Plixer

Pocket-conditioned generative models for drug design all run into the same wall: there are only a few thousand non-redundant protein–ligand complexes in the Protein Data Bank, which is nowhere near enough data to teach a model both where atoms should sit inside a binding site and what counts as a chemically valid molecule. Plixer, from Jude Wells and Brooks Paige at University College London, addresses this by splitting those two problems apart and training them on different data at very different scales.

The model works in two stages. A 3D convolutional U-Net treats ligand design as a voxel inpainting problem: given the atomic density grid of an empty pocket, it predicts the density grid of the ligand that belongs there. A separate encoder–decoder network then translates that soft density into SMILES strings. Because the decoder only ever sees voxelized ligands, it can be pretrained on millions of conformers from ZINC20 rather than the small supply of co-crystal structures — so chemical grammar is learned from a corpus orders of magnitude larger than the structural one, and only the spatial reasoning stage depends on complex data.

A useful side effect of the design is that Plixer is not only a generator. Because the decoder defines an explicit conditional likelihood over SMILES given a pocket-derived voxel grid, any arbitrary molecule can be scored against a target, which most 3D generative models cannot do. The work was released as a preprint in July 2025 and accepted to the Generative AI for Biology workshop at ICML 2025.

#Key Features

• Decoupled spatial and chemical learning: The pocket-to-ligand and voxel-to-SMILES stages are trained independently, letting ligand-only corpora compensate for the scarcity of experimentally resolved protein–ligand complexes.
• Likelihood-based ranking: The SMILES decoder assigns a calibrated log-likelihood to any candidate compound, so the same model can both generate binders and screen existing libraries far faster than docking.
• One-shot generation from a PDB file: Inference takes a receptor structure and a pocket location and emits SMILES directly, with no per-target retraining, fine-tuning, or sampling loop.
• Multiple chemotypes per pocket: Sampling repeatedly from the decoder for a single voxel hypothesis yields several chemically valid molecules, accommodating pockets that admit diverse scaffolds.
• Released checkpoints: Both trained components are published on the Hugging Face Hub.

#Technical Details

The first stage, PocVox2MolVox, is a 3D U-Net that maps a 4-channel protein grid (carbon, oxygen, nitrogen, sulfur) to a 9-channel ligand grid adding chlorine, fluorine, iodine, bromine, and an "other" channel. Both grids are 24 Å cubes at 0.75 Å resolution, giving 32³ voxels, with atoms rendered as Gaussians scaled by van der Waals radius; training uses equally weighted binary cross-entropy and Dice loss, plus random rotations and translations of up to 6 Å. The second stage, Vox2Smiles, is a Vision Transformer encoder of 8 layers and 8 attention heads over 4×4×4 voxel patches, feeding a GPT-2-style autoregressive decoder with a 76-token SMILES vocabulary. The inpainting network is trained on HiQBind, a filtered set of PDB complexes; the decoder is pretrained on ZINC20 conformers and fine-tuned on a mixture of 85% ZINC20 and 15% filtered PocVox2MolVox outputs.

Evaluation uses chronological, PLINDER pocket-cluster, and 30% sequence-identity test splits. On the strict PLINDER split Plixer reaches a mean AutoDock Vina score of -7.52 against -7.51 for Pocket2Mol and -6.58 for DrugGen, and raises the rate of generated molecules resembling the true binder by a factor of 5.5 over the decoy background, where both baselines score zero or near-zero enrichment. As a ranker it reaches 0.67 ROC-AUC on the chronological split and 0.58 on PLINDER, ahead of DrugGen but behind AutoDock Vina's 0.62. Pocket2Mol remains ahead on drug-likeness, LogP range, and diversity.

#Applications

Plixer suits early-stage hit finding against a structurally characterized target, and particularly targets with no known actives to seed a ligand-based approach. Two workflows are supported directly: generating novel SMILES for a pocket, and scoring a purchasable or in-house library by likelihood as a fast pre-filter ahead of docking. Because inference needs only a PDB file and a pocket centre, it slots into a computational chemistry pipeline without target-specific data.

#Impact

Plixer is evidence that voxel CNNs with heavy augmentation remain competitive with SE(3)-equivariant graph diffusion approaches such as Apo2Mol and FLOWR.root, and that the binding constraint on this problem is data rather than architecture. The likelihood-scoring capability is the more distinctive contribution, since it lets one model serve both generation and screening. The limits are stated plainly: soft densities are less interpretable than explicit coordinates, the deterministic generator can average incompatible chemotypes into non-physical densities, the receptor is treated as rigid, and all reported results are computational — no generated molecule has been experimentally assayed. The preprint has not yet been peer reviewed.