Deep learning-based high-information-content graph representation of early stage bacterial biofilms

Bacterial biofilms are complex, spatially organized microbial communities that exhibit enhanced resistance to antibiotics and contribute to chronic infections. Understanding their structure, especially during early formation stages, is critical for developing effective intervention strategies. Here, we present a high-resolution computational framework that models biofilms as undirected interaction graphs, where individual bacterial cells are vertices and predicted intercellular interactions are edges. Combining microscopy visualization and deep learning, we developed a pipeline that integrates Mask R-CNN for cell segmentation and a custom neural network (BINet) for interaction prediction. The described graph-based representation enables quantitative analysis of bio lm growth, identification of recurrent structural motifs, and classification of substrate-specific colonization patterns. We demonstrate the utility of this approach in predicting both the developmental stage and material type from image-derived graph features. This study provides the tool for revealing nonobvious patterns of biofilm organization and describes a scalable, high-information-content approach for the automated analysis of microbial communities, opening new possibilities for systems-level microbiological research.