Artificial neural network (ANN)-based algorithms for high light stress phenotyping of tomato genotypes using chlorophyll fluorescence features

Abstract

High light (HL) is a common environmental stress directly imposes photoinhibition on the photosynthesis apparatus. Breeding plants for tolerance against HL is therefore highly demanded. Chlorophyll fluorescence (ChlF) is a sensitive indicator of stress in plants and can be evaluated using OJIP transients. In this study, we compared the ChlF features of plants exposed to HL (1200 μmol m^-2 s^-1) with that of control plants (300 μmol m^-2 s^-1). To extract the most reliable ChlF features for discrimination between HL-stressed and non-stressed plants, we applied three artificial neural network (ANN)-based algorithms, namely, Boruta, Support Vector Machine (SVM), and Recursive Feature Elimination (RFE). Feature selection algorithms identified multiple features but only two features, namely the maximal quantum yield of PSII photochemistry (F_V/F_M) and quantum yield of energy dissipation (ɸ_D0), remained consistent across all genotypes in control conditions, while exhibited variation in HL. Therefore, considered reliable features for HL stress screening. The selected features were then used for screening 14 tomato genotypes for HL. Genotypes were categorized into three groups, tolerant, semi-tolerant, and sensitive genotypes. Foliar hydrogen peroxide (H₂O₂) and malondialdehyde (MDA) contents were measured as independent proxies for benchmarking selected features. Tolerant genotypes were attributed with the lowest change in H₂O₂ and MDA contents, while the sensitive genotypes displayed the highest magnitude of increase in H₂O₂ and MDA by HL treatment compared to the control. Finally, a F_V/F_M higher than 0.77 and ɸ_D0 lower than 0.24 indicates a healthy electron transfer chain (ETC) when tomato plants are exposed to HL.

Keywords: Artificial neural network; Bruta; Feature selection; Light stress; OJIP; Photosynthesis; RFE; SVM.