README.md

---
widget:
- output:
    url: images/stableprot_abstract_graphic_nature.png
  text: '-'
base_model: westlake-repl/SaProt_650M_AF2
instance_prompt: null
license: cc-by-nc-nd-4.0
license_name: strictrestrictedlicense
license_link: LICENSE
tags:
- biology
- proteins
- protein-language-models
- thermal-stability
- structure-aware
- saprot
---

# StableProt

<Gallery />

## Model Overview

**StableProt** is a structure-aware deep learning framework for predicting protein melting temperature ($T_m$) and host organism optimal growth temperature (OGT). 

By coupling **SaProt 650M** Foldseek 3Di structural tokens with parameter-isolated dual neural projection heads and observation-level uncertainty estimation, StableProt delivers continuous point estimates alongside calibrated 95% predictive intervals.

- **GitHub Repository:** [https://github.com/Bibhuprasadbehera/StableProt](https://github.com/Bibhuprasadbehera/StableProt)
- **Base Backbone:** `westlake-repl/SaProt_650M_AF2`

---

## Repository Contents

This repository hosts the weights for the 5-seed production ensemble:
- `seed1/` to `seed5/`: PyTorch state dicts (`model_tm.pt` and `model_ogt.pt`).
- `config.json`: Architecture specifications and projection head hyperparameters.
- `normalization_stats.json`: Ground-truth z-score normalization parameters (`tm_mean`, `tm_std`, `ogt_mean`, `ogt_std`).

---

## Loading Model Weights

```python
import torch
import json

# 1. Load normalization parameters
with open("normalization_stats.json", "r") as f:
    norm_stats = json.load(f)

# 2. Load trained weights for a specific seed (e.g. seed 1)
tm_state = torch.load("seed1/model_tm.pt", map_location="cpu", weights_only=False)
ogt_state = torch.load("seed1/model_ogt.pt", map_location="cpu", weights_only=False)

print("StableProt model weights loaded successfully!")

## Download model


[Download](/bibhu111/StableProt/tree/main) them in the Files & versions tab.