Research map / Image generation / Autoregressive and latent generation
CarbonNovo: Joint Design of Protein Structure and Sequence Using a Unified Energy-based Model
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 CarbonNovo, a unified energy-based model that jointly generates protein backbone structure and sequence via SE(3) score-based diffusion plus an MRF sequence energy, with a protein language model fed in through recycling.
Problem
Two-stage de novo protein design (structure diffusion followed by sequence design) suffers from sequence-model overfitting to crystal structures and no feedback from sequence errors to structure generation.
Approach
Combines an SE(3) score-based diffusion model for structure energy with a Markov Random Field sequence energy, and lets the structure and sequence modules communicate at every diffusion step with network recycling; a pretrained protein language model (ESM-2) is integrated as evolutionary constraint.
Key results
- CarbonNovo reaches 81.38% designability fraction (ESMFold) vs 69.81% for RFdiffusion and 25.62% for Chroma, with scRMSD 1.943 Å.
- Novelty of 43.15% (ESMFold) vs 36.50% for RFdiffusion; CarbonNovo-MPNN variant with two-stage design reaches 73.16% designability.
- Rosetta energy -2.83 (lowest among methods) and log-likelihood -2.44 under ProGen2; removing the PLM drops designability fraction to 51.75%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| jointly generate protein sequence and structure rather than use ProteinMPNN for sequence design (ablation) | CarbonNovo · De novo protein design | Fraction: 73.16% / 70.15% -> 81.38% / 77.38% | Helped |
| include the protein language model (ablation) | CarbonNovo · De novo protein design | Fraction: 51.75% -> 81.38% | Helped |
| include the MRF sequence design module (ablation) | CarbonNovo · De novo protein design | Fraction: 74.55% -> 81.38% | Helped |
| pre-train the model (ablation) | CarbonNovo · De novo protein design | Fraction: 74.68% -> 81.38% | Helped |
| retain the listed PLM, MRF, and pre-training components (ablation) | CarbonNovo · De novo protein design | Fraction: 45.16% -> 81.38% | Helped |
| include auxiliary training loss (ablation) | CarbonNovo · De novo protein design | Fraction: 35.55% -> 81.38% | Helped |
Surprising findings
- Removing the pretrained language model causes the largest drop in designability among ablations, larger than removing the MRF or pretraining.
- Chroma's designability fraction in the reported table (25.62%) is far below RFdiffusion's, despite being cited as a strong baseline.
Limitations
- Evaluation limited to in silico metrics; no wet-lab validation.
- Focuses on monomer design; complexes and conditional design (binders) left for future work.
Experimental setup
Datasets: FrameDiff structure dataset PDB (pre-Aug 2021, ProteinMPNN split)
Benchmarks: designability (scRMSD, scTMscore, fraction) with ESMFold and OmegaFold novelty via Foldseek against PDB diversity (pairwise TM-score) Rosetta energy sequence log-likelihood (ProGen2)
Metrics: scRMSD scTMscore designability fraction novelty diversity Rosetta energy log-likelihood
Baselines: RFdiffusion Chroma FrameDiff-ICML FrameDiff-Improved FrameFlow Genie ProteinMPNN
Builds on: FrameDiff CarbonDesign AlphaFold2 Discrete Metropolis-Hastings Langevin
Closest papers by meaning
- Proteina: Scaling Flow-based Protein Structure Generative Models (ICLR 2025 · similarity 0.88)
- ProteinBench: A Holistic Evaluation of Protein Foundation Models (ICLR 2025 · similarity 0.87)
- Diffusion on Language Model Encodings for Protein Sequence Generation (ICML 2025 · similarity 0.87)
- DPLM-2: A Multimodal Diffusion Protein Language Model (ICLR 2025 · similarity 0.87)
- Protein Design with Dynamic Protein Vocabulary (NeurIPS 2025 · similarity 0.87)
- Steering Generative Models with Experimental Data for Protein Fitness Optimization (NeurIPS 2025 · similarity 0.87)
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.