Proteina Complexa Agent Skill

Proteina-Complexa Backbone Generation

Plain-language role: Use this skill when you want a flow-based backbone generator with fold-class conditioning, especially for exploratory de novo design.

Source Notes

• Public upstream reference: NVIDIA-Digital-Bio/proteina
• Publicly described as a large-scale flow-based protein backbone generator with hierarchical fold class conditioning
• Upstream setup and weights may change over time, so verify the current README and license before running
• Check the upstream NVIDIA license before commercial use or redistribution of model artifacts

Prerequisites

| Requirement | Minimum | Recommended |

|-------------|---------|-------------|

| Python | 3.10+ | 3.11 |

| CUDA | 12.0+ | 12.1+ |

| GPU VRAM | 24GB | 40GB+ |

| Environment manager | conda | mamba or micromamba |

How to Run

Option 1: Upstream Proteina environment

git clone https://github.com/NVIDIA-Digital-Bio/proteina.git
cd proteina
mamba env create -f environment.yaml
conda activate proteina_env
pip install -e .

Create a .env file in the repository root:

echo "DATA_PATH=/path/to/proteina-data" > .env

Additional files

The upstream project documents extra data and weight bundles that must live under DATA_PATH.

At minimum, verify:

• metric feature files
• model weights
• CATH label mapping files
• dataset index files if you plan to train or evaluate

Recommended Use Pattern

1. Start from backbone generation

Use Proteina-Complexa when the main task is generating diverse backbones, not sequence optimization.

2. Prefer fold-conditioned exploration

The upstream model is especially useful when you want:

• hierarchical fold control
• long-chain generation
• comparison against diffusion-based backbone generators

3. Hand off to sequence design

After generating promising backbones:

• use proteinmpnn for general inverse folding
• use solublempnn when expression robustness matters more

4. Validate and filter

After sequence design:

• use chai1-structure-prediction, boltz-structure-prediction, or alphafold2-multimer
• use protein-design-qc for filtering and ranking

Typical Workflow

Target goal
  -> Proteina-Complexa backbone generation
  -> ProteinMPNN / SolubleMPNN sequence design
  -> Chai / Boltz / AlphaFold validation
  -> Protein Design QC

When to Prefer This Over Other Tools

| Need | Prefer |

|------|--------|

| Maximum backbone diversity with established community recipes | rfdiffusion |

| All-atom generation with side-chain awareness | boltzgen |

| Flow-based backbone generation with fold conditioning | proteina-complexa |

| End-to-end integrated binder pipeline | bindcraft |

Key Ideas to Preserve

• Keep fold-conditioning choices explicit
• Record which checkpoint and config produced each backbone batch
• Separate backbone-generation artifacts from downstream sequence-design artifacts
• Treat generated backbones as candidates that still require validation and QC

Common Mistakes

• Treating Proteina-Complexa as a sequence-design tool
• Skipping required upstream weight and data bundles
• Comparing outputs against RFdiffusion or BoltzGen without matching length and conditioning settings
• Moving generated backbones directly to experiments without refolding validation

Troubleshooting

| Error | Likely cause | Fix |

|-------|--------------|-----|

| Missing DATA_PATH files | Required upstream bundles not downloaded | Re-check upstream setup and place files under the documented directory tree |

| CUDA OOM | Backbone length or batch too large | Reduce batch size or use a larger GPU |

| Config mismatch | Wrong checkpoint/config pair | Keep checkpoint, config, and conditioning mode aligned |

| Weak downstream foldability | Backbone exploration too unconstrained | Tighten fold conditioning and validate more aggressively |

Inputs

• A backbone-generation objective such as fold-conditioned sampling, long-chain exploration, or de novo backbone discovery.
• A configured Proteina-style environment with checkpoints, configs, and required data bundles available under the configured data path.
• Optional fold-class or topology guidance for controlled generation.

Outputs

• Generated protein backbone candidates suitable for downstream inverse folding.
• Run metadata describing checkpoint choice, conditioning mode, and generation settings.
• Backbone batches ready for sequence design with proteinmpnn or solublempnn.

Next Step

Send promising backbones to proteinmpnn or solublempnn, then validate them structurally and filter with protein-design-qc.