DARPinMPNN

Point a general inverse-folding model at a fixed DARPin backbone and it will pack the target-contact positions with glycine. Measured zero-shot and unconstrained, glycine reaches 40.6% of the β positions — the very residues the DARPin design rules exclude — because a small, flexible side chain is the cheapest way to satisfy a rigid backbone. The damage then hides: 25 of the 33 residues in each internal repeat are invariant framework, so the designed monomer still folds cleanly (ESMFold pLDDT above 85, backbone Cα-RMSD 0.31 Å from the native scaffold) even after the paratope has been destroyed. The authors name the first failure the "Glycine Trap" and the second scaffold buffering.

DARPinMPNN, developed at Weill Cornell Medicine, removes both by refusing to let the generative model leave the validated DARPin design space. Its sequence generator is LigandMPNN — the atom-context successor to ProteinMPNN — fine-tuned once on roughly 50,000 ESMFold-predicted DARPin monomers and then frozen and reused across every target. At decode time a strict logit-bias mask penalizes any residue outside the position-specific pools of the DARPin consensus ruleset, the "Greek Grammar", by a factor of 10⁹, so grammar compliance is guaranteed rather than encouraged. Constraining generation this way eliminated the Glycine Trap without measurably distorting the interface backbone.

The second half of the framework is a validation hierarchy rather than a single scorer. A chimeric multiple sequence alignment lets AlphaFold-Multimer serve as a high-throughput sieve over synthetic sequences that have no evolutionary relatives, and AlphaFold 3 then re-scores survivors from raw sequence alone. The two predictors disagreed on 150 of 153 candidates, which is the paper's most transferable result: a single structure predictor is not enough to validate a designed interface.

#Key Features

• Grammar-constrained decoding: A continuous logit-bias mask restricts each of the eight variable positions per repeat to its curated amino-acid pool, so every sampled sequence is a legal DARPin rather than a backbone-satisfying artifact.
• Scaffold-specialized fine-tune: Training on DARPin monomers alone, with a 3× loss weight on the variable positions and the first encoder layers frozen, pushes the network's capacity toward interface chemistry instead of framework reconstruction.
• Chimeric alignment for evolutionary orphans: The receptor is given a deep homolog stack while the designed binder is padded with gaps, avoiding both the single-sequence collapse and the co-evolutionary "Misfolding Trap" a paired MSA induces.
• Independent structural oracles: AlphaFold-Multimer sees the chimeric alignment and AlphaFold 3 does not, so agreement between them is genuine corroboration rather than two runs of the same evidence.
• Per-target statistical baselines: Interface confidence is read against a noise floor built from thousands of random grammar-compliant designs, which differs sharply by antigen.

#Technical Details

The base checkpoint ligandmpnn_v_32_010_25.pt was fine-tuned with Adam (lr 1×10⁻⁴, batch size 24, 0.3 Å backbone noise) on a 95%/5% split of ESMFold-predicted DARPin structures filtered at pLDDT > 90, reaching its lowest validation loss (0.3782) at the final epoch; the Methods give the schedule as 50 epochs in one place and 100 in another. Inference samples at temperature 0.5. Screening ran 15,000 candidates — 5,000 each against mesothelin (a compact 59-residue domain), the 94-residue VEGF-A monomer, and the 591-residue HER2 extracellular domain — through an ESMFold foldability gate at pLDDT > 80 (2,503 of a 3,500-candidate subset passed, 71.5%), an AlphaFold-Multimer sieve, and 25 AlphaFold 3 predictions per candidate (five seeds × five diffusion samples). Eleven candidates cleared a ranking score of 0.70, a 0.07% pipeline hit rate. The champion mesothelin binder reached an AlphaFold 3 ipTM of 0.83 with a median inter-chain PAE of 4.7 Å, against 0.89 for the crystallographically validated 90 pM G3 control. Hit rates fell with target size — 44% for mesothelin, 20% for VEGF-A, 4% for HER2 — which the authors attribute to attention dilution in large multi-domain complexes.

#Applications

The framework is aimed at groups that want target-specific DARPin candidates without building and screening a physical display library, whose ~10⁸–10¹⁰ transformants cover a vanishing fraction of the 2.13×10²⁰ grammar-compliant sequences a three-repeat DARPin admits. Because the constraint layer is the scaffold's own design ruleset, the same recipe extends to other engineered binding scaffolds with consensus rules — leucine-rich repeat proteins, fibronectin domains, affibodies. The chimeric-alignment trick is independently useful to anyone folding a designed sequence with no natural homologs.

#Impact

DARPinMPNN's most durable contribution is diagnostic: it names and measures three failure modes — the Glycine Trap, scaffold buffering, and confidence inflation from a native homodimer interface — that other scaffold-constrained design efforts will hit. Its own ablation is candid about the neural component's contribution: PSSM-biased sampling from the fine-tuned model gave no measurable advantage over uniform sampling within the grammar, implying the constraints, not the learned priors, carry the result. All findings are in silico; no surface plasmon resonance, calorimetry, or cell-binding data accompany the preprint, the code is available on request rather than released, and the fine-tuned checkpoint is described but never distributed.