ByteDance has recently unveiled Protenix-v1, a cutting-edge model that aims to replicate the accuracy of AlphaFold3 in biomolecular structure prediction. This comprehensive model, released with code and model parameters under the Apache 2.0 license, is designed to excel in predicting the structures of proteins, DNA, RNA, and ligands while maintaining an open and extensible framework for both research and production purposes.
Protenix-v1, also known as ‘Protenix: Protein + X,’ serves as a foundational model for high-precision biomolecular structure prediction. It is capable of predicting detailed all-atom 3D structures for complex biomolecules, including proteins, nucleic acids (DNA and RNA), and small-molecule ligands. The model is a faithful reproduction of AlphaFold3, implementing its diffusion architecture for all-atom complexes and presenting it in a trainable PyTorch codebase.
The Protenix project encompasses a full stack of components, including training and inference code, pre-trained model weights, data and MSA pipelines, and a browser-based Protenix Web Server for interactive usage. This comprehensive approach ensures that users have access to all necessary tools and resources for effective biomolecular structure prediction.
One of the key highlights of Protenix-v1 is its ability to achieve AlphaFold3-level performance under matched constraints. By aligning with critical parameters such as the training data cutoff, model scale, and inference budget of AlphaFold3, Protenix-v1 has emerged as the first fully open-source model to outperform AlphaFold3 across diverse benchmark sets. This alignment enables fair comparisons and validates the model’s high accuracy in biomolecular structure prediction.
To support the claims of Protenix-v1’s performance, ByteDance has introduced PXMeter v1.0.0, an evaluation toolkit and dataset suite for transparent benchmarking on over 6,000 complexes. This toolkit includes a curated benchmark dataset, time-split and domain-specific subsets, and a unified evaluation framework for computing metrics such as complex LDDT and DockQ across different models. The research paper accompanying PXMeter, titled ‘Revisiting Structure Prediction Benchmarks with PXMeter,’ compares Protenix, AlphaFold3, Boltz-1, and Chai-1 on curated tasks, highlighting the impact of dataset designs on model ranking and performance evaluation.
In addition to Protenix-v1, ByteDance has developed a small ecosystem of related projects that complement the biomolecular structure prediction capabilities of Protenix. These projects include PXDesign, Protenix-Dock, and Protenix-Mini, each focusing on specific aspects such as binder design, protein-ligand docking, and lightweight model variants for reduced inference costs. Together, these components cover a wide range of structure prediction, docking, and design tasks, sharing interfaces and formats for seamless integration into downstream pipelines.
In conclusion, Protenix-v1 represents a significant advancement in biomolecular structure prediction, offering high accuracy and transparency in a fully open-source model. By aligning with AlphaFold3’s constraints and introducing tools like PXMeter for benchmarking, ByteDance has established Protenix-v1 as a reliable and effective solution for researchers and practitioners in the field of computational biology.
Be the first to comment