AlphaFold Stability Change Prediction

0. Quick Start

0.1 Project Structure

/data/store-data/yeyh/scripts/AF2stability/
├── develop.md                      # This document (Chinese)
├── README.md                       # English version
├── alphafold2/                     # AlphaFold2 source (modified for return_representations)
│   ├── alphafold/model/model.py   # Modified: added return_representations parameter
│   └── afdb -> /data/AFDB         # Database symlink
├── fasta_input/                    # Input FASTA files
│   ├── wt_*.fasta                  # 142 wild-type sequences
│   └── mut_*.fasta                 # 2050 mutant sequences
├── af2_output/                     # AlphaFold2 prediction outputs
├── fireprotdb_data_train.csv       # Training data (1640 samples)
├── fireprotdb_data_test.csv        # Test data (205 samples)
├── fireprotdb_data_validation.csv  # Validation data (205 samples)
├── prepare_fireprotdb.py          # Data preparation script
├── af2_stability_pipeline.py       # Stability prediction pipeline (18-dim features)
├── stability_pipeline.py           # Simple pipeline
├── run_4gpu.sh                     # 4 GPU parallel prediction script
├── run_af2_docker.sh               # Single GPU Docker script
├── extract_all_representations.py  # Extract Single Representation (384-dim)
├── train_model.py                 # MLP training (1152→512→512→1)
├── check_progress.py              # Monitor prediction progress
├── sequence_alignment.py          # Handle sequence length alignment
└── visualize_results.py           # Visualize results

0.2 File Index

File	Type	Description
Shell Scripts
run_4gpu.sh	Shell	4 GPU parallel prediction (distributes to GPU 0-3)
run_af2_docker.sh	Shell	Single GPU Docker script
Python Scripts - Data
prepare_fireprotdb.py	Python	Download and process FireProtDB from HuggingFace
Python Scripts - Pipeline
af2_stability_pipeline.py	Python	Complete pipeline with 18-dim features
stability_pipeline.py	Python	Simple pipeline
Python Scripts - Representation
extract_all_representations.py	Python	Extract 384-dim Single Representation
Python Scripts - Training
train_model.py	Python	MLP model training for ΔΔG prediction
sequence_alignment.py	Python	Handle different sequence lengths
Python Scripts - Tools
check_progress.py	Python	Monitor prediction progress and completeness
visualize_results.py	Python	Visualize results (scatter, distribution, residuals)
Data Files (CSV)
fireprotdb_data_train.csv	CSV	Training data (1640 mutations)
fireprotdb_data_test.csv	CSV	Test data (205 mutations)
fireprotdb_data_validation.csv	CSV	Validation data (205 mutations)
Directories
fasta_input/	Dir	Input FASTA files (wt_.fasta 142, mut_.fasta 2050)
af2_output/	Dir	AlphaFold2 prediction outputs
alphafold2/	Dir	AlphaFold2 source code
results/	Dir	Model outputs and visualizations

0.3 AlphaFold2 Output Files

Each prediction directory (e.g., af2_output/wt_0/) contains:

File	Description
features.pkl	Input features (sequence, MSA, etc.)
result_model_*_pred_*.pkl	Model predictions (pLDDT, coordinates)
ranked_*.pdb	Ranked predicted structures
unrelaxed_model_*.pdb	Unrelaxed structures
relaxed_model_*.pdb	Amber-relaxed structures
confidence_model_*.json	Confidence scores
msas/	Multiple Sequence Alignments (reusable)
single_representation.npy	384-dim representation (requires extract_all_representations.py)

0.4 Quick Start

cd /data/store-data/yeyh/scripts/AF2stability
bash run_4gpu.sh &
docker ps
nvidia-smi
ls af2_output/ | wc -l

# After predictions complete, extract representations
python extract_all_representations.py
# Train model to predict ΔΔG
python train_model.py

0.5 AlphaFold2 Source Key Files Modified

alphafold2/ directory contains AlphaFold2 core code:

File/Directory	Description
alphafold/model/model.py	Added return_representations parameter

---

References

1. Zhang et al. "Applications of AlphaFold beyond Protein Structure Prediction" bioRxiv 2021
2. McBride et al. "AlphaFold2 can predict single-mutation effects" arXiv:2204.06860, Phys Rev Lett 2023
3. Pak et al. "Using AlphaFold to predict the impact of single mutations on protein stability and function" PLOS ONE 2023
4. "AlphaFold-predicted deformation probes changes in protein stability" bioRxiv 2023
5. "Stability Oracle: a structure-based graph-transformer framework" Nature Communications 2024