.png){kind=icon}
.png){kind=icon}
Return
Machine learning-based prediction of morphophysiological, ionic, and biochemical responses of Lolium multiflorum to salinity stress
.png){kind=icon}
OA
AI
E
O
A
S
F
A
M
S
DOI：10.1186/s12870-026-09748-4.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Salinity is a major constraint on the establishment and productivity of annual ryegrass (Lolium multiflorum Lam.). This study evaluated the effects of NaCl stress on the germination, growth, mineral composition, and biochemical responses of four annual ryegrass cultivars (‘Grasslands Bill’, ‘Koga’, ‘Trinova’, and ‘Vivaro’) and assessed the ability of machine-learning algorithms to predict cultivar performance across different salinity levels. Germination experiments were conducted at 0, 5, 10, 15, and 20 dS m⁻¹ NaCl, while pot experiments used 0, 2.5, 5, 7.5, and 10 dS m⁻¹. Cultivar, salinity, and their interaction significantly affected most measured traits. Increasing salinity markedly reduced shoot and root lengths, seedling fresh weight, shoot fresh and dry weights, tiller number, and plant height. Crude protein, chlorophyll, proline, and antioxidant enzyme activities showed trait-specific and non-linear responses to salinity. Proline was generally higher under saline treatments, SOD activity tended to decline, and POD activity reached its maximum at 7.5 dS m⁻¹. Na⁺ concentration increased substantially with salinity, whereas K⁺ concentration declined. Trinova produced the highest shoot fresh weight (23.68 g), shoot dry weight (9.44 g), and plant height (46.98 cm). Koga had the highest K⁺ concentration (23,050 mg kg⁻¹) and crude protein content (7.80%), while Vivaro showed the highest proline content (8.50 mg g⁻¹ FW) and peroxidase activity (10,892 EU g⁻¹ FW). Repeated grouped training–test evaluation with 100 repetitions and five-fold nested cross-validation demonstrated strongly trait-dependent predictive performance. Mean test R² values ranged from 0.51 to 0.95 for germination traits and from 0.19 to 0.95 for pot-experiment traits. Predictions were comparatively consistent for shoot length, mean germination time, shoot fresh weight, plant height, and Na⁺ concentration, whereas seedling dry weight, chlorophyll content, superoxide dismutase activity, and peroxidase activity showed lower or less stable performance. No single algorithm was consistently superior, with XGBoost, random forest, support vector regression, and Gaussian process regression each performing best for different target traits. Annual ryegrass cultivars differed substantially in their physiological and biochemical responses to salinity, indicating trait- and cultivar-specific response patterns. Machine learning complemented conventional statistical analyses by providing trait-specific predictions across the tested salinity gradients. However, because the models were developed using only four cultivars and 20 unique cultivar × salinity combinations per experiment, the findings should be considered exploratory. Independent validation with additional genotypes, environments, and field-based datasets is required before broader predictive or crop-management applications.
Keywords:
Salinity stress
Osmotic adjustment
Ion homeostasis
Antioxidant enzymes
Proline
Shiny
Journal
IF:
4.8
Papers:
1.0W
Citations:
3.0W
Organization
Cited Papers