Research map / Image generation / Autoregressive and latent generation
Sampling 3D Molecular Conformers with Diffusion Transformers
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 DiTMC, a diffusion transformer for molecular conformer generation with graph-geodesic pair conditioning and interchangeable non-equivariant (aPE, rPE) or SO(3)-equivariant (PE(3)) attention, trained via conditional flow matching.
Problem
Adapting diffusion transformers to molecular conformer generation requires integrating discrete molecular graph information with continuous 3D geometry, handling Euclidean symmetries, and conditioning that generalizes across molecule sizes.
Approach
Conditioning tokens from atom-wise GNN features and pair-wise graph geodesic distances are injected into DiT blocks via AdaLN/AdaScale, with a modular choice of positional embedding and attention (absolute, relative, or SO(3)-equivariant) to trade accuracy against compute.
Key results
- On GEOM-QM9, DiTMC+rPE-B achieves COV-P mean 95.7% and AMR-P mean 0.080 Å, outperforming prior methods on precision and AMR metrics.
- On GEOM-DRUGS, DiTMC+rPE-L reaches COV-P mean 78.1% and AMR-P mean 0.466 Å, while DiTMC+aPE-L gives the best physical-property MAE among compared methods (e.g., E 0.16 kcal/mol vs 0.18 for ET-Flow).
- On GEOM-XL (102 molecules), DiTMC+aPE-L achieves AMR-R mean 1.88 Å vs 1.97 Å for MCF-L with about 8x fewer parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use relative positional embeddings instead of absolute positional embeddings (ablation) | DiTMC-B · GEOM-QM9 | AMR-P Mean (Å): 0.085 -> 0.080 | Helped |
| Use SO(3)-equivariant positional embeddings and attention instead of absolute positional embeddings (ablation) | DiTMC-B · GEOM-QM9 | AMR-P Mean (Å): 0.085 -> 0.089 | Hurt |
| Use SO(3)-equivariant positional embeddings and attention instead of non-equivariant variants (ablation) | DiTMC · GEOM-QM9 | sample fidelity at low RMSD thresholds | Helped |
| Use SO(3)-equivariant positional embeddings and attention instead of absolute positional embeddings (ablation) | DiTMC-B · GEOM-DRUGS | training time: 19.2 -> 70.0 ms | Hurt |
| Add pair-wise geodesic graph conditioning to atom-wise conditioning (ablation) | DiTMC+aPE-B · GEOM-DRUGS conditioning ablation | COV-P Mean (%): 55.6 -> 76.5 | Helped |
| Add atom-wise and pair-wise graph conditioning (ablation) | DiTMC+aPE-B · GEOM-DRUGS conditioning ablation | COV-P Mean (%): 5.15 -> 76.5 | Helped |
| Add pair-wise geodesic graph conditioning to atom-wise conditioning (ablation) | DiTMC+aPE-B · GEOM-DRUGS conditioning ablation | AMR-P Mean (Å): 0.762 -> 0.500 | Helped |
| Scale DiTMC+aPE from base to large (ablation) | DiTMC+aPE · GEOM-DRUGS | AMR-P Mean (Å): 0.500 -> 0.470 | Helped |
| Scale DiTMC+aPE from base to large (ablation) | DiTMC+aPE · GEOM-XL, 102 molecules | AMR-P Mean (Å): 2.90 -> 2.81 | Helped |
| Use a harmonic prior instead of a Gaussian prior (ablation) | DiTMC · GEOM-QM9 and GEOM-DRUGS | conformer generation performance | Helped |
| Vary the noise scaling parameter between 0.05 and 0.5 (ablation) | DiTMC · GEOM-QM9 and GEOM-DRUGS | conformer generation performance | Mixed |
Surprising findings
- Simple non-equivariant aPE/rPE models perform competitively with the SO(3)-equivariant PE(3) variant while being about 3x faster at inference.
- Pair-wise geodesic-distance conditioning on all atom pairs gives a large gain over atom-wise conditioning alone (COV-P mean 76.5% vs 55.6% on GEOM-DRUGS).
Limitations
- Evaluation restricted to small and medium-sized molecules; larger flexible compounds left for future work.
- Training depends on high-quality ground-truth conformers, which may be unavailable.
- Attention cost scales quadratically with number of atoms; the conditioning strategy is tailored to molecular graphs and does not directly extend to de novo generation.
Experimental setup
Datasets: GEOM-QM9 GEOM-DRUGS GEOM-XL
Benchmarks: GEOM-QM9 GEOM-DRUGS GEOM-XL
Metrics: COV-R COV-P AMR-R AMR-P median absolute error of ensemble properties
Baselines: GeoMol GeoDiff Torsional Diffusion MCF DMT ET-Flow E-NFs CGCF
Builds on: DiT stochastic interpolants flow matching MeshGraphNet Graphormer
Closest papers by meaning
- Equivariant Blurring Diffusion for Hierarchical Molecular Conformer Generation (NeurIPS 2024 · similarity 0.91)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.90)
- Geometric-Facilitated Denoising Diffusion Model for 3D Molecule Generation (AAAI 2024 · similarity 0.90)
- Enhancing Generalizability in Molecular Conformation Generation with METRIZATION-Informed Geometric Diffusion Pretraining (AAAI 2025 · similarity 0.89)
- Graph Diffusion Transformers for Multi-Conditional Molecular Generation (NeurIPS 2024 · similarity 0.89)
- All-atom Diffusion Transformers: Unified generative modelling of molecules and materials (ICML 2025 · 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.