Predicting photosynthetic pathway from anatomy using machine learning

Abstract

Plants with Crassulacean acid metabolism (CAM) have long been associated with a specialized anatomy, including succulence and thick photosynthetic tissues. Firm, quantitative boundaries between non-CAM and CAM plants have yet to be established - if they indeed exist. Using novel computer vision software to measure anatomy, we combined new measurements with published data across flowering plants. We then used machine learning and phylogenetic comparative methods to investigate relationships between CAM and anatomy. We found significant differences in photosynthetic tissue anatomy between plants with differing CAM phenotypes. Machine learning-based classification was over 95% accurate in differentiating CAM from non-CAM anatomy, and had over 70% recall of distinct CAM phenotypes. Phylogenetic least squares regression and threshold analyses revealed that CAM evolution was significantly correlated with increased mesophyll cell size, thicker leaves, and decreased intercellular airspace. Our findings suggest that machine learning may be used to aid the discovery of new CAM species and that the evolutionary trajectory from non-CAM to strong, obligate CAM requires continual anatomical specialization.

Keywords: Asparagaceae; Crassulacean acid metabolism; Portullugo; machine learning; photosynthesis.