BioGAN: Enhancing Transcriptomic Data Generation with Biological Knowledge.

Area of Science:

• Computational Genomics
• Bioinformatics
• Machine Learning

Background:

• Computational genomics drives data-driven disease prediction and precision medicine.
• Challenges include data scarcity, privacy concerns, and inherent biases.
• Synthetic data generation is a potential solution, but current AI methods lack biological grounding.

Purpose of the Study:

• To introduce BioGAN, a novel generative framework for synthetic transcriptomic data.
• To address limitations of existing methods by incorporating biological knowledge.
• To improve the realism and utility of generated transcriptomic profiles.

Main Methods:

• Developed BioGAN, integrating Graph Neural Networks (GNNs) within a Generative Adversarial Network (GAN).
• Leveraged gene regulatory and co-expression networks to guide data generation.
• Validated on *E. coli* and human gene expression datasets using unsupervised and supervised metrics.

Main Results:

• BioGAN successfully preserves biological properties in synthetic transcriptomic data.
• On human data, BioGAN improved precision by 4.3% and correlation with real data by up to 2.6%.
• Synthetic data from BioGAN enhanced downstream disease and tissue classification performance by an average of 5.7%.

Conclusions:

• Integrating a priori biological knowledge into generative models is effective for enhancing synthetic data quality and utility.
• BioGAN demonstrates robustness and strong predictive utility, outperforming state-of-the-art models.
• The framework offers a promising solution for data scarcity and bias issues in genomics research.