A thorough understanding of computational models of evolution is necessary to face the rising obstacles of antibiotic resistance, viral immune evasion, protein engineering, and the impact of dynamic evolutionary pressures. After validating our sequence evolution with epistatic contributions (SEEC) model’s propensity to generate diverse proteins with in vivo functionality, we continued to investigate the varying input parameters and their influence on variant performance. Here, we rationally direct enzyme evolution utilizing SEEC toward specified functionality. Following numerous fine-tuned simulations, we present a methodology that reliably outputs well-characterized proteins derived from multiple statistical inference methods, environmental pressures, and various lengths of evolutionary time all guided by epistasis inferred from extant protein families. These experiments investigate β-lactamase variants and their efficacy against not only native antibiotics (ampicillin) but also novel combinations of aztreonam and cefotaxime, aiming to explore the potential for evolved extended-spectrum resistance. Additionally, we are developing tools to visualize these paths of sequence evolution using our latent generative landscape (LGL), which combines a Potts Hamiltonian model with unsupervised variational autoencoder learning to create a latent manifold of sequence diversity. Furthermore, we seek to introduce an innovative process of computational evolution that employs both the generative landscape itself followed by sequence optimization using SEEC. Altogether, these methodologies present valuable tools by which we can continue to explore protein evolution and engineering.
Alvarez et al. (Sun,) studied this question.