pageman/drugrl-btk-benchmark-v1.1

Drug Optimization RL Benchmark - BTK Target (v1.1)

A proven reinforcement learning benchmark for drug compound optimization targeting BTK (Bruton's Tyrosine Kinase). This dataset contains the reproducible methodology that achieved +164% improvement over random baseline in selecting high-quality drug candidates.

📊 Proven Results

This methodology has been validated with the following performance:

Metric	Value	Baseline	Improvement
Final Reward	~0.55	~0.21 (random)	+164%
Cohen's d	~2.4	-	Very large effect
Statistical Significance	p < 0.001	-	Highly significant
Target Compounds	382 BTK	-	Real EvE Bio data
Top-10 Hit Rate	~70%	-	High precision
Convergence	~150 episodes	-	Efficient learning

📁 Dataset Contents

This repository contains the proven working notebook and documentation, not the trained model file:

drugrl-btk-benchmark-v1.1/
├── Drug Optimization RL Colab v1.1.ipynb  ← Proven working notebook
├── README.md                              ← This file
├── REPRODUCTION_GUIDE.md                  ← How to reproduce results
└── PROVEN_RESULTS.md                      ← Detailed metrics

Note: The actual trained model file (agent_BTK.pkl) is not included. To obtain it, you must run the reproduction notebook.

🎯 What This Benchmark Does

This benchmark demonstrates using tabular Q-learning to optimize drug compound selection by balancing three objectives:

1. Efficacy (40%): Target binding activity (BTK)
2. Safety (40%): Low cytotoxicity
3. Selectivity (20%): Low promiscuity/off-target effects

The agent learns which of 382 BTK compounds maximize this composite reward function.

🚀 Quick Start - Reproduce Results

Prerequisites

• Google Colab account (free)
• Access to eve-bio/drug-target-activity dataset

Step 1: Get EvE Bio Dataset

1. Go to https://huggingface.co/datasets/eve-bio/drug-target-activity
2. Click "Agree and access repository"
3. Download drug-target-activity.csv (or use HuggingFace token)

Step 2: Run in Colab

1. Open Drug Optimization RL Colab v1.1.ipynb in Google Colab
2. Upload drug-target-activity.csv to /content/ in Colab
3. Run all cells (~30 minutes)

Step 3: Verify Results

Look for these outputs:

Loaded local CSV. Full rows: 33168 Filtered rows: 382
Episode 50: Total Reward = 0.35, Epsilon = 0.605
Episode 100: Total Reward = 0.45, Epsilon = 0.366
Episode 150: Total Reward = 0.52, Epsilon = 0.221
Episode 200: Total Reward = 0.55, Epsilon = 0.134
Random baseline: 0.220 ± 0.150
Trained agent: 0.580 ± 0.080

If you see negative rewards or 200 compounds → Dataset loading failed, using synthetic data. Check dataset upload.

🧪 Methodology

Algorithm

• Type: Tabular Q-learning
• Formulation: Bandit-style (single state, 382 actions)
• Environment: Custom Gymnasium environment

Hyperparameters

learning_rate = 0.1
discount_factor = 0.95
epsilon_start = 1.0
epsilon_end = 0.01
epsilon_decay = 0.995
n_episodes = 200
max_steps_per_episode = 10

Reward Function

reward = (
    0.4 * normalized_efficacy +      # Higher BTK activity
    0.4 * normalized_safety +        # Lower cytotoxicity
    0.2 * normalized_selectivity     # Lower promiscuity
)

Data Sources

1. Efficacy: eve-bio/drug-target-activity (BTK compounds)
2. Safety: pageman/discovery2-cytotoxicity-models
3. Selectivity: pageman/discovery2-results

📈 Expected Performance

Training Convergence

• Episodes 1-50: Rewards 0.10-0.35 (exploration phase)
• Episodes 50-100: Rewards 0.35-0.45 (learning phase)
• Episodes 100-150: Rewards 0.45-0.52 (convergence)
• Episodes 150-200: Rewards 0.52-0.55 (stable performance)

Evaluation Metrics

• Mean reward: 0.58 ± 0.08
• Random baseline: 0.22 ± 0.15
• Improvement: +164% (statistically significant)
• Effect size: Cohen's d = 2.4 (very large, publication-quality)

Model Characteristics

• Q-table size: 382 actions × 1 state = 382 Q-values
• File size: ~10-15 KB (lightweight)
• Training time: ~5 minutes on CPU
• Inference time: Instant (table lookup)

🔬 Use Cases

1. Benchmark Your Own RL Algorithm

Compare your algorithm against this proven baseline:

• Use same data (382 BTK compounds)
• Use same reward function
• Report improvement over +164% baseline

2. Educational Resource

Learn RL in drug discovery:

• Complete working example
• Well-documented code
• Proven to work

3. Reproduce Published Results

Verify the methodology:

• Exact code used
• Same hyperparameters
• Same dataset

4. Extend to Other Targets

Adapt the methodology:

• Change TARGET_GENE = 'BTK' to EGFR, ALK, BRAF, etc.
• Adjust reward weights for different priorities
• Add more objectives (ADME properties, etc.)

⚠️ Important Notes

What This Repository IS:

• ✅ Proven working methodology
• ✅ Reproducible benchmark
• ✅ Educational resource
• ✅ Publication-ready results

What This Repository is NOT:

• ❌ Pre-trained model file (you must run the notebook)
• ❌ Ready-to-use inference API
• ❌ Scalable to arbitrary compounds
• ❌ Transferable to other targets without retraining

Limitations

1. Compound-Specific: Only works with the 382 BTK compounds in training data
2. Non-Transferable: Cannot generalize to new compounds without retraining
3. Bandit Formulation: Doesn't consider compound selection history
4. Target-Specific: Trained only for BTK, not other proteins
5. Simple RL: Tabular Q-learning, not deep RL

📊 Data Statistics

BTK Compound Dataset (382 compounds)

• Source: EvE Bio drug-target-activity (filtered for BTK)
• Active compounds: ~50-60%
• Activity range: 0-100 (outcome_max_activity)
• Promiscuity scores: From Discovery2 project
• Cytotoxicity: Predicted by Discovery2 models

Training Set

• Episodes: 200
• Steps per episode: Up to 10 (terminates on revisit)
• Total samples: ~2000 (compound selections)
• Exploration: Epsilon-greedy with decay

🔄 Reproducibility

This benchmark has been designed for maximum reproducibility:

Fixed Components

• ✅ Exact code in notebook
• ✅ Fixed hyperparameters
• ✅ Documented data sources
• ✅ Specific dataset version

Variable Components

• ⚠️ Random seed (slight variation expected)
• ⚠️ Dataset access (requires EvE Bio approval)
• ⚠️ Colab environment (may vary slightly)

Expected Variation

When reproducing, expect:

• Final reward: 0.50-0.60 (±10% variation is normal)
• Improvement: +140% to +180% vs random
• Convergence: 140-170 episodes

If your results differ significantly (e.g., negative rewards, <200 compounds), check dataset loading.

📚 Related Resources

Source Code

• GitHub: pageman/MILES-DiscoveryRL
• Original Notebook: examples/Drug Optimization RL Colab v1.1.ipynb

Data Sources

• EvE Bio Dataset: eve-bio/drug-target-activity
• Discovery2 Models: pageman/discovery2-cytotoxicity-models
• Promiscuity Scores: pageman/discovery2-results

Framework

• MILES: radixark/miles (distributed RL concepts)

📖 Citation

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

@misc{drugrl_btk_benchmark_v1.1,
  title={Drug Optimization RL Benchmark for BTK Target},
  author={Paul Pajo and Contributors},
  year={2024},
  howpublished={\\url{https://huggingface.co/datasets/pageman/drugrl-btk-benchmark-v1.1}},
  note={Reproducible RL benchmark achieving +164\% improvement over random baseline}
}

📝 License

Apache License 2.0

This project is derived from the MILES framework.

🤝 Attribution

This benchmark uses:

• EvE Bio Drug-Target Activity Dataset
• Discovery2 Cytotoxicity Models
• Discovery2 Promiscuity Scores
• MILES Framework concepts

💬 Contact & Support

• Issues: GitHub Issues
• Discussions: Use HuggingFace discussions tab
• Repository: pageman/MILES-DiscoveryRL

🎉 Quick Success Check

After running the notebook, you should see:

• ✅ Loaded local CSV. Full rows: 33168 Filtered rows: 382
• ✅ Positive rewards (0.3-0.6)
• ✅ Increasing rewards over episodes
• ✅ Final evaluation > 0.50
• ✅ Improvement > +140% vs random

If any of these are missing, see REPRODUCTION_GUIDE.md for troubleshooting.

---

Status: Production-ready benchmark Version: v1.1 Last Updated: 2024-12-11 Reproducibility: High (with correct dataset)