Modeling protein-small molecule conformational ensembles with PLACER
Ivan V. AnishchenkoiD, Yakov KipnisiD, Indrek KalvetiD, Guangfeng ZhouiD, Rohith KrishnaiD, Samuel J. PellockiD, Anna LaukoiD, Gyu Rie LeeiD, Linna AniD, Justas DauparasiD, Frank DiMaioiD, David A. BakeriD
Abstract
Modeling the conformational heterogeneity of protein-small molecule interactions is important for understanding natural systems and evaluating designed systems, but remains an outstanding challenge. We reasoned that while residue level descriptions of biomolecules are efficient for de novo structure prediction, for probing heterogeneity of interactions with small molecules in the folded state an entirely atomic level description could have advantages in speed and generality. We developed a graph neural network called PLACER (Protein-Ligand Atomistic Conformational Ensemble Resolver) trained to recapitulate correct atomic positions from partially corrupted input structures from the Cambridge Structural Database and the Protein Data Bank; the nodes of the graph are the atoms in the system. PLACER accurately generates structures of diverse organic small molecules given knowledge of their atom composition and bonding, and given a description of the larger protein context, builds up structures of small molecules and protein side chains for protein-small molecule docking. Because PLACER is rapid and stochastic, ensembles of predictions can be readily generated to map conformational heterogeneity. In enzyme design efforts described here and elsewhere, we find that using PLACER to assess the accuracy and pre-organization of the designed active sites results in higher success rates and higher activities; we obtain a preorganized retroaldolase with a kcat/KM of 11000 M-1min-1, considerably higher than any pre-deep learning design for this reaction. We anticipate that PLACER will be widely useful for rapidly generating conformational ensembles of small molecule and small molecule-protein systems, and for designing higher activity preorganized enzymes.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 38
- FWCI
- 9.44
- Field percentile
- top 1%
- Assay
- Research Focus
- Approach
In-depth summary
Claude Sonnet 4.6 · max effort · 8 Nov 2026 · from the full text of version 2 (27 Sept 2025) · CC BY 4.0
An atom-level graph neural network that denoises corrupted structures generates small-molecule and protein-ligand conformational ensembles, and its active-site pre-organization metric enriched for more active designed retroaldolases.
Residue-level predictors such as AF2, AF3 and RF are not built to rapidly sample atomic-level heterogeneity of small-molecule interactions in folded proteins, and MD or rotamer-based pre-organization estimates are costly or limited. PLACER takes backbone coordinates as input and stochastically generates ensembles quickly, which allows screening of active site accuracy and pre-organization.
Key results
ShownOn non-native protein structures from a standard drug-target docking set, PLACER with pRMSD scoring beat GALigandDock at the 2 Å threshold but trailed it at the 1 Å threshold.
65 targets, 1112 non-native structures; ligand RMSD < 2Å: 82.4% vs 73.6% for Rosetta GALigandDock; RMSD < 1Å: 41.8% vs 51.6%. Higher success than Vina, GOLD and GalaxyDock (values taken from their references). · Modeling protein-small molecule interactions…
ShownPredicted RMSD (pRMSD) selects better models from the sampled pool, and for two de novo designed binders the ligand starting position appeared to matter little. Rosetta-based minimization of the docks improved high-accuracy docking success.
Best pRMSD models gave ligand RMSD = 0.53Å (heme) and 0.77Å (cortisol). Minimizing PLACER docks with the Rosetta force field gave +7.3% docking success at <1Å, with no notable change at <2Å. · Modeling protein-small molecule interactions…; Figure legends
ShownTrained on crystal structures of small molecules from the Cambridge Structural Database, PLACER reproduces complex molecules, including macrocycles, and ablations of the iteration count, the loss and the bond-separation feature each degrade it.
Training/validation sets of 226,684 and 7,116 molecules; macrocycles with over 50 atoms modeled with sub-Å accuracy; the seven macrocycles shown are best-RMSD models from 1000 generated; validation error bars are standard deviations over 5 independent runs, one conformer per example. Ablation magnitudes are shown only graphically. · Tuning network architecture on experimental…; Figure legends
SuggestedPLACER ensembles for the RA95 retroaldolase series are diffuse for low-activity initial designs and become tighter for the more active evolved versions, so ensemble spread tracks pre-organization.
50 PLACER runs per reaction intermediate; mean pRMSD over lysine side-chain atoms (excluding backbone N, Cα, C, O); Fig 4C shows examples from three previously published enzymes; the trend is qualitative with no correlation statistic. · Assessment of enzyme active site design accuracy…; Figure legends
SuggestedIn a prospective design round, designs with the highest kcat/KM were more pre-organized by PLACER than low-activity designs, and the top design reached a kcat/KM above earlier pre-evolution computational designs.
320 designs screened; most active design cnRA-50 had kcat/KM of 11,000 M^-1min^-1 and one of the highest pre-organization levels; enrichment shown graphically without a reported test statistic. · Abstract; Assessment of enzyme active site design accuracy…; Figure legends
Methods, data and reproducibility
A three-track, SE3-Transformer-based graph network over heavy atoms (eight shared-weight iteration blocks) is trained with all-atom FAPE and confidence losses to denoise corrupted coordinates. Stochastic reruns yield ensembles, and pRMSD is used for ranking.
- Cambridge Structural Database organic small molecules (training and validation) for architecture tuning · n = 226,684 training; 7,116 validation
- PDB protein-ligand structures, resolution <2.5Å, deposited before January 12, 2023, for training and validation · n = 112,828 training; 7,090 validation
- Astex non-native docking test set: drug targets with co-crystal structures and non-native structures · n = 65 targets; 1112 non-native structures
- RA95 series of retroaldolases (initial designs and evolved versions) with crystal structures, analyzed by PLACER ensembles · n = 50 PLACER runs per intermediate
- New deep-learning-designed NTF2-like retroaldolase designs screened by IVTT, with top designs purified from E. coli and characterized · E. coli (expression host) · n = 320 designs
- Methods & tools
- PLACER, SE3-Transformer, all-atom FAPE loss, Chemical Component Dictionary, Rosetta GALigandDock, Vina, GOLD, GalaxyDock, RosettaMatch, LigandMPNN, IVTT screen
- Software & environment
- Retroaldolase designs were screened in an in vitro transcription translation system, then the most active were expressed and purified in E. coli. Software versions are not stated in the main text.
- Assay conditions & compute
- Azure compute credits and NERSC Perlmutter were used (funding statement). GPU type, counts and runtimes are not given; PDB training takes x10 longer than CSD training.
- Reproducibility gaps
- Zenodo is named for model files and designed sequences, but no URL is printed. Training details are deferred to Supplementary Methods, which are not in the provided text.
Limitations (5)
- PLACER trails Rosetta GALigandDock in the high-accuracy docking regime (RMSD < 1Å: 41.8% vs 51.6%). · stated by the authors
- Selection by pRMSD is not optimal: it still falls short of the ideal selection curve from the sampled pool. · stated by the authors
- PLACER does not predict protein backbone structure or model structure from sequence alone, so it needs backbone coordinates and a putative binding region as input. · stated by the authors
- The link between pre-organization and activity is shown graphically with no reported correlation or significance test. Retroaldolase results come from one reaction and a small set of enzymes. · noted in this summary
- Vina, GOLD and GalaxyDock docking numbers were taken from earlier publications, not rerun on the same inputs, so those comparisons may not be matched. · noted in this summary
Who should read:Read this if you want a fast atom-level ensemble generator for ligand docking or enzyme active-site pre-organization screening, with benchmark numbers against GALigandDock.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- G-PLIP: Knowledge graph neural network for structure-free protein-ligand bioactivity prediction
Simon Crouzet, Anja Maria Lieberherr, … Jitao David Zhang·Computational and Structural Biotechnology JournalPeer-reviewed·Jul 5, 2024
- An Energy Landscape Approach to Miniaturizing Enzymes using Protein Language Model Embeddings
Jakub Lála, Harsh Agrawal, … Stefano Angioletti‐Uberti·bioRxiv·Mar 5, 2026
- PLINDER: The protein-ligand interactions dataset and evaluation resource
Janani Durairaj, Yusuf O. Adeshina, … Luca Naef·bioRxiv·Jul 17, 2024
- Protein-Ligand Interaction Graphs: Learning from Ligand-Shaped 3D Interaction Graphs to Improve Binding Affinity Prediction
Marc A. Moesser, Dominik Klein, … Garrett M. Morris·bioRxiv·Mar 4, 2022
- Design in Voxel Space Decode in SMILES Space: Plixer Generates Drug-Like Molecules for Protein Pockets
Jude Wells, Brooks Paige·bioRxiv·Jul 15, 2025
- Combinatorial assembly and design of enzymes
R. Lipsh-Sokolik, O. Khersonsky, … S. J. Fleishman·bioRxiv·Sep 17, 2022