Machine-Learning-Based Proteomic Predictive Modeling with Thermally-Challenged Caribbean Reef Corals

Abstract

Coral health is currently diagnosed retroactively; colonies are deemed “stressed” upon succumbing to bleaching or disease. Ideally, health inferences would instead be made on a pre-death timescale that would enable, for instance, environmental mitigation that could promote coral resilience. To this end, diverse Caribbean coral (Orbicella faveolata) genotypes of varying resilience to high temperatures along the Florida Reef Tract were exposed herein to elevated temperatures in the laboratory, and a proteomic analysis was taken with a subset of 20 samples via iTRAQ labeling followed by nano-liquid chromatography + mass spectrometry; 46 host coral and 40 Symbiodiniaceae dinoflagellate proteins passed all stringent quality control criteria, and the partial proteomes of biopsies of (1) healthy controls, (2) sub-lethally stressed samples, and (3) actively bleaching corals differed significantly from one another. The proteomic data were then used to train predictive models of coral colony bleaching susceptibility, and both generalized regression and machine-learning-based neural networks were capable of accurately forecasting the bleaching susceptibility of coral samples based on their protein signatures. Successful future testing of the predictive power of these models in situ could establish the capacity to proactively monitor coral health.

Keywords

• artificial intelligence
• coral reefs
• dinoflagellates
• global climate change
• machine learning
• molecular biotechnology