AlphaFold Codec

Fine-tune AlphaFold 3, AlphaFold 2, Boltz-1, and Boltz-2 on downstream protein tasks. Algorithm notebooks and papers are the supporting layer.

Start here	Link
Fine-tuning guide	finetuning/FINETUNING_GUIDE.md
AF3 weight + LoRA design	alphafold3/AF3_WEIGHTS_FINETUNING_DESIGN.md
AF3 module	finetuning/af3/
Architecture figure	assets/architecture.svg

---

Fine-tuning (primary)

This repo is built around a shared finetuning/ framework. Official AlphaFold 3 v3.0.4 weights are now a public .bin.zst file (no request form). This repo loads that checkpoint, validates all 405 tensors, attaches LoRA, and exports adapter-only files so you never redistribute the base weights.

Architecture (fine-tune first)

Text / Mermaid fallback (if the SVG does not render)

flowchart LR
  W["Download af3.bin.zst"] --> V["Validate schema<br/>405 tensors / 368.4M"]
  V --> L["Attach LoRA<br/>freeze AF3 base"]
  L --> H["Train task head"]
  H --> A["Export adapter.npz<br/>deltas only"]

  subgraph framework ["finetuning/"]
    AF3["af3/ I/O + LoRA"]
    CFG["configs/"]
    MOD["modules/ LoRA Adapter"]
    HD["heads/ 15+"]
    TR["trainers/"]
    REG["registry.py"]
  end

  L --> AF3
  H --> HD
  A --> REG

Loading

Flow

1. Download / check AF3 weights (python -m finetuning.af3.weights)
2. Attach LoRA on Pairformer + Diffusion (AlphaFold3FineTuner)
3. Train a task head (affinity, antibody, enzyme, PPI, ...)
4. Save adapter.npz only (base AF3 weights stay frozen and are not written)

Then the four model families (alphafold2/, alphafold3/, boltz/, boltz2/) provide notebooks and papers.

Supported models and strategies

Model	Runtime	Fine-tune	Notes
AlphaFold 3	JAX / Haiku	Full, Head-only, LoRA	Public weights af3.bin.zst; use finetuning/af3
AlphaFold 2	JAX / Haiku	Full, Head-only, LoRA	32 algorithm notebooks
Boltz-1	PyTorch	Full, LoRA, Adapter	Open AF3-class interactions
Boltz-2	PyTorch	Full, LoRA, Adapter	Binding-affinity head

Strategy	Trainable params	When to use
LoRA	~0.1–3%	Default for AF3; small data
Adapter	~1%	Modular / multi-task
Head-only	~5%	New prediction task
Full	100%	Large data, max accuracy

AlphaFold 3 weights

DeepMind now publishes parameters at a public URL (compatible with any 3.0.x code):

• File: https://storage.googleapis.com/alphafold3/af3.bin.zst
• Terms (non-commercial, do not redistribute weights): WEIGHTS_TERMS_OF_USE.md
• Schema in this repo: 405 entries, 368,384,602 parameters (metadata only)

pip install numpy zstandard

python -m finetuning.af3.weights info
python -m finetuning.af3.weights download /path/to/weights --accept-terms
python -m finetuning.af3.weights check /path/to/weights

from finetuning.af3 import AlphaFold3FineTuner, AF3FineTuneConfig, LoRAConfig

config = AF3FineTuneConfig(strategy="lora", lora=LoRAConfig(rank=8, alpha=16.0))
tuner = AlphaFold3FineTuner.from_pretrained("/path/to/weights/", config)
print(tuner.parameter_summary().describe())

# Adapter-only export: does not write AF3 base weight values
tuner.save_adapter("./af3_lora_adapter.npz")

# Merged checkpoint is restricted; requires an explicit acknowledgement
# tuner.export_merged_weights("./merged.bin", acknowledge_weights_terms=True)

Task-registry quick start (Boltz / PyTorch)

from finetuning import TaskRegistry, create_finetuning_pipeline
from finetuning.modules import LoRAModule
from finetuning.heads import AffinityHead, AffinityHeadConfig
from finetuning import FineTuningConfig, Trainer

print(TaskRegistry.list_all_tasks())  # 50+ tasks
info = TaskRegistry.get_task_info("binding_affinity")

pipeline = create_finetuning_pipeline(
    task="binding_affinity",
    base_model=model,
    strategy="lora",
)

# Or assemble manually
lora_model = LoRAModule(model, rank=8, alpha=16.0)
head = AffinityHead(AffinityHeadConfig())
trainer = Trainer(lora_model, FineTuningConfig(strategy="lora", task="binding_affinity"), train_loader, val_loader)
trainer.train()
lora_model.save_lora_weights("./lora_weights.pt")

Downstream tasks (50+)

Drug discovery

Task	Outputs	Applications
Binding Affinity	pKd, pIC50, dG, Ki	Lead optimization, SAR
Virtual Screening	Hit probability, ranking	HTS prioritization
ADMET	Absorption, metabolism, toxicity	Compound triage

Protein engineering

Task	Outputs	Applications
Stability	ddG, Tm shift	Thermostabilization
Solubility	Expression score	Biomanufacturing
Mutation Effects	Fitness, pathogenicity	Variant analysis

Antibody design

Task	Outputs	Applications
Affinity Maturation	CDR binding, ddG	Therapeutic optimization
Humanization	Humanness score	Drug development
Developability	Aggregation, viscosity	Manufacturing

Enzyme / PPI / function / immunology / quality

Category	Outputs
Enzyme	kcat, Km, specificity, directed evolution
PPI	Kd, interface residues, hot-spot ddG
Function	GO terms, EC numbers, localization
Immunology	B/T epitopes, ADA risk
Structure quality	pLDDT, pAE, contacts, disorder

finetuning/ layout

finetuning/
├── af3/                 # AF3 binary I/O, schema, LoRA, CLI, FineTuner
├── configs/             # FineTuningConfig + task / LoRA presets
├── modules/             # LoRA, Adapter, prompt tuning
├── heads/               # Affinity, antibody, PPI, enzyme, GO, epitope, ...
├── trainers/            # Trainer, DistributedTrainer, callbacks
├── data/                # Datasets and transforms
├── registry.py          # TaskRegistry + create_finetuning_pipeline
├── tests/               # 101 unit tests for AF3 weight / LoRA path
└── FINETUNING_GUIDE.md

---

Repository layout

alphafold-notebooks/
├── finetuning/          # PRIMARY: shared fine-tuning framework
├── alphafold3/          # AF3 notebooks (23) + weight design doc
├── alphafold2/          # AF2 notebooks (32) + source/
├── boltz/               # Boltz-1 notebooks (20)
├── boltz2/              # Boltz-2 notebooks (10)
├── assets/architecture.svg
└── .gitmodules          # 14 reference submodules

git submodule update --init --recursive

Model	Notebooks	Index	Papers
AlphaFold 2	32	ALGORITHM_INDEX	AF2REFPAPERS
AlphaFold 3	23	ALGORITHM_INDEX	AF3REFPAPERS
Boltz-1	20	ALGORITHM_INDEX	BOLTZREFPAPERS
Boltz-2	10	ALGORITHM_INDEX	BOLTZ2REFPAPERS

---

Learning Source Availability

Papers

• Jumper, J., Evans, R., Pritzel, A. et al. Highly accurate protein structure prediction with AlphaFold. Nature (2021). https://doi.org/10.1038/s41586-021-03819-2

PPT

• My Public talk on Alphafold2 Paper Reading By Xingqiang,Chen .Key/.pptx in AF2-PPT file.
• Sergey Ovchinnikov talk on AF2 slides /.pptx in AF2-PPT file.

Learning by Code

📓 AlphaFold2 Algorithm Notebooks (32 Complete!)

We provide 32 Jupyter Notebooks covering every algorithm from the AlphaFold2 supplementary materials. Each notebook includes:

• Algorithm pseudocode/image reference
• Source code location mapping
• NumPy implementation
• Executable test cases with verification

👉 Full Algorithm Index

Quick Links by Category

Category	Algorithms	Notebooks
Data Preprocessing	MSA Block Deletion	Algorithm 1
Embedding	Input Embedder, relpos, one_hot	Alg 3, Alg 4, Alg 5
Evoformer	Stack, MSA Attention, Triangle Ops	Alg 6-15
Templates	Pair Stack, Pointwise Attention	Alg 16, Alg 17
Extra MSA	Stack, Global Attention	Alg 18, Alg 19
Structure Module	IPA, Backbone, Atom Coords	Alg 20-25
Losses	FAPE, Torsion, pLDDT	Alg 26-29
Recycling	Inference, Training, Embedder	Alg 30, Alg 31, Alg 32
Main Pipeline	Full Inference	Algorithm 2

📋 Complete Algorithm List (Click to Expand)

#	Algorithm	Notebook Link
1	MSA Block Deletion	algorithm-1-MSABlockDeletion.ipynb
2	Inference	algorithm-2-Inference.ipynb
3	Input Embedder	algorithm-3-InputEmbedder.ipynb
4	relpos	algorithm-4-relpos.ipynb
5	one_hot	algorithm-5-one_hot.ipynb
6	Evoformer Stack	algorithm-6-EvoformerStack.ipynb
7	MSA Row Attention with Pair Bias	algorithm-7-MSARowAttentionWithPairBias.ipynb
8	MSA Column Attention	algorithm-8-MSAColumnAttention.ipynb
9	MSA Transition	algorithm-9-MSATransition.ipynb
10	Outer Product Mean	algorithm-10-OuterProductMean.ipynb
11	Triangle Multiplication (Outgoing)	algorithm-11-TriangleMultiplicationOutgoing.ipynb
12	Triangle Multiplication (Incoming)	algorithm-12-TriangleMultiplicationIncoming.ipynb
13	Triangle Attention (Starting Node)	algorithm-13-TriangleAttentionStartingNode.ipynb
14	Triangle Attention (Ending Node)	algorithm-14-TriangleAttentionEndingNode.ipynb
15	Pair Transition	algorithm-15-PairTransition.ipynb
16	Template Pair Stack	algorithm-16-TemplatePairStack.ipynb
17	Template Pointwise Attention	algorithm-17-TemplatePointwiseAttention.ipynb
18	Extra MSA Stack	algorithm-18-ExtraMsaStack.ipynb
19	MSA Column Global Attention	algorithm-19-MSAColumnGlobalAttention.ipynb
20	Structure Module	algorithm-20-StructureModule.ipynb
21	Rigid from 3 Points	algorithm-21-rigidFrom3Points.ipynb
22	Invariant Point Attention	algorithm-22-InvariantPointAttention.ipynb
23	Backbone Update	algorithm-23-BackboneUpdate.ipynb
24	Compute All Atom Coordinates	algorithm-24-computeAllAtomCoordinates.ipynb
25	makeRotX	algorithm-25-makeRotX.ipynb
26	Rename Symmetric Ground Truth Atoms	algorithm-26-renameSymmetricGroundTruthAtoms.ipynb
27	Torsion Angle Loss	algorithm-27-torsionAngleLoss.ipynb
28	Compute FAPE	algorithm-28-computeFAPE.ipynb
29	Predict Per-Residue LDDT	algorithm-29-predictPerResidueLDDT.ipynb
30	Recycling (Inference)	algorithm-30-RecyclingInference.ipynb
31	Recycling (Training)	algorithm-31-RecyclingTraining.ipynb
32	Recycling Embedder	algorithm-32-RecyclingEmbedder.ipynb

📓 AlphaFold3 Algorithm Notebooks (NEW!)

We now include AlphaFold3 algorithm notebooks! AF3 introduces significant architectural changes including diffusion-based structure prediction.

👉 AlphaFold3 Algorithm Index

Key AF3 Components

Category	Key Algorithms	Notebooks
Input	MSA Features, Templates, Atom Features	Alg 1-4
MSA Module	Outer Product, MSA Attention	Alg 5-7
Pairformer	Triangle Ops, Single Attention	Alg 8-14
Diffusion	Diffusion Module, AdaLN, Transformer	Alg 15, Alg 16
Confidence	Distogram, Confidence, LDDT	Alg 20-23

AF3 Source Code Submodules

# Official AlphaFold3
alphafold3/ref-src/alphafold3-official/

# PyTorch Implementation (lucidrains)
alphafold3/ref-src/alphafold3-pytorch/

# Architecture Walkthrough
alphafold3/ref-src/alphafold3-walkthrough/

Weights, LoRA, and the fine-tuning CLI are documented at the top: Fine-tuning (primary).

📓 Boltz Algorithm Notebooks (NEW!)

We now include Boltz algorithm notebooks! Boltz is a family of models for biomolecular interaction prediction:

• Boltz-1: First fully open source model to approach AlphaFold3 accuracy
• Boltz-2: Adds binding affinity prediction, approaching FEP accuracy 1000x faster

👉 Boltz Algorithm Index

Key Boltz Components

Category	Key Algorithms	Notebooks
Input Processing	Input Embedder, Atom Encoder, RelPos	Alg 1-3
MSA Module	MSA Module, Outer Product, Pair Averaging	Alg 4-6
Pairformer	Pairformer, Triangle Ops, Attention	Alg 7-11
Diffusion	Diffusion Module, Transformer, Fourier	Alg 12-15
Confidence & Affinity	Confidence, Distogram, Affinity (Boltz-2)	Alg 16-18
Loss Functions	Diffusion Loss, Confidence Loss	Alg 19-20

Boltz Source Code Submodule

# Official Boltz Repository
boltz/ref-src/boltz-official/

Papers:

• Boltz-1: bioRxiv 2024.11.19.624167
• Boltz-2: bioRxiv 2025.06.14.659707

📓 Boltz-2 Specific Notebooks (NEW!)

Boltz-2 introduces binding affinity prediction - the first DL model approaching FEP accuracy while being 1000x faster.

👉 Boltz-2 Algorithm Index

Boltz-2 New Features

Category	Key Algorithms	Notebooks
Affinity Prediction	Affinity Module, Gaussian Smearing	Alg 1-2
Contact Guidance	Contact Conditioning	Alg 3
Enhanced v2 Modules	Input v2, Template v2, Diffusion v2	Alg 5-7
Improved Confidence	Confidence v2, B-Factor	Alg 8, 10

Boltz-2 Submodules

# Official Repository (contains both Boltz-1 and Boltz-2)
boltz/ref-src/boltz-official/

# Boltzina - Virtual Screening with Boltz-2
boltz/ref-src/boltzina/

Practice on Modeling Test of AF2

• https://github.com/sokrypton/ColabFold.git

MD+Alphafold2

• https://github.com/pablo-arantes/Making-it-rain

Fine-tuning APIs, AF3 weights, LoRA, and task heads live at the top of this README: Fine-tuning (primary). Full guide: finetuning/FINETUNING_GUIDE.md.

---

Blogs

• DeepMind: AlphaFold-Using-AI-for-scientific-discovery
• DeepMind: alphafold-a-solution-to-a-50-year-old-grand-challenge-in-biology
• DeepMind: putting-the-power-of-alphafold-into-the-worlds-hands

References

reference papers

• Reference papers list here and you can download them by Baidu Cloud Driver Link with the code 9w2p.
• Reference Papers' Source Codes are managed via git submodules in alphafold2/ref-src/

📦 AlphaFold2 Reference Source Code (Submodules)

# Official AlphaFold (DeepMind)
alphafold2/ref-src/alphafold-official/

# OpenFold (PyTorch implementation)
alphafold2/ref-src/openfold/

# ColabFold (Colab-friendly version)
alphafold2/ref-src/colabfold/

# MMseqs2 (Sequence search)
alphafold2/ref-src/mmseqs2/

# HH-suite (Template search)
alphafold2/ref-src/hh-suite/

# trRosetta2 (Predecessor model)
alphafold2/ref-src/trRosetta2/

# ESM (Facebook protein language model)
alphafold2/ref-src/esm/

# UniRep (Protein representations)
alphafold2/ref-src/unirep/

# SeqVec (Sequence embeddings)
alphafold2/ref-src/seqvec/

To initialize submodules after cloning:

git submodule update --init --recursive

Data availability

All input data are freely available from public sources.

Structures from the PDB were used for training and as templates (https://www.wwpdb.org/ftp/pdb-ftp-sites; for the associated sequence data and 40% sequence clustering see also https://ftp.wwpdb.org/pub/pdb/derived_data/ and https://cdn.rcsb.org/resources/sequence/clusters/bc-40.out).

Training used a version of the PDB downloaded 28/08/2019, while CASP14 template search used a version downloaded 14/05/2020. Template search also used the PDB70 data- base, downloaded 13/05/2020 (https://wwwuser.gwdg.de/~compbiol/data/hhsuite/databases/hhsuite_dbs/).

We show experimental structures from the PDB with accessions 6Y4F⁷⁶, 6YJ1⁷⁷, 6VR4⁷⁸, 6SK0⁷⁹, 6FES⁸⁰, 6W6W⁸¹, 6T1Z⁸², and 7JTL⁸³.

For MSA lookup at both training and prediction time,

we used UniRef90 v2020_01 (https://ftp.ebi.ac.uk/pub/databases/uniprot/previous_releases/release-2020_01/uniref/),

BFD (https://bfd.mmseqs.com), Uniclust30 v2018_08 (https://wwwuser.gwdg.de/~compbiol/uniclust/2018_08/),

and MGnify clusters v2018_12 (https://ftp.ebi.ac.uk/pub/databases/metagenomics/peptide_database/2018_12/). Uniclust30 v2018_08 was further used as input for constructing a distillation structure dataset.

Code and programmings availability

Source code

for the AlphaFold model, trained weights, and an inference script is available under an open-source license at https://github.com/deepmind/alphafold.

Neural networks

Neural networks were developed with

• TensorFlow v1 (https://github.com/tensorflow/tensorflow),
• Sonnet v1 (https://github.com/deepmind/sonnet),
• JAX v0.1.69 (https://github.com/google/jax/),
• Haiku v0.0.4 (https://github.com/deepmind/dm-haiku).

MSA search

For MSA search on

• UniRef90, MGnify clusters, and reduced BFD we used jackhmmer and for template search on the PDB SEQRES we used
• hmmsearch, both from HMMER v3.3 (http://eddylab.org/soft-ware/hmmer/).

For template search against PDB70, we used HHsearch from HH-suite v3.0-beta.3 14/07/2017 (https://github.com/soedinglab/hh-suite). For constrained relaxation of structures, we used OpenMM v7.3.1 (https://github.com/openmm/openmm) with the Amber99sb force field.

Docking analysis

Docking analysis on DGAT used

• P2Rank v2.1 (https://github.com/rdk/p2rank),
• MGLTools v1.5.6 (https://ccsb.scripps.edu/mgltools/)
• and AutoDockVina v1.1.2 (http://vina.scripps.edu/download/) on a workstation running Debian GNU/Linux rodete 5.10.40-1rodete1-amd64 x86_64.

Data analysis

Data analysis used

• Python v3.6 (https://www.python.org/),
• NumPy v1.16.4 (https://github.com/numpy/numpy),
• SciPy v1.2.1 (https://www.scipy.org/),
• seaborn v0.11.1 (https://github.com/mwaskom/seaborn),
• scikit-learn v0.24.0 (https://github.com/scikit-learn/),
• Matplotlib v3.3.4 (https://github.com/matplotlib/matplotlib),
• pandas v1.1.5 (https://github.com/pandas-dev/pandas),
• and Colab (https://research.google.com/colaboratory).
• TM-align v20190822 (https://zhanglab.dcmb.med.umich.edu/TM-align) was used for computing TM-scores.

Structure analysis

Structure analysis used Pymol v2.3.0 (https://github.com/schrodinger/pymol-open-source).