AFSE

AFSE enhances deep graph learning (DGL) models' prediction of ligand bioactivities by adversarially refining feature subspaces for improved molecular modeling and virtual screening.

Key Features:

• Dynamic Representation Generation: Employs bi-directional adversarial learning to dynamically generate abundant representations within new feature subspaces.
• Minimization of Molecular Divergence Loss: Minimizes the maximum loss associated with molecular divergence and bioactivity to enforce local smoothness in model outputs.
• Enhanced Generalization: Improves generalization of DGL models to mitigate overfitting on small or biased training datasets, including inner active cliff cases.
• Comprehensive Evaluation Metrics: Evaluates performance using enhancement factor (top-10%, 20%, and 30%), root mean square error (RMSE), and the coefficient of determination (r^2).

Scientific Applications:

• Virtual Screening: Facilitates virtual screening of compound databases to prioritize drug hits by improving predictive accuracy for ligand bioactivities.
• Benchmarking on GPCRs: Validates and benchmarks DGL model improvements across 33 G-protein-coupled receptor (GPCR) datasets and seven state-of-the-art open-source DGL models.
• Robustness to Dataset Bias: Addresses challenges posed by small-sized and biased training datasets and inner active cliff cases to increase reliability in pharmacological research.

Methodology:

Integrates bi-directional adversarial learning to generate abundant representations in new feature subspaces, minimizes the maximum molecular divergence and bioactivity loss to ensure local smoothness, and assesses performance using enhancement factor (top-10%, 20%, 30%), RMSE, and r^2 while benchmarking against seven open-source DGL models across 33 GPCR datasets.