Machine learning and Design of Experiments: Alternative approaches or complementary methodologies for quality improvement?

Machine Learning (ML), or the ability of self-learning computer algorithms to autonomously structure and interpret data, is a methodological approach to solve complicated optimization problems based on abundant data. ML is recently gaining momentum as algorithmic applications, computing potency, and available data sets increased manifold over the past two decades, providing an information-rich environment in which human reasoning can partially be replaced by computer reasoning. In this paper, we want to assess the implications of ML for Design of Experiments (DoE), a statistical methodology widely used in Quality Management for quantifying effects and interactions of factors with influence on the production quality or the process yield. We specifically want to assess the future role and importance of DoE: Will it remain unaltered by ML, will it be made obsolete, or will it be reinforced? With this, we want to contribute to the discussion of the future use of traditional Quality Management methodologies in production, as our ML assessment can in principle be applied to other statistical methodologies as well. While we are convinced that ML will heavily impact the field of Quality Management and its predominant set of statistical methodologies, we find reason to expect that this impact will be a mutual one. As this is the first paper addressing the joint force potential of the two methodologies ML and DoE, we expect a range of follow-up papers being written on the subject and a spark in specialized applications addressing DoE's ML-enhanced vital functionality for process improvements.     

Engineering a Highly Active Sucrose Isomerase for Enhanced Product Specificity by Using a “Battleship” Strategy

The sucrose isomerase SmuA from Serratia plymuthica efficiently catalyses the isomerisation of sucrose into isomaltulose, an artificial sweetener used in the food industry. However, the formation of a hygroscopic by-product, trehalulose, necessitates additional separation to obtain a crystalline product. Therefore, we have improved the product specificity of SmuA by first introducing a few exploratory amino acid exchanges around the active site and investigating their influence. Then, we devised a second set of mutations, either at promising positions from the preceding cycle, but with a different side chain, or at alternative positions in the vicinity. After seven iterative cycles involving just 55 point mutations, we obtained the triple mutant Y219L/D398G/V465E which showed 2.3 times less trehalulose production but still had high catalytic efficiency (k_cat/K_M=11.8 mM⁻¹ s⁻¹). Not only does this mutant SmuA appear attractive as an industrial biocatalyst, but our semirational protein-engineering strategy, which resembles the battleship board game, should be of interest for other challenging enzyme optimization endeavours.