Examining peptide–gold nanoparticle interactions through explainable machine learning

Peptides are promising tools for controlled nanomaterial creation. Peptides are essential for directing the process by which nanostructures develop as well as for affecting the final characteristics of the resultant nanomaterials. It is expected that integrating machine learning (ML) into the biomimetic workflow will speed up peptide discovery, increase resource efficiency, and reveal relationships between characteristics that might be helpful for peptide design. However, the need for interpretability in ML models is crucial, as it fosters deeper understanding about the underlying factors driving the predictions, thereby ensuring technical soundness and replicability. This work develops an explainable binary machine learning classifier using rough sets as the algorithm and amino acid composition as the features. The model was trained and evaluated on a dataset composed of 1720 peptides and their experimentally determined binding affinity for gold nanoparticles. The created model demonstrates reliable accuracy in distinguishing strong-binding peptides from weak binders. Specifically, the model identified sequence patterns important for peptide binding to gold, such as the residues categorized as tiny, small, and aromatic are greater than 15% in terms of composition. Similarly, strong binders were found to possess residues categorized as aliphatic to fall within 5–35%. Additionally, nonpolar amino acids were found to be greater than 15% in all instances. Taken together, using an explainable algorithm and intuitive features yielded an ML model that has provided straightforward insights regarding the composition of strong-binding peptides. This is valuable especially as rational peptide design for bionanotechnology applications remains a challenging endeavor.