Protenix

Model Introduction

Protenix is an AlphaFold 3-like model for predicting the structures of biomolecular complexes comprising proteins, nucleic acids, ligands, and other molecules. It predicts three-dimensional structures from molecular inputs described in JSON and locally generated MSA features, producing CIF structure files together with associated confidence scores.

Model Description

Protenix uses a Pairformer representation network and an atom-level diffusion Transformer to predict the structures of proteins, nucleic acids, ligands, and their complexes within a unified framework.

This Hugging Face package is designed for immediate use after download, rapid local validation, and automated execution in OneCode. All code, configuration files, example inputs, and pretrained weights are included in the model package directory.

Use Cases

Scenario Description
Complex structure prediction Takes a Protenix JSON file and locally generated MSA features as input and outputs predicted structures in CIF format and confidence scores in JSON format.
Full Hugging Face package validation Uses the included config/, models/, scripts/, examples/, and weight/ directories to run preflight checks and inference directly.
Fine-tuning pipeline validation Starts the fine-tuning workflow using the included weights and ft_datasets/finetune_subset.txt.
Training pipeline setup Supports single-GPU training once the complete Protenix dataset is provided.

Usage Guide

1. OneCode Usage

Try one-click AI4S development in the OneCode online environment:

Try one-click AI4S development

2. Manual Installation and Usage

Hardware Requirements

  • GPU or DCU is recommended.
  • A CPU can be used for basic connectivity checks, but execution will be slow.
  • DCU users must install DTK in advance. DTK 25.04.2 or later is recommended, or a OneScience-recommended version matching the current cluster.

Software Requirements

DCU users who need additional information about the adaptation may contact liubiao@sugon.com.

Environment Check

  • NVIDIA GPU:
nvidia-smi
  • Hygon DCU:
hy-smi

Quick Start

1. Install the Runtime Environment

conda create -n onescience311 python=3.11 -y
conda activate onescience311
pip install onescience[bio] -i http://mirrors.onescience.ai:3141/pypi/simple/  --trusted-host mirrors.onescience.ai

2. Download the Model Package and Dataset

hf download --model OneScience-Sugon/protenix --local-dir ./protenix
hf download --model OneScience-Sugon/protenix_dataset --local-dir ./protenix_dataset

Model Weights

The trained model weights are included in the weights/ directory and can be used immediately after the model package is downloaded.

3. Run Preflight Checks

To check only the model package, weights, and local imports, run:

python scripts/preflight.py --strict-weights --strict-imports

If the full dataset is already prepared:

export DATA_ROOT_DIR=../bio_protenix_dataset
python scripts/preflight.py --strict-weights --strict-imports --strict-data

4. Run Inference

export DATA_ROOT_DIR=../bio_protenix_dataset
bash scripts/inference_unified_demo.sh

Default output directory:

output_unified/7r6r/seed_101/predictions/

5. Training and Fine-Tuning

Training:

export DATA_ROOT_DIR=../bio_protenix_dataset
bash scripts/train_demo.sh

Fine-tuning:

export DATA_ROOT_DIR=../bio_protenix_dataset
bash scripts/finetune_demo.sh

OneScience Official Information

Citation & License

The Protenix project, including its code and model parameters, is available under the Apache License 2.0 and may be used free of charge for both academic research and commercial purposes.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support