ProDualNet

Inverse folding models take a backbone structure and propose sequences that fold into it. ProteinMPNN and its successors do this well for a single target, the right framing when a binder has one job. Much of therapeutic protein design does not fit it. Tirzepatide activates both the GIP and GLP-1 receptors; bispecific antibodies engage two antigens; allosteric regulators must accommodate two conformational states of the same partner. Satisfying two structural contexts with one sequence is a different optimization, and running a single-target designer twice does not solve it.

ProDualNet, from Shanghai Jiao Tong University, conditions sequence design on two receptor structures at once. It appeared as a bioRxiv preprint in March 2025 and in Briefings in Bioinformatics in July 2025. The architecture is a shared encoder–decoder over heterogeneous graphs built from both complexes, with evolutionary context from ESM-2. Its central practical problem is data: experimentally determined dual-target complexes number in the hundreds, against tens of thousands of single-target structures.

The answer is a two-stage recipe: pretrain on abundant single-target structures, then fine-tune on a mixture of real dual-target pairs and synthetic pairs manufactured by perturbing single-target structures with Gaussian noise. The resulting checkpoint applies directly to new receptor pairs, and the likelihood it assigns a sequence ranks experimentally measured dual-agonist potency without any task-specific training.

#Key Features

• Two structures, one decoder: Four heterogeneous graph layers, each pairing a message-passing block with a transformer block, encode both target complexes into a shared space; per-position features from the two conformations are mixed with a fixed weight before autoregressive decoding.
• NoiseMix fine-tuning: 466 real dual-target clusters are combined each epoch with 4,000 noise-augmented single-target proteins, lifting sequence recovery from 0.546 to 0.581 relative to fine-tuning on real pairs alone.
• Zero-shot functional ranking: Sequence likelihood under the trained model correlates with measured dual-agonist EC50 and with mutational fitness, so one checkpoint both generates and scores candidates.
• Coverage of hard dual-target geometries: Separate test sets isolate targets that change conformation on binding and targets whose two interfaces do not overlap.
• Released checkpoint with worked cases: The produalnet_02.pt weights ship with notebooks reproducing the GLP-1/GCGR dual-agonist study and the mutation-effect evaluation.

#Technical Details

Pretraining uses 20,137 protein clusters from the PDB at a 30% identity cutoff, with 1,662 clusters each for validation and test, under a cross-entropy objective that also weights neighbouring positions at α = 0.25. Fine-tuning runs 20,000 steps at a learning rate of 1e-5 on the NoiseMix mixture. ESM-2 features enter through a recycling scheme in which target sequence features begin at zero and are refined from the structure branch; one pass was optimal, and an ablation without ESM-2 still reached 0.54 recovery, so the language model supplements rather than carries the geometric signal.

On the primary test set of 159 dual-target pairs, ProDualNet reaches 0.581 sequence recovery, 0.627 interface recovery, and 3.9 perplexity, against 0.517, 0.541, and 4.8 for a multi-state ProteinMPNN baseline. The margin holds on 52 pairs with binding-induced conformational change (0.584 / 0.613) and 38 pairs with non-overlapping interfaces (0.556 / 0.571). Refolding the designs with AlphaFold 3 gives a mean ipTM of 0.728, PAE 8.65, and pLDDT 86.52. Two case studies quantify design yield: for ubiquitin thioesterase OTULIN, 37 of 100 designs met the success criteria versus 20 for the baseline, and for the 624-residue erbB-2 receptor, 6 of 100 versus none. In zero-shot scoring, likelihoods over 46 GLP-1/GCGR dual agonists of 29 residues correlate with combined potency at Spearman 0.78, ahead of an AlphaFold 3 structural metric at 0.70; on ProteinGym stability the model reaches 0.67 across 65 proteins, and 0.46 over 1,877 single-point and 675 multi-point mutations.

#Applications

The most direct use is designing dual agonists and bispecific binders where two receptor structures are known — incretin-receptor peptides for metabolic disease being the motivating example — and the OTULIN and erbB-2 cases show the method extends from short peptides to protein-scale scaffolds. The zero-shot scoring path is useful independently of generation: a group holding a library of dual-target candidates can rank them by likelihood before synthesis, and the mutation-effect results extend that score to affinity maturation across two partners at once.

#Impact

ProDualNet makes multi-state design a first-class training objective rather than a post-hoc filter over single-target samples, and its noise-augmentation strategy is a transferable answer to the scarcity of multi-state structural data. The evaluation is entirely computational — recovery rates, refolding metrics, and retrospective correlation with published assay data — with no wet-lab validation of a designed sequence, and the authors note in the repository that the mean-pooling fusion they ship is not always the best choice, a cross-attention variant performing better in some agonist design cases. The model covers natural amino acids only, limiting its reach into the modified-residue chemistry common in peptide drugs. Reproducibility is the sharper constraint: the repository ships the model definition and the evaluation notebooks but no training code at all, so the pretrain-then-NoiseMix recipe exists only as prose and cannot be rerun on a new corpus. Weights are distributed through a cloud drive link rather than a versioned artifact registry, and the repository carries no licence file, only a README statement prohibiting commercial use.