How PLACER Tackles Atomic‑Level Modeling of Protein Conformational Heterogeneity

Background

Protein–small‑molecule interactions are central to biological function and enzyme design, but atomic‑level modeling of their conformational heterogeneity remains difficult.

PLACER

PLACER (Protein‑Ligand Atomistic Conformational Ensemble Resolver) is a denoising graph‑neural network that receives the atomic composition and bond graph of a small molecule together with a perturbed protein structure and predicts full‑atom coordinates of the complex plus per‑atom uncertainty.

Datasets and Training

Small‑molecule training used >226 000 organic non‑polymer crystal structures from the Cambridge Structural Database (CSD); 7 116 were held out for validation. Random Gaussian noise was added to atom coordinates to train the model to recover precise structures.

Protein‑ligand complexes were drawn from the PDB, selecting entries with resolution < 2.5 Å. The training set contains ~113 000 complexes and 7 090 validation complexes. Water molecules were excluded, non‑biological small molecules were retained. Systems were trimmed to ≤600 heavy atoms and perturbed with Gaussian noise around a randomly chosen atom center.

Network Architecture

PLACER adopts a three‑track design inspired by RoseTTAFold:

1D track processes atom‑level features.

2D track encodes chemical‑graph edges and spatial proximity.

3D track updates atomic coordinates.

After initial 1D/2D embedding, an iterative block builds a neighbor graph for each atom (32 nearest neighbors, half from chemical edges, half from spatial proximity). A feed‑forward adapter projects 2D features to edge embeddings; together with 1D features, the neighbor graph, and current 3D coordinates they are fed to an SE3‑Transformer to update coordinates and embeddings. Eight shared‑weight iterative blocks are stacked.

Chirality is encoded via Type‑1 vector features and propagated through pair‑to‑pair updates. Separate confidence heads predict per‑atom and per‑pair uncertainty. The loss combines full‑atom Frame‑Aligned Point Error (FAPE) with confidence losses at atom and pair levels.

Results

Small‑Molecule Conformation Prediction

On the CSD test set, fully trained PLACER generates sub‑Å structures for complex molecules, including macrocycles with >50 atoms and peptidic macrocycles. Ablation experiments show that removing bond‑length information or reducing the number of iterative blocks markedly lowers accuracy, confirming the importance of the iterative, multi‑track design.

Protein‑Ligand Interaction Modeling

Multiple stochastic runs produce diverse ligand ensembles that are insensitive to the initial ligand placement; ensembles span the sampled space. Predicted ligand RMSD (pRMSD) correlates with experimental accuracy, enabling selection of near‑native models.

On a benchmark of 65 non‑native targets, PLACER’s success rate using pRMSD exceeds traditional docking tools (Vina, GOLD, GalaxyDock). Compared with Rosetta GALigandDock, PLACER achieves higher recall for RMSD < 2 Å (82.4 % vs 73.6 %) and slightly lower recall for RMSD < 1 Å (41.8 % vs 51.6 %).

Enzyme Active‑Site Design

In retro‑aldolase (RA95) redesign, 50 simulation repeats showed that low‑activity designs generate highly diverse conformational ensembles (low pre‑organization), whereas evolved high‑activity variants produce more ordered ensembles, indicating that lack of pre‑organization limits early designs.

A newly designed retro‑aldolase (cnRA‑50) evaluated with PLACER achieved kcat/KM = 11 000 M⁻¹·min⁻¹, surpassing earlier computational designs and approaching the performance of recent methods such as RFdiffusion and ProteinMPNN.

Implications

PLACER can rapidly generate atom‑level conformational ensembles for isolated small molecules and for molecules within protein environments, supporting evaluation of enzyme active‑site designs and protein‑small‑molecule binding projects.

References

Modeling protein‑small molecule conformational ensembles with PLACER – PNAS (2024).

Design of intrinsically disordered region binding proteins – Science (2025).

De novo Design of All‑atom Biomolecular Interactions with RFdiffusion3 – bioRxiv (2025).