RIBOSPAN: an RNA model that reads full transcripts and designs mRNA

RIBOSPAN was trained to design complete mRNA molecules, the instructions that cells use to make proteins

On August 24, the preprint authors introduced RIBOSPAN, an AI model for RNA sequences. It was trained to process sequences of up to 10 240 nucleotides while resolving each individual nucleotide.

mRNA is the molecule that a cell uses as a template to make a protein. The protein-coding sequence lies in the middle, but the surrounding regions also affect how the molecule functions. The region before the coding sequence helps initiate protein synthesis, while the region after it contributes to mRNA stability and regulation. Designing an mRNA therefore requires the protein-coding instructions and the interactions among all regions of the sequence to be considered together.

Earlier RNA models were usually trained on fragments of about one thousand nucleotides. Full-length mRNAs had to be truncated, and attempts to extend the supported sequence length reduced model performance. From the start of training, RIBOSPAN processed contexts of up to 10 240 nucleotides, which allowed it to relate distant regions within the same molecule.

The authors evaluated this design choice using their own tests. In a task that required models to reconstruct masked nucleotides, models trained natively on long contexts maintained their performance on sequences 10 240 positions long. Short-context models performed substantially worse when their input range was extended directly. In another test, the researchers moved a short mRNA fragment while preserving its nucleotide composition. This tested whether the model could recognize the fragment's new context while retaining its relationship with distant regions that had not changed.

The generative component of RIBOSPAN proposes new full-length mRNA sequences by considering the region before the coding sequence, the coding region, and the region after it together. Within the coding region, it can make synonymous substitutions by changing nucleotide triplets while preserving the encoded protein. This makes it possible to redesign the RNA sequence without changing the protein-coding instructions.

In this study, RIBOSPAN was evaluated on RNA sequence reconstruction and discrimination tasks. The authors plan to report experimental validation of the biological properties of the generated sequences in a subsequent journal article. At present, the model frames the computational task as the design of mRNA as one connected sequence whose three regions must be considered together.