bytedance/Protenix

Protenix – Open‑source high‑accuracy biomolecular structure prediction

What it is – Protenix is a deep‑learning system for predicting the three‑dimensional structures of proteins (and related biomolecules such as RNA‑protein complexes). It builds on the AlphaFold 3 architecture, reproducing and extending it with newer training data, diffusion‑based inference tricks, and optional constraints (e.g., atom‑level contacts or ligand pockets). The project ships a Python package, a command‑line interface, and a public web server for easy inference.

Why it matters – Accurate structure prediction is a cornerstone of modern computational biology, enabling drug discovery, enzyme engineering, and basic research. Protenix claims to surpass AlphaFold 3 on several benchmarks (especially antibody‑antigen complexes) while staying fully open‑source under Apache 2.0. It also provides lightweight “Mini” variants for faster, cheaper inference.

Quick start (for end‑users)

# Install the latest release from PyPI
pip install --upgrade protenix --index-url https://pypi.org/simple

# Predict a structure from a JSON description
protenix pred -i examples/input.json -o ./output \
    -n protenix_base_default_v1.0.0   # choose any released model name

The JSON format is documented in docs/infer_json_format.md and supports optional constraints such as contact maps or pocket definitions.

Main components

Component	Purpose	Status
protenix (Python package)	Core inference library, CLI, and utilities for data preprocessing.	Released on PyPI
Protenix Web Server	Hosted service (`https://protenix‑server.com`) that lets anyone upload a job through a browser.	Public
PXDesign	Down‑stream de‑novo protein‑binder design suite built on the Protenix foundation model.	Open‑source, separate repo
PXMeter	Benchmarking toolkit with curated datasets and evaluation scripts.	Open‑source
Protenix‑Dock	Classical (non‑ML) protein‑ligand docking implementation for rigid docking.	Open‑source

Model catalogue (as of the README)

Model name	Parameters	Training data cutoff	Features	Release
`protenix‑v2`	464 M	2021‑09‑30	MSA, RNA‑MSA, template support, diffusion inference	2026‑04‑08 (weights proprietary)
`protenix_base_default_v1.0.0`	368 M	2021‑09‑30	MSA, RNA‑MSA, template	2026‑02‑05
`protenix_base_20250630_v1.0.0`	368 M	2025‑06‑30	MSA, RNA‑MSA, template	2026‑02‑05
`protenix_base_default_v0.5.0`	368 M	2021‑09‑30	MSA only	2025‑05‑30

Note: The newest protenix‑v2 weights are not released under an open licence; they are listed as “proprietary” and cannot be redistributed.

Highlights & technical tricks

Diffusion‑based inference optimizations (v0.7.0): shared‑variable caching, kernel fusion, TF‑32 acceleration – dramatically lower wall‑time per token.
Constraint‑aware prediction: users can feed atom‑level contacts or pocket definitions to bias the model toward physically plausible conformations.
Mini model: a smaller variant that keeps most accuracy while cutting inference cost, useful for large‑scale screens.
Full training‑data pipeline: scripts for preparing MSAs, templates, and RNA alignments are open‑sourced (docs/prepare_training_data.md).
ColabFold‑compatible MSA search: one‑click local MSA generation for users familiar with the popular ColabFold workflow.

Benchmarks (as reported)

Protenix‑v2 improves DockQ > 0.23 success rates by 9‑13 percentage points over v1 on antibody‑antigen sets, and reaches comparable performance with only 5 random seeds versus 1000 seeds for v1.
Protenix‑v1 matches or exceeds AlphaFold 3 on the same data cutoff, model size, and inference budget, with additional gains when the sampling budget is increased.

Full benchmark tables are in docs/model_1.0.0_benchmark.md and the accompanying PDF reports.

Development & community

Contribution guide: CONTRIBUTING.md – use pre‑commit hooks for style checks.
Communication channels: Twitter, Slack, WeChat (links in the badge bar).
Related open‑source projects: PXDesign, PXMeter, Protenix‑Dock (all under the same ByteDance organization).
License: Apache 2.0 for code and released model parameters (except the proprietary v2 weights).

When to use Protenix

You need a free, open‑source alternative to AlphaFold 3 for academic or commercial projects.
You want to experiment with constraint‑driven folding (e.g., enforce known contacts or ligand pockets).
You are building downstream tools such as binder design (PXDesign) or benchmark suites (PXMeter).
You have limited compute and prefer the Mini variant for high‑throughput screening.

Getting help

Open an issue on the GitHub repo for bugs or feature requests.
For security concerns, contact ByteDance Security via the provided email (do not open a public issue).
Academic citations are provided for v1, v2, and the original reproduction paper; include them when publishing results.

Bottom line – Protenix is a genuine, actively maintained open‑source protein structure prediction platform that pushes the state of the art beyond AlphaFold 3 while offering a suite of auxiliary tools for design, evaluation, and docking.