PeptideArena

PeptideArena is a set of de novo designed peptide binder backbones against 12 protein targets, built to decouple sequence recovery from design-success metrics. Because the binders are generated rather than observed, there is no native sequence to recover: sequence-recovery numbers are undefined here, and evaluation relies on structural measures instead.

This repository holds the benchmark's peptide backbone templates — the inputs to an inverse-folding method.

Contents

templates/<TARGET>/len<L>/pick<N>_<TARGET>_len<L>.pdb

Each file is a two-chain complex:

• Chain A — the designed peptide backbone, len08, len16 or len24 residues. All residues are GLY and carry backbone atoms only (N, CA, C, O), so the file specifies a backbone and a pose, never a sequence.
• Chain B — the receptor, renumbered 1..N. Its sequence is identical across every pick and length of a target.

Files are plain ATOM records with a placeholder CRYST1 line, no hydrogens and no SEQRES header. Each complex sits in its own coordinate frame; the receptors of one target superpose to within a few tenths of an Å but are not pre-aligned to a common frame.

Quick use

1. Select a template from templates/<TARGET>/len<L>/.
2. Give the two-chain PDB to an inverse-folding method: design residues for chain A while using chain B as the fixed receptor context.
3. Refold the designed peptide–receptor complex with your structure predictor of choice (e.g., Boltz-2 or OpenDDE).
4. Assess target-aligned peptide self-consistency RMSD: align the predicted and intended complexes on chain B, then calculate chain A backbone RMSD without a second alignment on chain A.

Aligning on the target alone means scRMSD reports consistency of both the peptide backbone and its pose relative to the target, rather than peptide shape in isolation. This is the standard design–refold self-consistency principle, adapted to peptide–protein complexes.

Targets

Twelve targets spanning a diverse set of protein families, each taken from a solved peptide–protein complex and stripped to the receptor chain. The receptor is chain B because BoltzGen's de novo path assigns the designed binder to chain A.

PDB	UniProt	Protein
1FGL	P62937	Peptidyl-prolyl cis-trans isomerase A (PPIA / cyclophilin A)
3JQ5	P60045	Acidic phospholipase A2 3
3N00	P20393	Nuclear receptor subfamily 1 group D member 1 (NR1D1 / Rev-erbα)
3O2M	P45983	MAPK8 (JNK1)
4RRV	O15530	PDPK1 (PDK1)
4Y7R	P61964	WDR5
6CDG	Q8IVV7	GID4
6O33	Q13526	PIN1
6YOO	Q9H0R8	GABARAPL1
7OUN	Q9NZQ7	CD274 (PD-L1)
7YKH	Q9WVQ1	MAGI2 (PDZ0–GK fragment)
9CDT	Q07820	MCL1

How the backbones were made

Generation. Backbones were generated with BoltzGen (Stark et al., 2025; peptide-anything protocol) in de novo mode, conditioned on the receptor structure and a hotspot subset. For each length:target pair, 10 independent configurations were launched, each drawing a fresh subset of 3–6 hotspots constrained so all chosen residues are mutually reachable by a binder of the requested length (every pairwise Cα distance within a length-scaled cutoff). Binder length was fixed per run. This produced on the order of 10³ candidate backbones per length:target pair.

Hotspots. Binding-site residues were assigned per target by an automated literature-and-precedent pipeline that resolves the UniProt accession, retrieves related PDB structures and their literature, and scores every receptor residue by how much precedent supports it as a contact. Filtered sets were compared against the contacts made by the peptide in the original complex as a sanity check on site placement.

Selection. Candidates were clustered and filtered down to the 10 per length:target pair released here, giving 353 structures — 10 per pair except those that did not reach 10 distinct clusters (len08:1FGL with 7 and len16:7OUN with 6).

Evaluation

The steps in Quick use leave several choices open, and different choices give different numbers. This section records the evaluation behind the PeptideArena results in the AtomWeaver preprint, in enough detail to reproduce them.

Refolding. The whole complex was refolded: the structure predictor received the designed peptide sequence, the target protein (receptor) sequence, and the target MSA. The structures of both chains were then predicted. We used OpenDDE as the structure predictor.

Superposition. Each predicted complex was superposed onto its template on the backbone atoms (N, CA, C, O) of the receptor chain, over all receptor residues — not Cα alone, not all heavy atoms, and not an interface subset. The template was held fixed and the prediction was the mobile structure.

scRMSD. The receptor-derived rotation and translation were applied to the predicted peptide, and scRMSD is the peptide backbone self-consistency RMSD (N, CA, C, O) against the template peptide, with no second alignment on the peptide. Sidechains never enter the number, so canonical and non-canonical designs are scored on the same footing.

Residue correspondence. Residues were matched by position along the chain rather than by residue number, since predictors routinely renumber chains from 1. A predicted peptide therefore had to carry the same residue count as its template.

Unscorable designs. A design was unscorable if its predicted peptide had a different residue count from the template, or if any peptide residue was missing one of the four backbone atoms — the common case for non-canonical residues written with ligand-style atom naming, which have no identifiable backbone. Residue names were never compared between prediction and template: the sequence was redesigned, so disagreement is expected. Unscorable designs were counted as failures at every threshold and reported alongside the results, since dropping them silently inflates success counts.

Receptor diagnostic. The receptor backbone RMSD after superposition was reported as well. It separates a genuinely misplaced peptide from a prediction whose receptor was itself mispredicted, which is otherwise invisible in the peptide number.

Thresholds. Two were used: 2 Å as the strict criterion and 5 Å as the lenient one. The looser threshold reflects that peptides are floppier and less rigid than folded domains, so a design may recover the correct binding epitope while retaining conformational inconsistency that a strict cutoff rejects. Neither threshold is calibrated against binding data.

Reporting. Results were reported as raw counts per (target, length) cell — designs below each threshold, with unscorable designs counted as failures — rather than as a single pooled figure, together with the number of designs generated per template and any deduplication of identical sequences, which changes what a count is counting.

Scope and limitations

PeptideArena measures self-consistency, not affinity. A design that refolds to the intended backbone and pose is self-consistent with the template it was designed for. The set carries no binding data, and no experimental validation, so scRMSD pass rates should not be reported as hit rates or compared against measured affinities.

There are no peptide sequences, by construction. The backbones are de novo generated. Chain A is a poly-glycine placeholder carrying only backbone atoms. Sequence-recovery metrics are therefore undefined on PeptideArena.

The receptors are real proteins and may overlap your training data. Each receptor is taken from a solved complex, so a model may have seen it or a close homologue during training. The peptide backbones have no such counterpart, but a strong result on a given target should be read with that target's training overlap in mind. Overlap is specific to each model's corpus and is worth measuring against your own.

Report results stratified by target and length. Pooled results hide axes that was built to vary along: target difficulty and peptide length. We strongly recommend against pooling the results into a single statistic, as it is expected to be noisy.

Citation

To cite only PeptideArena itself:

@software{proteinqure2026peptidearena,
  author    = {Hostallero, David E. and Butterfoss, Glenn and Broom, Aron and Hwang, Sungwon and Layne, Elliot and Kanawaty, Ashlin and Kitaygorodsky, Alexander and Fingerhuth, Mark},
  title     = {{PeptideArena v1.0.0: De-Novo Peptide Inverse Folding}},
  version   = {v1.0.0},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22922294},
  url       = {https://doi.org/10.5281/zenodo.22922294}
}

PeptideArena was introduced in the AtomWeaver preprint:

@article{proteinqure2026atomweaver,
  author    = {Kitaygorodsky, Alexander and Hostallero, David Earl and
               Broom, Aron and Layne, Elliot and Hwang, Sungwon and
               Kanawaty, Ashlin K. and Babej, Tomáš and
               Butterfoss, Glenn L. and Fingerhuth, Mark},
  title     = {{AtomWeaver}: Multi-Component Flow Matching with a Structured
               Geometric Prior Facilitates Non-Canonical Peptide Design},
  journal   = {bioRxiv},
  year      = {2026},
  publisher = {openRxiv}
}