Optical flow reveals motility signatures for inferring pathogenic bacterial mixture compositions via temporal convolutional networks

Abstract

With the rapid expansion of global food demand, aquaculture has become a critical pillar for future food security. However, aquaculture systems remain highly vulnerable to pathogenic bacteria, and rapid identification of antagonistic microbes is essential for sustainable disease control. Conventional evaluation approaches rely on fluorescence labeling or post-culture assays, limiting the ability to quantify dynamic interactions in mixed microbial populations in a real-time and label-free manner. Here, we propose a computational framework for classifying the mixing ratio of Vibrio harveyi and environmental bacteria using time-series motion features extracted from microscopy videos. We defined 24 interpretable motility descriptors and employed a Temporal Convolutional Network (TCN) to learn their temporal structure. The proposed method achieved a classification accuracy of 93.3%, outperforming conventional static statistical approaches and alternative machine learning models. These findings indicate that mixture discrimination in microbial communities is governed not by absolute motility magnitude, but by collective alignment and its temporal stability. Our study establishes a time-resolved computational framework for quantifying dynamic collective order in mixed microbial populations and highlights its potential for label-free automated screening and robotic microbiological applications. Author summaryBacterial infections pose a major threat to aquaculture, and rapid identification of antagonistic microbes is essential for sustainable disease management. Existing screening approaches often require fluorescent labeling or post-culture analysis, making real-time evaluation of mixed bacterial populations difficult. In this study, we show that mixture ratios of Vibrio harveyi and environmental bacteria can be accurately classified from time-series motion features extracted from microscopy videos. By applying TCN to 24 interpretable motility descriptors, we achieved high classification accuracy without relying on fluorescent markers. Our analysis demonstrates that collective directional alignment and its temporal stability, rather than absolute swimming speed, are the key determinants of mixture discrimination. This work introduces a computational strategy for quantifying dynamic collective order in microbial communities and supports the development of label-free, automated screening platforms for microbiological applications.