration and PoSY deployment to achieve operational prediction accuracy due to its sensitivity to stress timing.

Overall, Fig. 7A and B highlights three significant insights that guide model selection in durum wheat breeding: (i) SL is the most reliable trait for both year-advance and in-season deployment due to strong genetic buffering during grain filling; (ii) NS supports stable selection when the environment is partially shared, consistent with its early developmental determination; (iii) prediction accuracy for PoSY is always superior than PoY.

Discussion

The MM-DL framework used in this study offers a more complete view of how durum wheat responds to changing environments, and the results reflect both the strengths and the biological constraints of predicting yield-related traits across seasons. One clear pattern that emerged is that the predictability of each trait is strongly tied to its developmental timing and to its sensitivity to stress at key physiological moments. This reinforces long-established wheat physiological knowledge: traits formed early and governed mainly by genetics behave differently—and are much easier to predict—than traits determined by short and stress-prone windows later in the season.

The contribution of each modality reflects the developmental timing and physiological sensitivity of the target trait. Traits such as GN and GW, which are strongly influenced by short-term metabolic and stress responses during reproductive and grain-filling stages, benefit more from NIRS and ECs, particularly under PoSY, where seasonal context is shared. In contrast, spike length is determined early and under strong genetic control, explaining the dominant role of genomic information and the limited added value of phenomic inputs.

Trait prediction comparison

Among the traits evaluated, SL showed the highest accuracy in both prediction scenarios. SL is determined mainly during stem elongation, a period that, in Mediterranean drylands, typically precedes the most damaging heat or drought episodes. This early, relatively sheltered development explains why SL behaves so consistently across environments and why MM-DL models capture it so well. Our predictions are comparable to those reported in T. aestivum by Xu et al. [30] and Hammers et al. [31]. In the two studies, conducted on panels of 314 lines and 594 accessions, respectively, the Random Forest (RF) model, enriched with Marker–Trait Associations (MTAs), and the multivariate models, which simultaneously integrate multiple traits, achieved levels of prediction accuracy consistent with those observed in the present work. Furthermore, Xu et al. [30], highlights how the integration of stable and major MTA is crucial to improve the genomic predictability of SL, underlining the importance of including informative covariates in prediction models.

A similar prediction pattern was observed for NS, another trait formed early in reproductive development. Both SL and NS benefited from apparent genetic determinism and exhibited the stability that breeders value for long-term selection. Spikelet number emerged as the second most predictable trait after spike length, a pattern that is fully consistent with previous findings in durum and bread wheat. NS is determined early, during the spike initiation and spikelet primordia formation phases, making it less sensitive to late-season heat and drought compared with GN or GW. This early developmental stability has been reported repeatedly in wheat architecture studies [32, 33], where NS showed stronger genetic control and higher cross-environment heritability than downstream yield components. Similar trends have been observed in prediction studies: genomic and multivariate models often rank NS among the best predicted traits due to its relatively low environmental plasticity [10, 34]. The high PoY and PoSY accuracies found here reinforce this view and highlight NS as a reliable upstream determinant of potential grain number. By contrast, GN—though physiologically downstream of NS—showed substantially lower predictability because of its intense sensitivity to short, stress-prone windows around anthesis [35, 36]. These comparative results underscore the developmental logic of wheat yield formation: traits fixed before major stress periods (SL, NS) offer much higher transferability across seasons than traits heavily shaped by stress timing (GN, SW).

In contrast, traits tied to more vulnerable phenological windows—especially grain number—showed much lower cross-season predictability. GN remains one of the most environmentally sensitive yield components in cereals because floret fertility depends heavily on temperature, vapor pressure deficit, water availability, and assimilate flow during a short period around booting and anthesis. Small shifts in stress timing can drastically alter final grain set, making GN one of the most challenging traits to predict across seasons. For instance, in a study of 287 spring wheat lines evaluated in 18 international environments, a comparison of seven GP models showed that the inclusion of G×E interactions and genomic/pedigree effects is crucial for accurate prediction of GN (r = 0.43) [34]. For this reason and especially for this yield component, the MM-DL consistently improved the predictive accuracy compared to other GP models, as highlighted by Yannam et al. [24] on 179 modern lines of common wheat (T. aestivum L.). In this study, and consistent with our results, the RF model achieved the highest predictive accuracy for the GN, equal to about 0.6 under different conditions (water stress, well-watered, and combined water regime), outperforming the ridge-regression best linear unbiased prediction model (RR-BLUP) [24].

GW displayed intermediate behavior: although buffered by post-anthesis assimilate dynamics, it is still influenced by late-season stress. SW, a composite of spike architecture and biomass allocation, also showed moderate predictability, reflecting its dependence on both early development and late assimilate partitioning. Together, these findings highlight that yield components differ not only genetically but also physiologically in ways that directly influence their predictability.

It is also informative to compare the predictability of the five traits evaluated here with other wheat traits that were not part of this study but are well documented in the genomic-prediction literature. For example, PH and HD typically exhibit very high genomic predictability (r ≈ 0.70–0.90) due to their strong Mendelian control by major genes such as Rht, Vrn, and Ppd [25, 32, 33]. Yield per se generally shows moderate to low cross-season predictability (often r < 0.40), reflecting its integration of multiple stress-sensitive components [9, 10]. Similarly, physiological traits such as canopy temperature and Normalized Difference Vegetation Index (NDVI), widely used in GP studies, are usually easier to predict within season than across years because they track real-time stress responses rather than stable genetic architecture [37, 38]. The trait hierarchy observed here mirrors these broader patterns: traits governed by early or genetically stable developmental processes tend to be the most predictable, whereas traits determined by short, stress-prone windows remain challenging for cross-environment forecasting.

Agreement with previous studies

Overall, the pattern of results obtained in this durum wheat study is highly consistent with findings reported across the genomic-prediction literature. Traits with strong genetic control and early developmental determination—such as SL and NS—regularly show the highest prediction accuracy across diverse studies in wheat [32, 33, 39, 40]. Likewise, GW typically shows moderate but stable predictability, reflecting both its heritability and sensitivity to grain-filling conditions, matching reports from multienvironment GP studies [10, 34].

Conversely, GN is almost universally cited as one of the most difficult yield components to predict across years because it depends on short, stress-prone windows surrounding anthesis—exactly the pattern observed here. Studies in both spring and durum wheat have documented the same challenge, even when sophisticated G×E or multivariate models are used [9, 36].

Thus, the hierarchy observed in this study not only makes physiological sense but is also consistent with trait-specific predictability reported by other authors. The agreement suggests that the MM-DL framework captures the underlying biological structure of yield formation in durum wheat in a way that is consistent with established wheat physiology and genomic-prediction evidence.

Role of ECs and NIRS

ECs played an essential role in improving model transferability, particularly under the PoY scenario, where the testing year’s ecological conditions were absent from the training data. By expressing environmental variation through eco-physiological descriptors aligned with growth stages, ECs enabled the model to capture reaction norms that would otherwise remain hidden. This was especially important for GN and SW, where stress-related variation drives much of the phenotypic diversity across seasons. These results echo earlier studies showing that prediction improves significantly when models have access to environmental descriptors that reflect the physiological processes shaping G×E.

NIRS and phenology covariates also contributed, but in complementary ways. NIRS was particularly helpful for traits associated with biochemical or metabolic adjustments under stress, while heading date and plant height provided context for development rate and canopy structure. Their contributions were smaller than those of environmental data but still meaningful, especially for traits with complex physiological backgrounds. The substantial contribution of NIRS in the PoSY scenario is attributable to the spectra’s ability to integrate the plant’s metabolic history; they provide a biochemical snapshot of the G × E interaction that occurred in the current season, thereby reducing the informational gap between genetic potential and final phenotypic performance.

The comparison between PoY and PoSY further illustrates the importance of shared climatic context. As expected, PoSY consistently produced higher accuracies across all traits because training and testing environments shared similar seasonal conditions. Years with comparable temperature patterns, stress profiles, and planting-to-heading trajectories enabled the MM-DL model to better reproduce genotype-specific responses. In contrast, PoY required extrapolation into unknown climatic conditions—one of the major bottlenecks in GP. Even so, multimodal integration helped to reduce this gap, showing that enviromic and phenomic covariates provide essential information when climatic uncertainty is high.

From a breeding perspective, these findings reinforce the idea that not all traits should be treated equally when designing predictive pipelines. Early-determined traits like SL and NS are reliable candidates for cross-year selection and can accelerate genetic gain with relatively low risk. GW and SW benefit from multimodal modeling and appear suitable for both within-season and cross-season predictions, depending on the environmental contrast. GN remains the most challenging trait for long-term forecasting, and breeders may obtain better results when prediction incorporates detailed environmental descriptors or when GN is used in PoSY-type frameworks.

Overall, the study shows that combining genomics, enviromics, phenomics, and phenology produces richer, more robust models for predicting yield components in durum wheat. Importantly, it demonstrates that climatic uncertainty, now a defining feature of Mediterranean agriculture—can be partially mitigated through integrative modeling approaches that explicitly consider the biological processes underlying genotype performance. This multimodal strategy provides a more faithful representation of the genotype–environment–physiology continuum and lays the foundation for more resilient, informed predictive breeding pipelines.

Added value of the MM-DL framework

The contribution of the MM-DL framework was central to the results of this study. By integrating genomics with ECs, NIRS, and phenology, MM-DL was able to capture complementary biological signals that traditional unimodal models often overlook. This was particularly evident for environmentally sensitive traits such as GN and SW, where MM-DL delivered clear gains by modeling stress timing, metabolic adjustments, and developmental context jointly. Importantly, the trait-ranking observed here (SL > NS > GW > SW > GN) closely matches findings from other wheat studies, confirming that traits determined early and under strong genetic control consistently benefit less from multimodal fusion, while traits shaped by short, stress-prone physiological windows gain the most (Fig. 7). This pattern is quantitatively supported by the full set of trait- and scenario-specific prediction results, where early-determined traits (SL, NS) show consistently high accuracy across both PoY and PoSY scenarios, whereas traits governed by narrow, stress-sensitive physiological windows (GN, SW) exhibit the largest relative gains from multimodal integration.

The improvement of PoY accuracy through ECs integration highlights the value of enviromics for cross-year generalization—a known limitation in GP. Overall, the results demonstrate that MM-DL provides a more faithful representation of the genotype–environment–physiology continuum and significantly enhances the robustness of predictions for yield components, especially under the climatic uncertainty typical of Mediterranean drylands.

A key operational assumption in this study was the replacement of ECs, NIRS, and TCs in the test environments with their training-set means, reflecting a realistic breeding situation in which future environmental conditions are not yet known at selection time. While this assumption avoids optimistic bias, it also implies that prediction accuracy—particularly for environmentally plastic traits such as GN and SW—may be sensitive to inaccuracies in environmental forecasting. Importantly, the MM-DL framework is not restricted to this assumption and can readily incorporate improved weather forecasts, early-season environmental data, or rapid phenotyping as these become available. Thus, the reported results represent a conservative baseline, and the value of multimodal integration is expected to increase as forecasting and sensing technologies advance.

It is also important to note that prediction accuracy was assessed using a LOEO validation scheme, which provides a stringent test of robustness across contrasting environments. Small numerical differences among top-ranked models should therefore not be over-interpreted, particularly when they fall within the variability observed across LOEO folds. Instead, the conclusions of this study are based on consistent and reproducible performance patterns across traits, data modalities, and prediction scenarios, which are more relevant for operational breeding decisions than marginal differences in correlation coefficients.

Conclusion

This study demonstrates that MM-DL provides a robust framework for predicting yield-related traits in durum wheat across environmentally contrasting growing seasons. By integrating genomic, enviromic, NIRS, and phenology-based information, the models consistently improved predictive accuracy relative to genomic-only baselines, particularly under conditions of strong G×E. The magnitude of these improvements varied across traits and reflected their physiological determinants. Early-developing, genetically controlled traits—such as spike length and spikelet number—showed the highest predictability and strongest cross-season stability, supporting their use for reliable year-advance selection. In contrast, highly environment-responsive traits such as grain number benefited more from enviromic integration and were best predicted within shared seasonal contexts (PoSY). Grain weight and spike weight showed intermediate behavior, with appreciable, yet trait-specific, gains from multimodal modeling.

Across all prediction scenarios, accuracies were consistently higher under PoSY than under PoY, highlighting the crucial role of a shared climatic context in capturing the physiological nuances of each genotype. Even so, incorporating enviromics into the models markedly reduced the gap between scenarios, showing how powerful environmental information can be for improving predictions from one year to another. NIRS and phenology covariables contributed complementary, physiologically grounded information, further enhancing prediction for selected traits.

Overall, the results highlight the value of combining genomic, environmental, and phenomic data streams to more comprehensively model the pathways linking genotype, environment, and phenotype. For durum wheat breeding in Mediterranean dryland conditions—where climate variability and heat stress frequently generate strong G×E—multimodal strategies represent a practical and biologically informed approach to increase prediction robustness and accelerate genetic gain. These findings provide a foundation for designing predictive breeding pipelines that are better aligned with the physiological and environmental complexities of yield formation in stress-prone agroecosystems.

Abbreviations

BO

Bayesian Optimization

DT

Developmental timing

ECs

Environmental covariates

EI

Expected Improvement

FT

Flowering time

G

Genomic markers

GN

Grain number per spike

GS

Genomic Selection

GP

Genomic Prediction

GW

Grain weight per spike

GXE

Genotype × environment

HD

Heading date

HT

Harvest time

LOEO

Leave-One-Environment-Out

LR

Landraces

MC

Modern cultivars

MM-DL

Multimodal deep learning

MSE

Minimizing mean squared

NIRS

Near-infrared spectroscopy

NS

Number of spikelets per spike

OC

Old cultivars

PH

Plant height

PoSY

Prediction of Sowing Year

PoY

Prediction of Year

RF

Random Forest

ResNets

Residual Neural Networks

SL

Spike length

SNV

Standard Normal Variate

ST

Sowing time

SW

Spike weight

TCs

Trait covariates

Authors’ contributions

**A.M.** : Writing-original draft, Writing-review & editing, Conceptualization, Data curation, Methodology; **D.P.** : Writing–review & editing, Figure preparation, Input data preparation and Data curation. **P.V.** : Writing-review & editing and input data preparation; **F.F.** : Writing-review & editing, Input data preparation and NIR data processing; **J.C.** : Writing-original draft, Writing-review & editing, Methodology, Supervision; **O.A.M.** : Writing-original draft, Writing-review & editing, Methodology, Supervision; **P.D.** : Writing-review & editing, Conceptualization, Funding acquisition, Resources, Supervision. All authors approved the final manuscript.

Funding

This work was supported by the Italian Ministry of Enterprises and Made in Italy - MIMIT in the framework of a project entitled “Sistemi avanzati di modellizzazione digitale per il miglioramento e la predizione di resa e qualità nella produzione cerealicola italiana” - GRANO.IT - Nr. F/360091/01-02/X75 (CUP B19J24000140005).

Data availability

Genotypic data (SNPs) are available in the Fighshare database (https://figshare.com/) with the following link: 10.6084/m9.figshare.28554947 [25].

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.