[en] Generative adversarial networks (GANs) struggle to capture position-specific nucleotide constraints that define splice site identity. We present a frequency-blended framework that guides GAN synthesis by linearly combining model predictions with conditional empirical priors. We evaluate synthetic donor and acceptor splice sites from Arabidopsis thaliana and Homo sapiens through direct biological assessments (3-mer context and sequence logo) and functional validation using SpliceRover, a state-of-the-art splice site classifier. Our results show that frequency blending substantially improves biological fidelity and predictive performance compared to unguided generation, recovering position-specific motifs that GANs systematically miss. Data augmentation with 50% real and 50% frequencyblended synthetic sequences achieves baseline-level predictive performance, effectively halving real genomic data requirements while maintaining state-of-the-art classification accuracy.
