The rapid rise of antimicrobial resistance (AMR) demands alternatives to conventional antibiotics. Antimicrobial peptides (AMPs)—with broad-spectrum activity and multi-target mechanisms—are a viable option. We integrate artificial intelligence (AI), molecular dynamics (MD), and state-of-the-art measurements to accelerate AMP discovery and optimization. Our pipeline couples a generative model (decoder-only transformer) for de novo peptide design with predictive models (encoder-only transformers) that classify antimicrobial activity and predict minimum inhibitory concentrations (MICs) against priority pathogens. We then apply reinforcement learning to optimize potency, selectivity, and solubility in a closed loop. Shortlisted peptides are validated experimentally for MICs and membrane-disruptive mechanisms, providing feedback to the AI models in iterative fashion. To provide physics-based insight, we initially perform high-throughput MD simulations based on the highly mobile membrane-mimetic (HMMM) model—which replaces lipid tails with an organic solvent phase to accelerate peptide-membrane association and insertion. The initial peptide structures were obtained using ESMFold. This way, we are able to observe membrane insertion within hundreds of nanoseconds (sub-μs), yielding actionable mechanistic readouts. The systems are subsequently converted to full atomistic membranes to refine interactions at all-atom resolution. All simulations were run in high throughput on HPC resources.
Mondal et al. (Sun,) studied this question.