Balanced Training\nSets Improve Deep Learning-Based\nPrediction of CRISPR sgRNA Activity

CRISPR-Cas systems have transformed the field of synthetic\nbiology\nby providing a versatile method for genome editing. The efficiency\nof CRISPR systems is largely dependent on the sequence of the constituent\nsgRNA, necessitating the development of computational methods for\ndesigning active sgRNAs. While deep learning-based models have shown\npromise in predicting sgRNA activity, the accuracy of prediction is\nprimarily governed by the data set used in model training. Here, we\ntrained a convolutional neural network (CNN) model and a large language\nmodel (LLM) on balanced and imbalanced data sets generated from CRISPR-Cas12a\nscreening data for the yeast <i>Yarrowia lipolytica</i> and\nevaluated their ability to predict high- and low-activity sgRNAs.\nWe further tested whether prediction performance can be improved by\ntraining on imbalanced data sets augmented with synthetic sgRNAs.\nLastly, we demonstrated that adding synthetic sgRNAs to inherently\nimbalanced CRISPR-Cas9 data sets from <i>Y. lipolytica</i> and <i>Komagataella phaffii</i> leads to improved performance\nin predicting sgRNA activity, thus underscoring the importance of\nemploying balanced training sets for accurate sgRNA activity prediction.