MARCEL: Molecular Representation and Conformer Ensemble Learning

Key Contribution

MARCEL contributes a large-scale dataset for molecular representation and conformer ensemble learning, facilitating advancements in drug discovery and cheminformatics.

Dataset Information

Format

SMILES RDKit mol objects 3D coordinates Statistical weights Experimental properties

Size

Type	Count
Conformers	722,193
Molecules	76,651
Reactions	6,787

Dataset Examples

Example conformer from Drugs-75K (SMILES: COC(=O)[C@@]1(Cc2ccc(OC)cc2)[C@H]2c3cc(C(=O)N(C)C)n(Cc4ccc(OC(F)(F)F)cc4)c3C[C@H]2CN1C(=O)c1ccccc1; IUPAC: methyl (2R,3R,6R)-4-benzoyl-10-(dimethylcarbamoyl)-3-[(4-methoxyphenyl)methyl]-9-[[4-(trifluoromethoxy)phenyl]methyl]-4,9-diazatricyclo[6.3.0.02,6]undeca-1(8),10-diene-3-carboxylate)

2D structure of Drugs-75K conformer above.

Example conformer from Kraken (ligand 10, conformer 0) in 2D

Example conformer from Kraken (ligand 10, conformer 0) in 3D

Example substrate from BDE in 3D (Pt_9.63).

Dataset Subsets

Subset	Count	Description
Drugs-75K	75,099	Drug-like molecules with at least 5 rotatable bonds
Kraken	1,552	monodentate organophosphorus (III) ligands
EE	872	Rhodium (Rh)-bound atropisomeric catalysts derived from chiral bisphosphine
BDE	5,195	Organometallic catalysts ML$_1$L$_2$

Results

Drugs 75k Property Prediction

Model	IP	EA	χ
1D - Random forest	0.4987	0.4747	0.2732
1D - LSTM	0.4788	0.4648	0.2505
1D - Transformer	0.6617	0.5850	0.4073
2D - GIN	0.4354	0.4169	0.2260
2D - GIN+VN	0.4361	0.4169	0.2267
2D - ChemProp	0.4595	0.4417	0.2441
2D - GraphGPS	0.4351	0.4085	0.2212
3D - SchNet	0.4394	0.4207	0.2243
3D - DimeNet++	0.4441	0.4233	0.2436
3D - GemNet	🥈 0.4069	🥈 0.3922	🥇 0.1970
3D - PaiNN	0.4505	0.4495	0.2324
3D - ClofNet	0.4393	0.4251	0.2378
3D - LEFTNet	0.4174	0.3964	0.2083
Ensemble - SchNet	0.4452	0.4232	0.2243
Ensemble - DimeNet++	0.4126	0.3944	0.2267
Ensemble - GemNet	🥇 0.4066	🥇 0.3910	🥈 0.2027
Ensemble - PaiNN	0.4466	0.4269	0.2294
Ensemble - ClofNet	0.4280	0.4033	0.2199
Ensemble - LEFTNet	0.4149	0.3953	0.2069

Kraken Property Prediction

Model	B₅	L	BurB₅	BurL
1D - Random forest	0.4760	0.4303	0.2758	0.1521
1D - LSTM	0.4879	0.5142	0.2813	0.1924
1D - Transformer	0.9611	0.8389	0.4929	0.2781
2D - GIN	0.3128	0.4003	0.1719	0.1200
2D - GIN+VN	0.3567	0.4344	0.2422	0.1741
2D - ChemProp	0.4850	0.5452	0.3002	0.1948
2D - GraphGPS	0.3450	0.4363	0.2066	0.1500
3D - SchNet	0.3293	0.5458	0.2295	0.1861
3D - DimeNet++	0.3510	0.4174	0.2097	0.1526
3D - GemNet	0.2789	0.3754	0.1782	0.1635
3D - PaiNN	0.3443	0.4471	0.2395	0.1673
3D - ClofNet	0.4873	0.6417	0.2884	0.2529
3D - LEFTNet	0.3072	0.4493	0.2176	0.1486
Ensemble - SchNet	0.2704	0.4322	0.2024	0.1443
Ensemble - DimeNet++	0.2630	🥈 0.3468	0.1783	🥈 0.1185
Ensemble - GemNet	🥈 0.2313	🥇 0.3386	🥇 0.1589	🥇 0.0947
Ensemble - PaiNN	🥇 0.2225	0.3619	🥈 0.1693	0.1324
Ensemble - ClofNet	0.3228	0.4485	0.2178	0.1548
Ensemble - LEFTNet	0.2644	0.3643	0.2017	0.1386

Ee Bde Property Prediction

Model	EE	BDE
1D - Random forest	61.2963	3.0335
1D - LSTM	64.0088	2.8279
1D - Transformer	62.0816	10.0771
2D - GIN	62.3065	2.6368
2D - GIN+VN	62.3815	2.7417
2D - ChemProp	61.0336	2.6616
2D - GraphGPS	61.6251	2.4827
3D - SchNet	17.7421	2.5488
3D - DimeNet++	14.6414	🥇 1.4503
3D - GemNet	18.0338	1.6530
3D - PaiNN	20.2359	2.1261
3D - ClofNet	33.9473	2.6057
3D - LEFTNet	19.7974	🥈 1.5328
Ensemble - SchNet	14.2238	1.9737
Ensemble - DimeNet++	🥈 12.0259	1.4741
Ensemble - GemNet	🥇 11.6142	1.6059
Ensemble - PaiNN	13.5570	1.8744
Ensemble - ClofNet	13.9647	2.0106
Ensemble - LEFTNet	18.4189	1.5276

Strengths

More than just drugs, also considers organometallics and catalysts
Flexible drug-like molecules
Ensemble-based dataset
DFT-level accuracy for conformer energies
Realistic scenarios (BDE, with lack of DFT-computed conformers)

Limitations

Only regression tasks
Does not cover the full breadth of chemical space
Computational cost of working with large conformer ensembles
EE appears to be proprietary-only

Technical Notes

Drugs-75K

Begins with SMILES from GEOM-Drugs.
Excludes any molecules with less than 5 rotatable bonds.
Limits to atomic species: H, C, N, O, F, Si, P, S, and Cl.
Computes DFT-level conformers and energies (higher accuracy than GEOM-Drugs).

Kraken

Uses the original dataset directly, but only focuses on 5 of the 78 properties.

EE

Uses Q2MM to simulate conformer ensembles of 872 catalyst-substrate pairs.
Seeks to predict enantiomeric excess (EE).
Not available publicly (as of September 2025).

BDE

Metallic centers with two flexible ligands.
Seeks to predict electronic dissociation energies by difference of bound and unbound states.
OpenBabel to generate initial conformers with further geometric optimization.
Actual energies computed with DFT.
Cannot compute DFT for conformers as a single search with DFT would take upwards of 2-3 days (as of September 2025).

Related Datasets

Dataset	Relationship	Link
GEOM	Contains Subset	📄 View Details
Kraken	Contains	N/A
EE	Contains	N/A
BDE	Contains	N/A

Dataset Details
Authors	Yanqiao Zhu, Jeehyun Hwang, Brock Anton Stenfors, Yuanqi Du, Olexandr Isayev, Keir Adams, Jatin Chauhan, Connor W. Coley, Yizhou Sun, Zhen Liu, Bozhao Nan, Olaf Wiest, Wei Wang
Paper Title	Learning over Molecular Conformer Ensembles: Datasets and Benchmarks
Institutions	UCLA, MIT, CMU, Notre Dame, Cornell
Published In	International Conference on Learning Representations
Category	Computational Chemistry
Format	SMILES RDKit mol objects 3D coordinates Statistical weights Experimental properties
Size	Conformers: 722,193 Molecules: 76,651 Reactions: 6,787
Date	September 2025
Year	2024
Links	📊 Dataset • 📄 Paper