Protenix
1. Introduction
Protenix1 is an open-source model launched by ByteDance, which reproduces AlphaFold3 using PyTorch. It focuses on biomolecular structure prediction and features performance, methodology, data, and openness at the forefront of the industry:
- Functionality: Capable of accurately predicting the three-dimensional structures of various biomolecules, including protein–ligand and protein–nucleic acid interactions. It can also be used to analyze complex antibody–antigen binding structures, providing critical structural information support for biological and pharmacological research.
- Excellent Performance: In benchmark tests such as PoseBusters V2, Protenix demonstrates outstanding performance in structure prediction across different molecular types, surpassing models like AF32, AF2.3, and RF2NA, placing it at the industry’s cutting edge.
- Method Optimization: Based on the AF3 description, Protenix re-implements the method, correcting ambiguous steps and errors, and adjusts the confidence head architecture to improve prediction accuracy.
- Rich Data: The training data includes PDB experimental structures, AlphaFold2, and OpenFold3 predicted protein monomer structures. It provides MSA search results after extensive computation, as well as mappings between PDB IDs and processed data, making it convenient for researchers to use.
- Fully Open Source: The model weights, inference, and training code are publicly available, along with detailed documentation, enabling researchers to deeply explore, reproduce results, and extend applications.
Currently, AbSeek has deployed the version of Protenix is v0.5.0, whose performance is shown in the figures below:
Figure 1. Protenix performance
Figure 2. Protenix performance with constraints
2. Parameter Description
This tool supports multiple molecule types, including proteins, DNA, RNA, and small molecules. You can select the type via the Molecule Type parameter. Currently, only standard amino acids, standard bases, and small molecules in SMILES format are supported.
Click the Add Sequence button to add a new sequence. Chains will be automatically labeled using uppercase letters A–Z based on the order of input.
Click the Add Constraints button to incorporate different types of constraints, such as:
- Inter-residue/atomic distance constraints（contact）
- Pocket distance constraints（pocket）
Note that constraints are not strictly enforced—the final output structure is determined by the prediction algorithm.
For more detailed information on constraints, please refer to the official documentation: https://github.com/bytedance/Protenix/blob/main/docs/infer_json_format.md
3. Result Description
Five predicted structure .cif files will be returned, along with a .csv file containing prediction metrics for each structure. Below is an explanation of the key metrics:
| Metric | Full Name | Value Range | Focus Level | Typical Threshold/Interpretation | Main Purpose |
|---|---|---|---|---|---|
| ranking score | ranking score | 0–1 | Whole Complex | Confidence in predicted ranking | Evaluates overall confidence |
| pLDDT | predicted Local Distance Difference Test | 0–100 | Residue/Atom | ≥ 90 Very high confidence; < 50 Low confidence | Identifies disordered regions, local resolution |
| pTM | predicted Template Modeling score | 0–1 | Single Chain/Whole Complex | ≥ 0.5 Correct fold in general | Evaluates global topology reliability |
| ipTM | interface predicted TM-score | 0–1 | Interface | ≥ 0.8 High-quality docking | Evaluates inter-chain interactions |
| PDE | Predicted Distance Error | ≥ 0 (Å) | Residue Pair | Lower is better | Predicts local distance error |
| GPDE | Global Predicted Distance Error | ≥ 0 (Å) | Whole Chain/Whole Complex | Lower is better | Macro comparison of overall error |
4. References
[1] ByteDance AML AI4Science Team. Protenix - Advancing Structure Prediction Through a Comprehensive AlphaFold3 Reproduction. bioRxiv 2025.01.08.631967. https://doi.org/10.1101/2025.01.08.631967
[2] Abramson, Josh, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel, Olaf Ronneberger, Lindsay Willmore, Andrew J Ballard, Joshua Bambrick, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 2024, 630(8016):493-500. https://doi.org/10.1038/s41586-024-07148-1
[3] Ahdritz, Gustaf, Nazim Bouatta, Christina Floristean, Sachin Kadyan, Qinghui Xia, William Gerecke, Timothy J O’Donnell, Daniel Berenberg, Ian Fisk, Niccolò Zanichelli, et al. OpenFold: Retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization. Nature Methods, 2024, 21(8):1514-1524. https://doi.org/10.1038/s41592-024-02219-9
[4] Mirdita, Milot, Konstantin Schütze, Yoshitaka Moriwaki, Lim Heo, Sergey Ovchinnikov, and Martin Steinegger. ColabFold: making protein folding accessible to all. Nature methods, 2022, 19(6):679-682. https://doi.org/10.1038/s41592-022-01488-1