Metalorian

Heavy metal contamination is a persistent environmental problem, and one remediation strategy is to build short peptides that chelate a specific toxic metal. Natural metalloproteins such as metallothioneins bind metals well but are awkward to repurpose, being large and hard to retune for specificity. Computational alternatives are also constrained: metal-binding site predictors typically require a resolved structure, and existing design pipelines lean on fixed scaffolds or directed evolution, which narrow the accessible sequence space.

Metalorian, released as a bioRxiv preprint in July 2025 by the Chatterjee lab at the University of Pennsylvania with collaborators at Duke University and Duke-NUS Medical School, attacks the problem directly in sequence space. It is a conditional diffusion model that generates peptides de novo in the latent space of protein language model embeddings, steered toward a user-specified metal class by MetaLATTE, a companion multi-label classifier built on ESM-2. Staying in sequence space matters here: metal-binding peptides are cysteine- and histidine-rich and largely disordered without a bound metal, a regime where structure predictors return low-confidence models. The approach complements sequence-level metalloprotein generators such as sm_protgpt2, which fine-tunes a language model on synthetic metalloprotein sequences rather than conditioning generation on a target ion.

#Key Features

• Metal-conditioned generation: Sampling is conditioned on a target class drawn from the 14 transition and heavy metals MetaLATTE covers, with peptides demonstrated for copper, zinc, cadmium, cobalt, and nickel.
• Coupled continuous and discrete diffusion: A continuous diffusion process over ESM-2 embeddings runs alongside a multinomial diffusion over metal labels, with the two branches conditioning on each other at every denoising timestep.
• Controllable length: A mask-based mechanism restricts generation to a chosen window, typically 30–80 residues, producing peptides of substantially lower molecular weight than natural metalloproteins.
• Two sampling regimes: Progressive verification sampling handles well-represented metals, while gradient-guided classifier sampling with dynamic scaling covers classes with sparse training data.
• Paired wet-lab readout: A SUMO-fusion expression and cleavage workflow feeds a metal-coated sandwich ELISA, giving quantitative binding measurements without biopanning or structural characterization.

#Technical Details

MetaLATTE fine-tunes ESM-2-650M — unfreezing the last two layers and adding attention pooling with rotary position embeddings — on metal-binding proteins from the MbPA database, with non-binders from Mpbipred as negatives. Training runs in two stages: class-balanced focal, F1, and reconstruction losses, then triplet contrastive learning against BLOSUM62- and alanine-substituted negatives at annotated binding sites. It reaches AUROC of 0.86–0.99 across the 14 metal classes, with recall of 0.55 and F1 of 0.57, exceeding an XGBoost baseline on ESM-2 embeddings, and it separates Cu-, Cd-, and dual-binding metallothionein isoforms from Helix pomatia sharing 75.4% sequence identity.

Metalorian adapts the CoDi co-evolving diffusion framework on top of those latents. The continuous branch denoises 1280-dimensional ESM-2 embeddings with the last ten layers unfrozen; the discrete branch runs multinomial diffusion over a 15-way label vector (14 metals plus non-binding) using a TabularUnet backbone. The two are trained jointly with diffusion losses plus contrastive triplet terms that penalize mismatched sequence–label pairs, using AdamW at a learning rate of 2 × 10⁻⁴ and batch size 140 on seven A100 GPUs. Generated peptides show elevated cysteine, histidine, and phenylalanine content relative to wild-type references. Molecular dynamics with MM/PBSA decomposition gave comparable or lower backbone RMSD and radius of gyration than wild-type controls, with stronger electrostatic interaction energies. In ELISA assays, the designed binder MTLrn_Cu_2 bound copper at low-nanomolar concentrations, outperforming a characterized wild-type copper metalloprotein, and MTLrn_Zn_1 bound zinc at mid-nanomolar concentrations.

#Applications

The immediate target is environmental bioremediation: designing short, cheaply expressed chelators for specific toxic metals in contaminated water or soil, where low molecular weight is itself an advantage for chelation activity. The same machinery suits metal-selective biosensors, affinity reagents for metal capture and recovery, and engineered metalloenzyme scaffolds. Because the conditioning signal is simply a class vector, the pipeline can be retargeted to additional metals as annotated binding data accumulates, and the ELISA workflow gives wet-lab teams a low-overhead way to triage candidates.

#Impact

Metalorian is one of the few generative peptide design efforts aimed at metal chelation rather than protein or small-molecule binding, and pairing classifier-guided latent diffusion with in vitro nanomolar Cu and Zn binders shows that continuous protein embedding spaces can support function-conditioned design. The scope is narrow: generation was demonstrated for five metals, only copper and zinc were tested experimentally, training used single-ion binding proteins so multi-metal specificity is not addressed, and rare classes such as lead and vanadium remain thinly represented in training and held-out data. The work is a preprint that has not been peer reviewed and is released under a CC-BY-NC-ND license; the HuggingFace repository designated in the paper as the home for Metalorian's weights and code hosts documentation and figures only, so the trained model is not yet runnable outside the authors' group.