Protenix
Summary¶
Protenix is an open-source reproduction of AlphaFold 3 that predicts structures containing proteins, small-molecule ligands, DNA and RNA. Its release includes model weights, inference and training code, processed datasets and documentation, with reported benchmark performance broadly competitive with the systems evaluated in its 2025 bioRxiv preprint [1].
Overview¶
Protenix was developed to make the methodology described for AlphaFold 3 more reproducible and modifiable. The implementation follows the published architecture from Google DeepMind, resolves ambiguities and reported typographical errors, and introduces targeted changes to data processing, the confidence head, parameter initialization and training efficiency.[1]
The model supports joint structure prediction for:
- proteins and protein complexes;
- protein-ligand complexes;
- protein-DNA and protein-RNA complexes;
- RNA structures and mixed biomolecular systems.
Data Pipeline¶
Training combines experimentally determined structures with predicted protein monomers. The reported mixture assigns weights of 0.60 to weighted Protein Data Bank structures, 0.35 and 0.03 to two protein-monomer distillation sets derived from MGnify sequences, and 0.02 to an OpenProteinSet subset.[1]
The pipeline uses multiple-sequence alignments for protein chains but not nucleic-acid chains, does not use structural templates, and applies molecule-aware cropping so ligands and non-standard residues remain intact. The released processed data include multiple-sequence-alignment results for wwPDB entries and mappings from Protein Data Bank identifiers to the model's processed representation.[1]
Model and Training¶
Protenix retains the central pair-representation and diffusion structure-generation design described for AlphaFold 3 while changing several implementation details. Reported optimizations include BF16 mixed-precision training, custom GPU kernels, optimized attention kernels, gradient checkpointing, tensor offloading and chunking.[1]
The preprint reports:
| Training property | Value |
|---|---|
| Model size | About 386 million parameters |
| Compute | 192 GPUs for approximately two weeks |
| Initial training | 75,000 steps with crop size 384 |
| Fine-tuning stage 1 | 15,000 steps with crop size 640 |
| Fine-tuning stage 2 | 4,000 steps focused on the confidence head |
Evaluation Findings¶
The authors evaluated ligand, protein-interface and nucleic-acid predictions using benchmark-specific structural metrics. In the reported comparisons, Protenix matched the confidence-ranked AlphaFold 3 ligand results closely, exceeded AlphaFold-Multimer 2.3 on overall protein-interface DockQ success rate, and produced nucleic-acid results similar to AlphaFold 3 and stronger than RoseTTAFold2NA. Protein-antibody top-ranked predictions were approximately on par with AlphaFold-Multimer 2.3, and the best generated ligand samples remained better than the model's confidence-selected samples, indicating room to improve ranking.[1]
These comparisons are not uniformly out-of-distribution. The model used a later structural-data cutoff than one AlphaFold 3 benchmark model, and some well-predicted examples had close training-set analogues. The authors therefore caution that memorization and dataset construction may contribute to apparent gains.[1]
Availability¶
The Protenix repository provides model weights, inference code, trainable code, processed datasets, an input-customization notebook and documentation. The source paper is a preprint rather than a peer-reviewed publication, so its performance claims should be interpreted at medium confidence pending independent validation.[1]