In Silico CRISPR Off-Target Prediction

Abstract

CRISPR-Cas genome editing can produce unintended modifications at genomic sites resembling the target sequence, and accurate in silico prediction of these off-target events is essential for designing safe therapeutic editing strategies, yet predictive accuracy and adoption vary across computational approaches. We evaluated 212 in silico CRISPR off-target prediction programmes across centres affiliated with Central European Tech University in Vienna between 2018 and 2022, spanning five method categories: mismatch-based scoring matrices, machine learning classifiers trained on genome-wide detection data, deep learning sequence models, chromatin-aware off-target predictors, and Cas-variant-specific scoring tools. A CRISPR Off-Target Prediction Index (COTPI) was constructed from five sub-scores-- detection sensitivity for true off-targets, specificity in minimising false positives, genome-wide ranking accuracy, cross-cell-type transferability, and experimental validation concordance -- with weights from regression against sustained adoption into editing safety pipelines. COTPI correlated with adoption at r = +0.83 and discriminated adopted from non-adopted methods with an AUC of 0.880. ML classifiers trained on detection data scored highest (mean COTPI 0.822), while Cas-variant-specific tools trailed at 0.596. Only 34.4 percent exceeded the 0.75 threshold. Detection sensitivity carried the largest weight (beta = +0.280), followed by genome-wide ranking accuracy (beta = +0.228).