Integrating machine learning with fruit trait and biochemical profiling of local European plum (Prunus domestica L.) germplasm from the Black Sea region

Understanding the relationships between fruit morphological traits and biochemical characteristics is essential for the efficient utilization of plum (Prunus domestica L.) genetic resources in breeding programs. In this study, fifteen local European plum varieties collected from the Central and Eastern Black Sea Region of Türkiye were evaluated during [2022 and 2023 year] to characterize their morphological and biochemical diversity and to identify the key biochemical determinants of antioxidant activity. Machine learning models were developed and evaluated using cultivar-level cross-validation to explore the relationships between antioxidant capacity and fruit morphological and biochemical traits. Although Random Forest showed relatively better performance than the other evaluated models, its negative R² under leave-one-cultivar-out cross-validation indicated limited predictive ability and poor generalization to previously unseen cultivars. Therefore, the machine learning results should be considered exploratory rather than as evidence of robust prediction or causal relationships among the measured traits. Nevertheless, the observed variation in fruit quality and biochemical characteristics provides useful information for the characterization and evaluation of local plum genetic resources and may contribute to future studies on their conservation and potential utilization in breeding programs. Substantial phenotypic and biochemical variation was detected among the evaluated local plum varieties. Fruit weight ranged from 7.06 to 36.72 g, while fruit width varied between 20.67 and 36.54 mm. DPPH radical scavenging activity reached a maximum value of 83.20%. Total phenolic content ranged from 11.15 to 210.52 mg/g fresh weight, whereas vitamin C content varied between 5.29 and 82.03 ppm. Several machine learning algorithms were evaluated to explore the potential of predicting antioxidant activity from the measured biochemical parameters. Under cultivar level leave one-cultivar-out cross-validation (LOCO-CV), Random Forest yielded the highest R² (-0.251) and the lowest RMSE (33.91), whereas SVR-RBF resulted in a lower MAE (25.24) than Random Forest (28.99). Nevertheless, the negative R² values obtained across all evaluated models indicate limited predictive performance and poor generalization to previously unseen cultivars. Feature-importance analysis indicated that total phenolic content and vitamin C contributed to the model predictions; however, these findings should be interpreted cautiously given the limited number of independent cultivars and the exploratory nature of the analysis. SHAP-based interpretation did not provide sufficient evidence to support positive or synergistic effects of total phenolic content and vitamin C on antioxidant activity. The results revealed considerable morphological and biochemical diversity among local European plum varieties, highlighting their potential value as genetic resources for plum improvement. Total phenolic content and vitamin C were identified as the principal biochemical determinants of antioxidant activity, while the integration of machine learning with biochemical characterization provided an effective approach for predicting antioxidant capacity and elucidating complex trait relationships. Beyond their value for characterizing existing germplasm, these findings indicate that locally adapted plum varieties may represent important reservoirs of functional diversity that can be exploited to broaden the genetic basis of breeding materials. Their incorporation into conservation and breeding strategies could contribute to the development of plum cultivars combining desirable agronomic traits with enhanced antioxidant potential and nutritional quality.