P-323 - Enhanced Citrus Pathogen Detection via Machine Learning of Combined K-mer and Alignment Features

High throughput sequencing (HTS) is increasingly used to screen for graft-transmissible diseases. However, most bioinformatic tools detecting pathogens in HTS data require expert interpretation. This study explores the application of machine learning (ML) to enhance HTS-based citrus diagnostics by integrating Kraken2 and Bowtie2 genomic analyses. ‘Candidatus liberibacter asiaticus’, citrus leaf blotch virus (CLBV), and citrus exocortis viroid were used as proof of concept for a citrus ML prototype (CMLP). Logistic Regression (LR), Random Forest (RF), and Support Vector Machine (SVM) were trained using features of read counts and alignment-specific metrics in CMLP. Results show the Principal Coordinate Analysis (PCoA) revealed two clusters, positives and negatives. PCo1 explained 78.22% of the dissimilarities, reflecting the primary factors driving the separation, while PCo2 accounted for 14.45%. In model evaluation, data were split into 70% training and 30% testing sets. For CLBV, LR achieved 87% accuracy and 100% specificity but failed to detect positives (0% sensitivity). RF showed 100% accuracy, specificity, and sensitivity but risked overfitting. SVM provided a balanced performance with 100% sensitivity and 60% specificity, resulting in 95% accuracy. CMLP preliminary findings demonstrate the strengths and limitations of Kraken2 and Bowtie2, emphasizing the need for multi-feature integration. RF and SVM demonstrated strong diagnostic potential, with RF excelling in accuracy but prone to overfitting, while SVM offered a balanced approach for the citrus virus detections.