A reference of 'AlphaFold2 Codec' include everything of AlphaFold2.

---

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
AF3-Ref-src/alphafold3-official/

# PyTorch Implementation (lucidrains)
AF3-Ref-src/alphafold3-pytorch/

# Architecture Walkthrough
AF3-Ref-src/alphafold3-walkthrough/

📓 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 Framework (NEW!)

We provide a comprehensive fine-tuning framework for adapting protein structure prediction models to downstream tasks.

👉 Full Fine-tuning Guide

Supported Models

Model	Framework	Fine-tuning Support
AlphaFold2	JAX/Haiku	✅ Full, Head-only, LoRA
AlphaFold3	JAX/Haiku	✅ Full, Head-only, LoRA
Boltz-1	PyTorch	✅ Full, LoRA, Adapter
Boltz-2	PyTorch	✅ Full, LoRA, Adapter

Fine-tuning Strategies

Strategy	Trainable Params	Use Case
LoRA	~0.1%	Small datasets, efficient fine-tuning
Adapter	~1%	Modular, multiple tasks
Head-only	~5%	New prediction tasks
Full	100%	Large datasets, maximum performance

Supported Tasks (50+ Task Types)

We support comprehensive task coverage inspired by production platforms like ProteinBase.com:

💊 Drug Discovery

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

🔬 Protein Engineering

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

🧫 Antibody Design

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

⚗️ Enzyme Engineering

Task	Outputs	Applications
Activity	kcat, Km, kcat/Km	Catalyst design
Specificity	Substrate profiles	Industrial enzymes
Directed Evolution	Fitness landscapes	Protein engineering

🔗 Protein-Protein Interactions

Task	Outputs	Applications
PPI Binding	Kd, interface stability	Complex analysis
Interface Prediction	Contact residues	Structure analysis
Hot Spot Detection	ΔΔG per residue	PPI drug targets

🧬 Function Prediction

Task	Outputs	Applications
GO Terms	MF, BP, CC	Annotation
EC Numbers	Enzyme classification	Function discovery
Localization	Subcellular compartment	Systems biology

🛡️ Immunology

Task	Outputs	Applications
B-cell Epitopes	Epitope probability	Vaccine design
T-cell Epitopes	MHC binding	Immunotherapy
Immunogenicity	ADA risk	Drug safety

📊 Structure Quality

Task	Outputs	Applications
Confidence	pLDDT, pAE, pTM	Model validation
Disorder	IDR prediction	Structure analysis
Contacts	Distance maps	Validation

Quick Start

from finetuning import TaskRegistry, create_finetuning_pipeline
from finetuning.modules import LoRAModule
from finetuning.heads import AffinityHead

# Option 1: Use Task Registry (Recommended)
# List all available tasks
print(TaskRegistry.list_all_tasks())  # 50+ tasks

# Get task info and recommendations
info = TaskRegistry.get_task_info("binding_affinity")
print(f"Recommended LoRA rank: {info.recommended_rank}")

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

# Option 2: Manual Setup
from finetuning import FineTuningConfig, Trainer

# 1. Load pretrained model
model = load_pretrained_boltz2()

# 2. Apply LoRA (only ~0.1% parameters trainable)
lora_model = LoRAModule(model, rank=8, alpha=16.0)

# 3. Add task-specific head
affinity_head = AffinityHead(AffinityHeadConfig())

# 4. Train
config = FineTuningConfig(
    strategy="lora",
    task="binding_affinity",
    lora_rank=8,
)
trainer = Trainer(lora_model, config, train_loader, val_loader)
trainer.train()

# 5. Save lightweight LoRA weights
lora_model.save_lora_weights("./lora_weights.pt")

Module Overview

finetuning/
├── configs/           # Configuration classes
│   ├── base_config.py      # FineTuningConfig, ModelConfig, TrainingConfig
│   ├── lora_config.py      # LoRA-specific configuration
│   └── task_config.py      # 25+ task configurations (ProteinBase-style)
├── modules/           # Fine-tuning modules
│   ├── lora.py             # LoRA implementation (PyTorch & JAX)
│   ├── adapter.py          # Adapter modules
│   └── prompt_tuning.py    # Prompt tuning
├── heads/             # Task-specific prediction heads (15+ specialized heads)
│   ├── affinity_head.py    # Binding affinity (Boltz-2 style)
│   ├── property_head.py    # Protein property prediction
│   ├── contact_head.py     # Contact prediction
│   ├── antibody_head.py    # Affinity maturation, humanization, developability
│   ├── ppi_head.py         # PPI binding, interface, hot spots
│   ├── enzyme_head.py      # Activity, specificity, evolution
│   ├── function_head.py    # GO terms, EC numbers, localization
│   └── epitope_head.py     # B-cell, T-cell epitopes, immunogenicity
├── trainers/          # Training utilities
│   ├── trainer.py          # Main trainer class
│   ├── distributed_trainer.py  # Multi-GPU training
│   └── callbacks.py        # Training callbacks (EarlyStopping, Wandb, etc.)
├── data/              # Data utilities
│   ├── datasets.py         # 10+ dataset classes for all task types
│   └── transforms.py       # Data augmentation (rotation, MSA dropout)
├── examples/          # Tutorial notebooks
│   └── finetuning_tutorial.ipynb  # Complete walkthrough
├── registry.py        # Task registry and factory pattern
└── utils/             # Utility functions
    ├── checkpoint.py       # Model checkpointing
    └── metrics.py          # Evaluation metrics (lDDT, TM-score, AUROC, etc.)

---

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 AF2-Ref-src/

📦 AlphaFold2 Reference Source Code (Submodules)

# Official AlphaFold (DeepMind)
AF2-Ref-src/alphafold-official/

# OpenFold (PyTorch implementation)
AF2-Ref-src/openfold/

# ColabFold (Colab-friendly version)
AF2-Ref-src/colabfold/

# MMseqs2 (Sequence search)
AF2-Ref-src/mmseqs2/

# HH-suite (Template search)
AF2-Ref-src/hh-suite/

# trRosetta2 (Predecessor model)
AF2-Ref-src/trRosetta2/

# ESM (Facebook protein language model)
AF2-Ref-src/esm/

# UniRep (Protein representations)
AF2-Ref-src/unirep/

# SeqVec (Sequence embeddings)
AF2-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).