You can deploy this model and start earning money today!

geestaltt/
ddg-protherm-refiner-sota

Name: ddg-protherm-refiner-sota
Availability: InStock

variant-effect-prediction

other

Views

No views yet

DDG ProTherm Refiner: VAE-Guided Stability Predictor

The DDG ProTherm Refiner is a state-of-the-art specialist model for predicting the change in protein stability ($\Delta\Delta G$) upon single-point mutations. It uses a high-capacity Variational Autoencoder (VAE) to learn a topological map of mutations, which is then refined by an MLP to achieve superior predictive accuracy.

Model Description

Architecture: Specialist VAE (ProTherm) + MLP Residual Refiner.
Latent Space: 32-dimensional embedding space capturing functional mutation clusters.
Dataset: Trained on the curated ProTherm dataset (high-quality calorimetry measurements).

Scientific Discoveries

Linear Stability Manifold: 94.7% of the variance in $\Delta\Delta G$ is explained by a single learned direction in the latent space.
Cross-Protein Transfer: Discovers mutations that are functionally similar across different protein families, despite low sequence identity.

Performance

Metric	Value
Spearman $
ho$	0.89
Pearson $r$	0.88
MAE	0.53 kcal/mol

Usage

python

import torch
from ddg_vae import DDGVAE
from ddg_mlp_refiner import DDGMLPRefiner

# 1. Load VAE base
vae = DDGVAE.create_protherm_variant(use_hyperbolic=False)
vae_ckpt = torch.load("vae_protherm.pt", map_location="cpu")
vae.load_state_dict(vae_ckpt["model_state_dict"])

# 2. Load MLP Refiner
refiner = DDGMLPRefiner(latent_dim=32, hidden_dims=[64, 64, 32])
ref_ckpt = torch.load("pytorch_model.bin", map_location="cpu")
refiner.load_state_dict(ref_ckpt["model_state_dict"])

# 3. Predict
vae.eval(); refiner.eval()
with torch.no_grad():
    vae_out = vae(mutation_features)
    mu = vae_out["mu"]
    vae_pred = vae_out["ddg_pred"]
    refined = refiner(mu, vae_pred)
    ddg = refined["ddg_pred"]

print(f"Predicted DeltaDeltaG: {ddg.item():.4f} kcal/mol")

Citation

bibtex

@software{ddg_refiner_2026,
  author = {AI Whisperers},
  title = {DDG ProTherm Refiner: VAE-Guided Stability Predictor},
  year = {2026},
  url = {https://huggingface.co/ai-whisperers/ddg-protherm-refiner-sota}
}

geestaltt/ddg-protherm-refiner-sota