Protenix-v1 - An open-source biomolecular structure prediction model from the ByteSeed team. - AiBoss

What is Protenix-v1?

Protenix-v1 is an open-source biomolecular structure prediction model from ByteDance's Seed team. It is the first fully open-source model to achieve performance equal to or even surpass AF3 under strict alignment with the AlphaFold 3 data cutoff date (September 30, 2021), model size, and inference budget. The model replicates AF3's inference-time scaling capability, and increasing the sampling budget leads to log-linear performance improvements. The antibody-antigen prediction DockQ success rate increases from 36% to 47.68%. The model employs a dual-version strategy: a standard version for fair academic comparisons, and a data-expanded version (June 30, 2025) for practical drug discovery scenarios. The model also supports new features such as protein templates and RNA MSA, and a PXMeter evaluation toolkit is provided to address benchmarking issues.

Main features of Protenix-v1

Complex structure predictionIt supports high-precision three-dimensional structure prediction of various biomolecular complexes, such as protein-protein, antibody-antigen, protein-nucleic acid, and protein-small molecule complexes.
RNA MSA supportIt integrates RNA multiple sequence alignment functionality, which can capture RNA evolutionary conservation and secondary structure features, significantly improving the accuracy of protein-RNA complex prediction.
Template information integrationIntroducing known protein structure templates enhances prediction reliability through homology information and stabilizes the convergence of the model training process.
Extended capabilities during reasoningIt supports performance improvement by increasing the number of sampling seeds, allowing users to flexibly balance computational cost and prediction accuracy, making it particularly suitable for complex docking tasks.
Physical constraints introducedIt provides atomic-level contact constraints and combined pocket constraints, allowing the incorporation of prior experimental data or physical knowledge to guide structure generation.
Dual-version model strategySimultaneously, an academic benchmark version (data as of September 30, 2021) and an application-optimized version (data as of June 30, 2025) are released to balance fair comparison with practical needs.

Technical principles of Protenix-v1

InfrastructureBased on AlphaFold 3, the end-to-end diffusion architecture includes an MSA encoder, a Pairformer relation modeling module, and a diffusion generation module, which directly outputs atomic coordinates through an iterative denoising process.
Inference-Time ScalingThe core innovation lies in replicating the inference-time extension characteristics of AF3. 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 increases logarithmically with computational input.
Template integration mechanismWe use tools such as Kalign to search for homologous templates and inject template features into the pair representation of Pairformer to improve the ability to identify conservative folding families and alleviate instability in the early training stage.
RNA feature engineeringExtend the MSA module to support RNA sequence alignment, encode nucleotide evolution information and covariation signals, and make up for the shortcomings of traditional protein-centric models in nucleic acid interaction modeling.
Lightweight design (Protenix-Mini)It reduces computational complexity by using a linear attention mechanism, introduces a protein language model to achieve single-sequence reasoning, eliminates MSA database retrieval overhead, and is suitable for large-scale virtual screening scenarios.

Project address for Protenix-v1

GitHub repositoryhttps://github.com/bytedance/Protenix
Technical Papers: https://github.com/bytedance/Protenix/blob/main/docs/PTX_V1_Technical_Report_202602042356.pdf

Application scenarios of Protenix-v1

Drug discovery and developmentAntibody-antigen complex prediction is a core strength of Protenix-v1. Its inference-time scalability can significantly improve the accuracy of structural modeling for challenging immunotherapy targets, providing a reliable foundation for antibody engineering and epitope design.
Protein designBy specifying binding pockets and key contact residues through constraint functions, 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 researchIt provides predictive models for experimental structural biology, assists in the analysis of low-resolution crystal structures or cryo-electron microscopy density maps, reduces experimental trial-and-error costs, and accelerates target structure characterization.
RNA-related researchThe protein-RNA complex prediction function is applicable to cutting-edge fields such as RNA interference, CRISPR systems, and RNA drug delivery, making up for the shortcomings of traditional tools in nucleic acid-protein interaction modeling.
Large-scale virtual screeningThe 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-fine refinement" workflow with the full-precision model.