Conditional GAN (cGAN)

The core mechanism of a cGAN involves feeding a condition vector y (e.g., a one-hot encoded cell type or disease state) into both the generator and discriminator. The generator learns the conditional distribution P(x|y), producing synthetic genomic sequences that are not just realistic but also statistically consistent with the specified label. This transforms the generator from an unsupervised sampler into a steerable model capable of producing data for a specific cell type or regulatory context on demand.

In genomic applications, a dedicated auxiliary classifier is often integrated into the discriminator network to enforce label fidelity. This classifier predicts the condition y from the generated sequence, adding a classification loss term to the adversarial objective. This dual loss structure ensures that the synthetic sequence not only fools the discriminator into believing it is real but also strongly exhibits the molecular signature of the target condition, such as a specific transcription factor binding profile.

cGANs facilitate the learning of disentangled representations where the latent noise vector z captures stochastic biological variation independent of the condition. This allows for controlled experimentation:

• Style mixing: The condition can be changed while keeping the noise vector fixed, generating the 'same' genome under a different disease state.
• Interpolation: Smooth transitions between conditions can be generated by interpolating the condition vector, revealing intermediate phenotypic states.

A primary use case in genomics is augmenting datasets for rare diseases where real samples are scarce. By conditioning on a pathogenic variant label, a cGAN can generate thousands of synthetic genomes carrying a specific mutation. This augmented dataset can then be used to train downstream variant calling models, significantly improving their sensitivity and robustness to rare alleles without violating patient privacy constraints.

The condition can be injected into the network at various architectural levels:

• Input concatenation: The condition vector is simply concatenated with the input noise or sequence embedding.
• Conditional Batch Normalization (CBN): The condition modulates the scale and shift parameters of normalization layers, allowing the label to control feature map statistics at every level of the generator.
• Projection-based conditioning: In the discriminator, the condition is often incorporated via an inner product with the feature vector, a method derived from the projection discriminator framework for improved stability.

Standard GAN evaluation metrics are insufficient for cGANs. Quality is assessed using conditional fidelity metrics:

• Conditional Frechet Genomic Distance: Measures the distance between real and generated distributions for each specific condition class.
• Train-Synthetic-Test-Real (TSTR) by Class: A predictive model is trained on synthetic data and evaluated on real data, with performance stratified by the conditioning label to ensure per-class utility.
• Label Recovery Rate: A pre-trained classifier is used to verify that the generated sequence can be correctly assigned to its intended condition.