Metalorian: De Novo Generation of Heavy Metal-Binding Peptides with Classifier-Guided Diffusion Sampling

Abstract

Heavy metal contamination is a severe and ongoing global problem that demands selective and efficient chelation strategies for environmental remediation. As a solution, we introduce Metalorian, a conditional diffusion model that generates de novo heavy metal-binding peptides, guided by MetaLATTE, a multi-label classifier fine-tuned to include underrepresented metal classes. Leveraging both continuous protein embeddings and discrete metal-binding constraints, Metalorian produces peptides with controllable lengths and user-defined binding specificity while preserving essential physicochemical properties. Crucially, our data augmentation strategy enables robust classification and generation even with sparse training data for certain metals. We demonstrate the effectiveness of our approach through peptides designed for copper (Cu), cadmium (Cd), and zinc (Zn), which retain key features such as net charge and hydrophobicity, while significantly reducing sequence length and molecular weight compared to known metal-binding proteins. Metalorian-generated peptides exhibit stable conformations and favorable binding energetics, as confirmed through molecular dynamics simulations. To validate binding in vitro, we developed a streamlined SUMO-fusion expression and cleavage system coupled with ELISA-based quantification, confirming robust Cu and Zn binding by multiple Metalorian-generated peptides. Overall, our work establishes a foundational platform for engineering heavy metal-binding peptides tailored to diverse and underrepresented targets, and highlights the potential of well-trained continuous latent spaces for diffusion-based de novo peptide design.