Learning from the Rhizosphere: Data-Driven Approaches to Understand How Plant Traits are Linked to Microbial Communities

Abstract

Understanding how rhizosphere microbial communities influence plant phenotype is central to developing strategies for sustainable agriculture, but systematic study is difficult because microbes function within complex, interacting communities. Machine learning methods offer a way to navigate the large combinatorial space of possible microbial communities and support future efforts to model how community composition relates to plant phenotype. This thesis addresses technical hurdles required for future active-learning experiments pairing five-member synthetic microbial communities with \textit{Arabidopsis thaliana} plant phenotypes. First, we adapted an existing plant image-segmentation model and root-analysis software to efficiently extract phenotypic descriptors and root graph objects from plant image data, making scalable plant phenotyping more feasible for large-scale experiments. Second, I developed and evaluated descriptor-based and machine-learned representations of plant phenotype using 1,645 time-lapse images from 291 plants across 85 agar plates. Biological, topological, and shape-based descriptors were screened for variation and redundancy to create an interpretable validation set. I then trained a GINE graph neural network with contrastive learning to generate unsupervised plant-graph embeddings. UMAP visualizations and K-nearest neighbor agreement analyses showed that the learned embeddings organized plants along established biological descriptors, graph-topological features, and additional shape-based traits. Biological and topological descriptors showed strong neighborhood agreement, while shape-based descriptors were preserved more moderately but still meaningfully. Together, these results demonstrate that unsupervised graph-based plant representations can recover standard biological measures of phenotype while also capturing additional morphological structure. This work lays the foundation for active-learning experiments that efficiently explore how microbial community composition shapes plant phenotype.