Integrating Machine Learning and In Vitro Screening to Evaluate Drought and Temperature Stress Responses for Vicia Species

Abstract

Drought and temperature extremes are major abiotic stressors limiting legume productivity worldwide. This study investigates the germination and early seedling responses of six cultivars belonging to three Vicia species (V. sativa, V. pannonica, and V. narbonensis) under varying levels of polyethylene glycol (PEG)-induced drought and temperature conditions (12 °C, 18 °C, and 24 °C) in vitro. Significant cultivar-dependent differences were observed in the germination rate (GR), shoot and root length (SL and RL), fresh and dry weight (FW and DW), and vigor index (VI). The Ayaz cultivar exhibited superior performance, particularly under severe drought (10% PEG) and optimal temperature (24 °C), while Özgen and Balkan were most sensitive to stress. Principal component and correlation analyses revealed strong associations between the vigor index, shoot height, and fresh and dry weight, particularly in high-performing genotypes. To further model and predict stress responses, four machine learning (ML) algorithms—Random Forest (RF), k-Nearest Neighbors (k-NNs), Multilayer Perceptron (MLP), and Support Vector Machines (SVMs)—were employed. Based on model performance metrics, and considering high R² values along with low RMSE and MAE values, the MLP model demonstrated the most accurate predictions for the GR (R² = 0.95, RMSE = 0.06, MAE = 0.05) and VI (R² = 0.99, RMSE = 0.02, MAE = 0.01) parameters. In contrast, the RF model yielded the best results for the SL (R² = 0.98, RMSE = 0.02, MAE = 0.02) and DW (R² = 0.93, RMSE = 0.06, MAE = 0.04) parameters, while the highest prediction accuracy for the RL (R² = 0.83, RMSE = 0.09, MAE = 0.07) and FW (R² = 0.97, RMSE = 0.05, MAE = 0.03) parameters was achieved using the SVM model. Comparative analysis with recent studies confirmed the applicability of ML in stress physiology and genotype screening. This integrative approach offers a robust framework for genotype selection and stress tolerance modeling in legumes, contributing to developing climate-resilient crops.

1. Introduction

The genus Vicia, commonly called vetches, represents a diverse assemblage of leguminous species encompassing approximately 210–240 taxa. These species are predominantly distributed across temperate regions of the Northern Hemisphere, with the Mediterranean basin recognized as their evolutionary epicenter and primary diversification hub [1,2]. Vicia sativa L. (common vetch) holds significant agronomic value due to its broad adaptability, pivotal role in soil enhancement, and extensive utilization in sustainable agriculture as a cover crop and forage species [3]. Other economically important vetch species, such as Vicia pannonica (Hungarian vetch), Vicia narbonensis (Narbon vetch), and Vicia villosa (hairy vetch), have demonstrated remarkable resilience across varying environmental conditions [2,3].

Vetches are remarkably esteemed for their high protein content and substantial nutritional value, rendering them an economical and sustainable resource for livestock feed [4]. Furthermore, their ability to engage in symbiotic nitrogen fixation with Rhizobium bacteria substantially enhances soil fertility, mitigates dependence on synthetic fertilizers, and contributes to improved soil structure and health [5,6]. Despite these agronomic advantages, global vetch production remains relatively constrained. According to the Food and Agriculture Organization (FAO), the worldwide harvested area and total production of vetch amounted to approximately 0.33 million hectares and 0.67 million tons, respectively, in 2023 [7].

Drought stress poses one of the most drastic environmental limitations to vetch productivity, significantly impeding plant growth, reducing biomass accumulation, and disrupting critical morphophysiological functions vital for crop performance and survival [8]. The escalating challenges of climate change, coupled with rapid industrial expansion and urbanization, have intensified global water scarcity, reducing the availability of irrigation resources for agricultural systems [9]. In arid and semi-arid regions, erratic precipitation patterns further exacerbate water deficits, making drought tolerance a key breeding objective for sustainable vetch cultivation [2]. At physiological and cellular levels, drought stress perturbs membrane integrity, alters metabolic pathways, and compromises overall plant fitness [10]. The severity of drought-induced stress is modulated by environmental variables such as temperature fluctuations, light intensity, and precipitation patterns [11,12]. Given its widespread impact on agricultural productivity, drought stress is recognized as a major constraint that limits genetic potential, reduces yields, and leads to substantial economic losses on a global scale [13,14].

Among leguminous forage crops, vetch exhibits considerable adaptability under fluctuating climatic conditions, despite the high sensitivity of legumes to abiotic stress [15,16,17]. Specifically, Vicia sativa has demonstrated resilience under diverse environmental stresses, positioning it as a promising candidate for drought-tolerance improvement programs [4,18,19]. Given the increasing frequency of extreme drought events worldwide, strategic research efforts are imperative to enhance the drought resilience of vetch species.

Temperature is another fundamental environmental determinant that governs seed germination, early seedling development, and overall plant physiological responses [20,21]. As climate models predict increased heatwaves and prolonged drought episodes, the interplay between temperature stress and water deficits becomes increasingly relevant for plant adaptation and survival. Moisture availability and temperature collectively regulate germination dynamics, with extreme thermal conditions and low soil water potential adversely affecting seedling establishment and population sustainability [22,23].

Empirical studies on the interactive effects of temperature and drought stress suggest that seed germination is significantly inhibited under polyethylene glycol (PEG)-simulated drought conditions due to osmotic stress-induced reductions in water potential [24,25,26,27,28]. For instance, Apeiba tibourbou seeds exhibited a sharp decline in germination rates at a water potential of −0.2 MPa, decreasing from 51% at 30 °C to 37% at 25 °C. However, a substantial proportion of ungerminated seeds remained viable and germinated uniformly once water stress was alleviated [24]. Given the complexities associated with evaluating drought responses under natural field conditions, PEG has been extensively employed in laboratory studies as an osmotic stress agent to simulate drought conditions [29,30,31,32]. Unlike natural drought, PEG-induced stress does not introduce ionic toxicity, making it a reliable method for assessing plant drought tolerance at both germination and seedling growth stages [33,34]. Investigating the interactive effects of PEG-mediated osmotic stress and varying thermal conditions can yield critical insights into the adaptive strategies of vetch species under abiotic stress, facilitating the development of more resilient genotypes for sustainable agricultural systems.

The connection between machine learning (ML) and plant tissue culture has generated substantial interest due to the potential to optimize complex biological processes efficiently [35,36,37]. ML, a subset of artificial intelligence, leverages algorithmic techniques to analyze datasets and extract meaningful patterns, enabling precise predictions and enhanced decision-making capabilities [36,37]. Among the various ML methodologies—supervised, unsupervised, and reinforcement learning—supervised learning is predominantly applied within plant tissue culture studies [37]. This approach utilizes labeled datasets, either quantitative parameters such as shoot regeneration rates or qualitative outcomes, to develop predictive models classified into regression and classification algorithms [38]. Regression-based ML algorithms have shown remarkable efficacy in forecasting and refining protocols across diverse tissue culture applications, including seed germination [39], stress studies [40], and micropropagation [35]. Advanced machine learning algorithms, notably Random Forest (RF), k-Nearest Neighbors (k-NNs), Multilayer Perceptron (MLP), and Support Vector Machines (SVMs), have demonstrated substantial efficacy in trait prediction across various research contexts, which is crucial for enhancing decision-making processes, accelerating breeding programs, and optimizing resource allocation in complex biological systems [35,39,40,41].

The primary objective of this study is to investigate the effects of polyethylene glycol (PEG)-induced drought stress and varying temperature conditions on the germination and early seedling development of different Vicia species. By examining key growth parameters such as germination percentage, shoot and root elongation, fresh and dry biomass accumulation, and vigor index, this research aims to determine how drought and temperature interact to influence seedling vigor and physiological responses. Additionally, statistical analyses, including analysis of variance, correlation analysis, and Principal Component Analysis (PCA), are employed to identify the most influential factors contributing to drought tolerance in vetch species. Machine learning models are integrated to evaluate the relationships between different experimental conditions and seedling performance to enhance predictive accuracy and gain deeper insights into plant responses under abiotic stress. By combining traditional statistical methods with data-driven modeling, this study seeks to provide a comprehensive understanding of drought adaptation mechanisms in Vicia species, ultimately supporting the selection of resilient genotypes for sustainable agricultural applications.

2. Materials and Methods

2.1. Plant Materials

In this study, six vetch genotypes representing three distinct species were selected for evaluation under different stress conditions. Specifically, two genotypes each of Common vetch (Vicia sativa L.), Hungarian vetch (Vicia pannonica Crantz), and Narbon vetch (Vicia narbonensis L.) were used as experimental materials. These species were selected due to their agricultural relevance, genetic diversity, and varying adaptability to abiotic stress conditions. Detailed information regarding specific genotypes is displayed in

Table 1. Vetch species, cultivars, and their source institutions used in the study.

Species		Cultivar	Institution
Common Vetch	Vicia sativa	Ayaz08	Ankara Field Crops Central Research Institute ‘2008’
		Cumhuriyet99	Ege Field Crops Central Research Institute ‘1999’
Hungarian Vetch	Vicia pannonica Crantz	Atom	Yonca Agriculture Products ‘2020’
		Altınova2002	General Directorate of Agricultural Enterprises ‘2007’
Narbon Vetch	Vicia narbonensis	Balkan	Transitional Zone Agricultural Research Institute ‘2011’
		Özgen	Dicle University Faculty of Agriculture ‘2009’

. The study was carried out in Erciyes University, Faculty of Agriculture Laboratories, under controlled conditions.

2.2. Sterilization Process

To minimize microbial contamination and maintain aseptic conditions during germination experiments, seeds were initially rinsed under running tap water for 10 min to remove surface debris. Subsequently, they were surface sterilized by immersion in 70% ethanol (Merck, Darmstadt, Germany) for 1 min, followed by treatment with 15% commercial bleach solution containing sodium hypochlorite (2.5% NaOCl) for 15 min. After sterilization, the seeds were thoroughly rinsed three times with sterile distilled water inside a laminar flow hood to remove residual sterilizing agents and ensure safe transfer to the culture medium.

2.3. Germination Medium Preparation and Culture Conditions

Sterilized seeds were cultured on a nutrient-rich medium prepared with Murashige and Skoog (MS) basal salts at a concentration of 4.4 g L⁻¹, supplemented with 30 g L⁻¹ sucrose (3% w/v) as a carbon source, and solidified with 7.5 g L⁻¹ agar (0.8% w/v). The pH of the medium was carefully adjusted to a range of 5.5–5.7 to optimize nutrient availability and seedling development. A total of 35 mL of the prepared medium was dispensed into sterile Petri dishes (75 × 15 mm).

2.4. Drought Stress Application Using PEG-6000

To simulate drought conditions under controlled in vitro conditions, polyethylene glycol (PEG-6000, Sigma-Aldrich, St. Louis, MO, USA) was incorporated into the germination media at concentrations of 5% (50 g PEG-6000 L⁻¹) and 10% (100 g PEG-6000 L⁻¹), besides control medium without PEG (0% PEG). PEG-6000 is an osmotic agent commonly used in plant stress studies due to its ability to lower water potential without being absorbed by plant tissues [26,27]. This approach allowed the examination of genotype-specific responses to water deficit stress during early growth stages.

2.5. Data Collection and Experimental Design

In this study, a germination experiment was conducted using six vetch cultivars (‘Ayaz’, ‘Cumhuriyet’, ‘Atom’, ‘Altınova’, ‘Balkan’, and ‘Özgen’), under three temperature regimes (12 °C, 18 °C, and 24 °C) and three polyethylene glycol (PEG 6000) concentrations (0%, 5%, and 10%). The experiment was arranged in a factorial design with six replications. A total of 324 Petri dishes were used, and 10 seeds were placed in each Petri dish, resulting in a total of 3240 seeds evaluated. After the seeds were placed in Petri dishes, they were left to germinate for fourteen days, and then the morphological parameters were examined. Germination percentage was calculated based on data from two replications (20 seeds) selected for each treatment combination. Additionally, five seedlings in total were randomly selected from these two replications to measure shoot length (cm), root length (cm), fresh weight (mg), and dry weight (mg). All measured data were digitally recorded, with 15 observations (n = 15) per treatment combination. The vigor index was calculated by multiplying the germination rate and shoot length values.

Figure 1 presents seedlings of the Ayaz cultivar under in vitro conditions and after drought and temperature treatments.

2.6. Statistical Analysis

All experiments were conducted using a factorial arrangement within a completely randomized design. The collected data were subjected to analysis of variance (ANOVA) to determine the significance of treatment effects. In large datasets (n > 30), the central limit theorem supports the robustness of ANOVA even when the normality assumption is not strictly met. Although the Shapiro–Wilk test indicated deviations from normality, it is well documented that such tests can be overly sensitive in large samples, often flagging minor departures that do not meaningfully affect the validity of parametric tests [42,43]. Schmider et al. [42] reported that the empirical Type I and II error rates of ANOVA remain stable under non-normal distributions, while Blanca et al. [43] demonstrated that the F-test maintains robustness across a wide range of skewed and kurtotic distributions. Based on these findings and considering the relatively large sample size in our study (n = 810), ANOVA was deemed appropriate for detecting both main and interaction effects in this multifactorial design. For completeness and to ensure robustness, non-parametric Kruskal–Wallis tests were also conducted, and both methods yielded consistent and statistically significant results (p < 0.05). Pairwise comparisons were performed using Fisher’s Least Significant Difference (LSD) test. Statistical groupings (a, b, c…) indicating significant differences among means are presented in Supplementary Tables S1–S7. All statistical analyses were carried out using R (v4.3.1), with the agricolae package used for ANOVA and post hoc tests. Additionally, correlation analysis (corrplot package) and Principal Component Analysis (factoextra package) were employed to explore interrelationships among measured variables and to reduce data dimensionality.

2.7. Machine Learning Approaches

In this study, machine learning (ML) models were developed and implemented using the R programming language (v4.3.1), with the aid of the ‘caret’ package, which provides a unified framework for data preprocessing, model training, and performance evaluation. The dataset used for model development consisted of 810 observations generated from factorial combinations of six vetch cultivars, three polyethylene glycol (PEG 6000) concentrations (0%, 5%, and 10%), and three temperature regimes (12 °C, 18 °C, and 24 °C). Each data entry included both the experimental conditions and a set of corresponding morphological response variables: germination rate, shoot height, root length, fresh weight, dry weight, and vigor index.

Prior to model training, the dataset was randomly partitioned into training and testing subsets using a 10-fold cross-validation procedure. This approach involved dividing the data into ten equal parts, where in each fold, 80% of the data was used for model training and the remaining 20% for validation. This procedure ensured that each data point was used for both training and testing across multiple iterations, thereby minimizing overfitting and enhancing model generalizability [35].

Four supervised machine learning algorithms were employed and comparatively evaluated: Multilayer Perceptron (MLP), Random Forest (RF), k-Nearest Neighbors (k-NNs), and Support Vector Machine (SVM). The MLP model, a type of artificial neural network, consists of interconnected layers of neurons capable of modeling complex and non-linear relationships [41]. The RF algorithm, an ensemble method based on the construction of multiple decision trees, improves prediction accuracy through aggregation and is particularly robust against overfitting [35,44]. The k-NN algorithm is a distance-based method that predicts outcomes based on the nearest training instances in the feature space, providing a simple yet effective approach for classification and regression [37,45]. Finally, the SVM algorithm constructs optimal hyperplanes in a high-dimensional space to distinguish between classes or predict continuous values, and it is especially effective in handling non-linear patterns through kernel transformations [35].

To evaluate model performance, a set of standard regression metrics was used [35]. The coefficient of determination (R²) was employed to quantify the proportion of variance in the observed outcomes that is explained by the model predictions (Equation (1)). In addition, the root mean square error (RMSE) was calculated to assess the average magnitude of the errors between predicted and actual values, giving greater weight to larger deviations (Equation (2)). Finally, the mean absolute error (MAE) was used to measure the average absolute difference between predicted and observed values, providing a direct and interpretable measure of prediction accuracy (Equation (3)). Together, these metrics enabled a comprehensive and comparative assessment of the predictive capabilities of each machine learning model under varying experimental conditions.

$${\mathrm{R}}^2=1-{\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{n}}{(\mathrm{Y}}_{\mathrm{i}}-{\widehat{\mathrm{Y}}}_{\mathrm{i}}{)}^2}{\sum _{\mathrm{i}=1}^{\mathrm{n}}{(\mathrm{Y}}_{\mathrm{i}}-\tilde{\mathrm{Y}}{)}^2}}$$

(1)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{\left(\sum _{\mathrm{i}=1}^{\mathrm{n}}{(\mathrm{Y}}_{\mathrm{i}}-{\widehat{\mathrm{Y}}}_{\mathrm{i}}{)}^2\right)}{\mathrm{n}}}}$$

(2)

$$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}|{\mathrm{Y}}_{\mathrm{i}}-{\widehat{\mathrm{Y}}}_{\mathrm{i}}|$$

(3)

where ${\mathrm{Y}}_{\mathrm{i}}$ = actual value, ${\widehat{\mathrm{Y}}}_{\mathrm{i}}$ = predicted value, $\tilde{\mathrm{Y}}$ = mean of the actual values, and n = sample count.

3. Results

3.1. Variance Analysis

The analysis of growth and vigor parameters in vetch revealed statistically significant differences among cultivars, polyethylene glycol (PEG) concentrations, and temperature treatments (p < 0.0001) (Figure 2).

As presented in Figure 2, the ‘Atom’ cultivar exhibited the highest germination rate (96.7%) among all varieties. The greatest shoot height was recorded in the ‘Altınova’ cultivar (6.8 cm), while the lowest shoot heights were observed in ‘Balkan’ and ‘Özgen’ (both 1.4 cm). In terms of root length, ‘Ayaz’ had the longest roots (7.7 cm). Regarding biomass accumulation, ‘Ayaz’ also showed the highest values for both fresh weight (377 mg) and dry weight (45.9 mg). The highest vigor index was recorded in the ‘Altınova’ cultivar (620), whereas the lowest value was found in ‘Özgen’ (98).

Figure 2 presents the effects of polyethylene glycol (PEG)-induced drought stress on vetch growth and vigor parameters. Under control conditions (0% PEG), the highest values were recorded across all measured traits, including the germination rate (85.6%), shoot height (6.16 cm), root length (7.8 cm), fresh weight (277 mg), dry weight (31 mg), and vigor index (456). At 5% PEG concentration, reductions were observed in all parameters, with the shoot height decreasing to 4.1 cm and the vigor index to 357. At the highest PEG concentration (10%), the lowest values were recorded for all traits: the germination rate (73.1%), shoot height (2.49 cm), root length (3.0 cm), fresh weight (94 mg), dry weight (19 mg), and vigor index (197). All differences among treatments were statistically significant (p < 0.0001).

Figure 2 shows the effects of temperature treatments on growth and vigor parameters in vetch. The highest values across all measured traits were observed at 24 °C, including the germination rate (89.1%), shoot height (6.36 cm), root length (7.6 cm), fresh weight (283 mg), dry weight (35 mg), and vigor index (593). At 18 °C, moderate reductions were recorded in all parameters, while the lowest values were obtained at 12 °C: the germination rate (70%), shoot height (2.24 cm), root length (3.0 cm), fresh weight (58.6 mg), dry weight (14 mg), and vigor index (169). Statistical analysis confirmed that all temperature-related differences were highly significant (p < 0.0001).

The results presented in Figure 3 show the interactions between vetch varieties and PEG-induced drought stress treatments. Under control conditions (0% PEG), the ‘Atom’ cultivar exhibited the highest germination rate (98.3%) and a high vigor index (818). The ‘Altınova’ cultivar showed a slightly lower germination rate (93.3%) but the highest vigor index (947), with a greater shoot height (9.92 cm) and notable fresh and dry biomass values. The ‘Ayaz’ cultivar had the longest root length (12.06 cm) and the highest fresh (511.5 mg) and dry weights (58.49 mg) under the same conditions. When exposed to 5% and 10% PEG concentrations, all cultivars showed reductions in germination and growth-related parameters. At 5% PEG, ‘Altınova’ and ‘Ayaz’ maintained relatively higher values in shoot height, root length, and biomass compared to the other varieties. At 10% PEG, all cultivars demonstrated significant declines, but ‘Ayaz’ retained higher biomass values (fresh weight: 222.8 mg; dry weight: 32.48 mg). The ‘Özgen’ cultivar showed the lowest performance across all PEG concentrations, especially at 10% PEG, with the lowest values recorded for the shoot height, root length, and vigor index.

Figure 3 presents the effects of temperature on growth and vigor parameters across different vetch varieties. At 24 °C, the highest values for the germination rate, shoot height, root length, fresh and dry biomass, and vigor index were observed across most cultivars. For example, the ‘Altınova’ cultivar recorded the highest vigor index (949) and fresh weight (341 mg), while ‘Ayaz’ exhibited the longest root length (9.8 cm) and the highest dry weight (67.5 mg) at the same temperature. At 12 °C, reductions were recorded in all measured parameters for each variety. In particular, ‘Özgen’ and ‘Balkan’ showed lower germination rates (approximately 50–67%), lower fresh and dry biomass values, and reduced vigor indices at this temperature. At 18 °C, intermediate values were recorded for all varieties, with ‘Ayaz’ showing a higher fresh and dry biomass accumulation and vigor index than other cultivars.

Figure 3 presents the interaction effects of in vitro PEG-induced drought stress and temperature on vetch growth parameters. All interactions between PEG concentration and temperature were statistically significant (p < 0.0001). Under control conditions (0% PEG), the highest values for all traits were recorded at 24 °C, including the germination rate (94.7%), shoot height (9.4 cm), root length (10.9 cm), fresh weight (443.7 mg), dry weight (44.3 mg), and vigor index (905.4). At intermediate (18 °C) and low temperatures (12 °C), reductions were observed across all parameters. At 5% PEG and 24 °C, the germination rate remained relatively high (90.3%), and biomass accumulation was moderate. In contrast, at 12 °C combined with 5% PEG, the germination rate decreased to 69.7%, and all other measured traits were lower than at higher temperatures. At 10% PEG, further reductions in the germination rate, shoot height, root length, biomass accumulation, and vigor index were recorded, with the lowest values observed at 12 °C.

Figure 4 and Figure 5 present the combined effects of cultivar, PEG-induced drought stress, and temperature on vetch growth and vigor parameters. All observed differences and interaction effects between cultivar, PEG concentration, and temperature were statistically significant (p < 0.0001). Under control conditions (0% PEG) and at 24 °C, the highest performance was observed across all varieties. The ‘Altınova’ cultivar showed the highest vigor index (1425.3), fresh weight (574.1 mg), dry weight (35.93 mg), and shoot height (14 cm). The ‘Atom’ and ‘Ayaz’ cultivars also recorded high vigor values (1310 and 1252.7, respectively).

At 5% PEG and 24 °C, germination rates remained between 95% and 100% for ‘Atom’, ‘Altınova’, and ‘Ayaz’, with ‘Ayaz’ having the highest fresh weight (564.1 mg) and dry weight (65.80 mg). At 10% PEG and 24 °C, all cultivars showed reductions, though ‘Ayaz’ maintained higher values for fresh weight (381.0 mg), dry weight (53.71 mg), and vigor index (341.9). In contrast, ‘Balkan’ and ‘Özgen’ showed low shoot height (1.2 cm), low biomass, and low vigor indices under the same conditions.

At lower temperatures (18 °C and 12 °C), further reductions were observed across all cultivars. At 12 °C, ‘Özgen’ and ‘Cumhuriyet’ exhibited germination rates as low as 45%, shoot and root lengths ranging from 0.3 to 2 cm, and vigor index values between 15.4 and 93.9. At 18 °C and 5% PEG, ‘Atom’ and ‘Altınova’ recorded vigor index values between 570 and 614.

3.2. Correlation and PCA Analysis

Correlation analysis and PCA were conducted to examine interrelationships among various growth traits (germination rate, shoot height, root length, fresh weight, dry weight, and vigor index) of vetch varieties. The correlation analysis revealed strong yet cultivar-specific trait associations. For instance, the ‘Altınova’ cultivar exhibited significant positive correlations between the germination rate and vigor index (0.86), as well as fresh and dry weight (0.91), indicating that robust germination and early biomass notably influence plant vigor. The ‘Atom’ cultivar showed an extremely high correlation between shoot height and vigor index (0.99), highlighting shoot length as an effective indicator of seedling vigor. In the ‘Ayaz’ cultivar, fresh weight was strongly correlated with dry weight (0.96), reflecting consistent biomass accumulation. Conversely, the ‘Balkan’ cultivar demonstrated a high correlation between germination rate and root length (0.87), emphasizing root growth as an essential outcome of germination, but a moderate correlation between root length and dry weight (0.48), suggesting variability in biomass allocation. The ‘Cumhuriyet’ cultivar presented consistently strong correlations among all measured traits, with the shoot height notably linked to the vigor index (0.99), reinforcing shoot development as a reliable vigor indicator. The ‘Özgen’ cultivar had moderately high correlations, particularly between the germination rate and vigor index (0.84), but showed greater variability in biomass partitioning, illustrated by a lower correlation between root length and dry weight (0.56) (Figure 6).

PCA further corroborated these correlation results by effectively summarizing and reducing trait complexity, explaining approximately 86.1% of the total variability among the traits (PC1: 13.4%; PC2: 72.7%). The PCA biplot (Figure 7) visually confirmed strong relationships by closely grouping the germination rate, shoot height, and vigor index and aligning fresh and dry weight closely together. Root length was positioned separately, aligning with its observed moderate correlation and indicating distinctive variability. Thus, these analyses collectively clarified key trait interrelationships, emphasizing the predictive significance of early growth parameters, such as germination performance and shoot development, in determining overall vigor and biomass production in vetch varieties.

3.3. Machine Learning

The performance of machine learning models (Random Forest [RF], k-Nearest Neighbor [k-NN], Multilayer Perceptron [MLP], and Support Vector Machine [SVM]) was evaluated to predict plant growth parameters including the germination rate, shoot length, root length, fresh weight, dry weight, and vigor index. The statistical evaluation criteria included R² (coefficient of determination), RMSE (root mean squared error), and MAE (mean absolute error), applied separately for training and testing datasets (

Table 2. Performance results of models under different parameters.

	Training Set			Testing Set
	R2	RMSE	MAE	R2	RMSE	MAE
Germination Rate
RF	0.96	0.05	0.04	0.90	0.08	0.06
k-NN	0.97	0.04	0.02	0.93	0.08	0.06
MLP	0.96	0.05	0.04	0.95	0.06	0.05
SVM	0.97	0.04	0.03	0.94	0.07	0.06
Shoot Length
RF	0.99	0.01	0.007	0.98	0.02	0.02
k-NN	0.94	0.05	0.03	0.93	0.05	0.04
MLP	0.98	0.03	0.02	0.97	0.03	0.03
SVM	0.98	0.03	0.02	0.95	0.05	0.04
Root Length
RF	0.84	0.08	0.07	0.81	0.10	0.08
k-NN	0.86	0.08	0.06	0.76	0.11	0.09
MLP	0.85	0.09	0.07	0.82	0.10	0.07
SVM	0.87	0.08	0.06	0.83	0.09	0.07
Fresh Weight
RF	0.95	0.05	0.03	0.91	0.06	0.05
k-NN	0.97	0.04	0.02	0.90	0.07	0.05
MLP	0.94	0.06	0.04	0.92	0.06	0.05
SVM	0.96	0.04	0.03	0.97	0.05	0.03
Dry Weight
RF	0.95	0.04	0.03	0.93	0.06	0.04
k-NN	0.94	0.03	0.02	0.86	0.07	0.05
MLP	0.93	0.05	0.04	0.88	0.06	0.05
SVM	0.95	0.04	0.03	0.91	0.05	0.04
Vigor Index
RF	0.99	0.01	0.006	0.98	0.03	0.02
k-NN	0.96	0.04	0.03	0.91	0.06	0.04
MLP	0.99	0.009	0.006	0.99	0.02	0.01
SVM	0.98	0.03	0.02	0.97	0.04	0.02

ML: Machine Learning; RF: Random Forest; k-NN: k-Nearest Neighbor; MLP: Multilayer Perceptron; SVMs: Support Vector Machines; R²: coefficient of determination; MAE: mean absolute error; RMSE: root mean square error; MSE: mean squared error.

).

All models performed effectively on the training dataset for the germination rate, demonstrating high accuracy. Specifically, k-NN and SVM models exhibited slightly superior performances, achieving an R² value of 0.97, indicating excellent predictive capability with minimal error (RMSE = 0.04, MAE between 0.02 and 0.03). The RF and MLP models also performed strongly, but their results (R² = 0.96) were slightly below k-NN and SVM, along with marginally higher errors. The predictions for shoot length indicated exceptionally high training accuracy for RF (R² = 0.99) with the lowest RMSE (0.01) and MAE (0.007) among all traits. Similarly, MLP and SVM models showed high accuracies (R² = 0.98), suggesting strong learning capabilities. However, the k-NN model performed less robustly (R² = 0.94), accompanied by relatively higher errors, indicating potential limitations in learning complex patterns related to shoot length. For root length, the accuracy of models during training was moderately high but notably lower compared to other traits. Among them, SVM yielded the highest performance (R² = 0.87), closely followed by k-NN (R² = 0.86), MLP (R² = 0.85), and RF (R² = 0.84). The higher error metrics for all models (RMSE between 0.08 and 0.09) suggest inherent variability or complexity in root growth, making it challenging for models to fully capture variations from available inputs. The prediction results for fresh weight were consistently strong across all models, with k-NN achieving the highest accuracy (R² = 0.97), suggesting that this model captures patterns within the data exceptionally well for this trait. SVM (R² = 0.96) and RF (R² = 0.95) also exhibited high accuracies, followed closely by MLP (R² = 0.94). In estimating dry weight, the models maintained strong accuracy, especially RF and SVM (both R² = 0.95), indicating very good predictive potential. The k-NN and MLP models performed slightly lower but were still very effective, achieving R² values of 0.94 and 0.93, respectively. Lastly, for the vigor index, RF and MLP models provided exceptional predictive performance, each achieving the highest accuracy observed among all traits (R² = 0.99), coupled with extremely low error metrics (RMSE ≤ 0.01). These results suggest that both RF and MLP models have learned the patterns in this data highly effectively. SVM and k-NN also provided strong predictive performances (R² = 0.98 and 0.96, respectively), confirming their robustness for predicting vigor.

Analysis based on training dataset performance reveals that RF, SVM, k-NN, and MLP models all provided consistently strong predictions across the examined plant growth parameters. Particularly notable was the excellent predictive ability of RF for shoot length and vigor index, k-NN for germination rate and fresh weight, and SVM for root length. These training results indicate the high capability of these models to capture underlying data patterns during model development. Also, the predictive accuracy and consistency of the models described above were visually confirmed by scatter plots illustrating the measured versus estimated values for the germination rate (Figure 8), shoot length (Figure 9), root length (Figure 10), fresh weight (Figure 11), dry weight (Figure 12), and vigor index (Figure 13).

4. Discussion

Seed germination and early seedling development are critical phases in the plant life cycle, directly influencing crop establishment and yield potential, particularly under abiotic stress conditions such as drought and temperature extremes. In this study, the interactive effects of PEG-induced drought stress and temperature on the germination and early growth traits of six Vicia cultivars from three species (V. sativa, V. pannonica, and V. narbonensis) were systematically evaluated under controlled in vitro conditions. The results demonstrated that PEG concentration and temperature had statistically significant and cultivar-dependent effects on all measured parameters, including the germination rate, shoot and root lengths, fresh and dry weights, and vigor index.

The present findings demonstrate that vetch growth and vigor are significantly influenced by genotype, drought stress level, and temperature, both independently and interactively. Among the evaluated cultivars, ‘Ayaz’, ‘Altınova’, and ‘Atom’ consistently exhibited superior performance across multiple parameters, including biomass accumulation, root development, and vigor indices, particularly under control conditions and at 24 °C. This suggests that these cultivars possess favorable physiological traits that support high metabolic activity and resource allocation when environmental conditions are optimal. Notably, ‘Ayaz’ maintained relatively high levels of fresh and dry biomass even under moderate (5%) and severe (10%) PEG-induced drought stress, especially when combined with higher temperatures, indicating a potentially enhanced drought tolerance mechanism. These results imply that ‘Ayaz’ may be a promising candidate for cultivation in drought-prone, warm environments.

Conversely, cultivars such as ‘Balkan’ and ‘Özgen’ displayed limited growth potential under both control and stress conditions, with significantly reduced shoot and root development and lower vigor indices, particularly under low temperature and high PEG concentration. Their pronounced sensitivity to combined drought and cold stress highlights the importance of environmental compatibility in genotype selection for optimal performance.

Temperature effects were also critical, with 24 °C consistently associated with the highest levels of growth and vigor, whereas exposure to 12 °C markedly restricted development across all cultivars. These results reflect a clear physiological limitation of vetch to cold conditions, even in the absence of osmotic stress. The negative interaction between low temperature and drought was especially evident at 10% PEG and 12 °C, where growth parameters deteriorated most severely. This interaction underscores the compounding impact of multiple abiotic stresses on legume crops and the necessity of identifying genotypes capable of maintaining productivity under such combined constraints.

Overall, the three-way interaction among cultivar, drought stress, and temperature clearly emphasizes the need for genotype-specific management strategies. Cultivars like ‘Ayaz’ and ‘Altınova’, which showed resilience under moderate stress conditions, may offer valuable genetic resources for breeding programs targeting drought and temperature adaptability in vetch. In contrast, susceptible cultivars such as ‘Özgen’ and ‘Balkan’ may require exclusion from stress-prone environments or use in more favorable climates to avoid yield instability.

Correlation and PCA analyses obtained in this study show that the relationships between germination parameters of vetch cultivars are specific to the cultivars and that they develop different adaptation strategies against environmental stress factors. Strong positive correlations between the germination rate and vigor index (r = 0.86) and fresh dry weight (r = 0.91) in the ‘Altınova’ cultivar indicate that early germination and biomass accumulation are important determinants of plant vigor. Similarly, a high correlation between the shoot length and the vigor index in the ‘Atom’ cultivar (r = 0.99) emphasizes that shoot length is an effective indicator of seedling vigor. These findings show that different cultivars develop different adaptation strategies against environmental stresses, and these strategies are reflected by the relationships between growth parameters [2,6]. Aydınoğlu et al. [46] reported significant correlations between the germination rate, shoot and root length, and vigor index in their study on germination and seedling development of common vetch (Vicia sativa) seeds under salt stress.

The observed negative impact of PEG on seedling performance aligns with previous reports indicating that PEG-simulated drought stress reduces water availability by lowering osmotic potential, thereby inhibiting water uptake during germination [31]. Notably, the germination rate and seedling vigor significantly declined with increasing PEG concentrations, particularly at 10%, where reductions were drastic across all cultivars. These findings are consistent with the results of Çifçi and Açıkbaş [47], who reported similar drought sensitivity patterns among Turkish vetch cultivars, with PEG levels above 5% leading to marked declines in germination indices and biomass traits.

Temperature effects were equally prominent, with 24 °C identified as the optimal condition for all varieties, promoting the highest germination rates, shoot and root elongation, and biomass accumulation. This thermal preference is supported by earlier work by Sharavdorj et al. [21] and Guo et al. [23], emphasizing the crucial role of ambient temperature in enhancing enzymatic activity, cellular expansion, and metabolic rates during germination. At 12 °C, severe limitations were observed across all growth traits, suggesting that low temperatures impose both metabolic and mechanical constraints on seedling emergence. These results also mirror those of Zeng et al. [48], who noted that reduced temperatures compromised the growth-promoting effects of external treatments like methane pulse spray in vetch.

Significant inter-cultivar differences were recorded, revealing distinct tolerance patterns among the evaluated genotypes. The ‘Ayaz’ (V. sativa) cultivar consistently showed superior performance across PEG and temperature gradients, especially under high-stress conditions (10% PEG at 24 °C), maintaining high fresh and dry biomass and vigor indices. This suggests that ‘Ayaz’ may possess inherent physiological mechanisms for osmotic adjustment and efficient water use efficiency, as hypothesized in earlier studies on stress-tolerant vetch lines [18,19]. In contrast, Özgen (V. narbonensis) and Balkan (V. narbonensis) showed high sensitivity to combined drought and temperature stresses, with severely reduced growth metrics and germination under suboptimal conditions, indicating limited adaptability.

These genotypic differences echo findings from Tilhou et al. [49], who demonstrated that genetic and maternal environmental factors significantly influence vigor traits such as seed size and biomass in Vicia villosa. Larger seed size was linked to improved seedling vigor, a trait indirectly supported by the strong correlation between biomass traits and the vigor index in this study’s high-performing cultivars, like ‘Altınova’ and ‘Ayaz’.

Beyond PEG and temperature, other abiotic stress mitigation approaches have shown promise in enhancing vetch performance. For example, Odat et al. [50] demonstrated that chitosan seed priming effectively improved germination and early growth under salinity stress in V. sativa. While our study did not include priming agents, the superior performance of certain cultivars under moderate PEG stress suggests that integrating biostimulants such as chitosan could improve drought resilience in sensitive genotypes like ‘Cumhuriyet’ and ‘Özgen’.

Similarly, Shi et al. [51] applied high-voltage corona discharge treatments to stimulate root growth and enhance germination in vetch seeds under high-altitude stress, highlighting the potential of physical mutagenic technologies in abiotic stress management. While these methods differ in mechanism from osmotic stress simulation, they underscore the diversity of approaches available to modulate seed responses to environmental limitations.

Another notable intervention is the application of complex biofertilizers. Skamarokhova et al. [52] reported that biofertilizer-treated winter vetch exhibited significantly improved germination energy and seedling vigor, suggesting enhanced metabolic priming and microbial interaction. Integrating such treatments in future PEG-temperature factorial designs could provide further insights into the synergistic effects of biostimulants under drought.

The correlation analysis in this study revealed robust positive associations between the shoot height, fresh weight, and vigor index, particularly in ‘Atom’ and ‘Cumhuriyet’ cultivars, underscoring shoot elongation as a reliable indicator of early seedling vigor. Similar patterns were reported by Zeng et al. [48], who demonstrated that maximum seedling growth coincided with optimal stress modulation under controlled methane pulse and irrigation treatments. Interestingly, root length, although a key trait for water acquisition, showed weaker correlations with vigor in some cultivars, indicating that shoot traits might be more critical indicators under mild to moderate stress. In contrast, root development may be more prominent during severe drought, especially in field conditions.

The PCA results further reinforced the key role of fresh and dry biomass traits in explaining overall variation across treatments. The alignment between the germination rate, shoot height, and vigor index in the PCA biplot emphasizes their integrated contribution to early growth performance. These findings echo those of Skamarokhova et al. [52], who also observed that vigor-related traits clustered together in response to biofertilizer treatment.

Integrating machine learning (ML) algorithms into plant stress physiology research has significantly enhanced the capacity to model and predict plant responses under complex abiotic conditions. The present study adopted a comprehensive ML-based approach to evaluate the germination and early seedling performance of six Vicia cultivars under different levels of PEG-induced drought and temperature treatments. This approach allowed for the identification of the most significant traits influencing stress tolerance, with models such as Random Forest (RF), Gaussian Process (GP), and Extreme Gradient Boosting (XGBoost) demonstrating high predictive power for the germination rate, vigor index, and biomass traits. A notable distinction in this study lies in its three-way factorial design, integrating PEG-induced drought, temperature variation, and genotypic diversity (six vetch cultivars) into a single predictive framework. This multifactorial experimental approach allowed for a more realistic simulation of site-relevant stress interactions and enabled the modeling of cultivar-specific response profiles.

In the study conducted by Zarbakhsh et al. [53] on pomegranate (Punica granatum), the Generalized Regression Neural Network (GRNN) algorithm was found to be the most effective model in predicting antioxidant enzyme activities. Moreover, optimal levels of exogenous GABA application were determined through optimization with the NSGA-II algorithm. These findings highlight that machine learning approaches can be effectively used not only for predictive modeling but also for optimizing stress-mitigating treatments. Therefore, although there may be variability across plant species and stress parameters, the comparative use of multiple ML algorithms offers a robust and broadly applicable framework for modeling both biotic and abiotic stress conditions.

Recent studies on the application of machine learning in plant stress physiology have demonstrated high accuracy in the early detection of stress and prediction of physiological parameters. Within this context, Okyere et al. [54] integrated hyperspectral imaging (HSI) data with both newly developed and existing vegetation indices (VIs) to create a machine learning model aimed at detecting drought stress in wheat under varying nitrogen conditions. Classification models, including Support Vector Machines (SVMs), Random Forest (RF), and deep neural networks (DNNs), achieved over 94% accuracy using the newly proposed indices. Particularly, RF regression performed best in predicting the stomatal conductance and photosynthetic rate. These findings are consistent with our study, where the RF algorithm exhibited excellent predictive accuracy for complex growth parameters such as the shoot length and vigor index. Furthermore, a study by Okyere et al. [54] underscores the power of machine learning in modeling combined stress scenarios (e.g., drought + nitrogen deficiency), offering valuable insight into the generalizability of our in vitro PEG and temperature stress models to more complex field conditions. Hence, algorithms like RF and SVM can be considered as versatile tools not only for modeling individual stress types but also for capturing their interactive effects.

Machine learning models have proven effective not only in predicting stress responses but also in supporting tissue culture-based propagation systems. Kirtis et al. [55] conducted an in vitro regeneration study in Cicer arietinum (chickpea), using various types and concentrations of cytokinins, and employed SVR, GPR, XGBoost, RF, and MLP models to predict shoot number and length. Among these, the RF algorithm demonstrated the highest performance, with R² = 0.99 and MSE = 2.86 for shoot count and R² = 0.98 and MSE = 0.29 for shoot length. These results align with our findings, where the RF model also yielded high predictive accuracy for the shoot length (R² = 0.99) and vigor index (R² = 0.99). Both studies demonstrate the RF model’s ability to effectively capture complex structural variation in biological data under in vitro conditions. This supports the broader applicability and reliability of RF as a robust predictive tool for tissue culture-based plant growth modeling.

In strawberry cultivar analysis under PEG stress, Şimşek [41] found that RF achieved the highest accuracy (91.16%) across morphometric traits, with MLP and GP models contributing to trait-specific predictions. Our results similarly point to the utility of RF in accurately modeling trait interactions under stress, particularly in genotypically diverse datasets like Vicia.

When evaluating more complex stress conditions such as heavy metal toxicity, several studies have demonstrated the efficacy of ML in optimizing in vitro responses. For instance, Tütüncü et al. [40] observed that in myrtle, the MLP model achieved R² values of 0.87 for shoot height and 0.99 for root length under cadmium exposure, outperforming RF and XGBoost. Likewise, in goji berry, MLP and RF provided up to 0.98 accuracies for shoot/root traits under Cd stress [44]. These results highlight the consistent superiority of MLP and RF across species and stress types, underscoring their generalizability in tissue culture prediction environments.

A broader evaluation by Shwetabh and Ambhaikar [56] in common bean regeneration found that RRNN, a deep learning variant, surpassed traditional ML models in predicting shoot and callus responses to BAP and CuSO₄, indicating a move toward more dynamic architectures in plant modeling. Although our study did not employ deep learning, the current results suggest that future modeling efforts in Vicia species could benefit from RNN-based architectures to capture temporal dependencies during seedling development.

Finally, Palaz et al. [35] demonstrated the successful use of ML in optimizing in vitro propagation in olive, where XGBoost and SVR were instrumental in refining sterilization protocols and growth media optimization. These findings indicate ML’s growing role in stress physiology and the design of reproducible, high-throughput plant tissue culture systems. This approach aligns with our study’s objective of cultivar-specific optimization under drought conditions.

This study reinforces the utility of ML in decoding complex interactions between genotype and environment, offering rapid and accurate insights for breeding programs. Future work may involve hybrid modeling (e.g., integrating ANN with optimization algorithms like GA or PSO), using image-based ML for real-time phenotyping, or incorporating omics data to enhance trait prediction. Moreover, deploying deep learning models like convolutional or recurrent networks could further improve the modeling of dynamic physiological responses under fluctuating abiotic stresses.

5. Conclusions

This study demonstrates that cultivar, PEG-simulated drought stress, and temperature variations significantly influence Vicia species’ germination and early seedling development. Among the tested cultivars, ‘Ayaz’ (V. sativa) displayed the highest resilience under combined abiotic stresses, suggesting its suitability for cultivation in arid environments. In contrast, Özgen and Balkan exhibited pronounced sensitivity, indicating the need for targeted improvement strategies. Integrating machine learning algorithms provided powerful predictive capabilities, with MLP and RF emerging as reliable models for estimating complex growth traits under stress. These findings not only contribute to the understanding of genotype-specific stress responses but also highlight the value of combining traditional phenotyping with AI-based modeling in legume improvement programs. Future research should explore genomic integration and field-level validations to further enhance the application of this integrative screening platform.