APAS-SB: Advanced Protein Analysis System with Structure-Based Learning

A hybrid implementation combining multi-task learning with Boltz-2 datasets for comprehensive protein property prediction.

🎯 Overview

APAS-SB extends the PEARL (Protein structure prediction) model with:

• Multi-task learning for diverse biochemical properties
• Boltz-2 datasets for state-of-the-art binding affinity prediction
• ΔΔG prediction for protein mutations
• Uncertainty quantification for confidence estimation
• Density-aware training for improved structure quality

🚀 Key Features

1. Multi-Task Predictions

• Binding Affinity: Protein-ligand interactions (ChEMBL, BindingDB, PDBbind)
• Protein-Protein ΔΔG: Interaction energy changes (SKEMPI 2.0)
• Enzyme Catalysis: kcat predictions (BRENDA)
• Fitness Scores: Deep mutational scanning (ProteinGym)

2. Hybrid Dataset Integration

• 7.25M training examples across 11 datasets
• Original multi-task datasets: 2.6M examples
• Boltz-2 datasets: 4.6M examples
• 200 GB processed data with embeddings

3. Advanced Training Strategies

• Uncertainty-aware losses with confidence estimation
• Density-aware training with electron density maps
• Multi-task learning with shared representations
• Three-phase training (structure → confidence → affinity)

📊 Dataset Summary

Dataset	Examples	Task	Source
ChEMBL	600K	Binding affinity	Boltz-2
BindingDB	600K	Binding affinity	Boltz-2
PubChem HTS	2.0M	Binary classification	Boltz-2
ProteinGym	2.5M	Fitness scores	Original
BRENDA	100K	Enzyme kcat	Original
PDBbind	20K	Binding affinity	Original
SKEMPI 2.0	8K	Protein-protein ΔΔG	Original
Others	1.4M	Various	Boltz-2

🛠️ Installation

# Clone the repository
git clone https://github.com/acadev/APAS-SB.git
cd APAS-SB

# Install dependencies
pip install -r pearl/requirements.txt

🎓 Quick Start

1. Test All Boltz-2 Datasets

python scripts/test_all_boltz2_datasets.py

2. Test Boltz-2 Loss Functions

python scripts/test_boltz2_losses.py

3. Test MD Trajectory Loaders

python scripts/test_md_loaders.py

4. Download Datasets

python scripts/download_datasets.py --datasets mdcath atlas chembl bindingdb

5. Train on Oracle Cloud (64 H100 GPUs)

# Phase 2A: Baseline Training (48 GPUs, 13 days)
torchrun --nproc_per_node=8 --nnodes=6 \
    scripts/train_oracle_cloud.py --config scripts/oracle_cloud_config.yaml --phase 2a

# Phase 2B: Multi-task Training (56 GPUs, 17 days)
torchrun --nproc_per_node=8 --nnodes=7 \
    scripts/train_oracle_cloud.py --config scripts/oracle_cloud_config.yaml --phase 2b

# Phase 2C: Uncertainty-Aware Training (64 GPUs, 20 days)
torchrun --nproc_per_node=8 --nnodes=8 \
    scripts/train_oracle_cloud.py --config scripts/oracle_cloud_config.yaml --phase 2c

📈 Expected Performance

Binding Affinity (vs. Boltz-2)

• FEP+ (OpenFE): 0.64-0.66 Pearson R (Boltz-2: 0.62)
• CASP16: 0.66-0.68 Pearson R (Boltz-2: 0.65)
• MF-PCBA: 0.026-0.028 AP (Boltz-2: 0.0248)

Multi-Task (New Capabilities)

• Protein-Protein ΔΔG: 0.55-0.60 Pearson R
• Enzyme kcat: 0.45-0.50 Pearson R
• Fitness Scores: 0.50-0.55 Spearman ρ

💰 Cost Estimates

Phase	Time	Cost	Description
Data Preparation	40 days	$8K	Download + processing
Training (Pretrained)	24 days	$20M	32× A100 GPUs
Training (From Scratch)	75 days	$87M	64× A100 GPUs
Storage	-	$10/month	200 GB processed data

📚 Documentation

Essential Documents

• Implementation Complete Summary - Latest implementation status (Steps 1-5)
• Development Roadmap - 85-day training plan for Oracle Cloud (64 H100 GPUs)
• Documentation Index - Complete documentation navigation

Quick Links

• Quick Start Guide - Getting started
• Deployment Guide - HPC deployment
• Architecture Overview - Technical design documents
• Cost Analysis - Training cost estimates

🏗️ Architecture

APAS-SB/
├── pearl/                          # Core library
│   ├── models/                     # Model architectures
│   │   ├── pearl.py               # Base PEARL model
│   │   ├── multitask_pearl.py     # Multi-task extension
│   │   ├── ddg_predictor.py       # ΔΔG prediction
│   │   └── ...
│   ├── training/                  # Training utilities
│   │   ├── losses.py              # Loss functions
│   │   ├── boltz2_losses.py       # Boltz-2 loss functions (NEW)
│   │   └── ...
│   └── data/                      # Data loaders
│       ├── multitask_datasets.py  # 11 dataset loaders (NEW)
│       ├── mdcath_loader.py       # mdCATH MD trajectories (NEW)
│       ├── atlas_loader.py        # ATLAS MD trajectories (NEW)
│       └── ...
├── scripts/                       # Training & testing scripts
│   ├── train_oracle_cloud.py      # Oracle Cloud training (NEW)
│   ├── download_datasets.py       # Dataset downloader (NEW)
│   ├── test_all_boltz2_datasets.py # Dataset tests (NEW)
│   ├── test_boltz2_losses.py      # Loss function tests (NEW)
│   ├── test_md_loaders.py         # MD loader tests (NEW)
│   └── ...
└── docs/                          # Organized documentation
    ├── guides/                    # User guides
    ├── architecture/              # Technical architecture
    ├── summaries/                 # Cost & scaling analysis
    └── archive/                   # Historical documents

🔬 Research Applications

1. Drug Discovery: Binding affinity prediction, hit-to-lead optimization
2. Antibody Design: Protein-protein interaction engineering
3. Metabolic Engineering: Enzyme catalysis optimization
4. Protein Design: Fitness-guided directed evolution

📝 Citation

If you use this code in your research, please cite:

@software{apas_sb_2024,
  title={APAS-SB: Advanced Protein Analysis System with Structure-Based Learning},
  author={acadev},
  year={2024},
  url={https://github.com/acadev/APAS-SB}
}

📄 License

[Add your license here]

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📧 Contact

For questions or issues, please open an issue on GitHub.

---

✅ Implementation Status

Current Status: ✅ Steps 1-5 Complete - Ready for Oracle Cloud deployment

Completed Features

• ✅ 7 Boltz-2 datasets implemented (ChEMBL, BindingDB, PubChem, CeMM, MIDAS, Decoys)
• ✅ 3 Boltz-2 loss functions (Huber, Pairwise Ranking, Focal)
• ✅ Download infrastructure for all datasets (mdCATH, ATLAS, etc.)
• ✅ MD trajectory loaders (mdCATH: 135K trajectories, ATLAS: 4K trajectories)
• ✅ Oracle Cloud training scripts (64 H100 GPUs, 3-phase training)
• ✅ All components tested with synthetic data

Next Steps

1. Download real datasets (Days 1-12 of roadmap)
2. Test with real data
3. Launch Phase 2A training on Oracle Cloud (48 GPUs, 13 days)
4. Scale to Phase 2B and 2C (56-64 GPUs)

Last Updated: December 2024