Research map / Image generation / Autoregressive and latent generation
Enhancing Generalizability in Molecular Conformation Generation with METRIZATION-Informed Geometric Diffusion Pretraining
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.
Proposes MiGDiff, which pretrains a GeoDiff-style geometric diffusion model on METRIZATION-sampled distance-geometry conformations and then fine-tunes on real data, improving cross-dataset generalization of molecular conformation generation.
Problem
Diffusion-based molecular conformation generators trained on one dataset (e.g., GEOM-QM9) generate poor conformations for out-of-distribution molecules (e.g., GEOM-DRUGS).
Approach
Inject distance-geometry constraints into diffusion pretraining by pretraining the geometric diffusion backbone on conformations sampled with the METRIZATION algorithm (plus Schoenberg embedding and coordinate refinement), then fine-tuning on real conformations ("metrization-driven pretraining + data-driven finetuning").
Key results
- Trained only on GEOM-QM9 and tested on GEOM-DRUGS, MIGDIFF-P + FF reaches COV-R 89.85% and MAT-R 0.8056 Å, versus GeoDiff's 18.93% and 2.6392 Å.
- On the in-distribution GEOM-DRUGS benchmark (trained with GEOM-DRUGS supervision), MIGDIFF achieves COV-R 89.76% and COV-P 63.61%, slightly above GeoDiff (89.13% and 61.47%).
- In OOD property prediction, MIGDIFF's average energy MAE is 0.1707 versus GeoDiff's 0.4395 on the OOD setting for the gap metric in Table 4 (reported per setting).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add METRIZATION pretraining without fine-tuning (ablation) | GeoDiff · GEOM-DRUGS test (OOD) | COV-R mean (%): 18.93 -> 73.51 | Helped |
| Add METRIZATION pretraining without fine-tuning (ablation) | GeoDiff · GEOM-DRUGS test (OOD) | MAT-R mean (Å): 2.6392 -> 1.0732 | Helped |
| Add METRIZATION pretraining followed by real-world data fine-tuning | GeoDiff · GEOM-DRUGS test (OOD) | COV-R mean (%): 89.13 -> 89.76 | Helped |
| Add METRIZATION pretraining followed by real-world data fine-tuning | GeoDiff · GEOM-DRUGS test (OOD) | MAT-P mean (Å): 1.1712 -> 1.1203 | Helped |
| Add METRIZATION pretraining followed by real-world data fine-tuning | GeoDiff · Property prediction OOD | Mean energy MAE: 0.3210 -> 0.1707 | Helped |
| Add METRIZATION pretraining followed by real-world data fine-tuning | GeoDiff · Property prediction OOD | Average HOMO-LUMO gap MAE: 0.4395 -> 0.3296 | Helped |
| Add coordinate refinement (ablation) | GeoDiff · GEOM-DRUGS test (OOD) | COV-R mean (%): 18.93 -> 23.99 | Helped |
| Add force-field post-processing (ablation) | MIGDIFF-P · GEOM-DRUGS test (OOD) | COV-R mean (%): 73.51 -> 89.85 | Helped |
| Add force-field post-processing (ablation) | MIGDIFF-P · GEOM-DRUGS test (OOD) | COV-P mean (%): 58.34 -> 81.31 | Helped |
| Add force-field post-processing (ablation) | MIGDIFF-P · GEOM-DRUGS test (OOD) | MAT-P mean (Å): 1.2755 -> 1.0374 | Helped |
Surprising findings
- The METRIZATION sampling alone performs only averagely (COV-R 45.71%), yet a model pretrained only on METRIZATION data (MIGDIFF-P) substantially outperforms it.
Limitations
- METRIZATION sampling alone yields only average-quality conformations and relies on bounds derived from empirical information.
- Generalization is evaluated only on the QM9-to-DRUGS transfer setting and a small property-prediction set of 30 molecules.
Experimental setup
Datasets: GEOM-QM9 GEOM-DRUGS
Benchmarks: GEOM-QM9 GEOM-DRUGS
Metrics: COV-R MAT-R COV-P MAT-P average energy error average HOMO-LUMO gap error
Baselines: GeoDiff GeoDiff-CR GraphDG CGCF ConfVAE GeoMol ConfGF RDKit METRIZATION CVGAE
Builds on: GeoDiff METRIZATION Schoenberg's theorem
Closest papers by meaning
- Geometric-Facilitated Denoising Diffusion Model for 3D Molecule Generation (AAAI 2024 · similarity 0.91)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.90)
- Efficient Molecular Conformer Generation with SO(3)-Averaged Flow Matching and Reflow (ICML 2025 · similarity 0.89)
- Sampling 3D Molecular Conformers with Diffusion Transformers (NeurIPS 2025 · similarity 0.89)
- Equivariant Blurring Diffusion for Hierarchical Molecular Conformer Generation (NeurIPS 2024 · similarity 0.89)
- DiSCO: Diffusion Schrödinger Bridge for Molecular Conformer Optimization (AAAI 2024 · similarity 0.89)
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.