What is Protenix-v1
Protenix-v1 is a biomolecular structure prediction model open-sourced by the ByteDance Seed team. It is the first fully open-source model that achieves or even surpasses AF3 performance under conditions strictly aligned with the AlphaFold 3 data cutoff date (2021-09-30), model scale, and inference budget. The model reproduces AF3's Inference-Time Scaling capability; increasing the sampling budget can bring logarithmic-linear performance improvements, and the antibody-antigen prediction DockQ success rate can increase from 36% to 47.68%. The model adopts a dual-version strategy: a standard version for fair academic comparison, and a data-extended version (2025-06-30) for practical drug discovery scenarios. The model also supports new features such as protein templates and RNA MSA, and is accompanied by the release of the evaluation toolkit PXMeter to address the confusion in benchmarking.
Main features of Protenix-v1
- Complex structure prediction: Supports high-precision 3D structure prediction for multiple types of biomolecular complexes, including protein-protein, antibody-antigen, protein-nucleic acid, and protein-small molecule.
- RNA MSA support: Integrates RNA multiple sequence alignment functionality, can capture RNA evolutionary conservation and secondary structure features, and significantly improves the prediction accuracy of protein-RNA complexes.
- Template information integration: Introduces known protein structure templates, enhances prediction reliability through homologous information, and simultaneously stabilizes the convergence of the model training process.
- Inference-time scaling capability: Supports performance improvement by increasing the number of sampling seeds. Users can flexibly trade off between computational cost and prediction accuracy, which is especially suitable for highly difficult docking tasks.
- Introduction of physical constraints: Provides atomic-level contact constraints and binding pocket constraints, allowing the incorporation of prior experimental data or physical knowledge to guide structure generation.
- Dual-version model strategy: Simultaneously releases an academic benchmarking version (data cutoff 2021-09-30) and an application-optimized version (data cutoff 2025-06-30), balancing fair comparison and practical needs.
Technical principles of Protenix-v1
- Base architecture: Based on AlphaFold 3's end-to-end diffusion architecture, including an MSA encoder, a Pairformer relational modeling module, and a diffusion generation module, directly outputting atomic coordinates through an iterative denoising process.
- Inference-Time Scaling: The core innovation lies in reproducing AF3's inference-time scaling property. By sampling hundreds of candidate structures in parallel and combining them with a scoring function to select the optimal solution, the prediction success rate of flexible docking tasks such as antibody-antigen grows logarithmically-linearly with computational investment.
- Template integration mechanism: Uses tools such as Kalign to search for homologous templates, injects template features into the Pairformer's pair representation, improves the ability to recognize conserved fold families, and alleviates instability in the early training stage.
- RNA feature engineering: Extends the MSA module to support RNA sequence alignment, encodes nucleotide evolutionary information and covariation signals, and compensates for the shortcomings of traditional protein-centric models in modeling nucleic acid interactions.
- Lightweight design (Protenix-Mini): Reduces computational complexity through a linear attention mechanism, introduces a protein language model to enable single-sequence inference, eliminates the overhead of MSA database retrieval, and is suitable for large-scale virtual screening scenarios.
Protenix-v1 project address
- GitHub repository：https://github.com/bytedance/Protenix
- Technical paper：https://github.com/bytedance/Protenix/blob/main/docs/PTX_V1_Technical_Report_202602042356.pdf
Application scenarios of Protenix-v1
- Drug discovery and development: Antibody-antigen complex prediction is a core advantage scenario of Protenix-v1. The inference-time scaling capability can significantly improve the structural modeling accuracy of highly challenging immunotherapy targets, providing a reliable foundation for antibody engineering and epitope design.
- Protein design: By constraining the binding pocket and key contact residues through functional specification, it guides the rational design of high-affinity molecules; the Protenix-Mini series can be used for rapid screening of massive candidate sequences.
- Structural biology research: It provides predictive models for experimental structural biology, assists in interpreting low-resolution crystal structures or cryo-EM density maps, reduces experimental trial-and-error costs, and accelerates target structure characterization.
- RNA-related research: The protein-RNA complex prediction capability is applicable to frontier fields such as RNA interference, CRISPR systems, and RNA drug delivery, filling the gap of traditional tools in modeling nucleic acid-protein interactions.
- Large-scale virtual screening: The lightweight Mini version supports high-throughput docking evaluation, enabling rapid filtering of compound libraries or protein variants in the early stages of drug discovery, forming a "coarse screening-refinement" combined workflow with the full-precision model.