🧬 ORACLE

Optimized Recombinant Assessment for Cellular Laboratory Expression

Predict protein solubility in E. coli using deep learning

Installation • Usage • Performance • How It Works • Citation

---

What is ORACLE?

ORACLE is a deep learning tool that predicts whether your protein will express as soluble or form inclusion bodies in E. coli. It helps researchers prioritize which proteins to express experimentally, saving time and resources.

Why use ORACLE?

• Save weeks of lab work - Know beforehand which proteins are likely to fail
• State-of-the-art - Uses ESM-2 protein language model embeddings
• Beautiful visualizations - Understand why your protein may or may not express
• Actionable recommendations - Get suggestions to improve expression

---

📦 Installation

pip install git+https://github.com/61-Keys/oracle-solubility.git

Requirements

• Python 3.8+
• PyTorch 2.0+
• ~500MB disk space (for ESM-2 model)

---

Usage

Python API

from oracle import Oracle

# Initialize predictor (downloads ESM-2 model on first run)
predictor = Oracle()

# Predict solubility
sequence = "MSKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLK..."
result = predictor.predict(sequence)

# Access results
print(result.soluble)          # True or False
print(result.confidence)       # 0.0 to 1.0
print(result.solubility_score) # Probability of being soluble
print(result.recommendations)  # List of suggestions

# Get detailed summary
print(result.summary())

# Visualize results
result.visualize()          # Full dashboard
result.visualize("minimal") # Compact card view

Command Line Interface

# Basic prediction
oracle predict MSKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLK...

# With visualization
oracle predict MSKGEELFTGVVPILVELDGDVNGHK... --visualize

# Minimal visualization
oracle predict MSKGEELFTGVVPILVELDGDVNGHK... --visualize --minimal

# Model information
oracle info

---

📊 Performance

Benchmark Results

Metric	Value
Test Set Accuracy	66.8%
F1 Score	0.67
AUC-ROC	0.73

Real-World Protein Tests

Protein	Expected	Predicted	Correct
GFP	Soluble	✅ Soluble (60%)	✅
Ubiquitin	Soluble	✅ Soluble (75%)	✅
Thioredoxin	Soluble	✅ Soluble (86%)	✅
SUMO	Soluble	✅ Soluble (88%)	✅
MBP	Soluble	✅ Soluble (71%)	✅
GPCR (membrane)	Insoluble	✅ Insoluble (74%)	✅
Amyloid Beta	Insoluble	✅ Insoluble (82%)	✅
Alpha-Synuclein	Insoluble	✅ Insoluble (78%)	✅
Spider Silk	Insoluble	✅ Insoluble (91%)	✅
Prion Protein	Insoluble	✅ Insoluble (85%)	✅

Overall: 12/14 correct

---

How It Works

ORACLE combines two types of features:

1. ESM-2 Embeddings (320 features)

• Uses Facebook's ESM-2 protein language model
• Captures evolutionary and structural information
• Pre-trained on millions of protein sequences

2. Biophysical Features (14 features)

• Sequence length
• Hydrophobicity
• Charge distribution
• Disorder propensity
• Aggregation propensity
• Amino acid composition

Architecture

Input Sequence
      ↓
┌─────────────────┐    ┌──────────────────┐
│   ESM-2 Model   │    │ Sequence Features │
│  (320-dim emb)  │    │    (14 features)  │
└────────┬────────┘    └────────┬─────────┘
         │                      │
         └──────────┬───────────┘
                    ↓
            ┌───────────────┐
            │  Neural Net   │
            │ 512→256→128→2 │
            └───────┬───────┘
                    ↓
         Soluble / Insoluble

---

Training Data

ORACLE was trained on 60,000 proteins from TargetTrack (PSI:Biology):

• 35 research centers worldwide
• Real experimental outcomes (not computational predictions)
• Balanced dataset (30,000 soluble, 30,000 insoluble)
• 17 years of structural genomics data

Label Definitions

Status	Label	Meaning
In PDB	Soluble	Successfully crystallized
Purified	Soluble	Expressed and purified
Soluble	Soluble	Confirmed soluble expression
Work Stopped	Insoluble	Failed to express/purify
Cloned	Insoluble	Never made it past cloning

---

🔧 API Reference

Oracle(device="auto", verbose=True)

Initialize the predictor.

Parameter	Type	Default	Description
device	str	"auto"	Device to use: "auto", "cpu", "mps", "cuda"
verbose	bool	True	Print loading messages

Oracle.predict(sequence) -> PredictionResult

Predict solubility for a protein sequence.

Parameter	Type	Description
sequence	str	Protein sequence (single letter amino acids)

PredictionResult

Attribute	Type	Description
.soluble	bool	True if predicted soluble
.confidence	float	Prediction confidence (0-1)
.solubility_score	float	Probability of being soluble
.features	dict	Computed sequence features
.recommendations	list	Suggestions for improving expression
.visualize(style)	method	Show visualization ("full" or "minimal")
.summary()	method	Get text summary
.to_dict()	method	Convert to dictionary

---

💡 Recommendations

ORACLE provides actionable recommendations based on sequence analysis:

Issue	Recommendation
High hydrophobicity	Try MBP or SUMO fusion tags
Aggregation-prone	Express at 16-18°C
Contains cysteines	Use Origami/SHuffle strains
Large protein (>500 aa)	Consider domain truncations
Disordered regions	May need binding partner

---

📖 Citation

If you use ORACLE in your research, please cite:

@software{oracle2024,
  title={ORACLE: Protein Solubility Predictor},
  author={Rath, Asutosh},
  year={2024},
  url={https://github.com/61-Keys/oracle-solubility}
}

Data Source

@article{targettrack,
  title={TargetTrack: A resource for tracking targets in structural genomics},
  journal={Zenodo},
  doi={10.5281/zenodo.821654}
}

---

📄 License

MIT License - see LICENSE for details.

Made with 🧬 by Asutosh Rath