Abstract
Reference genomes are often treated as faithful representations of experimentally validated viral genomes, yet the relationship between historically curated reference sequences and infectivity is rarely tested experimentally. Here, we developed a cDNA-based reconstruction platform for the canonical RNA phage MS2 and used it to compare the current NCBI reference genome (RefSeq) with closely related published isolate sequences. We found that isolate-derived sequences reproducibly yielded infectious phage, whereas the current MS2 RefSeq-derived construct did not, showing that the present reference does not represent a single experimentally validated infectious genome but instead reflects sequence curation across multiple studies. We then compared conventional and AI-enabled approaches to identify minimal changes that restore infectivity to MS2 RefSeq; a human experimentalist correctly prioritized corrective changes, whereas the genome language model Evo2 did not. We also observed that closely related corrected reference-derived constructs showed a ~4-log difference in phage output, and subsequent analysis indicated that this difference was associated with an apparent replicase frameshift in the lower-output background. This suggests that the low output construct class represents rare mutations from genomes that are one mutational step away from true function, rather than uniform function of the dominant construct population. A complementary cell-free assay provided a lower-background orthogonal readout of construct-level function, yielding ~1 x106 PFU/mL from the high-output background within 2 hours while showing no detectable recovery from the low-output background. Together, these results establish a robust platform for RNA phage reconstruction and raise the possibility that historical reference genomes, especially for RNA viruses, may not always remain faithful to experimentally validated biological function. More broadly, these findings underscore the need to verify the infectivity of reference genomes, particularly when they were assembled non-contiguously or shaped by cumulative human curation. They also highlight the importance of clearly distinguishing historically curated reference sequences from experimentally validated infectious genomes when such data are used to train or evaluate AI/ML models.