How shut can an open mannequin get to AlphaFold3-level accuracy when it matches coaching knowledge, mannequin scale and inference finances? ByteDance has launched Protenix-v1, a complete AlphaFold3 (AF3) replica for biomolecular construction prediction, launched with code and mannequin parameters beneath Apache 2.0. The mannequin targets AF3-level efficiency throughout protein, DNA, RNA and ligand constructions whereas protecting all the stack open and extensible for analysis and manufacturing.
The core launch additionally ships with PXMeter v1.0.0, an analysis toolkit and dataset suite for clear benchmarking on greater than 6k complexes with time-split and domain-specific subsets.
What’s Protenix-v1?
Protenix is described as ‘Protenix: Protein + X‘, a basis mannequin for high-accuracy biomolecular construction prediction. It predicts all-atom 3D constructions for complexes that may embody:
- Proteins
- Nucleic acids (DNA and RNA)
- Small-molecule ligands
The analysis crew defines Protenix as a complete AF3 replica. It re-implements the AF3-style diffusion structure for all-atom complexes and exposes it in a trainable PyTorch codebase.
The undertaking is launched as a full stack:
- Coaching and inference code
- Pre-trained mannequin weights
- Information and MSA pipelines
- A browser-based Protenix Internet Server for interactive use
AF3-level efficiency beneath matched constraints
As per the analysis crew Protenix-v1 (protenix_base_default_v1.0.0) is ‘the primary absolutely open-source mannequin that outperforms AlphaFold3 throughout various benchmark units whereas adhering to the identical coaching knowledge cutoff, mannequin scale, and inference finances as AlphaFold3.‘
The essential constraints are:
- Coaching knowledge cutoff: 2021-09-30, aligned with AF3’s PDB cutoff.
- Mannequin scale: Protenix-v1 itself has 368M parameters; AF3 scale is matched however not disclosed.
- Inference finances: comparisons use related sampling budgets and runtime constraints.
On difficult targets corresponding to antigen–antibody complexes, rising the variety of sampled candidates from a number of to a whole bunch yields constant log-linear enhancements in accuracy. This provides a transparent and documented inference-time scaling conduct somewhat than a single fastened working level.
PXMeter v1.0.0: Analysis for 6k+ complexes
To assist these claims, the analysis crew launched PXMeter v1.0.0, an open-source toolkit for reproducible construction prediction benchmarks.
PXMeter offers:
- A manually curated benchmark dataset, with non-biological artifacts and problematic entries eliminated
- Time-split and domain-specific subsets (for instance, antibody–antigen, protein–RNA, ligand complexes)
- A unified analysis framework that computes metrics corresponding to complicated LDDT and DockQ throughout fashions
The related PXMeter analysis paper, ‘Revisiting Structure Prediction Benchmarks with PXMeter,‘ evaluates Protenix, AlphaFold3, Boltz-1 and Chai-1 on the identical curated duties, and reveals how completely different dataset designs have an effect on mannequin rating and perceived efficiency.
How Protenix suits into the broader stack?
Protenix is a part of a small ecosystem of associated tasks:
- PXDesign: a binder design suite constructed on the Protenix basis mannequin. It reviews 20–73% experimental hit charges and 2–6× larger success than strategies corresponding to AlphaProteo and RFdiffusion, and is accessible through the Protenix Server.
- Protenix-Dock: a classical protein–ligand docking framework that makes use of empirical scoring capabilities somewhat than deep nets, tuned for inflexible docking duties.
- Protenix-Mini and follow-on work corresponding to Protenix-Mini+: light-weight variants that cut back inference price utilizing architectural compression and few-step diffusion samplers, whereas protecting accuracy inside a number of % of the complete mannequin on commonplace benchmarks.
Collectively, these elements cowl construction prediction, docking, and design, and share interfaces and codecs, which simplifies integration into downstream pipelines.
Key Takeaways
- AF3-class, absolutely open mannequin: Protenix-v1 is an AF3-style all-atom biomolecular construction predictor with open code and weights beneath Apache 2.0, concentrating on proteins, DNA, RNA and ligands.
- Strict AF3 alignment for honest comparability: Protenix-v1 matches AlphaFold3 on crucial axes: coaching knowledge cutoff (2021-09-30), mannequin scale class and comparable inference finances, enabling honest AF3-level efficiency claims.
- Clear benchmarking with PXMeter v1.0.0: PXMeter offers a curated benchmark suite over 6k+ complexes with time-split and domain-specific subsets plus unified metrics (for instance, complicated LDDT, DockQ) for reproducible analysis.
- Verified inference-time scaling conduct: Protenix-v1 reveals log-linear accuracy positive aspects because the variety of sampled candidates will increase, giving a documented latency–accuracy trade-off somewhat than a single fastened working level.
Take a look at the Repo and Try it here. Additionally, be happy to comply with us on Twitter and don’t neglect to affix our 100k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
