Transfer learning has emerged as a powerful tool for enhancing predictive accuracy in complex tasks, particularly in scenarios where data is limited or imbalanced. This study explores the use of similarity-based pre-evaluation as a methodology to identify optimal source datasets for transfer learning, addressing the dual challenge of efficient source-target dataset pairing and off-target prediction in CRISPR-Cas9, while existing transfer learning applications in the field of gene editing often lack a principled method for source dataset selection. We use cosine, Euclidean, and Manhattan distances to evaluate similarity between the source and target datasets used in our transfer learning experiments. Four deep learning network architectures, i.e. Multilayer Perceptron (MLP), Convolutional Neural Networks (CNNs), Feedforward Neural Networks (FNNs), and Recurrent Neural Networks (RNNs), and two traditional machine learning models, i.e. Logistic Regression (LR) and Random Forest (RF), were tested and compared in our simulations. The results suggest that similarity scores are reliable indicators for pre-selecting source datasets in CRISPR-Cas9 transfer learning experiments, with cosine distance proving to be a more effective dataset comparison metric than either Euclidean or Manhattan distances. An RNN-GRU, a 5-layer FNN, and two MLP variants provided the best overall prediction results in our simulations. By integrating similarity-based source pre-selection with machine learning outcomes, we propose a dual-layered framework that not only streamlines the transfer learning process but also significantly improves off-target prediction accuracy. The code and data used in this study are freely available at: https://github.com/dagrate/transferlearning_offtargets.

CRISPR/Cas9 (Clustered Regularly Interspaced Short Palindromic Repeats and CRISPR-associated protein 9) is a popular and effective two-component technology used for targeted genetic manipulation. It is currently the most versatile and accurate method of gene and genome editing, which benefits from a large variety of practical applications. For example, in biomedicine, it has been used in research related to cancer, virus infections, pathogen detection, and genetic diseases. Current CRISPR/Cas9 research is based on data-driven models for on- and off-target prediction as a cleavage may occur at non-target sequence locations. Nowadays, conventional machine learning and deep learning methods are applied on a regular basis to accurately predict on-target knockout efficacy and off-target profile of given single-guide RNAs (sgRNAs). In this paper, we present an overview and a comparative analysis of traditional machine learning and deep learning models used in CRISPR/Cas9. We highlight the key research challenges and directions associated with target activity prediction. We discuss recent advances in the sgRNA-DNA sequence encoding used in state-of-the-art on- and off-target prediction models. Furthermore, we present the most popular deep learning neural network architectures used in CRISPR/Cas9 prediction models. Finally, we summarize the existing challenges and discuss possible future investigations in the field of on- and off-target prediction. Our paper provides valuable support for academic and industrial researchers interested in the application of machine learning methods in the field of CRISPR/Cas9 genome editing.

MOTIVATION: Off-target predictions are crucial in gene editing research. Recently, significant progress has been made in the field of prediction of off-target mutations, particularly with CRISPR-Cas9 data, thanks to the use of deep learning. CRISPR-Cas9 is a gene editing technique which allows manipulation of DNA fragments. The sgRNA-DNA (single guide RNA-DNA) sequence encoding for deep neural networks, however, has a strong impact on the prediction accuracy. We propose a novel encoding of sgRNA-DNA sequences that aggregates sequence data with no loss of information. RESULTS: In our experiments, we compare the proposed sgRNA-DNA sequence encoding applied in a deep learning prediction framework with state-of-the-art encoding and prediction methods. We demonstrate the superior accuracy of our approach in a simulation study involving Feedforward Neural Networks (FNNs), Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) as well as the traditional Random Forest (RF), Naive Bayes (NB) and Logistic Regression (LR) classifiers. We highlight the quality of our results by building several FNNs, CNNs and RNNs with various layer depths and performing predictions on two popular gene editing datasets (CRISPOR and GUIDE-seq). In all our experiments, the new encoding led to more accurate off-target prediction results, providing an improvement of the area under the Receiver Operating Characteristic (ROC) curve up to 35%. AVAILABILITY AND IMPLEMENTATION: The code and data used in this study are available at: https://github.com/dagrate/dl-offtarget. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.