henhaobio
Sign in

bioRxivBioinformaticsPreprint

Protenix - Advancing Structure Prediction Through a Comprehensive AlphaFold3 Reproduction

ByteDance AML AI4Science Team, Xinshi CheniD, Yuxuan ZhangiD, Chan Lu, Wangjing MaiD, Jiaqi GuaniD, Chengyue GongiD, Jincai YangiD, Hanyu ZhangiD, Ke ZhangiD, Shenghao Wu, Kuangqi Zhou, Yanping YangiD, Zhenyu LiuiD, Lan WangiD, Bo Shi, Shaochen Shi, Wenzhi XiaoiD

PDFDOI
In-depth full-text digest available · Jump to digest ↓

Abstract

In this technical report, we present Protenix, a comprehensive reproduction of AlphaFold3 (AF3), aimed at advancing the field of biomolecular structure prediction. Protenix tackles the challenges of predicting complex interactions involving proteins, ligands, and nucleic acids, while enhancing accessibility and reproducibility. Across diverse benchmarks, including PoseBusters V2, low-homology PDB sets, and CASP15 RNA, Protenix achieves state-of-the-art performance in protein-ligand, protein-protein, and protein-nucleic acid predictions. We also address limitations, such as potential memorization effects, and outline future directions for improvement. By open-sourcing Protenix, we aim to democratize advanced structure prediction tools and accelerate interdisciplinary research in computational biology and drug discovery.

In-depth summary

Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 1 (11 Jan 2025) · CC BY 4.0

An open AlphaFold3 reproduction, with weights, inference and training code, that matches AF3 on ligand and RNA benchmarks, beats AlphaFold-Multimer 2.3 on protein interfaces and flags possible memorisation.

AlphaFold3 raised accuracy on complexes of proteins with ligands and nucleic acids, but its code is unavailable and its written algorithms contain ambiguities and typographic errors, while earlier open efforts such as HelixFold3 and Chai-1 still lack comprehensive training code and preprocessed data. Protenix reimplements the model from that description, documents the corrections needed, and releases weights, inference and trainable code, and processed data including whole-wwPDB MSA searches.

Key results

  1. ShownOn the PoseBusters V2 ligand benchmark the reproduction edged past the AF3 model trained to a 2019 cutoff; a similarity analysis places the gain on common ligands, likely from extra training data, with equal performance on non-common ligands.

    Success is a pocket-aligned ligand RMSD of 2 A or less on 308 structures; Protenix is reported slightly above AF3-2019 on both RMSD success and PB-valid, with its oracle configuration highest at every RMSD threshold, but the rates sit in figures missing from the text. Protenix beat AF3-2019 on common ligands and high-homology targets, matched it on non-common ligands, and was slightly worse on other targets. · 2.2 Ligands; A.2.1 PoseBusters V2

  2. ShownProtein interface accuracy exceeded AlphaFold-Multimer 2.3 overall; antibody interfaces were the exception, where top-ranked predictions were only on par, pointing to room for improvement in the sample ranker.

    DockQ success rates on all protein-protein and on protein-antibody interfaces were compared for Protenix with 10 recycles against AF2.3 (5x5) with 20 recycles, reporting oracle, median and top-1 ranked predictions; Protenix was higher overall and on par with AF2.3 for top-ranked antibody cases. The numbers are in a table missing from the extracted text. Ranking used the chain-pair ipTM confidence. · 2.3 Proteins

  3. SuggestedOn nucleic acids the model tracked AF3 and clearly beat RoseTTAFold2NA, but trailed the human-assisted CASP15 entry.

    Average LDDT and TM-score over eight publicly available CASP15 RNA targets (recomputed with OpenStructure for Protenix, RF2NA and AIchemy_RNA2; single seed to match AF3 values taken from Bernard et al.) were similar to AF3, well above RF2NA, and below AIchemy_RNA2, which used human input. On curated complexes (25 RNA-protein, 38 dsDNA-protein) Protenix scored above RF2NA. Values are in missing figures. · 2.4 Nucleic Acids; A.2.3 Nucleic Acids Evaluation Set

  4. SuggestedThe authors' own check on AF3's showcase cases suggests part of the accuracy of both models may come from memorising closely similar training structures.

    Protenix also predicted AF3's showcase structures accurately, but the training set holds close counterparts: 3MZH for the protein-DNA complex 7PZB, and 5OAX for the protein-ligand complex 7XFA. The authors conclude that out-of-distribution test sets are needed to judge generalisation. · 2.5 Discussion

  5. ShownEngineering changes sped up training and eased memory bottlenecks, and a stereochemistry penalty in the ranking score raised physical validity.

    Training took about two weeks on 192 GPUs: 75K steps initially, 15K in fine-tuning stage 1 and only 4K in stage 2. A custom LayerNorm kernel gave a 30-50% end-to-end training speedup over PyTorch's, DS4Sci_EvoformerAttention another 10-20%, and BF16 mixed precision nearly doubled speed over FP32. The ranking penalty raised PB-valid success by over 10%, leaving roughly a 5% chirality failure rate. · 3.2 Model and Training; 2.2 Ligands

Methods, data and reproducibility

The team reimplemented AlphaFold3 from its published description, correcting ambiguities and typographic errors, revising the confidence head, zero-initialising several modules and rebuilding the data pipeline (MMseqs2/ColabFold MSAs, no templates, none for nucleic chains, modified cropping and permutation). One model was trained on PDB entries to 2021-09-30 plus distillation, then benchmarked.

  • Experimental structures from the Protein Data Bank used for training (WeightedPDB), cutoff 2021-09-30, PoseBusters V2 targets removed
  • Protein monomer distillation set: AlphaFold2 predictions for MGnify representative sequences over 200 residues, plus a subset of OpenProteinSet, filtered by pLDDT
  • PoseBusters Benchmark Set Version 2 for protein-ligand evaluation, with clashing chains removed per AF3's procedure · n = 308 structures
  • Low Homology Recent PDB set, released 2022-05-01 to 2023-01-12, under 40% sequence similarity and fewer than 2560 tokens · n = differs from AF3's PDB and cluster counts
  • CASP15 RNA targets public as of 2023-12-01 · n = 8 targets · R1116/8S95, R1117/8FZA, R1126, R1128/8BTZ, R1136/7ZJ4, R1138, R1189/7YR7, R1190/7YR6
  • Protein-nucleic acid complexes screened from the RecentPDB IDs in AF3's supplementary Table 14, capped at 1000 residues for RF2NA compatibility · n = 25 RNA-protein and 38 dsDNA-protein structures
Methods & tools
Protenix, AlphaFold3, AlphaFold-Multimer 2.3, RoseTTAFold2NA, AIchemy_RNA2, PoseBusters v0.2.7, DockQ, OpenStructure, MMseqs2 / ColabFold MSA pipeline, RDKit with pdbeccdutils v0.8.5, DS4Sci_EvoformerAttention (DeepSpeed4Science), OpenFold
Reagents & environment
CCD snapshot of 2024-06-08, pdbeccdutils v0.8.5, PoseBusters v0.2.7, Uniref100 for MSA pairing; BF16 training needs a GPU that supports it, such as NVIDIA Ampere or Hopper.
Experimental conditions
Roughly two weeks on 192 GPUs; 75K training steps, 15K in fine-tuning stage 1, 4K in stage 2. Inference generates 25 samples per entry from 5 seeds with 10 recycles.
Reproducibility gaps
The repository address for code, weights and processed data was lost in extraction, so no URL is listed. Most quantitative results (PoseBusters rates, DockQ tables, LDDT and TM-scores, the AF3 correction list, training-stage settings) are in absent figures and tables.
Limitations (6)
  • A single model was trained because of resource limits, so PoseBusters is evaluated against an AF3 model with an earlier (2019) data cutoff rather than a matched one. · stated by the authors
  • The evaluation sets could not be reproduced exactly: the low-homology set differs from AF3's in PDB and cluster counts, and the protein-nucleic set only matches AF3's counts, which may explain why RF2NA scores differ from those AF3 reported. · stated by the authors
  • Accuracy on AF3's showcase structures may partly reflect memorisation of similar training entries; the authors call for more out-of-distribution test sets to better assess generalisation. · stated by the authors
  • Chirality and steric clashes are still not reliably correct; a ranking penalty mitigates but does not remove them, and selected samples trail the best sample available, so the ranker loses accuracy. · stated by the authors
  • Fine-tuning stage 2 ran only 4K steps, which the authors say may be too few to train the confidence head adequately; the PAE head likewise saw only 4K steps. · stated by the authors
  • Comparisons rest on figures and tables absent from the extracted text, and differences from AF3 are described as slight without effect sizes or variability across seeds. · noted in this summary

Read this if you want an openly trainable AlphaFold3-class model for protein, ligand and nucleic-acid complexes, or a record of the corrections needed to reimplement AF3.

Citations
131

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISTeam '26Corley '25Saha '25Furui '25Follonier '26Tang '25
75% semantic matchBioinformatics
Protenix Team, Yuxuan Zhang, Chengyue Gong+8 more·bioRxiv·Feb 22, 2026· 19 citations