Machine Learning-Guided Directed Evolution of Candida antarctica Lipase B for the Sustainable Synthesis of Atorvastatin Intermediate: A Combined Computational and Experimental Strategy

Abstract

The pharmaceutical industry faces increasing pressure to adopt green chemistry principles in active pharmaceutical ingredient (API) manufacturing. Atorvastatin, a widely prescribed statin for cholesterol management, has a complex chiral synthesis involving the reduction of a β-keto ester intermediate. Here, we report the development of a machine learning-guided directed evolution platform for engineering Candida antarctica Lipase B (CALB) variants with dramatically improved activity for the enantioselective synthesis of the atorvastatin side-chain precursor, (3R,5R)-tert-butyl 7-[2-(4-fluorophenyl)-5-isopropyl-3-phenyl-4-(phenylcarbamoyl)-1H-pyrrol-1-yl]-3,5-dihydroxyheptanoate. We trained a graph neural network (GNN) on a curated dataset of 1,243 CALB variants with measured activities, achieving a predictive R2 of 0.81 on held-out data. Using the trained model, we prioritized 96 candidate mutations for experimental testing, of which 14 were validated as beneficial (>1.5-fold improvement in activity). Iterative rounds of machine learning-driven evolution (three rounds) yielded the variant CALB-M8, which contains 8 mutations (L144V, S160A, V210I, A225V, I233L, L278V, A282T, T293S) and exhibits a 38-fold improvement in catalytic efficiency (k_cat/K_M) and a 99.4% enantiomeric excess (ee) for the target reaction, with substantially enhanced thermostability (T_50 of 78°C vs 56°C for wild-type). Computational analysis revealed that the mutations collectively reshape the active site geometry, expanding the substrate-binding pocket to better accommodate the bulky atorvastatin intermediate. Through the application of CALB-M8 in a green synthesis employing isopropenyl acetate as an acyl donor, >99% conversion and 98% isolated yield of the desired (3R,5R)-isomer was achieved in 8 hours at 40°C, with an E factor of 5.2, which represents a 92% reduction compared to conventional chemical synthesis. The industrialization of this process has successfully resulted in consistent results at a 50 L pilot scale, demonstrating the potential for practical use of the engineered enzyme in sustainable pharmaceutical manufacturing.