Research map / Image generation / Autoregressive and latent generation
Exploit Your Latents: Coarse-Grained Protein Backmapping with Latent Diffusion 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 LatCPB, a latent diffusion model with discrete multimodal-prior autoencoder and latent contrastive learning for backmapping coarse-grained proteins to all-atom structures, evaluated on four PED intrinsically disordered protein structures.
Problem
Recovering high-resolution all-atom conformations from coarse-grained protein representations (backmapping), where existing generative methods model all-atom features directly in continuous space and struggle with multimodal discrete/continuous atomic features and conformational diversity.
Approach
Encode all-atom structures into discrete latent embeddings aligned with a learnable multimodal discrete prior (avoiding posterior collapse), run the conditional diffusion process in the continuous latent space conditioned on the CG structure, and add a latent contrastive loss to separate molecules and conformations.
Key results
- LatCPB achieves lowest RMSD on PED00055 (1.695 Å), PED00151 (1.539 Å) and PED00218 (1.563 Å); GenZProt is best on PED00090 RMSD (0.029 Å).
- LatCPB obtains the lowest GED on all four test structures (e.g., 0.015 on PED00151 vs 0.019 for GenZProt).
- Ablation shows discrete latent space plus contrastive learning gives the best RMSD/GED/KL across all four structures.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add discrete latent space (DL) (ablation) | LatCPB · PED00055, PED00090, PED00151, PED00218 | RMSD: 1.743 -> 1.736; 1.810 -> 1.806; 1.557 -> 1.549; 1.620 -> 1.595 | Helped |
| Add contrastive learning (CL) to DL (ablation) | LatCPB · PED00055, PED00090, PED00151, PED00218 | RMSD: 1.736 -> 1.695; 1.806 -> 1.758; 1.549 -> 1.539; 1.595 -> 1.563 | Helped |
| Add discrete latent space (DL) (ablation) | LatCPB · PED00055, PED00090, PED00151, PED00218 | GED: 0.042 -> 0.062; 0.108 -> 0.144; 0.010 -> 0.022; 0.033 -> 0.044 | Hurt |
| Add contrastive learning (CL) to DL (ablation) | LatCPB · PED00055, PED00090, PED00151, PED00218 | GED: 0.062 -> 0.039; 0.144 -> 0.055; 0.022 -> 0.015; 0.044 -> 0.020 | Helped |
| Add discrete latent space (DL) (ablation) | LatCPB · PED00055, PED00090, PED00151, PED00218 | KL: 0.212 -> 0.229; 0.942 -> 0.443; 0.212 -> 0.167; 0.094 -> 0.168 | Mixed |
| Add contrastive learning (CL) to DL (ablation) | LatCPB · PED00055, PED00090, PED00151, PED00218 | KL: 0.229 -> 0.153; 0.443 -> 0.124; 0.167 -> 0.108; 0.168 -> 0.087 | Helped |
| Use LatCPB instead of GenZProt | protein backmapping models · PED00055, PED00151, PED00218 | RMSD: 1.871 -> 1.695; 1.711 -> 1.539; 1.727 -> 1.563 | Helped |
| Use LatCPB instead of GenZProt | protein backmapping models · PED00090 | RMSD: 0.029 -> 1.758 | Hurt |
| Use LatCPB instead of GenZProt | protein backmapping models · PED00055, PED00090, PED00151, PED00218 | GED: 0.054 -> 0.039; 0.070 -> 0.055; 0.019 -> 0.015; 0.030 -> 0.020 | Helped |
| Use LatCPB instead of CGVAE | protein backmapping models · PED00055, PED00090, PED00151, PED00218 | KL: 0.369 -> 0.153; 0.424 -> 0.124; 0.316 -> 0.108; 0.331 -> 0.087 | Helped |
Surprising findings
- GenZProt achieves a much lower RMSD (0.029 Å) on PED00090 than LatCPB (1.758 Å), contrary to the overall trend.
- Discrete latent space alone yields only limited improvement; the gains depend on combining it with contrastive learning.
Limitations
- Relies on corresponding CG structures as input and evaluates on only four PED IDP structures.
- Unclear whether generated all-atom structures have reasonable binding energies.
- Terminal residues are masked and cannot be reconstructed.
Experimental setup
Datasets: PED protein dataset
Benchmarks: PED00055 PED00090 PED00151 PED00218
Metrics: RMSD GED KL divergence
Baselines: CGVAE GenZProt
Builds on: GenZProt CGVAE latent diffusion models AlphaFold2
Closest papers by meaning
- Generative Modeling of Full-Atom Protein Conformations using Latent Diffusion on Graph Embeddings (NeurIPS 2025 · similarity 0.89)
- Equivariant Blurring Diffusion for Hierarchical Molecular Conformer Generation (NeurIPS 2024 · similarity 0.87)
- All-atom Diffusion Transformers: Unified generative modelling of molecules and materials (ICML 2025 · similarity 0.86)
- Full-Atom Peptide Design with Geometric Latent Diffusion (NeurIPS 2024 · similarity 0.86)
- Steering Masked Discrete Diffusion Models via Discrete Denoising Posterior Prediction (ICLR 2025 · similarity 0.86)
- Unlocking hidden biomolecular conformational landscapes in diffusion models at inference time (NeurIPS 2025 · similarity 0.85)
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.