High-throughput interrogation of programmed ribosomal frameshifting in human cells

Martin Mikl; Yitzhak Pilpel; Eran Segal

As described in the main text and illustrated in Figure S2A The copyright holder for this preprint (which was not peer-reviewed) is the author/funder.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 filtering, a single variant was randomly chosen for all subsequent steps to avoid biases resulting from 896 having duplicated sequences. We created sets of variants corresponding either to the full rationally 897 designed library ("full library"), the -1 PRF sites from the main set of events tested positive in our 898 assay combined ("frameshifting events") or to all variants based on one of these PRF events in 899 isolation. 10% of the variants in the individual sets (20% in case of the HIV variants) were put aside 900 and used only for evaluation of models built using the other 90% (80%). We chose Gradient Boosting 901 For predictions based on amino acid identity, information on the type of amino acids around the 915 slippery site (from position -3 relative to the site of frameshifting until the end of the slippery site) was 916 used ("unpolar", "polar" or "charged"). 917 . CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder.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 systematic and combinatorial fashion using 10-fold cross-validation for each set of variants and 919 features independently. Typically around 100 tests were performed and the best set of 920 hyperparameters used for subsequent steps. The code used to process the data, generate the figures and train and test the predictive models will 943 be made available as a GitHub repository (https://github.com/martinmikl/). The copyright holder for this preprint (which was not peer-reviewed) is the author/funder.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 Supervision: Y.P. and E.S. 958