Research map / Image generation / Autoregressive and latent generation
PolyConf: Unlocking Polymer Conformation Generation through Hierarchical Generative Models
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 PolyConf, a hierarchical generative model that generates polymer conformations by sampling repeating-unit conformations with a masked autoregressive model plus diffusion, then assembling them via SO(3) diffusion rotations, and introduces PolyBench, an MD-derived polymer conformation dataset.
Problem
Generating 3D conformations of polymers is hard: existing small-molecule and protein conformation generators do not scale to polymers' size, flexibility and chemical diversity, and polymer conformation datasets are scarce.
Approach
Decompose a polymer into repeating-unit conformations generated in random order by a masked autoregressive model with a per-unit torsion-space diffusion head, then generate each unit's rotation with an SO(3) diffusion model and assemble units by aligning overlapping atoms.
Key results
- On PolyBench, PolyConf reaches S-MAT-R mean 35.021 vs 53.210 for the best baseline TorsionalDiff, and E-MAT-R mean 0.933 vs 2.605.
- In the doubled-length scalability test, PolyConf gets S-MAT-R mean 65.040 vs 119.289 for TorsionalDiff.
- Average generation time is 0.40 min for PolyConf vs 0.45 min for TorsionalDiff and 3.54 min for GeoDiff.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use hierarchical PolyConf instead of TorsionalDiff for polymer conformation generation | PolyConf · PolyBench | S-MAT-R mean: 53.210 -> 35.021 | Helped |
| Use hierarchical PolyConf instead of TorsionalDiff for polymer conformation generation | PolyConf · PolyBench | S-MAT-P mean: 70.679 -> 46.861 | Helped |
| Use hierarchical PolyConf instead of TorsionalDiff for polymer conformation generation | PolyConf · PolyBench | E-MAT-R mean: 2.605 -> 0.933 | Helped |
| Use hierarchical PolyConf instead of TorsionalDiff for polymer conformation generation | PolyConf · PolyBench | E-MAT-P mean: 8.402 -> 6.191 | Helped |
| Use PolyConf with doubled repeating units instead of TorsionalDiff | PolyConf · PolyBench, doubled repeating units | S-MAT-R mean: 119.289 -> 65.040 | Helped |
| Use PolyConf with doubled repeating units instead of TorsionalDiff | PolyConf · PolyBench, doubled repeating units | S-MAT-P mean: 146.816 -> 84.626 | Helped |
| Use PolyConf with doubled repeating units instead of TorsionalDiff | PolyConf · PolyBench, doubled repeating units | E-MAT-R mean: 5.219 -> 1.259 | Helped |
| Use PolyConf with doubled repeating units instead of TorsionalDiff | PolyConf · PolyBench, doubled repeating units | E-MAT-P mean: 11.692 -> 5.785 | Helped |
Surprising findings
- MCF gives energy errors above 10^10 on PolyBench, indicating it fails on polymer-scale systems.
- TorsionalDiff was given the reference initial polymer structure as input, a biased advantage the authors acknowledge.
Limitations
- Only linear homopolymers are studied; copolymers and blends are out of scope.
- The dataset is force-field MD based (GAFF) rather than DFT, so energies are approximate.
- The benchmark uses PET-based polymers as the main dataset.
Experimental setup
Datasets: PolyBench (MD-derived polymer conformations, PET-based)
Benchmarks: PolyBench
Metrics: S-MAT-R S-MAT-P E-MAT-R E-MAT-P S-COV-R S-COV-P inference time
Baselines: GeoDiff TorsionalDiff MCF ET-Flow
Builds on: MAR (masked autoregressive diffusion) GeoDiff TorsionalDiff MolCLR Uni-Mol
Closest papers by meaning
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.87)
- Enhancing Generalizability in Molecular Conformation Generation with METRIZATION-Informed Geometric Diffusion Pretraining (AAAI 2025 · similarity 0.85)
- Protein Conformation Generation via Force-Guided SE(3) Diffusion Models (ICML 2024 · similarity 0.85)
- DiSCO: Diffusion Schrödinger Bridge for Molecular Conformer Optimization (AAAI 2024 · similarity 0.85)
- Efficient Molecular Conformer Generation with SO(3)-Averaged Flow Matching and Reflow (ICML 2025 · similarity 0.84)
- Equivariant Blurring Diffusion for Hierarchical Molecular Conformer Generation (NeurIPS 2024 · similarity 0.84)
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.