manumartinm/bbb-geo

BBB Geo EGNN — exp09_struct_egnn_noise

Geometry-only EGNN that predicts BBB permeability from 3D coordinates (p_geo). Used as a differentiable guidance signal in BoltzGen diffusion.

Model summary

Field	Value
Experiment	exp09_struct_egnn_noise
Architecture	struct_egnn_geo
EGNN hidden	64
EGNN layers	3
PR-AUC (calibrated)	0.8409463785514206
Guidance gate	hybrid (gate_pass=True)

Files

• checkpoints/best.ckpt — PyTorch weights
• calibrators/calibrator.pkl — isotonic calibration
• train_metadata.json — experiment + struct config
• metrics.json, metrics_multisigma.json, guidance_gate.json

Usage

cd TFG/bbb_models && uv sync
hf download manumartinm/bbb-geo --local-dir ./bbb-geo

Structural prediction (requires coords manifest or coords_path column):

uv run python scripts/geo/predict.py \
  --run-dir ./bbb-geo \
  --input candidates.parquet \
  --manifest ../dataset/data/processed/peptides_struct_manifest.parquet \
  --output scored.parquet

BoltzGen diffusion guidance:

boltzgen run target.yaml \
  --protocol peptide-anything \
  --config design guidance.bbb_ckpt=/abs/path/bbb-geo/checkpoints/best.ckpt \
  --config design guidance.bbb_weight=0.3 guidance.membrane_weight=0.7

Dataset

Trained on Boltz-folded peptide structures. Companion dataset: manumartinm/bbb-peptides.

Citation

If you use this model, cite the TFG BBB structural guidance pipeline.