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

Abstract

Heavy metal contamination threatens food production, ecosystems, and human health, creating a need for chelators that remove target contaminants without depleting essential metals. Existing chelators often lack this selectivity, while natural metal-binding proteins remain difficult to repurpose because of their size, folding complexity, and metal promiscuity. To address these limitations, we introduce Metalorian, a conditional diffusion model that co-evolves protein language model representations and metal labels to generate de novo metal-binding peptides. We guide generation with MetaLATTE, a multi-label sequence classifier trained across 14 metal classes. MetaLATTE achieves class-specific AUCROC values ranging from 0.86 to 0.99, enabling Metalorian to generate compact Cu-, Cd-, Zn-, and Ni-binding peptides enriched in metal-coordinating residues. Isothermal titration calorimetry confirmed target-metal binding for selected designs, with buffer-corrected K_D values of 19 nM to 0.96 μM and a 68-fold Zn-over-Ni preference for Zn-directed peptides. Finally, in an independent screen, ten chemically distinct Ni-directed designs displaced Ni from the PAR metal indicator; all ten showed stronger apparent competition for Ni over Co and nine did so relative to Zn. Together, these results connect conditional peptide generation with experimental affinity measurements and cross-metal competition screening, supporting a strategy for designing selective peptide chelators for environmental remediation and metal recovery.

Figure 1: In silico workflow

Figure 2: In vitro workflow

Code, checkpoints, and data

The computational release associated with Zenodo record 22960654 is available in this repository under Computational/.

See the computational README for setup and usage instructions.

The archived computational release and experimental raw data are available at Zenodo.