DeepCRISTL: deep transfer learning to predict CRISPR/Cas9 functional and endogenous on-target editing efficiency

Area of Science:

• Genomics
• Bioinformatics
• Molecular Biology

Background:

• CRISPR/Cas9 gene editing relies on guide RNAs (gRNAs) for targeting specific DNA sequences.
• Predicting gRNA editing efficiency is crucial but challenging due to dataset limitations.

Purpose of the Study:

• To develop a deep-learning model, DeepCRISTL, for accurate prediction of CRISPR/Cas9 on-target editing efficiency.
• To address the limitations of small functional and endogenous datasets by utilizing transfer learning.

Main Methods:

• Developed DeepCRISTL, a deep-learning model incorporating multi-task and ensemble techniques.
• Pre-trained the model on a large high-throughput dataset (DeepHF) and fine-tuned it using transfer learning on functional and endogenous datasets.
• Employed gradual learning as the optimal transfer learning approach.

Main Results:

• DeepCRISTL achieved state-of-the-art performance on high-throughput datasets, with up to 0.89 Spearman correlation.
• The model significantly outperformed existing methods on functional and endogenous editing datasets.
• Saliency maps were used to identify and compare key features learned by the model.

Conclusions:

• DeepCRISTL effectively predicts CRISPR/Cas9 on-target editing efficiency by integrating high-throughput and biologically relevant datasets.
• Transfer learning is a powerful strategy for improving model performance on smaller, specialized datasets.
• DeepCRISTL offers a valuable tool for enhancing CRISPR/Cas9 editing applications.