piCRISPR: Physically Informed Deep Learning Models for CRISPR/Cas9 Off-Target Cleavage Prediction

Abstract

CRISPR/Cas programmable nuclease systems have become ubiquitous in the field of gene editing. With progressing development, applications in in vivo therapeutic gene editing are increasingly within reach, yet limited by possible adverse side effects from unwanted edits. Recent years have thus seen continuous development of off-target prediction algorithms trained on in vitro cleavage assay data gained from immortalised cell lines. Here, we implement novel deep learning algorithms and feature encodings for off-target prediction and systematically sample the resulting model space in order to find optimal models and inform future modelling efforts. We lay emphasis on physically informed features, hence terming our approach piCRISPR, which we gain on the large, diverse crisprSQL off-target cleavage dataset. We find that our best-performing model highlights the importance of sequence context and chromatin accessibility for cleavage prediction and outperforms state-of-the-art prediction algorithms in terms of area under precision-recall curve.

Learning to quantify uncertainty in off-target activity for CRISPR guide RNAs

“…Deep learning models could effectively handle large volumes of complex data, capturing intricate patterns and achieving superior performance in predicting off-target activity. The development of deep learning-based prediction models also enabled utilization of physical features modelling the off-target activity problem with more depth [17]. Accurately quantifying uncertainty in off-target activity predictions is a crucial next milestone this study aims to address.…”

“…While conventional machine learning models have shown promising results, recent studies using deep learning techniques have demonstrated even better performance [16]. These models utilize novel sequence encoding strategies, feature engineering approaches by introducing physical features [17], class rebalancing techniques, and attention mechanisms [18] to improve prediction performance. While several models have been developed, most of them are primarily dedicated to the classification task, aiming to predict the activity status of the sgRNA-target interface.…”

“…NuPoP scores refer to a Hidden Markov Model (HMM), trained to estimate the nucleosome affinity and occupancy at single base-pair resolution. For a more detailed discussion on the effects of these features to CRISPR-based cleavage activity, we refer to Störtz et al[17].Architecture and training of crispAI. crispAI is an end-to-end multi-output CNN and bi-LSTM fusion neural network specifically designed to quantify the uncertainty in off-target cleavage activity of sgRNA-target pairs for CRISPR/Cas9 system.…”

Generating, modeling, and evaluating a large-scale set of CRISPR/Cas9 off-target sites with bulges

“…Integrating them, either by data combination or by a meta-predictor that considers the predictions of models trained on each data separately, may improve prediction performance. Second, the efficiency of assays for measuring the genome-wide activity of CRISPR/Cas9 nucleases in cellula is limited by chromatin accessibility and other epigenetic factors (31; 58; 44). Therefore, including such information in our models may provide more insights regarding their potential for future improvements.…”

“…We trained three deep-neural-network architectures (inspired by pi-CRISPR (44) and DeepHF (45) architectures) for off-target classification and regression tasks, leveraging multilayer-perceptron (MLP), gated-recurrent-unit (GRU), and embedding techniques. The first architecture, which we denote by MLP-Emb , begins with an embedding layer of size 44, which transforms each 25-long one-hot-encoded aligned nucleotide pair into a 44-long real vector representation.…”