Geometry-Informed PEFT Boosts Blood-Brain Barrier Permeability Prediction.

Predicting whether a molecule can cross the blood-brain barrier (BBBP) is a vital but challenging step in central nervous system drug discovery, complicated by limited and imbalanced datasets and the critical role of molecular structure. While Graph Neural Networks (GNNs) are powerful for molecular representation, and pre-trained GNNs offer transferable knowledge, traditional full fine-tuning is often inefficient and prone to overfitting. Existing parameter-efficient fine-tuning (PEFT) methods typically focus on 2D covalent graphs or node features, overlooking crucial 3D geometry and higher-order interactions. To overcome this, the new BBBP-GeoPEFT framework integrates geometry-informed insights. It constructs distance-based graphs from molecular conformers at various cutoffs and their corresponding line graphs to capture spatial atom and second-order edge interactions. Lightweight auxiliary geometric graph encoders generate cutoff-specific representations, which are then integrated into each pre-trained layer using node-wise cutoff attention and gated residual connections. This design effectively preserves pre-trained knowledge while introducing permeability-relevant geometric information with a minimal increase in trainable parameters (only 10.1% of the model). Experiments show BBBP-GeoPEFT achieves competitive or superior performance compared to full fine-tuning and other PEFT baselines.