The field of gene editing has been revolutionized by CRISPR/Cas9 technology. Cas9 proteins can target particular genomic regions for editing thanks to guide RNAs (gRNAs). Nevertheless, as gRNAs differ in their editing efficiency, computational techniques were created to forecast editing efficiency for any given gRNA of interest. In order to train machine learning models to predict editing efficiency, high-throughput datasets of Cas9 editing efficiencies were created. Nevertheless, there is no association between these high-throughput datasets and functional and endogenous datasets, which are too tiny to be used for fine-tuning machine-learning models.The field of gene editing has been revolutionized by CRISPR/Cas9 technology. Cas9 proteins can target particular genomic regions for editing thanks to guide RNAs (gRNAs). Nevertheless, as gRNAs differ in their editing efficiency, computational techniques were created to forecast editing efficiency for any given gRNA of interest. In order to train machine learning models to predict editing efficiency, high-throughput datasets of Cas9 editing efficiencies were created. Nevertheless, there is no association between these high-throughput datasets and functional and endogenous datasets, which are too tiny to be used for fine-tuning machine-learning models.
The authors created a deep learning model called DeepCRISTL to forecast the editing effectiveness in a particular cellular setting. Using high throughput datasets, DeepCRISTL learns broad patterns of gRNA editing efficiency and then applies endogenous or functional data to fine-tune the model to a particular cellular setting. The authors evaluated two cutting-edge models for editing efficiency prediction, the newly enhanced DeepHF and CRISPRon, paired with different transfer-learning techniques. These models were trained on high-throughput datasets. The optimal outcome was achieved by combining CRISPRon with fine-tuning all model weights. On functional and endogenous datasets, DeepCRISTL beat state-of-the-art approaches in predicting editing efficiency in a particular cellular setting. The authors discovered and contrasted the salient elements that DeepCRISTL learned in various cellular settings using saliency maps. DeepCRISTL, in our opinion, will enhance prediction accuracy.
You may access DeepCRISTL at https://github.com/OrensteinLab/DeepCRISTL.
Reference:
Shai Elkayam et. al. (2024). DeepCRISTL: deep transfer learning to predict CRISPR/Cas9 on-target editing efficiency in specific cellular contexts. Bioinformatics 40(8):btae481