Loading…
Structure-based drug design model building 3D ligands atom by atom inside a protein pocket, with explicit hydrogens and fragment completion.
No providers recorded yet. Browse all providers
A sentence has a first word; a molecule has no first atom. Autoregressive models that build 3D molecules one atom at a time have worked around this by imposing a canonical atom ordering or designating a "focus atom" — conventions that bake an arbitrary sequence into every next-atom prediction. NEAT-POCKET removes the convention: it treats the partial molecule as an unordered set of atoms and trains the model to reproduce the entire boundary of that set — every 1-hop neighbour that would be a legitimate continuation — rather than one blessed successor, a training signal the authors call neighborhood guidance.
That mechanism came from NEAT, an unconditional 3D generator from the same group. NEAT-POCKET is its pocket-conditioned extension, released in September 2026 by Roxane Axel Jacob, Daniel Rose, Thierry Langer and Johannes Kirchmair at the University of Vienna. The conditioning is what makes it a design tool rather than a molecular sampler. NEAT-POCKET encodes the pocket as an all-atom point cloud and feeds it into the generator through cross-attention and adaptive layer normalization, so every atom is placed with the surrounding protein in view.
The practical argument for the autoregressive route is cost. Transport-based generators such as TargetDiff and DrugFlow re-evaluate an equivariant graph network over the whole ligand–pocket system at every denoising step, so the pocket is processed hundreds of times before one molecule emerges. NEAT-POCKET encodes the context once per added atom.
The best configuration pairs a 12-layer, 12-head set transformer of hidden dimension 768 with a 1–4–1 layer pocket encoder of the same width and an MLP flow-matching head carrying six adaptive layer-norm blocks at hidden dimension 1536 — 242M parameters, with a guidance factor of 0.5 and a clash penalty weight of 4.0. Training starts from a NEAT checkpoint pretrained for 1,000 epochs on GEOM-Drugs, then fine-tunes separately on CrossDocked (94,981 training pairs) and SPINDR (35,333 pairs), yielding two released checkpoints. Bonds are not generated directly: a separate predictor classifies edges of a 2.5 Å radius graph, and an integer linear program resolves them under atom-valence constraints.
On the 100 CrossDocked test pockets, NEAT-POCKET reaches 69.2% PoseBusters validity, the lowest clash count among the evaluated baselines at 6.6 protein–ligand clashes, and the closest aggregate physicochemical match to the training distribution (rank 2.48), with generations averaging 289 Da. It samples 100 molecules per pocket in about 4 seconds against 88 seconds for the next-fastest baseline. On the crystal-structure-derived SPINDR benchmark it records 80.3% validity and 3.7 clashes, sampling roughly nine times faster than FLOWR. Conditioning on the largest BRICS-derived fragment of a reference ligand lifts validity to 81.8% and 88.8% respectively, while structural uniqueness falls from 92.8% to 42.4% — the expected trade-off as the prefix constrains accessible chemistry.
The intended workflow is early-stage structure-based design: point the model at a crystallographic pocket and sample candidate ligands, or hand it a validated binding motif and let it elaborate. The released tutorial runs from a PDB entry and a reference ligand that defines the site, extracting all protein atoms within 6 Å. The fragment mode is the one most directly useful to medicinal chemists — in the paper's PRMT5:MEP50 case study the supplied prefix keeps its crystallographic pose atom-for-atom while new substituents reach into adjacent subpockets, which is what scaffold decoration asks for.
NEAT-POCKET makes a concrete case that autoregressive construction, long treated as the weaker alternative to diffusion and flow matching for 3D molecule generation, is competitive once the atom-ordering problem is handled properly — and that it buys an order-of-magnitude reduction in sampling cost. The accompanying benchmark, which re-evaluates Pocket2Mol, TargetDiff, DiffSBDD, DrugFlow and FLOWR under one hydrogenation and scoring pipeline, is a contribution in its own right, since inconsistent hydrogen handling has made published numbers here hard to compare. The authors are candid about the limits: sequential generation can propagate errors into larger compounds, CrossDocked's cross-docked poses carry geometric artifacts a model trained on them may reproduce, pockets are rigid, and waters are not modeled. The work is a preprint awaiting peer review.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.