Cross-Adversarial Learning for Molecular Generation in Drug Design

Area of Science:

• Computational chemistry
• Artificial intelligence in drug design

Background:

• Molecular generation is crucial for drug design but faces challenges in validity and property optimization.
• Existing deep learning methods for molecular representation often struggle with generation accuracy and semantic label information.

Purpose of the Study:

• To propose a novel cross-adversarial learning method, CRAG, for molecular generation.
• To integrate the strengths of Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs) for improved molecular property prediction and diversity.
• To address the limitations of existing methods in terms of generation validity and semantic information.

Main Methods:

• CRAG employs an adversarially regularized encoder-decoder to convert Simplified Molecular Input Linear Entry Specification (SMILES) into discrete variables.
• Discrete variables are trained for property prediction and adversarial sample generation using projected gradient descent.
• The model is trained using a cross-adversarial pattern, combining VAE-based facticity with GAN-based diversity.

Main Results:

• Extensive experiments on two benchmarks demonstrate CRAG's effectiveness across multiple metrics.
• The proposed method shows significant improvements in molecular generation validity and property prediction.
• A novel metric, Novel/Sample, was used to effectively measure overall generation performance.

Conclusions:

• CRAG offers a robust framework for molecular generation, overcoming key limitations of prior methods.
• The integration of VAE and GAN principles enhances both the accuracy and diversity of generated molecules.
• CRAG shows significant promise for advancing AI-based molecular design in various chemical applications.