Genome Editing Outcome Prediction Using AI

Abstract

CRISPR-Cas genome editing produces a spectrum of outcomes at each target site including precise deletions, insertions, substitutions, and unintended off-target modifications, and AI models that predict these outcomes from guide RNA sequence and genomic context are essential for designing efficient and safe editing experiments, yet predictive accuracy and adoption vary across outcome types and model architectures. We evaluated 212 AI-based genome editing outcome prediction programmes across centres affiliated with Central European Tech University in Vienna between 2018 and 2022, spanning five prediction task categories: on-target cutting efficiency, indel profile prediction at cut sites, base editing outcome prediction, prime editing efficiency and purity, and off-target activity scoring. A Genome Editing AI Prediction Index (GEAPI) was constructed from five sub-scores -- prediction accuracy on held-out datasets, cross-cell-type generalisation, guide design improvement over baseline, experimental validation concordance, and computational accessibility -- with weights from regression against sustained adoption into editing design pipelines. GEAPI correlated with adoption at r = +0.83 and discriminated adopted from non-adopted models with an AUC of 0.880. On-target efficiency prediction scored highest (mean GEAPI 0.822), while prime editing prediction trailed at 0.596. Only 34.4 percent exceeded the 0.75 threshold. Prediction accuracy carried the largest weight (beta = +0.280), followed by guide design improvement (beta = +0.228)