Reducing the vicissitudes of heterologous prochiral substrate catalysis by alcohol dehydrogenases through machine learning algorithms

Abstract

Alcohol dehydrogenases (ADHs) encompass three distinctive superfamilies: medium-chain dehydrogenase (MDRs), short-chain dehydrogenase (SDRs), and iron-containing alcohol dehydrogenases. MDRs and SDRs have pivotal roles in converting pro-pharmaceutical ketones into chiral alcohols, serving as active pharmaceutical intermediate catalysts in industries. Nevertheless, the pursuit of aligning enzymes to a one enzyme-one substrate paradigm is intricate and resource-intensive. This study undertook the catalysis of 284 antibacterial ketone intermediates, yielding distinct patterns for both MDRs and SDRs. We endeavored to leverage machine learning to discriminate between compounds subjected to the influence of these superfamilies. To this end, we curated a dataset comprising 33 features, encompassing 4 descriptors for each compound. Subsequently, an ensemble of machine learning techniques, including Partial Least Squares (PLS) regression, k-Nearest Neighbors (kNN) regression, and Support Vector Machine (SVM) regression, was harnessed. Furthermore, the assimilation of Principal Component Analysis (PCA) augmented precision and accuracy, refining compound classification and intensifying differentiation and demarcation within diverse compound classes. As such, this classification strategy emerged as a potent approach for discerning substrates amenable to diverse alcohol dehydrogenases, thereby mitigating the reliance on high-throughput screening in identifying the optimal enzyme for specific substrate preferences.