How to Use Protenix (AlphaFold3)

Protenix is an open AlphaFold3-class model for all-atom biomolecular structure prediction across proteins, nucleic acids, ligands, and mixed complexes. The ByteDance technical report presents it as a faithful open implementation of multimodal diffusion-based structure prediction, making it relevant for researchers who want modern complex modeling without relying on closed infrastructure.

On Neurosnap, a single job can combine Input Sequences with inline CCD modifications and Input Molecules, which makes the service useful for protein monomers, protein-ligand systems, nucleic-acid complexes, and more heterogeneous assemblies. The value of the workflow is not only coordinates, but also the confidence decomposition needed to decide whether a model is globally reliable, locally informative, or still too ambiguous for downstream use.

Methodologically, Protenix follows the AlphaFold3-style all-atom formulation in which folding and docking are treated as one multimolecular structure problem. The report emphasizes implementation details that matter in practice, including open reproduction of multimodal structure reasoning together with support for MSA pairing, pocket-aware conditioning, unified cropping strategies, and a confidence stack tuned for complex interpretation.

On Neurosnap, MSA Mode, Number Recycles, Sampling Steps, and Diffusion Samples expose the main tradeoffs between speed, refinement depth, and structural diversity. Input Sequences can cover proteins and nucleic acids, while Input Molecules and inline CCD modifications within Input Sequences make it possible to keep cofactors, ligands, or chemically modified residues inside the same prediction problem.

Researchers generally review Protenix through residue- and interface-level confidence rather than a single rank alone. pLDDT is useful for local geometry, pTM and ipTM summarize fold and interface reliability, and aligned-error views help determine whether chain placement or ligand context should be trusted before moving into docking, mutagenesis, or experimental design.

• Multimodal biomolecular input: Protenix accepts proteins, nucleic acids, ligands, and inline CCD modifications in one workflow, which better matches real experimental systems than protein-only predictors.
• Open AlphaFold3-class modeling: The method brings all-atom complex prediction into an openly accessible setting suitable for reproducible academic and applied research.
• Scientist-facing control surface: MSA Mode, recycling, and diffusion sampling controls let researchers decide how much depth and diversity to explore for a given target.
• Confidence-based triage: Residue, interface, and aligned-error summaries help separate credible assemblies from hypotheses that still need stronger evidence.