An Attention-Gated Graph Spiking Neural Membrane System for Structure-Activity Relationship Prediction

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking Neural P (SNP) systems offer event-driven computation and temporal modeling.
• Existing SNP models struggle with long-range dependencies due to fixed or local information propagation.

Purpose of the Study:

• Introduce the Attention-Gated Spiking Neural membrane system (AGSNP) to enhance SNP capabilities.
• Address limitations in capturing contextual interactions in complex structured data.

Main Methods:

• Incorporate an attention-guided gating mechanism directly into spiking neuron dynamics.
• Embed attention signals into nonlinear spiking updates and memory regulation.
• Instantiate AGSNP within a graph-based learning framework for Structure Activity Relationship (SAR) prediction.

Main Results:

• AGSNP consistently outperforms representative baseline methods on benchmark datasets.
• Achieved significant improvements (2.0-5.7% AUC on Tox21, 3.5-17.0% on MUV) under limited data and class imbalance.
• Demonstrated effective adaptive information propagation across distant structural components.

Conclusions:

• AGSNP offers a novel approach to enhance biologically inspired neural networks.
• The model's embedded attention mechanism improves performance on complex predictive tasks like SAR.
• AGSNP shows promise for applications requiring robust modeling of structured data with limited or imbalanced information.