Research map / Image generation / Autoregressive and latent generation
Protein Autoregressive Modeling via Multiscale Structure Generation
Yanru Qu, Cheng-Yen Hsieh, Zaixiang Zheng, Ge Liu, Quanquan Gu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add noisy context learning (ablation) | PAR (teacher forcing) · 60M PAR unconditional | sc-RMSD: 2.20 -> 1.58 | Improved |
| add scheduled sampling on top of noisy context learning (ablation) | PAR + NCL · 60M PAR unconditional | sc-RMSD: 1.58 -> 1.48 | Improved |
| add noisy context learning (ablation) | PAR (teacher forcing) · 60M PAR unconditional | FPSD vs AFDB: 37.64 -> 23.69 | Improved |
| add SDE at first scale and ODE at later scales with 2 steps (ablation) | PAR all-SDE · length 150 | time (s): 312 -> 67 (designability 97% -> 97%) | Improved |
| use ODE sampling at all scales instead of SDE at first scale (ablation) | PAR S/O/O · length 150 | designability (%): 96% -> 28% | Hurt |
| halve single-scale SDE steps from 400 to 200 (ablation) | Proteina · length 150 | designability (%): 97% -> 89% | Hurt |
| define scales by length instead of by ratio (ablation) | PAR 60M · unconditional | FPSD vs PDB: 310.64 -> 279.63 | Improved |
| increase scales from two to three (ablation) | PAR 60M · unconditional | designability (%): 83.0 -> 85.0 | Improved |
| increase scales from three to four (ablation) | PAR 60M · unconditional | designability (%): 85.0 -> 77.8 | Hurt |
| encode prior scales with AR transformer instead of direct input (ablation) | PAR 60M · prompted generation (64,128] | RMSD vs reference: 3.38 -> 2.72 | Improved |
| increase AR transformer size from 60M to 400M (ablation) | PAR (decoder 60M) · unconditional | designability (%): 92.60 -> 87.80 | Hurt |
| increase decoder size from 60M to 400M (ablation) | PAR (AR 60M) · unconditional | designability (%): 92.60 -> 96.00 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Proteina (200M) — Proteina: Scaling Flow-based Protein Structure Generative Models (ICLR 2025) · better on 5 of 5 shared results
- Genie2 (16M) — GENIE: Higher-Order Denoising Diffusion Solvers (NeurIPS 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- TEMPO: Temporal Multi-scale Autoregressive Generation of Protein Conformational Ensembles (NeurIPS 2025 · similarity 0.89)
- Simultaneous Modeling of Protein Conformation and Dynamics via Autoregression (NeurIPS 2025 · similarity 0.88)
- Proteina: Scaling Flow-based Protein Structure Generative Models (ICLR 2025 · similarity 0.88)
- ProtPainter: Draw or Drag Protein via Topology-guided Diffusion (ICLR 2025 · similarity 0.87)
- ProtComposer: Compositional Protein Structure Generation with 3D Ellipsoids (ICLR 2025 · similarity 0.86)
- Steering Masked Discrete Diffusion Models via Discrete Denoising Posterior Prediction (ICLR 2025 · similarity 0.86)
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.
