Predicting Scale-Up of Microbial Production Performance from 96-well plates to Benchtop Fermentors

Amyris is a renewable chemicals company that engineers microbes to produce commodity chemicals in industrial-scale fermentations. Using both rational and random (mutagenesis) approaches, we create genetic diversity in our microbial strains and screen tens of thousands of distinct genotypes each week in high-throughput 96-well plates. One of the challenges of adapting microbes to produce a chemical at high flux and titer are fitness perturbations resulting from redox and cofactor imbalances and product-cell association. It is therefore necessary to select not just for product titer, but also microbial strain health and genetic stability, resulting in almost a dozen primary measurements with different variance profiles per strain. A second challenge is to accurately map strain performance in plates to performance in 0.5L or larger bioreactors.  To address these challenges, we have developed multivariate and machine learning algorithms that improve our ability to predict performance at the next screening tier from R2=0.5 to R2>0.9. We generate roughly 20,000 data points daily and make weekly decisions on which of thousands of strains to subsequently test in nine-day fermentor runs. Because the fermentor runs are relatively costly in terms of staffing and equipment, it is crucial to have an optimized predictive model. Our data analysis pipeline represents a robust automated method for rapidly analyzing large volumes of biological data, with which we have demonstrated improved hit-picking rates, better process control, and decreased false positive rates.