Plastics smaller than five millimeters, termed microplastics or nanoplastics (MNPs), are a concerning pollutant. Millions of tons of MNP pollution pervade the environment. Humans and many other organisms unintentionally consume MNPs, which recent work suggests may harm health. Because the impacts of MNP pollution on the environment and human health worsen as MNP concentration increases, it is important to remediate MNP pollution. Here, we discover plastic-binding peptides (PBPs) for multiple plastics (polyethylene, polypropylene, polystyrene, and PET) that could facilitate MNP detection, capture, and degradation. We conducted multiple iterations of PBPs design using a combination of biophysical modeling, optimization techniques, and machine learning. Since there is no experimental data on peptide affinity for any plastic, our first round of peptide design was based solely on biophysical modeling. We used Peptide Binding Design (PepBD), an algorithm that pairs simulated annealing and atomistic force fields to optimize peptide affinity over the space of amino acid sequences and adsorbed conformations to a given plastic. We confirmed with molecular dynamics (MD) simulations that the PepBD designs had high affinity for plastic. This led us to our second round of PBP design, where we trained two ML modules (long short-term memory network, or ESM2 language model plus convolutional neural network) to predict the PepBD data score for arbitrary amino acid sequences, then used a generative ML module (Monte Carlo tree search or generative adversarial network) to discover new peptides. Crucially, this approach allows for optimization of not just peptide affinity for a single plastic, but also affinity for multiple plastics or physiochemical properties like aqueous solubility. Evaluation of the PBPs in MD simulations and single molecule force experiments show that the best ML peptides have higher affinity for their target plastic(s) than the best PepBD peptides. A third round of PBP design was conducted by replacing the noisier PepBD training data with MD training data, then using Gaussian process regression to design even higher affinity peptides. We envision that the PBP designs will be of great help in addressing MNP pollution, and that the design procedure will be useful when experimental data is sparse or non-existent.
