Predict E. coli protein solubility from sequence alone.
Frozen ESM2-650M embeddings + trained MLP head. 2.8 MB weights. Runs on CPU.

---

Quick Start

# Install
pip install git+https://github.com/61-Keys/oracle-sol.git

# Predict a protein
oracle predict MSKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLK

# Batch from FASTA
oracle predict designs.fasta --rank

# Compare against known references (GFP, MBP, insulin, etc.)
oracle predict MVKVYAPASS... --compare

# Per-residue solubility heatmap
oracle predict MVKVYAPASS... --residues

Why

You design a protein with RFdiffusion + ProteinMPNN + AlphaFold. It looks perfect in silico. You send it to the wet lab. Three weeks and $2,000 later: insoluble aggregate. Failure rates are 30-70%.

ORACLE-Sol predicts whether your protein will actually be soluble when expressed in E. coli, before you spend time and money on synthesis. It runs in seconds on a laptop.

How It Works

Your sequence
      |
      v
 ESM2-650M (frozen)     # 650M parameter protein language model
      |                  # extracts deep evolutionary features
      v
 CLS embedding [1280]   # single vector per protein
      |
      v
 MLP head (2.8 MB)      # trained on 70K real experimental outcomes
      |
      v
 P(soluble) score        # 0-100% with confidence level

The model was trained on the UESolDS dataset (70,031 E. coli proteins with real wet-lab solubility labels from TargetTrack), not synthetic or computed labels.

Python API

from oracle_sol import predict_solubility

results = predict_solubility([
    "MSKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLK",
    "DAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVIA",
])

for r in results:
    print(f"{r.name}: {r.score_pct} ({r.label}, {r.confidence} confidence)")

CLI Commands

Command	What it does
oracle predict SEQUENCE	Predict solubility for a single protein
oracle predict file.fasta	Batch prediction from FASTA
oracle predict file.fasta --rank	Ranked table sorted by score
oracle predict SEQ --compare	Compare against 8 reference proteins
oracle predict SEQ --residues	Per-residue contribution heatmap
oracle predict file.fasta -o results.csv	Export to CSV
oracle predict SEQ --quiet	Minimal output (pipeable)
oracle info	Model card + reference proteins

Performance

Metric	ORACLE-Sol	PLM_Sol (2024 SOTA)
Accuracy	71.9%	73.0%
MCC	0.450	0.469
Training data	70K proteins	70K proteins
Model size	2.8 MB head	not published
Backbone	ESM2-650M (frozen)	ESM2-650M (frozen)

The accuracy ceiling across all published methods is approximately 77% / 0.50 MCC, constrained by label noise in the experimental data — not by model capacity. Every method plateaus here.

How to Read the Output

Score (0-100%): Probability of soluble expression in E. coli. Higher = more likely soluble.

Confidence levels:

• HIGH — score is far from the decision boundary (>80% or <20%). Strong signal.
• MEDIUM — moderate certainty (65-80% or 20-35%).
• LOW — near the boundary (35-65%). Treat with caution.

Residue heatmap: Shows which regions of your protein contribute to or detract from predicted solubility. Green = stabilizing, red = destabilizing. Useful for identifying problematic regions in designed proteins.

Requirements

• Python 3.10+
• ~3 GB RAM (for ESM2-650M)
• CPU works fine (~2s per protein). GPU optional for speed.
• No internet needed after first run (ESM2 downloads once from HuggingFace Hub).

License

MIT. Use it however you want.

Citation

If ORACLE-Sol is useful in your research:

@software{oracle_sol,
  author = {Rath, Asutosh},
  title = {ORACLE-Sol: Protein Solubility Prediction with ESM2},
  year = {2026},
  url = {https://github.com/61-Keys/oracle-sol}
}

---

_{Built by 61-Keys}