An efficient high-throughput screening of high gentamicin-producing mutants based on titer determination using an integrated computer-aided vision technology and machine learning

The ‘design-build-test-learn’ (DBTL) cycle has been adopted in rational high-throughput screening for obtaining high-yield industrial strains. However, the mismatch between build and test slows the DBTL cycle due to the lack of high-throughput analytical technologies. In this study, a highly-efficient, accurate, and non-invasive detection method of gentamicin (GM) was developed, which can provide timely feedback for the high-throughput screening of high-yield strains. Firstly, a self-made tool was established to obtain datasets in 24-well ​based on the coloring of cells. Subsequently, the random forest (RF) algorithm was found to have the highest prediction accuracy with 98.5% for the training and 91.3% for verification. Finally, a stable genetic high-yield strain (998U/mL) was successfully screened out in 3005 mutants, which was verified to improve the titer by 72.7% in a 5 L bioreactor. Moreover, the verified new datasets were updated to the model database in order to improve learning ability of DBTL cycle.