De Novo Design of Multiple Microplastic-Binding Peptides with a Protein Language Model-Guided Generative Adversarial Network

Microplastics are heterogeneous pollutants that pose significant risks to ecosystems and human health. Innovative mitigation strategies are urgently needed. Plastic-binding peptides represent a promising eco-friendly approach for detecting or capturing microplastic pollution. Since real-world microplastic pollution consists of multiple types of plastic, it would be particularly useful to have peptides that bind to multiple plastics. However, there are no known peptides with this property. We present a generalizable AI-driven framework for the de novo design of plastic-binding peptides with a high affinity for multiple plastics. The framework integrates a pretrained protein language model (PLM), fine-tuned on biophysical modeling data of peptide adsorption to plastics generated by the PepBD algorithm, that guides peptide design with a generative adversarial network (GAN). The PLM provides appropriate embeddings of peptide physicochemical features that lead to accurate predictions of peptide affinity for a given plastic. The GAN model is trained via a modular split-training strategy to ensure stability, sequence diversity, and the ability to optimize peptide affinity to any desired combination of plastics. We use this framework to design peptides with high affinity for polyethylene, polypropylene, and poly(ethylene terephthalate). Molecular dynamics simulations confirm that the generated peptides exhibit strong multiplastic binding, having average adsorption free energies to the three plastics that are ∼30% more favorable than those of peptides previously designed using biophysical methods. Steered molecular dynamics simulations reveal that one peptide has an exceptionally high affinity for both polyethylene and polypropylene. These findings highlight the potential of AI-driven peptide design for addressing microplastic pollution and broader applications in peptide engineering.