oteam/alphafold2-multimer

AlphaFold 2

This repository contains oteam/alphafold2-multimer: model_1_multimer_v3, converted from the official parameter archive.

Native PyTorch implementation of AlphaFold 2 monomer, monomer-pTM, and AlphaFold-Multimer for OFoldX, including the released five-model CASP14 eight-ensemble protocol and the monomer/pTM initial-guess evaluator used by PXDesign.

Disclaimer

This is an unofficial OFoldX implementation of AlphaFold 2 by John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A. A. Kohl, Andrew J. Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stanislav Nikolov, and Rishub Jain, and AlphaFold-Multimer. The upstream authors did not write, review, or endorse this implementation or model card. The official implementation is maintained by Google DeepMind.

Model Details

AlphaFold 2 predicts protein three-dimensional structure from amino-acid sequence, multiple-sequence alignments, and optional structural templates. Its Evoformer jointly updates MSA and residue-pair representations before an invariant-point-attention Structure Module produces backbone frames, side-chain angles, atomic coordinates, and confidence estimates. AlphaFold-Multimer extends the same design with paired MSAs and chain-relative features for protein complexes.

Variants

• oteam/alphafold2: AlphaFold 2 monomer model with pLDDT confidence.
• CASP14 five-model Pipeline: The fixed released model_1 through model_5 monomer ensemble. Every checkpoint uses eight input ensembles; one shared processed input is evaluated sequentially, all samples are stable-ranked by mean pLDDT, and the concrete Pipeline owns optional Amber relaxation of none, the best prediction, or all predictions.
• oteam/alphafold2-ptm: Monomer model with predicted aligned error and pTM confidence.
• AF2 initial guess: Distinct monomer/pTM Model--Processor--Pipeline pair using model_1_ptm weights. It seeds the first recycled atom37 prev_pos state from an in-memory target--binder structure without initializing Structure Module frames. Its all-GAP template retains backbone/CB geometry for every chain and masks target--binder template pair features.
• oteam/alphafold2-multimer: AlphaFold-Multimer for multi-chain protein complexes.

Model Specification

Variant	Evoformer Blocks	Extra-MSA Blocks	Structure Blocks	MSA Size	Pair Size	Single Size	Recycling Steps	Num Parameters (M)
Monomer	48	4	8	256	128	384	4	93.229
Monomer pTM	48	4	8	256	128	384	4	93.237
Multimer	48	4	8	256	128	384	21	93.224

Links

• Code: ofoldx.models.alphafold2
• Paper: Highly accurate protein structure prediction with AlphaFold
• Multimer Paper: Protein complex prediction with AlphaFold-Multimer
• Developed by: John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A. A. Kohl, Andrew J. Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stanislav Nikolov, Rishub Jain, Jonas Adler, Trevor Back, Stig Petersen, David Reiman, Ellen Clancy, Michal Zielinski, Martin Steinegger, Michalina Pacholska, Tamas Berghammer, Sebastian Bodenstein, David Silver, Oriol Vinyals, Andrew W. Senior, Koray Kavukcuoglu, Pushmeet Kohli, and Demis Hassabis
• Model type: MSA- and template-conditioned protein structure predictor with an Evoformer trunk and invariant-point-attention structure module
• Original Repository: google-deepmind/alphafold

Feature and geometry contract

All public model coordinate boundaries consume an explicit finite numerical projection. Unavailable values are selected out under the declared mask before arithmetic or dtype conversion and have zero coordinate derivatives. Active nonfinite values are invalid requests; finite values outside the model dtype range are an unsupported numeric range. Normalization returns the same input carrier type without changing canonical membership, source masks, or source precision. Decode uses prepared atom correspondences. APIs without an availability mask require complete finite model fields.

The model's atom37/atom14 slots and pseudo-beta convention belong to its residue vocabulary. Template observations and initial model coordinates have separate masks. Physical backbone frames additionally require available, noncollinear defining atoms. Relaxation and its linear-neighbor diagnostics admit only the actual closed standard linear protein graph; sequence labels alone do not certify this chemistry. The clash calculation exempts the admitted peptide C–N pair and has no blanket exemption for slot 5 or putative disulfides.

Usage

The model depends on the ofoldx library. Install it with pip:

pip install ofoldx

Structure Prediction

Load the concrete folding Pipeline:

import ofoldx

from ofoldx.io import read_structure

structure, report = read_structure("input.cif")
pipeline = ofoldx.pipeline("fold", model="oteam/alphafold2")
output = pipeline.fold(structure)

Every prediction is decoded onto the input Structure: each residue's atom37 atoms get predicted coordinates in a new frame, and atoms the model does not predict (ligands, hydrogens, a terminal OXT) have none.

The Pipeline is callable with the same arguments as fold. For staged execution:

prepared = pipeline.prepare(structure)
raw = pipeline.forward(prepared.input, num_samples=1)
result = pipeline.postprocess(prepared, raw)

An ordinary Pipeline returns raw AlphaFold2Output values from forward. The CASP14 ensemble returns checkpoint-labelled predictions; its postprocess ranks them and accepts models_to_relax, platform, and platform_properties. Stages describe the operation and do not prescribe CPU or GPU placement.

Load the CASP14 five-model ensemble through the same high-level API:

pipeline = ofoldx.pipeline("fold", model="oteam/alphafold2-casp14")
result = pipeline.fold(structure, num_samples=1, models_to_relax="best", platform="CPU")
best = result.best

result.ranked retains every prediction, its official checkpoint name, zero-based sample index, and any Amber relaxation result. Equal mean-pLDDT values retain checkpoint and sample generation order.

The initial-guess variant consumes a sequence-assigned target--binder Structure and an explicit binder instance:

initial_guess_pipeline = ofoldx.pipeline("fold", model="oteam/alphafold2-initial-guess")
result = initial_guess_pipeline.fold(structure, binder_instance_id="binder", num_recycles=4)

Amber relaxation also remains an independently callable family operation over a decoded structure.

Install OpenMM and PDBFixer in the same environment before requesting Amber relaxation, including models_to_relax="best":

conda install -c conda-forge openmm pdbfixer

These packages are required for the requested relaxation operation. The folding Model and the default unrelaxed prediction use the ordinary OFoldX installation.

Training Details

Training Data

Refer to the AlphaFold 2 and AlphaFold-Multimer papers and supplementary materials for the upstream training corpora and data construction. OFoldX does not redistribute those corpora, sequence databases, or template databases.

Training Procedure

The upstream monomer models use recycling, MSA and template evidence, and the released auxiliary losses; the pTM variants were fine-tuned with the predicted-aligned-error head. AlphaFold-Multimer adds paired-MSA and chain-relative features and interface confidence.

OFoldX preserves a differentiable final recycle and exposes a masked-MSA reconstruction loss covering that recycle. It is not the complete released structural and auxiliary-loss program; MSA, template, and sequence-database generation remain caller-owned.

Citation

@article{jumper2021highly,
  author = {Jumper, John and Evans, Richard and Pritzel, Alexander and others},
  title = {Highly accurate protein structure prediction with AlphaFold},
  journal = {Nature},
  volume = {596},
  pages = {583--589},
  year = {2021},
  doi = {10.1038/s41586-021-03819-2}
}

For AlphaFold-Multimer predictions, also cite Protein complex prediction with AlphaFold-Multimer.

Contact

Please use OFoldX GitHub issues for questions about this implementation or model card. Questions about the original models should be directed to the upstream AlphaFold authors.

License

AlphaFold 2 parameters are Copyright 2022 DeepMind Technologies Limited, distributed under CC BY 4.0. Original parameters: Google DeepMind AlphaFold; upstream parameter licensing. The parameter license allows redistribution and adaptation, including commercial use, subject to attribution and the other license conditions.

The OFoldX implementation is separately licensed under AGPL-3.0-or-later; see license.md and license-faq.md.

OFoldX conversion remaps tensor names, layouts and vocabulary axes into the native Model and Processor format. These are modified parameter representations; the original weight terms remain applicable. No upstream endorsement is implied.

Upstream source revision: c77e5d2a8961d1a353632c462914ff0a32a950f6.