High-throughput interrogation of programmed ribosomal frameshifting in human cells

Martin Mikl; Yitzhak Pilpel; Eran Segal

To determine to what extent a model trained on one PRF event could predict frameshifting potential 448 and -1 PRF signal in a different context, we trained our model (based on the full set of features) on 449 sequence variants derived from one event and predicted on variants of another event (Fig 5F, Fig 450 S10B ). While performance is usually best for variants of the same PRF event, this approach revealed 451 functional similarities that allowed relatively accurate prediction on "foreign" PRF events (ROC AUC 452 up to 0.94, Pearson r up to 0.69). There is considerable overlap with other attempts of functional 453 clustering, e.g. based on sensitivities to mutations (Fig 2A) , with SRV1 and HERV-K10 PRF sites 454