Prot2RNA: Discrete Diffusion and Reinforcement Finetuning for Coding-Sequence Design

A single protein can be encoded by many synonymous mRNA coding sequences (CDSs), whose composition affects RNA structure, stability, and protein output, making CDS design an attractive target for computational optimization. We introduce Prot2RNA, a protein-conditioned discrete diffusion model that learns a prior over natural human CDSs and generates sequences that closely follow natural human coding-sequence distributions. We then perform reinforcement finetuning of Prot2RNA toward normalized minimum free energy (MFE) with Group Relative Policy Optimization, both globally across many proteins (Prot2RNA-RFT) and through Test-Time Reinforcement Learning for a specific target (Prot2RNA-TTRL). Finally, we experimentally evaluate selected NanoLuc reporter designs, where protein output is quantified by luminescence. The TTRL-selected candidate gives the highest measured output in the tested panel (3.30$\times$ the Promega reference), compared with 2.38$\times$ for GEMORNA and 1.93$\times$ for LinearDesign. Within the tested Prot2RNA lineage, higher normalized MFE coincides with higher measured output, while across the full panel normalized MFE does not reproduce the experimental ranking, showing that a structural proxy alone does not determine protein output.