iPPIGAN

QED, the standard quantitative estimate of drug-likeness, was fitted to approved oral drugs, and by its scoring a working protein-protein interaction inhibitor looks like a bad molecule: heavier, greasier and more aromatic than the rules reward. The inhibitors in iPPI-DB average a QED of 0.43. Filter a molecular generator on QED and it discards the very chemistry that disrupts an interface.

iPPIGAN answers that in two moves. It builds its training corpus with QEPPI, a drug-likeness index refitted to compounds reported to act at PPI interfaces. And it represents each compound in three dimensions rather than as a string, because the features that characterise PPI inhibitors are molecular shape and aromatic bonding. A molecule is voxelised into shape and pharmacophore channels; a generative adversarial network over 3D convolutional features turns that shape plus a latent code into new molecular representations; and a captioning network — a 3D CNN feeding an LSTM — decodes each back into a SMILES string. Generation is seeded: supply the 3D characterisation of one compound, or one peptide, and vary the latent code to expand it into a library.

The work was led from Yonsei University and the Bioinformatics and Molecular Design Research Center in Incheon, with collaborators at Shanghai Jiao Tong University, Soongsil University, China University of Petroleum and the University of Turbat, and published in Briefings in Bioinformatics in July 2022. Its code descends from Skalic et al.'s shape-based generator LigDream. The same group's later GENiPPI takes the opposite tack, conditioning on the protein complex interface itself; iPPIGAN has no protein input at all.

#Key Features

• Seed-driven generation from released weights: Training and generation are separate stages, and the repository ships the trained encoder, decoder and generator at 120,000 iterations, so a new seed compound is an input rather than a reason to retrain.
• QEPPI-filtered training corpus: The training set is a ChEMBL subset built around PPI-targeted drug-likeness rather than conventional drug-likeness, which is what steers output toward interface chemistry.
• Voxel and pharmacophore encoding: Each conformer is written into 1 Å cubic voxels over a 24 Å box in five channels, with point-like pharmacophore features beside the atoms carrying them.
• Peptide seeds as well as ligand seeds: Because the input is a 3D characterisation rather than a molecular graph, a binding peptide can seed small-molecule mimics.

#Technical Details

ChEMBL 29 supplied 2,078,892 molecules; standardisation and filtering — an atom-type whitelist, 10 to 35 heavy atoms, QEPPI above 0.5, and no ring assembly of more than eight smallest rings — left 933,559 as the training set. Conformers came from RDKit optimised with MMFF94, randomly rotated and translated before voxelisation. The adversarial and captioning networks were trained jointly for 120,000 iterations at batch size 128 with Adam, at learning rates of 10⁻⁴ and 10⁻²; the run took roughly seven days on one NVIDIA GeForce RTX 3090.

Benchmarked on MOSES over 300,000 preprocessed leads and sampling 30,000 SMILES in each of five independent runs, iPPIGAN reached 0.989 ± 0.005 validity, 1.000 uniqueness at 1k, 0.999 at 10k and 0.990 novelty, with Fréchet ChemNet distances of 5.879 against the test set and 6.171 against the scaffold-split test set — lower than the CharRNN, VAE and LatentGAN baselines. The generated library also scores higher on QED and QEPPI and lower on synthetic accessibility, and a UMAP projection of MACCS fingerprints shows it overlapping iPPI-DB inhibitors while extending into unoccupied regions between them.

#Applications

The paper works two design campaigns end to end. Seeding with the p53 peptide produced 10,000 molecules; a LightGBM regressor trained on 2,920 MDM2-p53 bioactivity records predicted activity, a QEPPI above 0.5 and synthetic accessibility below 4 left 243 candidates, and semi-flexible DOCK6.8 against the 1T4F structure returned 10 that beat the clinical reference idasanutlin on both docking score and QEPPI at the Phe19, Trp23 and Leu26 pocket. The Bcl-2 campaign seeded with obatoclax, drew on 3,217 bioactivity records, and left 44 candidates of which 15 outscored obatoclax against 6O0K. That is the shape of the tool: take one peptide or one known binder, expand it into a focused library, and triage on predicted activity, PPI drug-likeness and docking.

#Impact

iPPIGAN is, by its authors' account, the first deep generative model applied to de novo design of PPI inhibitors, and the first to use QEPPI to evaluate a generative model rather than merely to filter a library. Its limits are stated in the paper's own future work: no information about the protein-protein complex enters the model, so PPI targeting is a property of the training corpus and the seed rather than of a modelled interface — the gap GENiPPI went on to close. All evaluation is computational, with no synthesis or binding assay. Source code, the ChEMBL identifier list behind the training set, and the trained model files sit in one repository, which makes the pipeline reproducible and a starting point for later PPI-focused generative work.