PepFlow

What holds a peptide onto its receptor is mostly side-chain chemistry: a salt bridge from a lysine, a hydrophobic tuck of a leucine, a hydrogen bond donated by a serine hydroxyl. Yet most generative protein design models emit only the four backbone heavy atoms and hand the sequence off to a separate inverse-folding step, so the atoms that actually make the contacts are never present while the binder is being generated, and the sequence is chosen after the shape it is supposed to explain.

PepFlow generates the whole peptide at once. Every residue is represented as four coupled quantities — a position in R³, an orientation in SO(3), five torsion angles, and an amino acid identity — and each one is given its own conditional flow on the manifold it naturally lives on: Euclidean flow for the position, a geodesic flow on SO(3) for the frame, a toric flow on the 5-torus for the angles, and a simplex flow, constructed in logit space with a logistic-normal prior, for the residue type. A single network regresses all four vector fields jointly, conditioned on the target receptor, and Euler integration of the resulting ODE produces a full-atom bound-state peptide with its sequence already determined.

The model comes from Helixon Research, the Institute for AI Industry Research at Tsinghua University, and UIUC, and was published at ICML 2024. It shares a senior-author lineage with PepHAR, which attacks the same target-conditioned problem from an anchor-and-extend direction rather than a flow.

#Key Features

• One flow per geometric modality: Position, orientation, torsion and residue type each get an analytically derived flow on their own manifold, so the periodicity of dihedrals and the curvature of the rotation group are respected rather than approximated in Euclidean space.
• Receptor-conditioned generation: The binding pocket is encoded once by a geometric equivariant encoder and fed into every decoding step, so designs are shaped by the target rather than sampled unconditionally and docked afterwards.
• Three tasks from one set of weights: Holding some modalities at their ground-truth values and sampling the rest turns the same trained model into a fix-backbone sequence designer (freeze positions and orientations) or a side-chain packer (freeze everything but torsions).
• Full-atom output: Side-chain angles are generated alongside the backbone, so the result is a complete bound-state structure whose interface contacts can be scored directly.

#Technical Details

The encoder produces hidden residue representations and pair embeddings for the receptor; a time-dependent decoder built mainly on invariant point attention takes the interpolated peptide state and the timestep and predicts the recovered peptide, from which the four vector fields are reparameterized. Training minimizes the weighted sum of the four conditional flow-matching objectives plus auxiliary backbone and torsion losses. The benchmark was assembled from PepBDB and Q-BioLiP: duplicates removed, resolution held under 4 Å, peptide length restricted to 3–25 residues, and complexes clustered by mmseqs2 at 40% peptide sequence identity, giving 8,365 non-orphan complexes across 292 clusters, with 10 clusters holding 158 complexes reserved for testing. Sixty-four peptides are generated per target for evaluation. On sequence–structure co-design the full-atom variant reaches 2.07 Å Cα RMSD, 86.89% binding-site ratio and 21.37% affinity, against 4.17 Å, 26.71% and 16.53% for RFdiffusion with sequences from ProteinMPNN; the baselines retain the edge on Rosetta stability and on ESMFold-scored designability. On fix-backbone sequence design PepFlow recovers 56.40% of native residues versus 53.28% for ProteinMPNN and 43.51% for ESM-IF, and its side-chain packing is correct within 20° for 62.79% of angles.

#Applications

The immediate use is therapeutic peptide binder design: given a receptor structure and a chosen pocket, generate candidates whose sequences and side-chain conformations are already consistent with the interface they must occupy, then triage them by Rosetta energy before synthesis. The partial-sampling modes cover two adjacent jobs a peptide modeler usually reaches for separate tools to do — redesigning a sequence onto a fixed backbone, and repacking side chains onto a known complex. The authors note the same construction should carry to other full-atom design problems such as antibody and enzyme design.

#Impact

PepFlow's contribution outlasted its leaderboard position: its curated PepBDB/Q-BioLiP assembly and its in-silico metric suite became shared evaluation ground for target-conditioned peptide design. THFlow trains on the dataset PepFlow assembled, GPFlow adapts its architecture for the peptide co-design setting, and it appears as a standing baseline in work such as CPSDE and PepMirror. The authors are candid about two limits: the deterministic ODE sampler produces less diverse ensembles than a stochastic one, and the model offers no mechanism for property-guided generation, so optimizing toward a target property means filtering rather than steering. The two checkpoints and the packaged dataset are distributed through a Google Drive folder rather than a versioned model hub, and the repository notes that its generation and evaluation pipelines remain rough.