Harnessing Deep Learning Models for Guide RNA Optimization and Off-Target Prediction in CRISPR Systems

CRISPR (Clustered Regularly Interspaced Short Palindromic Repeats)-based genome and transcriptome editing technologies have emerged as powerful tools for therapeutic, agricultural, and industrial applications. However, their broader clinical and translational use remains limited by variable guide RNA (gRNA) or single-guide RNA (sgRNA) efficiency and unintended off-target activity, which may lead to genotoxic effects and major safety concerns. To address these challenges, recent research has increasingly shifted from heuristic scoring approaches and traditional machine learning (ML) methods toward deep learning (DL) models capable of learning complex sequence-function relationships from large-scale experimental datasets generated by assays such as GUIDE-seq (Genome-wide Unbiased Identification of Double-stranded Breaks Enabled by Sequencing), CIRCLE-seq (Circularization for In Vitro Reporting of Cleavage Effects by Sequencing), and CHANGE-seq (Cumulative and Homology-independent Analysis of Nuclease Genome-wide Effects by Sequencing). This review critically examines recent advances in DL approaches for gRNA optimization and off-target prediction in CRISPR systems. We discuss the development of convolutional neural networks (CNNs), recurrent neural networks (RNNs), transformer-based architectures, and foundation models designed to improve prediction accuracy, specificity, and generalizability across diverse biological contexts.