Skip to content

PROJECT_LOCKED.md

Latest commit

 

History

History
69 lines (49 loc) · 1.66 KB

PROJECT_LOCKED.md

File metadata and controls

69 lines (49 loc) · 1.66 KB

PROJECT LOCKED

BBB Permeability Predictor - Stereo-Aware GNN v1.0

Status: COMPLETED AND LOCKED Lock Date: December 20, 2025

Final Performance

Metric Value
Mean AUC 0.8968 ± 0.0156
Mean Accuracy 85.04%
Baseline Improvement +6.52%

Project Summary

  • Model: StereoAwareEncoder (GATv2 + Transformer)
  • Features: 21 dimensions (15 atomic + 6 stereo)
  • Pretraining: 322,594 ZINC stereoisomer graphs
  • Fine-tuning: BBBP dataset (2,050 molecules)
  • Web App: Streamlit UI with name/formula/SMILES input

Key Files (DO NOT MODIFY)

models/
  pretrained_stereo_full.pth      # Pretrained encoder
  bbb_stereo_fold1_best.pth       # Fine-tuned models
  bbb_stereo_fold2_best.pth
  bbb_stereo_fold3_best.pth
  bbb_stereo_fold4_best.pth       # Best fold (AUC 0.9111)
  bbb_stereo_fold5_best.pth

data/
  zinc_stereo_graphs.pkl          # 322k preprocessed graphs (1.3 GB)
  bbbp_dataset.csv                # Training data

Core Scripts:
  zinc_stereo_pretraining.py      # StereoAwareEncoder architecture
  pretrain_full_stereo.py         # Pretraining script
  finetune_bbb_stereo.py          # Fine-tuning script
  bbb_webapp.py                   # Web application
  TECHNICAL_SUMMARY.md            # Documentation

Version Tag

StereoGNN-BBB-v1.0-FINAL

This project is complete. Do not modify core model files. For improvements, create a new project directory.

Citation

If using this model, reference:

  • Architecture: Stereo-Aware GATv2 + TransformerConv
  • Features: 21-dim (atomic + R/S chirality + E/Z geometry)
  • Pretraining: Self-supervised on ZINC stereoisomers