Motion as a Language: Transformer-Based Classification of Antimicrobial Peptide Conformational Dynamics

Antimicrobial peptides (AMPs) represent a promising alternative to traditional antibiotics against which many bacteria are rapidly gaining resistance. Today, databases containing tens of thousands of AMPs, along with their properties and biological activities, can be screened to select lead candidates for a given application. The conformational plasticity of AMPs has been proven crucial for the recognition of their targets. However, the volume, complexity, and recalcitrance to classification of conformational data, obtained from e.g. molecular dynamics (MD) simulations, prevent it from being included in databases, let alone used as a criterion for the screening of AMPs. This work applies the transformer neural network architecture (which powers large language models such as ChatGPT) to the detection of temporal and spatial context in time series of AMP conformations from MD simulations. It shows how the representation of AMP conformational space learned by the network can be leveraged for the unsupervised classification of AMP plasticity, which can subsequently be used alongside conventional properties for the screening of databases. Thus, it reveals how deep learning can pave the way toward restoring conformational dynamics to its legitimate importance within drug design pipelines.