A Paradigm of Computer Vision and Deep Learning Empowers the Strain Screening and Bioprocess Detection

Abstract

High-performance strain and corresponding fermentation process are essential for achieving efficient biomanufacturing. However, conventional offline detection methods for products are cumbersome and less stable, hindering the “Test” module in the operation of “Design-Build-Test-Learn” cycle for strain screening and fermentation process optimization. This study proposed and validated an innovative research paradigm combining computer vision with deep learning to facilitate efficient strain selection and effective fermentation process optimization. A practical framework was developed for gentamicin C1a titer as a proof-of-concept, using computer vision to extract different color space components across various cultivation systems. Subsequently, by integrating data preprocessing with algorithm design, a prediction model was developed using 1D-CNN model with Z-score preprocessing, achieving a correlation coefficient (R²) of 0.9862 for gentamicin C1a. Furthermore, this model was successfully applied for high-yield strain screening and real-time monitoring of the fermentation process and extended to rapid detection of fluorescent protein expression in promoter library construction. The visual sensing research paradigm proposed in this study provides a theoretical framework and data support for the standardization and digital monitoring of color-changing bioprocesses.