Generative Adversarial Networks (GANs) Based Synthetic Sampling for Predictive Modeling.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Ligand-based molecular modeling requires extensive datasets.
• Generative Adversarial Networks (GANs) offer potential for data generation.
• Mapping chemical space is crucial for identifying novel molecular properties.

Purpose of the Study:

• To evaluate the utility of GANs in mapping chemical structural space.
• To generate synthetic molecular samples for ligand-based modeling.
• To assess GANs' performance in data augmentation for bioactivity prediction.

Main Methods:

• Training GANs on subsamples of BACE-1 and Dengue Virus (DENV) inhibitory activity datasets.
• Generating synthetic molecular examples using trained GANs.
• Building and validating predictive classifiers using pooled original and synthetic data.

Main Results:

• Classifiers built with GAN-augmented data showed strong predictivity for BACE-1 (ACC=0.80, MCC=0.59) and DENV (BACC=0.81, MCC=0.70).
• Performance was comparable or superior to existing literature models.
• GANs effectively mapped chemical spaces and generated useful synthetic samples.

Conclusions:

• GANs demonstrate significant utility in mapping chemical space for molecular property profiles.
• Synthetic data generated by GANs can enhance computational modeling and drug discovery efforts.
• This methodology provides a valuable tool for data augmentation in cheminformatics.