Prompteina

Overview

Prompteina generates all-atom protein structures from natural-language prompts. For binder design, provide a target structure as well.

Choose the checkpoint for your task:

• prompteina-monomer.ckpt for monomer generation
• prompteina-binder.ckpt for binder design

See jleechung/prompteina for the code.

Download

hf download jleechung/prompteina --local-dir ./weights

Usage

Set the checkpoint directory, checkpoint filename, and autoencoder path:

export CKPT_PATH=./weights
export CKPT_NAME=prompteina-binder.ckpt
export AE_CKPT=./ckpts/complexa_ae.ckpt

The autoencoder is not included with these weights. Download it from the code repository with:

./env/download_startup.sh --complexa

Put your prompts in a CSV file with a description column, then run:

TEXTCOND_CSV=prompts.csv python -m proteinfoundation.generate \
    --config-name text_binder_generate generation.task_name=02_PDL1

To load a checkpoint in Python, pass strict=False:

from proteinfoundation.proteina import Proteina

model = Proteina.load_from_checkpoint(
    "weights/prompteina-binder.ckpt",
    strict=False,
    autoencoder_ckpt_path="ckpts/complexa_ae.ckpt",
)

The checkpoints do not contain the frozen Qwen3-Embedding-8B and Qwen3-8B text encoders. The code downloads them from the Hub when it builds the model.

These checkpoints are for inference only and cannot resume training. See SHA256SUMS for file checksums.

License

Both checkpoints are Derivative Models of NVIDIA Proteina-Complexa under the NVIDIA Open Model License.

Licensed by NVIDIA Corporation under the NVIDIA Open Model License