altpath-autophagy-gnn

AltPath Autophagy GNN

Predicting Chaperone-Mediated Autophagy (CMA) and endosomal Micro-Autophagy (eMI) substrates using structure-aware deep learning.

🎯 Overview

Existing tools like KFERQ-finder perform simple binary sequence pattern matching — they can't tell whether a KFERQ motif is surface-exposed and accessible to HSC70, or buried deep in the protein core. AltPath-GNN fixes this by considering the 3D structural context of each motif.

📊 Results

AUPRC 95% CI: [0.786, 0.958] | 5-fold CV

🏗️ Architecture

Protein Sequence → SaProt-650M (frozen) → 1280-dim embedding → 3-layer MLP → CMA probability
                                     ↗
                            Foldseek 3Di tokens

• Backbone: SaProt-650M — structure-aware protein language model
• Classifier: MLP [256→128→64] with LayerNorm + Dropout(0.4)
• Loss: Focal loss (α=0.75, γ=2.0)
• Training: 5-fold stratified CV, 73 proteins total

🚀 Quick Start

# Coming soon: pip install altpath-gnn
# For now, use the prediction script:

# Download and run
wget https://huggingface.co/vedatonuryilmaz/altpath-autophagy-gnn/resolve/main/predict.py
python predict.py P04406  # Predict GAPDH

📦 Dataset

• 55 validated CMA/eMI substrates from Cuervo lab, Kirchner 2019, 2020-2025 literature
• 18 negative controls including gold-standard negatives, eMI-only, and KFERQ+ non-substrates
• All sequences fetched from UniProt
• Full dataset card

🔬 Novelty

This is the first ML model for CMA substrate prediction that uses structural information. All existing tools (KFERQ-finder, PhilippKirchner/KFERQ_analysis) are sequence-only pattern matchers with high false-positive rates.

📚 Citation

@software{altpath_gnn_2026,
  author = {vedatonuryilmaz},
  title = {AltPath Autophagy GNN},
  year = {2026},
  url = {https://huggingface.co/vedatonuryilmaz/altpath-autophagy-gnn}
}

📄 License

MIT

<!-- ml-intern-provenance -->

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "vedatonuryilmaz/altpath-autophagy-gnn"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

For non-causal architectures, replace AutoModelForCausalLM with the appropriate AutoModel class.