mRNA therapy offers promising capabilities in biomedicine, although its efficacy and safety depend on off-target effects, including non-specific expression of the protein load.
5' and 3' UTRs affect mRNA translation efficiency and stability through interactions with miRNAs and RNA-binding proteins, potentially enabling rational design of sequences with cell-type-specific activity.
Here we employed a rich dataset obtained with a massively parallel reporter assay, MPRA, which profiled 5' and 3' UTR activity in multiple cell lines (MDA-MB-231, HepG2, Jurkat, NALM6, SW480) to train predictive and generative deep learning models and develop PARADE (Prediction And RAtional DEsign of UTRs). Running TF-MoDISco with PARADE along with classic motif discovery with MPRA data revealed hundreds of known and dozens of novel RNA motifs.
Next, we trained and compared generative deep learning models, including cGAN, Fast SeqProp, Flow Matching, and cold diffusion; the latter was chosen for validation alongside a classic genetic algorithm. The newly designed UTRs showed significant correlation between the desired and achieved activity and a major improvement in cell-type-specificity. The most successful designs were validated in vitro in traditional luciferase reporter assays and in vivo in mice injected with nanoparticles containing Fluc-encoding mRNAs bearing PARADE‑engineered UTRs.

Co-authors: Elizaveta Aristova, Matvei Khoroshkin, Arsenii Zinkevich, Hassan Yousefi, J. Winston Arney, Sean B. Lee, Tabea Mittmann, Karoline Manegold, Arsenii Rybakov, Semen Karaban, Irina Eliseeva, David R. Raleigh, Hani Goodarzi, Ivan Kulakovskiy