BlendSplice: A Frequency-Blended Generative Framework for In Silico Synthesis of Biologically Realistic Splice Site...

Generative models for biological sequences face challenges in balancing sequence realism with diversity. We investigated whether posttraining frequency blending, combining model-learned distributions with empirical nucleotide priors, can improve synthetic-sequence quality across diverse generative architectures. We present BlendSplice, a frequency-blended generative framework for the in silico synthesis of biologically realistic splice site sequences. Our approach combines position-specific empirical nucleotide priors with 3 generative architectures: a one-dimensional U-Net-based denoising diffusion probabilistic model, a generative adversarial network, and a variational autoencoder. By linearly blending model-learned distributions with empirical base frequencies through a tunable weight parameter lambda ( $\lambda$ ), we enable explicit control over the realism-diversity trade-off. We evaluate synthetic donor (5') and acceptor (3') splice sites from Arabidopsis thaliana, Homo sapiens, and Danio rerio, at 2 sequence lengths (402 and 2,002 bp), at $\lambda =0.0$ (no blending) and $\lambda =0.5$ (balanced blending of priors and model probabilities) using direct assessments (GC content, nucleotide conservation, sequence logos, $k$ -mer divergence, and polypyrimidine tract analysis) and indirect functional validation via state-of-the-art splice site classifiers (SpliceRover and Spliceator) and independent validation with SpliceAI, under 3 evaluation protocols: Train-Real/Test-Synthetic (biological realism), Train-Synthetic/Test-Real (transferability), and fixed-budget data augmentation with varying real/synthetic mixtures. Our results demonstrate that frequency blending at $\lambda =0.5$ substantially improves biological fidelity and downstream predictive performance for the generative adversarial network and diffusion models, whereas the variational autoencoder model generates high-quality sequences even without blending. Data augmentation with a 50% real and 50% synthetic mix achieves predictive performance comparable to the baseline across all 3 species, indicating that blending is an effective strategy for generating high-quality synthetic genomic data.