DOI: https://doi.org/10.36719/2663-4619/130/173-184
Machine Learning-Guided Optimization of CRISPR-Cas9 Enzyme Specificity Using Physicochemical DNA Descriptors
Aygun Nasibova1* , Yasmin Fatiyeva2
Abstract. The CRISPR-Cas9 endonuclease is a programmable enzyme widely used in genome editing, yet its clinical and biotechnological utility remains limited by off-target cleavage caused by imperfect enzyme-substrate recognition. This study develops a machine learning-guided framework for improving CRISPR-Cas9 enzyme specificity by classifying potential off-target sites using Random Forest models and physicochemical DNA sequence descriptors. Experimentally validated guide-target sequence data were encoded using conventional sequence features and biophysical descriptors, including GC content, nucleotide composition, melting temperature, duplex stability, and local structural characteristics. Three model configurations were compared: a logistic regression baseline using sequence-derived features, a sequence-only Random Forest model, and a Random Forest model integrating sequence and physicochemical descriptors. The preliminary validation results showed a clear performance increase across model complexity and descriptor richness, with mean validation scores of 0.71 for logistic regression, 0.78 for the sequence-only Random Forest, and 0.93 for the combined sequence and physicochemical descriptor Random Forest. The descriptor-enriched Random Forest improved performance by 0.15 relative to the sequence-only Random Forest and by 0.22 relative to logistic regression, indicating that physicochemical properties provide biologically meaningful information for modeling Cas9-DNA recognition. These findings support the use of interpretable ensemble learning as a computational enzyme-engineering strategy for prioritizing safer guide RNAs, reducing off-target screening burden, and improving the reliability of CRISPR-based gene therapy and bioprocess applications.
Keywords: CRISPR-Cas9, enzyme engineering, off-target prediction, Random Forest, physicochemical descriptors, guide RNA design, machine learning, genome editing