Flow matching model for de novo 3D peptide design that converges peptide position before conformation, mirroring the physical order of docking.
No providers recorded yet. Browse all providers
THFlow is a flow matching generative model for de novo 3D peptide design, developed by Dengdeng Huang and Shikui Tu at Shanghai Jiao Tong University and posted in February 2025. Given a target pocket, it generates a full-atom peptide — sequence, placement, orientation, and side-chain torsions — docked into it.
Peptide design models generally treat the peptide as a bundle of modalities and let them converge at once: position, rigid-body rotation, side-chain torsion, and amino acid identity are all denoised in parallel toward the target distribution. The paper's observation is that this is physically incoherent at the start of generation. Early in the trajectory the peptide sits far from the pocket and no interaction field exists, so any gradient applied to rotation or torsion optimizes a conformation against nothing. The authors name this the multimodal temporal inconsistency problem and argue it is a direct cause of the weak binding affinity seen in generated peptides.
The fix is a temporal hierarchy. Position is driven by a time-varying flow that front-loads its movement, pulling the peptide into the pocket first; the other modalities advance at a constant rate and are refined once an interaction field is actually present. The trajectory ends up resembling physical docking rather than a simultaneous collapse of all four modalities, which places THFlow alongside flow matching approaches to protein backbones such as ReQFlow — the same framework, applied to the coupling between modalities rather than to the rotation representation.
The peptide is represented per residue as a position, a rotation in SO(3), four side-chain torsion angles, and an amino acid type. Conditional flow matching normally uses linear interpolation between noise and data; THFlow replaces this for the position modality with a polynomial schedule whose coefficients still sum to one, reducing to the linear case at exponent one and using exponent two in practice. Two MLPs encode residue features and residue-pair relationships, and the vector field is learned by stacked invariant point attention blocks followed by transformer encoders. Sampling integrates the learned ODE with Euler steps, each modality's update depending on the current state of the others.
Training uses the dataset assembled for PepFlow from PepBDB and Q-BioLip: duplicates removed, a 4 A resolution threshold applied, and peptide length restricted to 3-25 residues, giving 10,348 complexes with 166 held out for testing. Training ran 65,000 iterations on a single Tesla P40 at batch size 12 under Adam. On sequence-structure co-design, evaluated by Rosetta energies, 12.33% of generated peptides show better stability than the reference complex and 17.81% better affinity, against 4.16% and 13.19% for PepFlow, 3.05% and 10.68% for PPFlow, and 5.34% and 12.21% for DiffPP, with novelty at 84.94% and diversity 0.651. Ablations isolate the contributions: removing the interaction features drops affinity to 11.11%, and removing the sequence-structure balance weight as well drops stability to 4.32%.
The target use is therapeutic peptide discovery, where a structure of the target pocket exists and the question is what short binder to make. Because generation is conditioned on the pocket rather than on a known binder, it applies to targets without a peptide starting point, and the co-design mode returns a sequence and a docked full-atom structure together — what a synthesis decision needs. The re-docking and side-chain packing modes make the same model useful for refining existing leads: hold the sequence and ask where it binds, or hold the backbone and ask how the side chains pack. Campaigns against protein-protein interfaces, where small molecules struggle and peptides are the natural modality, are the clearest fit.
THFlow's contribution is conceptual as much as empirical: it identifies a specific way the standard multimodal generative recipe violates the physics it is meant to imitate, and shows that correcting the relative timing of modalities more than doubles the fraction of designs that beat their reference on stability. That framing generalizes to any co-design problem where one modality must be resolved before another becomes meaningful. Two limitations bound the result. The preprint links no code repository or released weights, so the model cannot be run outside the authors' group, and all reported gains are in-silico Rosetta energies and structural similarity metrics, with no experimental binding data.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.