Optimization of CRISPR-Cas9 based genome editing systems using machine learning models

The CRISPR-Cas9 system, one of the most powerful tools in gene editing technologies, has enabled groundbreaking progress in the fields of biotechnology and molecular biology. Accurate prediction of the on-target efficiency of single-guide RNAs (sgRNAs), which are used in this system, is a critical factor directly affecting gene editing performance. The aim of this thesis is to predict sgRNA OnTargetScore values within the CRISPR/Cas9 system using machine learning approaches and to compare the performance of different models in order to determine the most effective one. In this context, three machine learning models XGBoost, LightGBM, and Random Forest were trained using the Arabidopsis thaliana dataset. The models were evaluated using 10-fold cross-validation and compared based on performance metrics including RMSE, MAE, R², Pearson, and Spearman correlation coefficients. Additionally, the generalizability of the models was tested and validated on an independent dataset from Solanum lycopersicum (tomato). The results showed that the XGBoost model achieved the highest predictive accuracy. While LightGBM produced similar outcomes, the Random Forest model performed relatively poorly. This thesis demonstrates that machine learning-based approaches can serve as reliable, scalable, and cross-species adaptable optimization tools for predicting on-target performance in CRISPR research and can reduce experimental workload by identifying efficient sgRNA candidates beforehand.