Structure-Aware Protein Thermostability ($T_m$) & Environmental Adaptation ($OGT$) AI

---

🔬 Graphical Abstract

---

🌟 Key Highlights

1. Dual-Track 3Di & Sequence Tokenization: Fuses primary amino acid sequences with Foldseek 3Di geometric structural tokens, giving the foundation model full structure-aware spatial intelligence directly from sequence.
2. Decoupled Thermodynamic & Ecological Heads: Disjoint neural pathways separate intrinsic protein thermal denaturation ($T_m$) from host organism optimum growth temperature ($OGT$), transferring evolutionary environmental priors ($\hat{y}_{\text{OGT}} \to T_m$).
3. Calibrated Predictive Confidence Intervals: Replaces single-point estimates with Gaussian probabilistic density profiles ($\mu \pm \sigma$), giving practitioners trustworthy uncertainty bounds.
4. Interactive Web Application & In-Silico Loop Engineering: Integrated interactive suite for single-sequence analysis, Chou–Fasman secondary structure & loop identification, and live mutant thermostability scoring.

---

📊 Benchmark Performance Summary

All evaluations use strict bidirectional homology audits ($<30%$ sequence identity) between training and test sets to prevent data leakage.

Benchmark Dataset	Metric	TemBERTure	ESMStabP	DeepSTABp	ThermoFormer	StableProt (Ours)
ProThermDB (In-Distribution, n=3,340)	MAE (°C)	5.76	6.54	7.12	6.89	6.16
	CRPS (°C)	5.76	6.54	7.12	6.89	4.52
FireProtDB (OOD Holdout, n=322, <30% id)	MAE (°C)	12.76	13.12	14.05	13.84	11.85
	CRPS (°C)	12.76	13.12	14.05	13.84	8.71
Bin-Balanced OGT (Brenda/BacDive, n=1,200)	MAE (°C)	9.42	9.80	10.15	8.94	6.48
Prospective Marine Carrageenases (n=13)	Acc (%)	61.5%	53.8%	46.2%	61.5%	81.8%

---

🚀 Quickstart & Web Interface

1. Environment Setup

# Clone the repository
git clone https://github.com/Bibhuprasadbehera/StableProt.git
cd StableProt

# Create and activate conda environment
conda create -n stableprot python=3.10 -y
conda activate stableprot

# Install dependencies
pip install -r requirements.txt

2. Launching the Interactive Web Suite

# Start the FastAPI + Jinja2 web application on port 8000
python -m uvicorn inference.main:app --host 0.0.0.0 --port 8000

Open http://localhost:8000 in your browser to access:

• Home: Architecture overview and benchmark highlights.
• Predict: Instant $T_m$ and $OGT$ predictions with calibrated thermometers and amino acid composition analysis.
• Design: Chou–Fasman loop detection with live in-silico mutation testing and trial history logging.
• About: Model details, terms of use, and citation info.

---

💻 Programmatic Python API

from inference.v9_predict import V9Predictor

# Initialize the 5-seed ensemble predictor
predictor = V9Predictor(models_dir="experiments/src/training/v9_disjoint/results")

# Predict on a protein sequence
sequence = "RPDFCLEPPYTGPCKARIIRYFYNAKAGLCQTFVYGGCRAKRNNFKSAEDCMRTCGGA"
result = predictor.predict_single(sequence)

print(f"Melting Temperature (Tm): {result['tm_pred']:.2f} ± {result['tm_conf']:.2f} °C")
print(f"Optimal Growth Temp (OGT): {result['ogt_pred']:.2f} ± {result['ogt_conf']:.2f} °C")
print(f"Thermal Tier: {result.get('thermal_tier', 'Mesophilic')}")

---

📖 Citation

If you find StableProt useful in your research, please cite our manuscript:

@article{behera2026stableprot,
  title={StableProt: Structure-Aware Deep Learning for Protein Thermostability ($T_m$) and Environmental Adaptation ($OGT$) Prediction with Calibrated Confidence Intervals},
  author={Behera, Bibhu Prasad and Daxit, Anshuman},
  journal={Working Manuscript},
  year={2026},
  publisher={iBRIC--Institute of Life Sciences},
  url={https://github.com/Bibhuprasadbehera/StableProt}
}

---

📬 Contact & Affiliation

Computational Biology and Bioinformatics Laboratory
iBRIC–Institute of Life Sciences (ILS), Bhubaneswar, Odisha, India

• Bibhu Prasad Behera: bibhu.prasad@ils.res.in
• Dr. Anshuman Dixit: anshuman@ils.res.in