High-throughput interrogation of programmed ribosomal frameshifting in human cells

Martin Mikl; Yitzhak Pilpel; Eran Segal

In order to predict frameshifting efficiency of naturally occurring variants of the HIV gag-pol 506 frameshifting site, we built a model based on the feature sets described above. Despite eliminating 507 the most important feature of the HIV gag-pol PRF site (the slippery sequence) by restricting our 508 analysis to clinical isolates with no sequence alterations in and around the canonical UUUUUUA and 509 despite the narrow range of frameshifting rates (Fig 6A) , we could accurately predict frameshifting 510 rates of unseen variants (Pearson r=0.6; Fig 6E, left) . Interestingly, removing the features based on 511 pairedness of downstream positions and minimum free energy led to similar agreement between 512 measured and predicted values (Fig 6E, right) . This is in line with the limited importance of the exact 513 downstream structure in comparison to other PRF events tested. Including designed HIV variants in 514 the training set yielded similar prediction accuracy (Fig S12E) , indicating that our model can learn the 515 The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/469692 doi: bioRxiv preprint relevant rules from our set of natural variants alone. Taken together, these results show that a model 516 based on data from our assay is sensitive enough to detect even subtle differences between HIV 517 variants and allows prediction of frameshifting rates with an accuracy of clinical relevance. approach entails particular limitations. While we show that our assay is highly sensitive to changes in 527 the PRF signal (e.g. in the testing of natural isolates of HIV) and therefore able to identify even subtle 528 effects of sequence alterations on frameshifting, our detection limit is higher than in assays using 529 amplification methods and/or overexpression of the reporter. 530