Pep2Mol

Pep2Mol is a diffusion-based generative model that designs 3D small molecules to inhibit protein-protein interactions (PPIs). PPIs govern most cellular processes, and small molecules that disrupt dysregulated interactions offer a route to targets long considered undruggable. Conventional structure-based drug design, however, is tuned for well-defined, deep small-molecule pockets and generalizes poorly to the large, shallow, and chemically complex interfaces that characterize PPI sites. Pep2Mol addresses this gap by generating molecules directly against orthosteric PPI sites.

The model's central idea is to condition generation not only on the receptor surface but on the natural binding partner. By explicitly incorporating a binding peptide or protein as structural guidance, Pep2Mol learns from how a known partner engages the interface and uses that information to place a small molecule at the same orthosteric site. This moves beyond the pocket-only conditioning used by most structure-based generators, which see the receptor but not the competitive binder that defines the functional interface.

Pep2Mol was introduced in June 2026 by Rongting Yue, Zekun Yang, Gustavo Seabra, Chenglong Li, and Yanjun Li at the University of Florida, spanning the Department of Medicinal Chemistry and the Department of Computer and Information Science and Engineering. It is described in a bioRxiv preprint awaiting peer review.

#Key Features

• PPI-interface targeting: Pep2Mol generates molecules against orthosteric PPI sites rather than classical deep pockets, addressing the large, shallow interfaces where conventional structure-based generators struggle.
• Binding-partner guidance: The model conditions on a binding peptide or protein as structural context, using the natural partner's engagement of the interface to steer molecule placement instead of relying on the receptor pocket alone.
• Dual SE(3)-equivariant encoders: Two SE(3)-equivariant graph neural networks separately encode protein-ligand and protein-peptide interactions, preserving the rotational and translational symmetry of the 3D binding site.
• Attention-based fusion: The two interaction representations are combined through attention-based conditioning that jointly guides the diffusion trajectory, integrating ligand- and peptide-side information at each step.
• State-of-the-art binding affinities: In the authors' evaluations, Pep2Mol produces chemically valid ligands with leading predicted binding affinities for the targeted interfaces.

#Technical Details

Pep2Mol is a denoising diffusion model for 3D molecule generation. Its architecture integrates two SE(3)-equivariant graph neural networks: one encodes protein-ligand interactions and the other encodes protein-peptide interactions. These representations are fused via attention-based conditioning that jointly guides the diffusion trajectory, so each denoising step reflects both the receptor-ligand geometry and the natural binder's engagement of the interface. To support model development and benchmarking, the authors curated a dataset of 10,956 experimentally resolved protein complex structure pairs, each pairing an orthosteric competitive ligand with a protein binder at overlapping receptor interfaces. Across the reported evaluations, generated molecules are chemically valid and achieve state-of-the-art binding affinities relative to the compared methods.

#Applications

Pep2Mol is aimed at computational chemists and drug-discovery teams pursuing PPI targets, where a validated peptide or protein binder exists but a tractable small-molecule starting point does not. By conditioning on that binder, the model generates candidate orthosteric inhibitors directly at the interface, supporting early hit generation against targets that classical pocket-based methods handle poorly. This is particularly relevant for the substantial class of disease-associated PPIs historically labeled undruggable.

#Impact

Pep2Mol reframes structure-based design for PPI interfaces by treating the natural binding partner as a source of structural guidance, an alternative to the pocket-only conditioning common in 3D molecule generators. Its curated set of 10,956 paired protein complex structures also provides a benchmark resource for the emerging problem of generating PPI inhibitors. As a recent preprint, its results await peer review and experimental validation, and no code or pretrained weights have been released. Pep2Mol is distinct from the similarly named Peptide2Mol, which converts peptide binders into peptidomimetics within a single pocket; Pep2Mol instead designs molecules at competitive PPI interfaces using paired complex structures.