GENiPPI

A protein-protein interaction interface is the wrong shape for most molecular generative models. Where a classical binding site is a 300–1,000 Ų concave pocket, a PPI interface spreads over 1,000–6,000 Ų of shallow, flat, often featureless surface, and compounds that successfully disrupt one tend to break the drug-likeness rules those models were trained to respect — higher molecular weight, higher logP, more rings. A pocket-conditioned generator such as TargetDiff is optimizing against a cavity that does not exist, in a chemical space that is not the one PPI modulators occupy.

GENiPPI makes the interface itself the condition. Graph attention networks encode the atoms and contacts of a protein-protein complex interface; convolutional networks encode candidate compounds in voxel and electron-density space; a conditional Wasserstein GAN emits molecular embeddings that a separate conditional network judges compatible with those interface features; and a captioning network decodes the embeddings into SMILES. Once trained, a new complex enters as an input condition steering sampling from the latent space, rather than requiring the network to be rebuilt.

The work was led from the Department of Integrative Biotechnology at Yonsei University with collaborators at Hunan University and elsewhere. It was posted to bioRxiv in October 2023 — as "An interface-based molecular generative framework for protein-protein interaction inhibitors" — and published in the Journal of Cheminformatics in December 2024. Its code descends from the same group's earlier iPPIGAN, and its interface representation from the Kihara lab's GNN_DOVE.

#Key Features

• Interface as a sampling-time condition: The complex interface is supplied as input to a model trained once over the full dataset, so a new PPI target needs no new training run.
• Double-graph interface encoding: The extracted interface becomes two graphs over the same nodes — one whose edges are covalent bonds within each subunit, one that also carries non-covalent contacts across the interface — read by graph attention networks.
• Voxel and electron-density compound representation: Compounds are encoded in three dimensions by a CNN operating on both occupancy voxels and electronic density, rather than as strings or 2D graphs.
• Explicit condition-matching network: Alongside the generator and discriminator, a third network scores whether a molecular embedding matches the interface features, tying chemistry to target.
• Few-shot adaptation: A target with only a handful of known disruptors can be addressed by fine-tuning on those examples, demonstrated here for Hsp90-Cdc37.

#Technical Details

Training used ten validated PPI drug targets with resolved interfaces — Mdm2/p53, Bcl-2/Bax, BAZ2B/H4, Bcl-xL/Bak, BRD4-1/H4, CREBBP/H4, EphA4/ephrin, Mcl-1/Bax, menin/MLL and XIAP/SMAC — paired with active and inactive compounds from ChEMBL, plus 250,000 randomly drawn ChEMBL compounds as additional inactives. The conditional Wasserstein GAN was trained for 50,000 iterations at batch size 8 with 65 steps per iteration, updating the discriminator every step and the generator every thirtieth, using RMSprop at a learning rate of 1×10⁻⁴ and gradient-penalty and condition-loss weights both set to 10. The conditional network was pre-trained with binary cross-entropy and frozen during adversarial training, and the whole run fits on a single NVIDIA A40.

On the MOSES benchmark, sampling 30,000 SMILES, GENiPPI reaches 0.999 validity, 0.998 uniqueness at 1k, 0.977 at 10k and 0.998 novelty, with Fréchet ChemNet distances of 7.450 against the test set and 7.884 against the scaffold-split test set. The GENiPPI-noninterface ablation matches on validity but is worse on FCD (7.653), and conditioning on MDM2/p53, Bcl-2/Bax and BAZ2B/H4 produces measurably different QED, QEPPI and Fsp³ distributions — the paper's evidence that the condition is doing work.

#Applications

The intended use is early-stage design of PPI modulators for targets where screening conventional libraries has been unproductive. The paper's worked example is the Hsp90-Cdc37 chaperone interaction, an oncology target with only seven published disruptors: fine-tuning on those seven produced compounds whose chemical space overlaps the known disruptors by t-SNE projection, whose pharmacophores match, and which score better in semiflexible DOCK 6.9 docking against the 1US7 complex than the reference triazine DCZ3112. Generate against an interface, filter on QEPPI, dock, triage for synthesis — that is the practical shape of the tool for a medicinal-chemistry group.

#Impact

GENiPPI is, by its authors' account, the first structure-based molecular generative model built around PPI interfaces rather than binding pockets, and it establishes that interface geometry is a usable conditioning signal. The scope limits are stated plainly: the framework has not been tested across a large number of receptor-ligand PPI pairs, so generalization beyond the ten training targets is unestablished, and it conditions on the interface without modelling the three-dimensional ligand-receptor interaction. All evaluation is computational — property distributions, chemical-space overlap and docking scores, with no synthesis or binding assay. Code, processed data and trained model files are released, making it a reproducible starting point for a target class generative chemistry has largely routed around.