PLACER

A docked ligand pose is one snapshot, and for a designed enzyme the question that matters is not where the substrate sits in the best snapshot but whether the catalytic side chains stay put across all of them. An active site whose lysine samples half the pocket between turnovers pays an entropic penalty on every one. Molecular dynamics answers that, but a microsecond trajectory per design does not scale to the hundreds of candidates a campaign produces.

PLACER — Protein-Ligand Atomistic Conformational Ensemble Resolver — answers it by dropping to the atom. The molecular system becomes a graph whose nodes are heavy atoms and whose edges are chemical bonds, with no residue-level tokens anywhere; coordinates are corrupted with Gaussian noise and the network trained to denoise them back to the deposited positions. Because the backbone is supplied rather than predicted the calculation is local and cheap, and because the denoising is stochastic, fifty to a hundred runs on one input yield an ensemble that maps conformational heterogeneity rather than a single pose.

The network comes from the Institute for Protein Design at the University of Washington, first posted in September 2024 under the name ChemNet and renamed PLACER in the 2025 revision, published in PNAS. It is deliberately not a structure predictor: unlike RoseTTAFold All-Atom or AlphaFold 3 it will not fold a chain, trading that for speed and generality over chemistry.

#Key Features

• One representation for all chemistry: Every node is an atom and every edge a bond, so the same network handles organic ligands, metals, covalent modifications, non-canonical amino acids and nucleic acids without per-class heads.
• Stochastic ensembles, not poses: Each run starts from a fresh random initialization of the modeled atoms, so repeated sampling produces a structurally diverse set of docks and side-chain arrangements.
• Per-atom confidence: Dedicated heads predict atomic deviations, which combine over a chosen subset into a predicted RMSD, alongside all-atom predicted lDDT from the 1D and 2D tracks.
• Explicit chirality handling: Chiral centers enter as ordered atom tuples, and gradients of a triple-product pseudoscalar supply biasing vectors that push atoms toward the intended R or S configuration.
• Preorganization as a design filter: The spread of the catalytic side chains across an ensemble gives a per-design estimate of preorganization that would otherwise require molecular dynamics.

#Technical Details

PLACER is a three-track network in the mould of RoseTTAFold, iteratively updating 1D atom embeddings, 2D pair features and 3D coordinates through eight iteration blocks with shared weights. Within each block an atom neighbor graph is built from each atom's 32 closest neighbors, picked in equal proportions by spatial and chemical-graph proximity, and an SE(3)-Transformer updates coordinates and 1D features; hydrogens are not modeled. Training used two corpora: 226,684 Cambridge Structural Database small molecules for architecture tuning (7,116 held out at 75% Tanimoto similarity), then 112,828 crops from Protein Data Bank entries better than 2.5 Å deposited before 12 January 2023 (7,090 held out at 30% sequence identity), cropped to at most 600 heavy atoms and corrupted with σ = 1.5 Å noise. The primary loss is an all-atom generalization of frame-aligned point error.

On the Astex non-native benchmark — 65 drug targets with 1,112 receptor structures not co-crystallized with the query ligand — selecting by predicted RMSD places 82.4% of complexes within 2 Å ligand RMSD, above Rosetta GALigandDock's 73.6%, though behind it below 1 Å (41.8% versus 51.6%); Vina, GOLD and GalaxyDock score lower on both. Rescoring PLACER's docks with GALigandDock adds 7.3 points at 1 Å. One model of a typical complex takes one to three seconds on a GPU.

#Applications

The immediate uses are docking a ligand into a pocket whose location is approximately known, generating conformers for molecules as complex as 50-atom macrocycles, and sampling side chains in apo pockets. The authors press hardest on enzyme design triage: across the RA95 retro-aldolase lineage, ensemble spread at the catalytic lysine tracked measured activity, and in a prospective campaign of 320 designs built with RosettaMatch and LigandMPNN, preorganization scores separated active designs from inactive ones. The best, cnRA-50, reached a kcat/KM of 11,000 M⁻¹min⁻¹ — higher than any pre-deep-learning design for this reaction.

#Impact

Code and weights ship in one repository under a BSD 3-Clause license covering both, and the same scoring approach raised hit rates in the institute's de novo serine hydrolase and metallohydrolase campaigns. The constraints are real: the ligand must already be present in the input coordinates, so a job cannot start from a SMILES string and an apo protein; crops are limited to the 600 heavy atoms seen in training; and blind docking across a whole protein surface is possible but off-label. The University of Washington has filed a provisional patent on the network, licensed to a company co-founded by the senior author.