A Capsule Network-Based Hybrid Deep Learning Model for Efficient Prediction of CRISPR-Cas9 Off-Target Effects

Abstract

CRISPR-Cas9 genome editing has transformed biomedical research and the development of therapies, yet the challenge of unintended off-target effects remains a significant obstacle to its clinical use. In this study, we present a new deep learning model that combines Capsule Networks with Transformer blocks, bidirectional LSTM layers, and CNNs. The model is further strengthened by incorporating k-mer encoded sequence features and biological rule checks to predict CRISPR-Cas9 off-target activity with greater accuracy. It processes guide and off-target DNA sequences through a hybrid pipeline, which includes convolution, temporal modelling, attention-based representation learning, and spatial hierarchy encoding using capsule layers. At the same time, the model extracts and analyses numerical features, such as mismatch counts, GC content, and PAM motif patterns. To improve reliability, we introduce a biological constraint layer that filters predictions based on well-established domain knowledge. The final predictions result from integrating these various feature representations. Our results show that this biologically-informed architecture significantly enhances both sensitivity and specificity in off-target prediction, indicating its potential to improve the safety and design of CRISPR  experiments.