PepHAR

Not every residue in a peptide binder does the same work. A small number sit deepest in the target's pocket and carry most of the interaction — the hot spots — while the remainder scaffold those anchors into the right relative geometry. Generative models that emit every residue in one shot treat the two classes identically, spending the same capacity on a linker as on the residue that makes the contact. A second problem compounds it: peptide bonds are partially double-bonded and planar, so adjacent residues cannot be placed independently, yet representing a backbone as free local frames invites exactly that.

PepHAR is organized around both constraints. It generates a peptide in three stages. A founding stage samples a handful of hot-spot residues on the target surface from a learned residue density. An extension stage grows a fragment outward from each hot spot one residue at a time, predicting the dihedral angles that connect the new residue to its neighbor rather than its absolute coordinates, so peptide-bond geometry holds by construction. A correction stage then reconciles the independently grown fragments into a single valid peptide by descending the gradient of a learned objective — optimizing the peptide's own coordinates, frames, dihedrals and residue types while the trained networks stay frozen.

The model was developed by researchers at Tsinghua University, the University of Washington, MIT, the University of Illinois Urbana-Champaign, ByteDance and Helixon, and published as an ICLR 2025 conference paper.

#Key Features

• Hot spots first: An energy-based density model over residue type, backbone position and orientation is sampled with Langevin dynamics on the SE(3) manifold, placing one to three anchor residues where the target is most likely to accept a contact.
• Dihedral-angle extension: Each fragment is extended to the left or the right by sampling backbone dihedrals from a product of von Mises distributions, a circular analogue of the Gaussian that keeps the angular geometry of the peptide bond intact.
• Fragment assembly by gradient: Independently grown fragments rarely meet cleanly. The correction objective combines a backbone self-consistency term with the learned dihedral likelihood, updating the sample until the seams close.
• Scaffold generation as a task: Alongside de novo design, the paper introduces a benchmark supplying known hot-spot residues that the model must link into a complete peptide — closer to how peptide leads are optimized than designing from scratch.

#Technical Details

Both networks use an Invariant Point Attention backbone. The density model is trained by noise contrastive estimation, distinguishing native peptide residues from spatially perturbed negatives; the prediction network is trained by maximum likelihood over the four von Mises parameters per direction, with directional attention masks that prevent a residue from seeing the side it is about to be extended toward. Training used 8,207 non-homologous peptide-target complexes drawn from PepBDB and Q-BioLip, with a held-out benchmark of 158 complexes spanning 10 mmseqs2 clusters.

On de novo peptide design with three hot spots, PepHAR reaches 84.91% secondary structure ratio, 86.74% binding site rate and 2.68 Å Cα RMSD against native peptides, alongside 79.11% novelty. Raising the hot-spot count from one to three improves geometry and energetics while reducing novelty and diversity — a trade-off between designs close to the native and designs that explore. On the scaffold generation benchmark with three given hot spots it reaches 2.15 Å RMSD, 83.02% SSR and 88.02% BSR, the strongest SSR and BSR among the compared methods, which include RFdiffusion with sequences recovered by ProteinMPNN, ProteinGenerator, PepFlow and PepGLAD. Ablations show the sampled hot spots matter most for validity, RMSD, SSR and BSR, while removing the von Mises parameterization or the correction stage degrades affinity and stability.

#Applications

Peptides occupy a useful middle ground between small molecules and antibodies, with roughly 80 peptide drugs marketed and hundreds more in development. PepHAR addresses both settings a peptide program encounters: proposing a binder from nothing but a target pocket, and completing a peptide around interaction residues already known from a structure, an alanine scan or a prior lead. Generation is conditioned on the target and the released checkpoints run without retraining, so a new target requires only its pocket structure.

#Impact

PepHAR argues that peptide-bond geometry and the uneven importance of residues are design constraints worth building into the generative process rather than penalties applied afterward, and its scaffold generation setting gives the field a benchmark matching how peptide programs actually proceed. Limitations are real: evaluation is entirely in silico, with success measured by AlphaFold 2-predicted ipTM rather than binding assays, and no designed peptide has been experimentally characterized. The authors report that both networks overfit the training set, so early training checkpoints were selected for evaluation. Code is MIT-licensed on GitHub, but the trained weights ship through a Google Drive folder under no stated license, and the sampling scripts carry absolute paths from the authors' cluster as defaults.