Multi-objective machine learning for designing enzyme cofactor specificity

Many current approaches in chemical engineering and biotechnology involve the construction of synthetic metabolic pathways and networks tailored to produce a desired product. This can often disrupt the natural regulation of cellular reducing power by the over or under consumption of electron-rich cofactors. To achieve optimal flux through these combinatorial pathways, many enzymes need to adopt a cofactor that differs from their native specificity in metabolism. The engineering task of cofactor switching has been attempted before, but has resulted in unintended consequences to an enzyme's catalytic efficiency. Previous methods have also been tedious, often relied on structural knowledge, or have required high-throughput screens. Here, we have developed a machine learning-guided approach to switch cofactor specificity in Rossmann folds—the largest known cofactor binding fold across protein evolution. We applied the model to alcohol-forming fatty acyl reductases (FARs) family to engineer the multimeric, multiple Rossmann fold-containing fatty acyl-CoA reductase (ACR) of M. aquaeolei to favor NADH over its natural preference for NADPH. Our approach leverages known structural information from known enzyme and cofactor interactions as well as protein sequence evolution data to infer high order contacts for cofactor and enzyme specificity. This combination of supervised and unsupervised machine learning models also allows for a sequence-based query, where structural inputs may not be necessary for the model to learn higher order interactions within a given protein's sequence. Using multi-objective optimization, we will optimize for variants with re-engineered cofactor specificity and the retaining of native catalytic activity. We will also expand this model to an increased pool of cofactors to allow for orthogonal cofactors in metabolic pathway design.