High-content image generation for drug discovery using generative adversarial networks.

Area of Science:

• Computational biology
• Bioimage analysis
• Machine learning in drug discovery

Background:

• High-content imaging in drug discovery generates vast data requiring automated analysis.
• Deep learning models excel at image interpretation but need large, high-quality datasets, often scarce in preclinical research.
• Generative modeling offers a solution to synthesize data for phenotypic profiling.

Purpose of the Study:

• To develop a computational framework using generative modeling for synthesizing high-content images.
• To enable phenotypic profiling of drug-induced perturbations using synthetic images.
• To address the challenge of limited high-quality data in preclinical drug discovery.

Main Methods:

• Investigated three Generative Adversarial Network (GAN) variants: Vanilla GAN, Deep Convolutional GAN (DCGAN), and Progressive GAN (ProGAN).
• Utilized a pre-trained Convolutional Neural Network (CNN) for feature extraction from real and synthetic images.
• Evaluated synthetic image quality by comparing feature distributions with real images and trained classification models on augmented datasets.

Main Results:

• Deep Convolutional GAN (DCGAN) demonstrated the highest efficiency in generating realistic synthetic images.
• The DCGAN-based framework successfully synthesized high-quality cellular and bacterial images.
• Augmenting real image data with synthesized images improved classification performance compared to using real images alone.
• Demonstrated application on bacterial images, showing distinct feature distributions for different drug treatments.

Conclusions:

• The proposed DCGAN-based framework effectively generates realistic synthetic high-content images.
• This approach enables robust analysis of drug-induced effects on cells and bacteria, even with limited real data.
• Synthetic image generation is a valuable tool for enhancing phenotypic profiling in drug discovery.