PocketXMol

Nearly every deep learning tool that touches a protein pocket is built for one job. Docking programs place a fixed ligand; structure-based drug design models grow molecules into a cavity; peptide binder pipelines chain a backbone generator to a sequence designer. Each uses its own representation — amino acid residues here, molecular graphs there — and none of the learned physics transfers between them, even though all of these problems are governed by the same atomic contacts.

PocketXMol collapses that fragmentation into a single model. Developed by researchers at Tsinghua University and Peking University with collaborators at the University of Illinois Urbana-Champaign, it represents every molecule — small molecule, peptide, or protein pocket — as a bare set of atoms and bonds, with no residue-level abstraction. What the user wants is expressed not by choosing a different model but by setting a task prompt: a group of binary indicators marking which atom coordinates, atom types, and bond types are fixed and which are to be generated. Docking, fragment linking, PROTAC design, and de novo cyclic peptide design all become different settings of the same prompt on the same weights.

First posted as a preprint in October 2024, the work was published in Cell in 2026. Across 13 generative tasks, PocketXMol was compared against 55 baselines using 51 metrics, leading on 11 of the 13 without any per-task fine-tuning.

#Key Features

• Task prompt mechanism: A binary prompt over atoms and bonds specifies what is fixed and what is generated, so one checkpoint covers structure prediction, design, and mixtures of the two.
• Exclusive atom-level representation: Peptides are generated atom by atom rather than residue by residue, which is what lets small-molecule and peptide training data reinforce one another.
• Non-Markovian denoiser: The network infers the noise type and scale from the noisy input and predicts the clean molecule directly, rather than the previous state on a diffusion trajectory, allowing per-atom noise schedules.
• Confidence-scored sampling: Each generated pose carries a self-confidence score used to rank the 100 samples drawn per pocket; a separately tuned ranker improves selection further.
• Experimental validation: Designed caspase-9 inhibitors suppressed PARP and caspase-3 cleavage, and 382 designed PD-L1 peptides yielded 76 binders at 10⁻⁷ M and 15 at 10⁻⁸ M.

#Technical Details

The backbone is a geometric neural network acting as a universal denoiser over atom types, atom coordinates, and bond types, with basic noise components drawn from Gaussian distributions on coordinates and translations, categorical distributions on atom and bond types, isotropic Gaussians on SO(3) for rotations, and circular normal distributions on torsion angles. Training assembled roughly two million 3D structures from PDBBind, Binding MOAD, CrossDocked2020, PepBDB, AlphaFold DB, ChEMBL, ZINC, GEOM-Drug, and CREMP, with test-set overlaps removed. The loss is a weighted sum of atom-type, coordinate, bond-type, pairwise-distance, dihedral, and confidence terms, optimized with AdamW for 180,000 steps on eight 80 GB A100 GPUs, about 54 hours.

On the 428-complex PoseBusters benchmark, 82.5% of top-ranked poses fall under 2 Å RMSD with self-ranking and 83.4% with the tuned ranker; requiring PoseBusters chemical and geometric validity as well gives 78.5% and 79.4%. On the 308-target PoseBusters V2 set it reaches 84.7%, comparable to the 84.4% reported for AlphaFold 3 with defined pockets. Oracle selection among the generated poses would reach 96.5%, showing that ranking, not generation, is the binding constraint. On peptide design against the PepBDB test set, sequence recovery is 37.2% versus 32.1% for an RFdiffusion plus ProteinMPNN pipeline.

#Applications

The single checkpoint serves small-molecule and peptide docking, molecular conformation generation, structure-based drug design, 3D molecule generation, fragment linking and growing, molecule optimization, PROTAC design, linear and cyclic peptide design, and peptide inverse folding. Medicinal chemists can start from a fragment and grow it, or from a hit and optimize it by resampling at a lower noise scale that preserves the input scaffold. Peptide engineers get binder backbones and sequences from one model, including designs with nonstandard amino acids. Three Colab notebooks cover docking, peptide design, and small-molecule design without local setup.

#Impact

PocketXMol is a demonstration that the transferability arguments behind language-model pretraining apply to atomic structure as well: a shared atom-level substrate lets peptide tasks benefit from small-molecule data and vice versa. Code is MIT-licensed, and model weights alongside both raw and processed training data are deposited on Zenodo under CC-BY-4.0 — an unusually complete release for a model of this scope. The main limitation is rigidity: the model places ligand atoms strictly according to the input pocket coordinates and does not relax them, so inaccurate or apo pockets degrade results. The authors address this with PocketXMol-PF, a variant fine-tuned on perturbed pockets for docking against predicted or apo structures.