HydraFlow: A Controlled Comparison of Guidance Mechanisms for 5'UTR Design

Designing 5'UTRs with desired translational properties requires searching a sequence space far larger than any assay can cover. Guided discrete generative models can steer samples toward a target property, but attaining the target does not establish that the samples resemble the sequences that naturally satisfy it. We present HydraFlow, a shared discrete-flow platform built on the RNA-pretrained HydraRNA encoder, and use it to conduct a controlled comparison of three classifier-free and three predictor-guided mechanisms for mRNA 5'UTR design. We test these methods on a 2D benchmark with a closed-form conditional and an MPRA library targeting mean ribosome load (MRL) and minimum free energy (MFE). An exact-posterior diagnostic shows that logit-level classifier-free guidance (Logit-CFG) and posterior-based discrete guidance matching (DGM-post) share the same ideal posterior at $\gamma=1$, yet DGM-post learns a correction closer to the oracle. In MPRA, predictor-guided methods achieve mean normalized Wasserstein-1 distances of 27% to 30% across MRL, MFE, and GC content, versus approximately 43% for classifier-free methods. These results show that the guidance intervention point materially affects distributional fidelity and computational efficiency in 5'UTR generation.