temisobodu/psichic-multitask-bbb

PSICHIC Multitask BBB Weights

Pre-trained PSICHIC weights for blood-brain barrier protein-ligand affinity prediction.

Files

• model.pt (23MB) — GNN model weights
• degree.pt (2.4KB) — Degree tensor for PNA aggregation
• config.json (1KB) — Model architecture configuration

Usage with BBB-Nuke

These weights are automatically downloaded by the BBB-Nuke pipeline when PSICHIC inference is needed for efflux transporter screening.

Model Details

• Architecture: PSICHIC (Protein-Sequence-Informed Chemical Interaction Classifier)
• Training: Multitask (regression + 3-class classification)
• Hidden channels: 200
• Layers: 3 (PNA with mean/min/max/std aggregation)
• Heads: 5 (attention)

Citation

If you use these weights, please cite the PSICHIC paper and the BBB-Nuke pipeline.