Conversion of renewable and low-cost materials, like plant-derived biomass, has become an increasingly important research area for sustainable biofuel and chemical production. Clostridium thermocellum is a thermophilic anaerobic bacterium that has risen in popularity for biofuel production because of its innate capability to both degrade cellulose and produce ethanol. As a complementary strategy to metabolic and process engineering approaches, using protein engineering to engineer key enzymes in C. thermocellum could enhance bioethanol production by this organism. This study focuses on applying a novel computational and experimental workflow that incorporates molecular dynamics (MD) simulation and machine learning (ML) to engineer cofactor specificity of an alcohol dehydrogenase (ADH) in C. thermocellum. CtADH natively uses NADH as a cofactor to interconvert between aldehydes/ketones and alcohols, but competing pathways that also use NADH reduce cofactor availability. An alternate cofactor is NADPH, which is already present in C. thermocellum and is otherwise underutilized in competing pathways. Altering cofactor specificity from NADH to NADPH is expected to alleviate stoichiometric limitations of NADH and lead to elevated ethanol production. Although experimental and rational engineering approaches exist for engineering ADH cofactor specificity, these methods do not explicitly take protein dynamics into account. Including dynamics information that provides mechanistic insight could allow us to design improved variants that might otherwise be overlooked, and provide rationale for certain mutation outcomes. Here, MD simulations are used to add dynamic and biophysical information, as well as to augment ML training data, by modeling molecular behavior at an atomistic level. Taking advantage of small amounts of existing experimental data, which map sequence to function, and incorporating MD trajectories, which map sequence to biophysical features related to function, allows us to efficiently train ML models that can uncover relationships between mutation and function to predict improved variants. Interestingly, neither NADH nor NADPH are explicitly modeled in our simulations; nevertheless, insight into key biophysical features related to cofactor specificity are extracted from this methodology. This generalizable MD-informed workflow can be applied to other protein engineering applications, such as improving binding affinity, thermal stability, or pH stability, regardless of whether the relevant mechanisms of action are previously understood.
