Design of β-sheet forming antimicrobial peptides using deep learning and molecular dynamics simulations

Antimicrobial peptides (AMPs) represent a potent alternative to traditional antibiotics treatments. They are typically short peptides created by the innate immune system that preferentially attach themselves to the cell walls of bacteria and lyse them, through a number of different biophysical mechanisms. Design of such peptides can be slow due to the length of time it takes to assess their properties, whether computationally or experimentally. Machine learning methods have recently emerged as powerful tools for efficient design of biomolecules. In this talk, I will discuss our efforts to instantiate a computational workflow for active learning design of AMPs through search space design and molecular dynamics assessment of points in the search space. To investigate deep learning for instantiation of continuous search spaces, we created different AMP search spaces produced by different deep learning model architectures. We assessed reconstruction accuracy, generative capability, and model interpretability and demonstrate that while most models are able to create a partitioning in their latent spaces into regions of low and high AMP sampling probability, they do so in different manners. In this way we demonstrated several benchmarks that must be considered for such models and suggest that for optimization of search space properties, an ensemble methodology is most appropriate for design of new AMPs. To assess qualities of points in the latent space, we focused on a synthetic beta-sheet forming antimicrobial peptide, GL13K, and performed molecular dynamics simulations of it and several variants. We show that in the presence of negatively-charged bacterial membranes, it undergoes aggregation followed by conformational rearrangement into larger disordered clumps of smaller clusters. Overall, our work provides an important step towards rational and efficient design of AMPs to combat growing antimicrobial resistance.