An Intelligent Deep Transfer Learning Framework with Domain Adaptation for Cross-Cell Line CRISPR-Cas9 Guide RNA Efficiency Prediction and Therapeutic Target Prioritization

Abstract

CRISPR/Cas9 has become an important tool for targeted genome editing, and the efficiency of the guide RNA (gRNA) is one of the factors that affects the outcome of gene editing experiments. However, gRNAs designed for the same target can show considerable variation in their editing activity. Testing a large number of candidate guides experimentally can therefore require considerable time and resources. In this work, a hybrid computational framework is proposed to predict gRNA editing efficiency and to identify promising guide sequences under different cellular conditions. The proposed framework uses one-hot encoded gRNA sequences together with sequence-derived characteristics, including GC content and k-mer frequencies. A CNN–BiLSTM network with Multi-Head Attention is used to learn relevant patterns from the input sequences. The features obtained from this network are combined with the sequence-based features, and Recursive Feature Elimination (RFE) is applied to reduce redundant features. XGBoost regression is then used to estimate gRNA editing efficiency. Since gRNA performance can vary between cell lines, the framework also considers cross-cell-line prediction. Transfer learning and domain adaptation are introduced using a Domain-Adversarial Neural Network (DANN) and CORAL to reduce differences between source and target datasets. Prediction uncertainty is estimated using Monte Carlo Dropout, which provides an indication of the confidence associated with the predicted efficiency values. The framework is evaluated using gRNA datasets from HEK293T, HeLa, and HL60 cell lines. Finally, SARS-CoV-2 sequences are used as a computational case study to examine the possible use of the proposed approach for viral targets. Candidate gRNAs are prioritized by jointly considering predicted efficiency, off-target risk, sequence conservation, and prediction uncertainty. The resulting ranking provides a computational basis for selecting promising candidates for subsequent experimental validation and does not by itself establish therapeutic efficacy.