Lumen’s AI Evaluation Frees Manufacturing Optimization from DOE Limits

By finishing up multivariate knowledge evaluation with machine studying (ML), Lumen Bioscience improved its therapeutic protein manufacturing methodology between 70 and 100%, whereas needing solely a small fraction of the runs that might have been needed below the standard “design of experiment (DOE)” strategy for a similar variety of variables, in accordance with firm officers.

The problem was to optimize the excessive progress price of spirulina, the photosynthetic microbe Lumen makes use of to provide recombinant proteins for its vaccines and therapeutics, whereas concurrently reaching excessive expression of the proteins.

To optimize these elements, the corporate collaborated with Google Accelerated Science (GAS) to analyze the relationships amongst manufacturing outcomes and 17 environmental variables, utilizing 96 photobioreactors. By designing an adaptive, iterative mannequin, the corporate was capable of run solely 245 configurations versus the 130,000 that might have been wanted for a two-level, full factorial DOE strategy.

That mannequin allowed assets to be shifted to extra promising areas and fashions to be refined as new knowledge grew to become accessible. “This adaptability helped de-risk our research design part, permitting for higher flexibility than many DOE-based research,” says Caitlin Gamble, PhD, director of informatics at Lumen Bioscience.

Optimum parameters

Moreover, ML revealed optimum parameters that doubtless wouldn’t have been recognized with DOE, based mostly on current preconceptions. The optimum temperature vary (which fell in a slender band between typical center and excessive factors) and the optimum pH vary (which fell beneath these reported within the literature) are two examples.

To attain such outcomes required a major inflow of correct knowledge, notes Gamble, “so we targeted on a high-throughput fluorescent protein assay for our readout. Making use of ML additionally required cautious consideration of tips on how to construction our extra complicated, light-schedule parameters and reward perform.

“ML has a method of arriving at options that you could be not have anticipated, so it was vital for us to completely contemplate whether or not optimization of our reward perform would result in sensible options. The strategy was comparatively strong to commentary noise and non-linear interactions, and it allowed us to discover a fancy subspace with gentle ramping and cyclic lighting schedules.

“Between weeks 5 and 15, we found a number of bioreactor setting configurations that roughly doubled productiveness. Through the course of, we added numerous controls and normalizations that we are going to proceed to use to assist guarantee noticed outcomes come up from outlined–fairly than hidden–variables.”

Lumen continues to discover a number of media parameters, however Gamble says it hasn’t wanted to introduce any main modifications to the pressure background. The group at present is introducing genetic and environmental variables for additional optimization.