MARCEL

Key Contribution

MARCEL provides a benchmark for conformer ensemble learning. It demonstrates that explicitly modeling full conformer distributions improves property prediction across drug-like molecules and organometallic catalysts.

Overview

The Molecular Representation and Conformer Ensemble Learning (MARCEL) dataset provides 722K+ conformations across 76K+ molecules spanning four diverse chemical domains: drug-like molecules (Drugs-75K), organophosphorus ligands (Kraken), chiral catalysts (EE), and organometallic complexes (BDE). MARCEL evaluates conformer ensemble methods across both pharmaceutical and catalysis applications.

Dataset Examples

Dataset Subsets

Benchmarks

Related Datasets

Strengths

• Domain diversity: Beyond drug-like molecules, includes organometallics and catalysts rarely covered in existing benchmarks
• Ensemble-based: Provides full conformer ensembles with statistical weights
• DFT-quality energies: Drugs-75K features DFT-level conformers and energies (higher accuracy than GEOM-Drugs)
• Realistic scenarios: BDE subset models the practical constraint of lacking DFT-computed conformers for large catalyst systems
• Comprehensive baselines: Benchmarks 18 models across 1D (SMILES), 2D (graph), 3D (single conformer), and ensemble methods
• Property diversity: Covers ionization potential, electron affinity, electronegativity, ligand descriptors, and catalytic properties

Limitations

• Regression only: All tasks evaluate regression metrics exclusively
• Chemical space coverage: The 76K molecules encapsulate a fraction of the expansive drug-like and catalyst chemical spaces
• Compute requirements: Working with large conformer ensembles demands significant computational resources
• Proprietary data: EE subset is proprietary (as of December 2025)
• DFT bottleneck: BDE demonstrates a practical limitation: single DFT optimization can take 2-3 days, making conformer-level DFT infeasible for large organometallics
• Uniform sampling baseline: The initial data augmentation strategy tested for handling ensembles samples conformers uniformly rather than by Boltzmann weight. This unprincipled physical assumption likely explains why the strategy occasionally introduces noise and fails to aid complex 3D architectures.
• Drugs-75K properties: The large-scale benchmark (Drugs-75K) specifically targets electronic properties (Ionization Potential, Electron Affinity, Electronegativity). As the authors explicitly highlight in Section 5.2, these properties are generally less sensitive to conformational rotations compared to steric or spatial interactions. This significantly confounds evaluating whether explicit conformer ensembles actually benefit large-scale regression tasks.
• Unrealistic single-conformer baselines: The 3D single-conformer models are exclusively evaluated on the lowest-energy conformer. This setup is inherently flawed for real-world application, as knowing the global minimum a priori requires exhaustively searching and computing energies for the entire conformer space.

Technical Notes

Data Generation Pipeline

Drugs-75K

Source: GEOM-Drugs subset

Filtering:

• Minimum 5 rotatable bonds (focus on flexible molecules)
• Allowed elements: H, C, N, O, F, Si, P, S, Cl

Conformer generation:

• DFT-level calculations for both conformers and energies
• Higher accuracy than original GEOM-Drugs (semi-empirical GFN2-xTB)

Properties: Ionization Potential (IP), Electron Affinity (EA), Electronegativity (χ)

Kraken

Source: Original Kraken dataset (1,552 monodentate organophosphorus(III) ligands)

Properties: 5 of 78 available properties

• $B_5$: Ligand sterics (buried volume)
• $L$: Ligand electronics
• $\text{Bur}B_5$: Buried volume variant
• $\text{Bur}L$: Electronic parameter variant

EE (Enantiomeric Excess)

Generation method: Q2MM (Quantum-guided Molecular Mechanics)

Molecules: 872 Rhodium (Rh)-bound atropisomeric catalysts from chiral bisphosphine

Property: Enantiomeric excess (EE) for asymmetric catalysis

Availability: Proprietary-only (closed-source as of December 2025)

BDE (Bond Dissociation Energy)

Molecules: 5,195 organometallic catalysts (ML₁L₂ structure)

Initial conformers: OpenBabel with geometric optimization

Energies: DFT calculations

Property: Electronic dissociation energy (difference between bound and unbound states)

Key constraint: DFT optimization for full conformer ensembles computationally infeasible (2-3 days per molecule)

Benchmark Setup

Task: Predict molecular properties from structure using different representation strategies (1D/2D/3D/Ensemble). The ground-truth regression targets are calculated as the Boltzmann-averaged value of the property across the conformer ensemble:

$$ \langle y \rangle_{k_B} = \sum_{\mathbf{C}_i \in \mathcal{C}} p_i y_i $$

Where $p_i$ is the conformal probability under experimental conditions derived from the conformer energy $e_i$:

$$ p_i = \frac{\exp(-e_i / k_B T)}{\sum_j \exp(-e_j / k_B T)} $$

Data splits: Datasets are partitioned 70% train, 10% validation, and 20% test.

Model categories:

1. 1D Models: SMILES-based (Random Forest on Morgan fingerprints, LSTM, Transformer).
2. 2D Models: Graph-based (GIN, GIN+VN, ChemProp, GraphGPS).
3. 3D Models: Single conformer (SchNet, DimeNet++, GemNet, PaiNN, ClofNet, LEFTNet). For evaluation, single 3D models exclusively ingest the lowest-energy conformer. This baseline setting often yields strong performance but is unrealistic in practice, as identifying the global minimum requires exhaustively searching the entire conformer space.
4. Ensemble Models: Full conformer ensemble processing via explicit set encoders. For each conformer embedding $\mathbf{z}_i$, three aggregation strategies are evaluated:

Mean Pooling: $$ \mathbf{s}_{\text{MEAN}} = \frac{1}{|\mathcal{C}|} \sum_{i=1}^{|\mathcal{C}|} \mathbf{z}_i $$

DeepSets: $$ \mathbf{s}_{\text{DS}} = g\left(\sum_{i=1}^{|\mathcal{C}|} h(\mathbf{z}_i)\right) $$

Self-Attention: $$ \begin{aligned} \mathbf{s}_{\text{ATT}} &= \frac{1}{|\mathcal{C}|} \sum_{i=1}^{|\mathcal{C}|} \mathbf{c}_i, \quad \text{where} \quad \mathbf{c}_i = g\left( \sum_{j=1}^{|\mathcal{C}|} \alpha_{ij} h(\mathbf{z}_j) \right) \\ \alpha_{ij} &= \frac{\exp\left((\mathbf{W} h(\mathbf{z}_i))^\top (\mathbf{W} h(\mathbf{z}_j))\right)}{\sum_{k=1}^{|\mathcal{C}|} \exp\left((\mathbf{W} h(\mathbf{z}_i))^\top (\mathbf{W} h(\mathbf{z}_k))\right)} \end{aligned} $$

Evaluation metric: Mean Absolute Error (MAE) for all tasks.

Key Findings

Ensemble superiority (task-dependent): Across benchmarks, explicitly modeling the full conformer set using DeepSets often achieved top performance. However, these improvements are not uniform:

• Small-Scale Success: Ensemble methods show large improvements on tasks like Kraken (Ensemble PaiNN achieves 0.2225 on $B_5$ vs 0.3443 single) and EE (Ensemble GemNet achieves 11.61% vs 18.03% single).
• Large-Scale Plateau: The performance improvements did not strongly transfer to massive subsets like Drugs-75K (Ensemble GemNet achieves 0.4066 eV on IP vs 0.4069 eV single). The authors posit the computational burden of processing all ensemble embeddings changes learning dynamics and increases training difficulty.

Conformer Sampling for Noise: Data augmentation (randomly sampling one conformer from an ensemble during training) improves performance and robustness when underlying conformers are imprecise (e.g., the forcefield-generated conformers in the BDE subset).

3D vs 2D: 3D models generally outperform 2D graph models, especially for conformationally-sensitive properties, though 1D and 2D methods remain highly competitive on low-resource datasets or less rotation-sensitive properties.

Model architecture: GemNet and PaiNN architectures consistently top-ranked across tasks.

Reproducibility Details

• Data: The Drugs-75K, Kraken, and BDE subsets are openly available via Google Drive links on the project’s GitHub repository. However, the EE dataset remains closed-source/proprietary (as of 2026), making the EE suite of the benchmark currently irreproducible.
• Code: The benchmark suite and PyTorch-Geometric dataset loaders are open-sourced at GitHub (SXKDZ/MARCEL) under the Apache-2.0 license.
• Hardware: The authors trained these models using Nvidia A100 (40GB) GPUs. Memory-intensive models (e.g., GemNet, LEFTNet) required Nvidia H100 (80GB) GPUs. Total computation across all benchmark experiments was approximately 6,000 GPU hours.
• Algorithms/Models: Hyperparameters for all 18 evaluated models are provided directly within the repository configuration files (benchmarks/params), ensuring clear algorithmic reproducibility. All baseline models use popular, publicly available frameworks (e.g., PyTorch Geometric, OGB, RDKit).
• Evaluation: Detailed evaluation scripts are provided in the repository with consistent tracking of Mean Absolute Error (MAE) and proper configuration of benchmark splits.

Citation

@inproceedings{zhu2024learning,
title={Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks},
author={Yanqiao Zhu and Jeehyun Hwang and Keir Adams and Zhen Liu and Bozhao Nan and Brock Stenfors and Yuanqi Du and Jatin Chauhan and Olaf Wiest and Olexandr Isayev and Connor W. Coley and Yizhou Sun and Wei Wang},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=NSDszJ2uIV}
}