Machine learning and model-agnostic methods to identify influential features in volatile profiles from tea plants infested by the tea aphid

Machine learning and model-agnostic methods to identify influential features in volatile profiles from tea plants infested by the tea aphid Zehua Qi first-author Huaguang Qin Shuyuan Wang Xinyi Hu Yuxi Wang Nuo Wang Hui Wang Huiwen Zhang Rui Shi Yinghong Hu Wuxuan Hong Dan Mu corresp mudansmile@126.com and Tingzhe Sun corresp confucian007@126.com School of Life Sciences, The Province Key Laboratory of the Biodiversity Study and Ecology Conservation in Southwest Anhui, Anqing Normal University,Anhui,Anqing,China,246133 School of Life Sciences, The Province Key Laboratory of the Biodiversity Study and Ecology Conservation in Southwest Anhui, Anqing Normal University,Anhui,Anqing,China,246133 The tea plant ( Camellia sinensis ), as a major economic crop in China, suffers greatly from infestation by the tea aphid Toxoptera aurantii (Boyer), which pierces and sucks the sap of tea shoots. In response to herbivore attack, tea plants can release volatiles with various functions. In the current work, volatile emission in tea plants upon tea aphid infestation was measured using gas chromatography-mass spectrometry (GC-MS). Significant differences were observed in the categories and relative concentrations of volatiles in infested tea shoots. Machine learning (ML) modeling and model-agnostic methods were employed to interpret the volatile profiles in healthy, mildly infested, and severely infested tea shoots. The volatile emission data were fitted using a random forest (RF) model. Since several features were highly correlated, the accumulated local effect (ALE) was used to unravel the relationship between volatile abundance and model prediction. & #x3B1;-Pinene, phytol, naphthalene, camphor, and nonanal were identified as critical factors based on the ALE plots. Notably, even the most influential volatile had a relatively low effect on predicted probability, suggesting that the determination of infestation status possibly requires integrating the impacts of at least a few volatiles. The volatile interactions were also investigated using two-dimensional ALE plots. The RF model with model-agnostic methods can obtain both prediction accuracy and model interpretability, which helps to identify the importance of volatiles and intrinsic volatile interactions associated with tea aphid infestation. This combined strategy has great potential in related fields to interpret volatile emission patterns. Tea plant Tea aphid Machine Learning Model-agnostic methods 10.1631/jzus.B2200618 2022-12-05 2023-03-19 2023-04-02 2023-04-02