DeepPrime7: Predicting PE7 Prime Editing Efficiency Across PAM Variants

Yumin Cheong, Yunyoung Choi, Hui Kwon Kim, Hyongbum Henry Kim

Keywords: Prime editing, PE7, Genome editing, Multi-task learning

TL;DR: DeepPrime7 is a PAM-aware predictor for PE7 that jointly models pegRNA sequence and editor choice. Trained on PAM-diverse screens, it accurately predicts efficiency and enables reliable PE-system selection beyond NGG.

Abstract:

Prime editing efficiency depends critically on pegRNA design and PAM compatibility, but the emergence of Prime Editor 7 (PE7) systems with engineered PAM variants has introduced an additional layer of complexity. Selecting the appropriate editor configuration for each target now requires more than pegRNA ranking. Existing predictors, largely trained on PE2/NGG data, do not explicitly model PE7-or PAM-dependent activity differences.
Here, we present DeepPrime7, a PE7-optimized predictor trained on PAM-diverse pooled screening data that jointly models pegRNA sequence features and PE7 system identity. Using a multi-task formulation and strict guide-disjoint evaluation, DeepPrime7 accurately predicts editing efficiency across NGG-, NRCH-, and NRTH-compatible PE7 systems and enables both PE-system selection and pegRNA prioritization.

Submission Number: 44