PocketGen

A ligand-binding pocket is two answers that have to agree: which residues line it, and where each of their side-chain atoms sits. Most design pipelines settle those separately — RFdiffusion generates a backbone, then ProteinMPNN or LigandMPNN reads a sequence off it — and nothing forces the returned sequence to be the one whose side chains would have produced that geometry against that ligand. PocketGen closes the seam by predicting residue types and full-atom coordinates in the same iterative refinement, conditioned on the target small molecule and on the surrounding protein scaffold, which is held fixed.

PocketGen was developed by Zaixi Zhang and Qi Liu at the University of Science and Technology of China with Wan Xiang Shen and Marinka Zitnik at Harvard Medical School. It was first posted in February 2024 as "PocketGen: Generating Full-Atom Ligand-Binding Protein Pockets" — the title under which much of the earlier literature cites it — and published in Nature Machine Intelligence in November 2024 as "Efficient generation of protein pockets with PocketGen".

It is a redesign model rather than a de novo backbone generator. Given an existing protein–ligand complex, it rebuilds the residues nearest the ligand — about eight of them at the default 3.5 Å cutoff — leaving the rest of the fold intact. That matches how protein engineers work on binding sites: mutate a few contact residues, keep the scaffold that holds the protein together.

#Key Features

• Sequence and structure co-design: One model emits the pocket residue sequence and the coordinates of all its atoms, backbone and side chain, so the two modalities are consistent by construction rather than by post-hoc filtering.
• Bilevel graph transformer: The protein–ligand complex is encoded as a graph of blocks, and bilevel attention operates simultaneously over atoms and over residue/ligand units, covering intra-protein and protein–ligand interactions at both granularities.
• Structural adapter on a frozen protein language model: Cross-attention adapter layers are inserted into ESM-2 so that evolutionary sequence priors inform the design. Only the adapter trains, giving PocketGen 7.9M trainable parameters against 82.9M for RFdiffusion All-Atom.
• The ligand pose moves too: Ligand coordinates update alongside the pocket during refinement, so the model can account for a redesigned pocket shifting how the molecule sits.
• Bulk sampling is cheap: 100 pockets for one ligand take 44.2 seconds, against 1,633 seconds for RFdiffusion, 2,210 for RFdiffusion All-Atom, and over 1,000 for physics-based PocketOptimizer.

#Technical Details

PocketGen trains on two benchmarks under a shared split: 100k CrossDocked protein–ligand pairs, filtered to poses under 1 Å RMSD and clustered at 30% sequence identity, and 40k Binding MOAD pairs split by Enzyme Commission class. Training runs 5k iterations with Adam, batch size 64, about 48 hours on a single A100, and generation uses three refinement rounds. The default language model is ESM-2 650M; swapping in larger ESM-2 checkpoints raises amino acid recovery from 54.58% to 66.61% between the 35M and 15B models, following a logarithmic scaling trend. On CrossDocked the published model reaches 63.40% amino acid recovery, a top-1 Vina score of −9.655 kcal/mol and a 97% success rate — the fraction of targets whose generated pocket beats the reference on affinity — versus 93% for RFdiffusion All-Atom, the strongest baseline. Designability is assessed by refolding designs with ESMFold and AlphaFold2 and measuring self-consistency RMSD and TM-score.

#Applications

The intended use is redesigning an existing binding site around a chosen ligand: tuning an antibody, an enzyme active site, or a small-molecule biosensor. The paper works through five such cases — a cortisol-binding antibody (PDB 8cby), Factor Xa with the anticoagulant apixaban (2p16), a fentanyl-binding protein, the designed binder PiB with rucaparib, and the luciferase luxsit with diphenylterazine. Generated pockets recover most of the non-covalent contacts seen in the crystal structures (13 of 15 for cortisol) and add plausible new ones, with predicted stability changes staying near neutral or positive (ΔΔG from 0.02 to 0.92 kcal/mol by DDMut). Sampling is cheap enough to generate hundreds of candidates per target and rank them, trading affinity against diversity via a temperature knob.

#Impact

PocketGen turned all-atom, ligand-conditioned pocket co-design into one feed-forward procedure instead of a generate-then-inverse-fold pipeline, and its adapter-on-a-frozen-pLM recipe showed pocket design benefits from language model scale without paying to fine-tune it. The code has drawn a few hundred GitHub stars and the model is now a standard baseline in later pocket and binder design work. The evidence remains computational: every affinity number is a docking or MM-GBSA score, structural validity is measured against structure predictors rather than experiment, and no designed pocket in the paper was expressed and assayed. The model also redesigns rather than invents — it needs a starting complex and a scaffold to hold fixed — and accuracy slips as the redesigned region grows past the default radius.