BiGAN

BiGAN predicts novel long non-coding RNA (lncRNA)–disease associations by using a bidirectional generative adversarial network to map latent feature representations for association inference.

Key Features:

• Bidirectional Generative Adversarial Network (BiGAN): The core architecture comprises an encoder, a generator, and a discriminator that jointly learn latent representations to predict unverified lncRNA–disease associations.
• Feature Construction: Utilizes disease semantic similarity, lncRNA sequence similarity, and Gaussian interaction profile kernel similarities computed for lncRNA–disease pairs.
• Predictive Performance: Demonstrates superior performance in cross-validation compared to other state-of-the-art approaches and identifies candidate lncRNAs associated with renal cancer and colon cancer.
• Validation and Case Studies: Approximately 70% of the top 10 predicted lncRNAs are verified by recent biological research.

Scientific Applications:

• Pathogenesis analysis: Elucidates lncRNA-linked disease pathogenesis, providing insights that can aid early diagnosis and prevention.
• Biomarker discovery: Predicts potential lncRNA biomarkers for various cancers and other diseases, with case examples including renal cancer and colon cancer.
• Functional annotation: Facilitates annotation of lncRNA functions and their implications in human health by associating lncRNAs with diseases.

Methodology:

Constructs feature vectors from disease semantic similarity, lncRNA sequence similarity, and Gaussian interaction profile kernel similarities for lncRNA–disease pairs; processes these features with a BiGAN (encoder, generator, discriminator) that maps inputs into latent spaces to predict associations, with performance assessed by cross-validation.