I-Corps: Translation Potential of an Artificial Intelligence (AI) Co-Scientist Platform for messenger Ribonucleic Acid (mRNA) Therapeutics Design

This I-Corps project is based on the development of an artificial intelligence platform for messenger ribonucleic acid (mRNA) drug and vaccine design. Synthetic mRNA sequences are used to direct cells to produce specific therapeutic proteins for new drugs and antigens for new vaccines. However, designing an mRNA drug or vaccine requires selecting among a large number of equivalent genetic codes for the same protein, and those choices govern whether the medicine is stable, productive, and safe. Developers currently use trial and error, with many designs failing in the laboratory. This technology gives drug and vaccine developers a way to narrow the search computationally before laboratory work begins. Faster and more reliable sequence design may shorten timelines and reduce costs for vaccines, protein-replacement therapies for rare diseases, and personalized cancer treatments, and can strengthen national preparedness against emerging infectious diseases. This I-Corps project utilizes experiential learning coupled with first-hand investigation of the industry ecosystem to assess the translation potential of a codon-level equivariant foundation model for messenger ribonucleic acid (mRNA) sequence design combined with an agentic software platform. The redundancy of the genetic code, in which several codons encode one amino acid, forms a group-theoretic structure that can be imposed as an architectural inductive bias rather than learned. Synonymous codon families are represented as cyclic subgroups of sulfur dioxide (SO(2)), and equivariance to these symmetries is built into the network through geometric deep learning. Pretrained on annotated protein-coding sequences, the model leads downstream benchmarks against established nucleotide-level and codon-level baselines; uses fewer parameters; and improves property-prediction accuracy, sequence-generation realism, and preservation of functional properties. Learned representations recover known biology, correlating with guanine and cytosine (GC) content and inversely with transfer RNA abundance, yielding interpretable rather than opaque predictions. The goal is for users to gain multi-objective optimization of stability, expression, and immunogenicity with fewer laboratory iterations, which may enable rapid, programmable development of vaccines and therapeutics. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.