BalancedDiff: Balanced Diffusion Network for High-Quality Molecule Generation.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Traditional drug discovery is costly and slow.
• Deep learning (DL) offers efficiency but struggles with sample bias and biochemical principles.
• Existing DL methods often overlook spatial arrangements and ADME properties.

Purpose of the Study:

• To develop a DL framework for efficient and high-quality molecule generation.
• To address limitations of existing DL methods in drug discovery.
• To enhance the reliability and applicability of generated molecules.

Main Methods:

• Proposed a Balance Loss to mitigate sample bias.
• Introduced a KAN-based Balanced Feature Filtering (KBFF) module integrating molecular features and spatial data.
• Developed a QikProp module for predicting ADME properties to filter molecules.

Main Results:

• The Balance Loss effectively addresses sample bias.
• KBFF module improves molecular quality by balancing chemical properties and spatial arrangements.
• QikProp module enhances molecule selection by filtering based on ADME properties.
• Experiments on CrossDocked2020 dataset show superior performance.

Conclusions:

• The proposed method significantly enhances molecule generation quality and efficiency.
• Integrating biochemical principles and property prediction improves drug-like characteristics.
• This approach offers a more reliable and practical DL solution for drug discovery.