Protenix-v2

Overview

Protenix-v2 is the second major release of ByteDance AI Lab's Protenix family of structure prediction and design models, posted to bioRxiv in April 2026. The 464-million parameter v2 substantially improves on Protenix-v1 (released January 2025) with 9 to 13 percentage-point gains in antibody-antigen complex prediction accuracy and adds target-conditioned VHH-Fc generative design capabilities, achieving up to 48% experimental hit rates against tested antigens.

Protenix-v2 is open-source under Apache 2.0, making it one of the most capable open-source antibody-design systems available and a meaningful counterpoint to closed proprietary platforms.

Key Features

• Antibody-antigen prediction gains: 9 to 13 percentage-point improvements in antibody-antigen complex prediction accuracy over Protenix-v1, surpassing AlphaFold 3 on antibody-specific subsets.
• Generative VHH design: Target-conditioned generation of single-domain antibody (VHH) sequences and structures, achieving up to 48% hit rates in experimental validation.
• Apache 2.0 open source: Fully open code and weights, enabling commercial use and downstream extension.
• Unified prediction + design: Same architecture handles both structure prediction and conditional design.
• 464M parameters: Sized for accessible inference on single high-end GPUs.

Technical Details

Protenix-v2 retains the broad diffusion-transformer architecture of Protenix-v1 with refinements targeting antibody-class structures: increased model capacity, antibody-enriched training data, and improved pair-representation modules. The bioRxiv preprint provides the full architectural specification, training corpus, and benchmark results.

The generative design capability is implemented as conditional sampling from the diffusion model with target structure as conditioning input. Experimental validation reports 48% hit rates against several targets in a yeast-display screening assay.

Applications

Protenix-v2 is well-suited for antibody discovery and engineering teams that need open-source tools matching or exceeding the capabilities of closed alternatives. The VHH focus is particularly relevant for therapeutic-VHH developers and for applications requiring small antigen-binding domains. As a structure-prediction model, it is also competitive with AlphaFold 3 for antibody-containing complexes.

Impact

Protenix-v2 represents the most capable open-source generative antibody-design system released to date and narrows the gap with closed proprietary platforms. The combination of SOTA antibody-antigen prediction, target-conditioned generative design, Apache 2.0 licensing, and competitive parameter count makes it a strong candidate for adoption by both academic and commercial antibody-discovery teams.