Research map / Graph learning / Graph self-supervised learning
Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces MARCEL, a benchmark of four datasets (Drugs-75K, Kraken, EE, BDE) for molecular representation learning over conformer ensembles, benchmarking 1D, 2D, and 3D models plus data-augmentation and set-encoder strategies.
Problem
Existing molecular representation learning models mostly encode a single 2D graph or a single 3D conformer, ignoring molecular flexibility; prior conformer-ensemble studies used small datasets, few tasks, and few architectures.
Approach
Benchmark learning over Boltzmann-weighted conformer ensembles across diverse molecules (organocatalysts, transition-metal catalysts, reactions), and test two simple strategies for injecting ensembles into 3D GNNs: random conformer sampling as augmentation and set encoders (mean, DeepSets, self-attention) over per-conformer embeddings.
Key results
- Ensemble strategies with set encoders improve the single-conformer 3D baseline in 48 of 54 experiments across 9 tasks and 6 base models (per the paper's reported counts).
- On Kraken B5, the best ensemble model (PaiNN with DeepSets/attention variants) reaches MAE 0.2225 versus 0.3443 for single-conformer PaiNN, with GemNet ensemble best on several Kraken and EE tasks (EE MAE 11.61 vs 18.03 for single-conformer GemNet).
- Random forest on fingerprints beats LSTM and Transformer on small datasets (Kraken, EE), while 3D models excel on EE (MAE ~12-20 vs ~61-64 for 1D/2D) and large Drugs-75K tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add an explicit conformer ensemble set encoder to single-conformer 3D models | 3D MRL models · MARCEL tasks (9 tasks, 6 base models) | test error | Helped |
| Add DeepSets conformer aggregation (ablation) | 3D MRL models · MARCEL tasks (9 tasks, 6 base models) | test error | Helped |
| Use mean pooling to aggregate conformer embeddings (ablation) | 3D MRL models · MARCEL tasks | test error | Hurt |
| Use self-attention to aggregate conformer embeddings (ablation) | 3D MRL models · MARCEL tasks | test error | Mixed |
| Train with randomly sampled conformers instead of a fixed conformer (ablation) | 3D MRL models · BDE | MAE: 1.4503 -> 1.4741 | Mixed |
| Train with randomly sampled conformers instead of a fixed conformer (ablation) | 3D MRL models · MARCEL tasks | test error | Mixed |
| Evaluate on a random conformer rather than the lowest-energy conformer (ablation) | 3D models trained with conformer sampling · MARCEL additional evaluation tasks | MAE | Hurt |
| Average predictions across all conformers at evaluation instead of using the lowest-energy conformer (ablation) | 3D models trained with conformer sampling · MARCEL additional evaluation tasks | MAE | Mixed |
| Use 3D models instead of 2D models | Molecular representation models · EE and BDE reaction tasks | MAE | Helped |
Surprising findings
- Mean pooling over conformer embeddings loses discriminative power and performs worse than DeepSets, while attention set encoders give mixed results despite being theoretically the most expressive.
- Some 2D models are competitive with 3D models on the large-scale Drugs-75K property tasks.
Limitations
- EE dataset is proprietary and restricted to review use at time of writing, so not fully public.
- Ensemble strategies do not universally improve performance; Drugs-75K gains are inconsistent.
- Datasets only contain regression tasks and do not cover all chemical space.
Experimental setup
Datasets: GEOM-Drugs Kraken EE BDE
Benchmarks: Drugs-75K Kraken EE BDE
Metrics: MAE
Baselines: Random forest LSTM Transformer GIN GIN-VN ChemProp GraphGPS SchNet DimeNet++ GemNet PaiNN ClofNet
Builds on: DeepSets GraphGPS GemNet LEFTNet GEOM-Drugs
Closest papers by meaning
- Structure-Aware E(3)-Invariant Molecular Conformer Aggregation Networks (ICML 2024 · similarity 0.91)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.89)
- 3D Denoisers Are Good 2D Teachers: Molecular Pretraining via Denoising and Cross-Modal Distillation (AAAI 2025 · similarity 0.88)
- A Systematic Evaluation of Co-folding Model Representations for Small-Molecule Learning (NeurIPS 2026 · similarity 0.88)
- Beyond Atoms: Enhancing Molecular Pretrained Representations with 3D Space Modeling (ICML 2025 · similarity 0.88)
- Empowering Active Learning for 3D Molecular Graphs with Geometric Graph Isomorphism (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.