PepFlow

A fifteen-residue peptide contains as many as 377 atoms and 1,131 degrees of freedom, and unlike a folded domain it has no single native structure worth predicting: its function depends on the range of shapes it visits. Single-state predictors answer a question peptides do not really pose, and molecular dynamics, which does sample that range, spends hours of wall time per sequence.

PepFlow attacks that gap directly: a transferable generative model that draws all-atom conformations for an arbitrary input sequence. It is trained as a score-based diffusion model over atomic coordinates, then run as the equivalent probability-flow ODE. That equivalence is load-bearing — a deterministic flow admits exact likelihood computation, which lets the model be trained by energy, minimising the divergence between its proposal distribution and the Boltzmann distribution of a molecular force field rather than fitting observed structures alone.

Generalised all-atom modelling is normally prohibitive because the couplings between degrees of freedom differ for every sequence. PepFlow's answer is a hypernetwork: an attention stack reads the sequence and emits a subset of the dynamics network's own weights, so the model specialises to each peptide in a forward pass instead of growing or refitting the network. Osama Abdin and Philip M. Kim developed it at the University of Toronto, releasing a preprint in 2023 and the peer-reviewed version in Nature Machine Intelligence in 2024. It shares its name with an unrelated peptide binder design model.

#Key Features

• All-atom generation in three stages: separate score networks produce backbone atoms and side-chain centroids, then side-chain heavy atoms, then hydrogens, so the joint likelihood factorises cleanly across the three.
• Sequence-specific weights from a hypernetwork: an attention-based network predicts parameters for a subset of the dynamics layers, adding expressivity for longer-range interactions without enlarging the network that runs at sampling time.
• Training by energy: the flow is optimised against an Amber ff99SB-ILDN force field with the GB-Neck2 implicit solvent model through OpenMM, which is what shifts it from a structure predictor toward an ensemble generator.
• Chain-aware noising: the backbone network denoises each atom relative to its N-terminal neighbour, so corruption degrades a peptide toward an ideal chain rather than a diffuse cloud.
• Constrained sampling for macrocycles: a Metropolis search in latent space finds conformations meeting a target bond distance, giving head-to-tail and side-chain cyclised peptides from a model trained only on linear ones.

#Technical Details

The dynamics networks are built from E(3)-equivariant graph neural network layers; the hypernetwork is three attention blocks with an embedding dimension of 128. Training runs in three phases: protein fragments from the PDB — structures deposited before September 2021, filtered at 3 Å resolution, clustered at 30% identity with MMseqs2 and cut into fragments of length 3 to 15, yielding roughly 40 million unique fragment sequences — then fine-tuning on solved peptide structures including NMR models, then training by energy on molecular dynamics ensembles from DBAASP peptides. Held-out test sets hold 167 linear and 46 cyclic peptides, and a thousand energy-training iterations on one NVIDIA A100 took about three days.

For single-structure prediction, 100 samples are clustered and the centroid of the largest cluster is the answer; this performs comparably to AlphaFold 2 and ESMFold and better in many cases, with failure modes largely orthogonal to AlphaFold 2's. On ensembles, 100 samples cover 22.8% of the conformations from a 400 ns simulation at 0.15 Å Cα RMSD per residue. Across 75 short linear motifs from the Protein Ensemble Database, 46.3% of generated conformations fall within the experimental ensembles and 46.5% of experimental conformations are recovered, while explicit-solvent simulation took 4.90 times the wall time of drawing a thousand PepFlow conformations.

#Applications

Peptide docking pipelines typically begin from a library of candidate conformations, and generating that library from sequence changes what is feasible to screen. Beyond docking, the model lets structural biologists characterise short linear motifs and other disordered segments where crystallography offers no single answer, and lets peptide therapeutic programmes assess how readily a sequence adopts a target interface conformation. Because side chains and protons are modelled explicitly, constrained sampling extends to macrocyclisation and stapling, modifications routinely used to improve peptide stability and permeability.

#Impact

PepFlow showed that a generalised Boltzmann generator can explicitly model every degree of freedom in a peptide and still transfer across sequences, where earlier Boltzmann generators handled one molecular system at a time. Its limitations are stated plainly by its authors: energy training uses implicit rather than explicit solvent, the model samples free peptides rather than ensembles bound to a partner protein, ranking relies on a clustering heuristic that leaves AlphaFold 2 more accurate on average for head-to-tail cyclic peptides, and lasso peptides — whose folds are enzymatically catalysed — are a clear failure case. Code is MIT-licensed, with pretrained parameters and processed training features distributed from the authors' own server.