Research map / Image generation / Autoregressive and latent generation
Generative Modeling of Full-Atom Protein Conformations using Latent Diffusion on Graph Embeddings
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 LD-FPG, a latent diffusion framework over ChebNet graph embeddings that generates all-atom conformational ensembles of the D2 dopamine receptor from MD data, comparing blind, sequential, and residue-based decoder pooling.
Problem
Generating diverse all-atom (including side-chain) conformational ensembles of dynamic proteins such as GPCRs, which existing generative models either simplify or restrict to backbone or static structures.
Approach
Encode MD frames with a ChebNet graph autoencoder into per-atom latents, pool them into a compact vector, train a DDPM on the pooled latents, and decode samples relative to a reference-structure latent conditioner back to Cartesian all-atom coordinates.
Key results
- Sequential pooling achieves all-atom lDDT 0.712 and backbone dihedral sum-JSD 0.0029 after diffusion on D2R-MD.
- Residue pooling gives the best side-chain dihedral sum-JSD (0.0224) and lowest average clashes (1145.6), close to the MD reference.
- BioEmu's generated A100 distribution (mean -17.19) diverges strongly from the D2R-MD reference (mean about -47.5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use sequential pooling instead of blind pooling | LD-FPG · D2R-MD | ΣJSD bb: 0.006582 -> 0.0029 | Helped |
| use sequential pooling instead of blind pooling | LD-FPG · D2R-MD | lDDT All: 0.719 -> 0.712 | Hurt |
| use residue pooling instead of sequential pooling | LD-FPG · D2R-MD | ΣJSD sc: 0.02895 -> 0.0224 | Helped |
| use residue pooling instead of sequential pooling | LD-FPG · D2R-MD | Avg. Clashes: 1220.5 -> 1145.6 | Helped |
| use residue pooling instead of blind pooling | LD-FPG · D2R-MD | lDDT All: 0.719 -> 0.6880 | Hurt |
| add dihedral fine-tuning (ablation) | LD-FPG blind pooling · D2R-MD | lDDT All: 0.719 -> 0.683 | Hurt |
| add dihedral fine-tuning (ablation) | LD-FPG blind pooling · D2R-MD | ΣJSD sc: 0.04185 -> 0.0409 | Helped |
| add dihedral fine-tuning (ablation) | LD-FPG blind pooling · D2R-MD | Avg. Clashes: 1350.5 -> 1340.9 | Helped |
Surprising findings
- Residue pooling achieves the best side-chain and landscape coverage despite the worst global backbone metrics after diffusion.
- Dihedral fine-tuning of blind pooling did not improve JSD and slightly reduced lDDT.
Limitations
- Generated structures show increased steric clashes compared to MD (about 1145-1350 vs about 1023).
- Residue pooling needs multi-epoch sampling to reach full diversity.
- Evaluated on a single protein system (D2R) with one MD trajectory.
Experimental setup
Datasets: D2R-MD
Benchmarks: D2R-MD
Metrics: lDDT TM-score JSD of dihedral angles steric clashes MSE A100 activation index
Baselines: BioEmu EnsembleVAE idpSAM ConfDiff P2DFlow MD-Gen Distributional Graphormer
Builds on: ChebNet DDPM Kabsch alignment
Closest papers by meaning
- Full-Atom Peptide Design with Geometric Latent Diffusion (NeurIPS 2024 · similarity 0.91)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.89)
- Few-Step Cofolding with All-Atom Flow Maps (NeurIPS 2026 · similarity 0.89)
- Exploit Your Latents: Coarse-Grained Protein Backmapping with Latent Diffusion Models (AAAI 2025 · similarity 0.89)
- Simultaneous Modeling of Protein Conformation and Dynamics via Autoregression (NeurIPS 2025 · similarity 0.88)
- 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.