Listeria Genome Identification Using DNABERT Embedding With LightGBM and SHAP-Based Explainable Classification.

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Food Safety Science

Background:

• Accurate whole-genome identification of Listeria monocytogenes is crucial for food safety and outbreak prevention.
• Current methods (culture-based, PCR, NGS) are often slow, labor-intensive, or use non-interpretable machine learning models.
• There is a need for efficient, accurate, and explainable genomic identification tools for pathogen surveillance.

Purpose of the Study:

• To develop an explainable genomic classification framework for distinguishing Listeria from non-Listeria bacterial genomes.
• To couple transformer-based DNA embeddings with gradient boosting for high-performance genome classification.
• To utilize SHapley Additive exPlanations (SHAP) for model interpretability and identification of genomic signatures.

Main Methods:

• Assembled and filtered a dataset of 700 bacterial genomes (350 Listeria monocytogenes, 350 non-Listeria).
• Encoded genomes using DNABERT for contextual DNA embeddings based on 6-mer tokenization.
• Classified embeddings using a LightGBM model and interpreted predictions with SHAP analysis.

Main Results:

• The DNABERT + LightGBM + SHAP pipeline achieved 95.00% corrected accuracy, classifying 665/700 genomes correctly.
• Achieved high performance metrics: 94.37% precision, 95.71% recall, 95.03% F1-score, and 0.9976 AUC.
• Outperformed conventional methods including k-mer based Random Forest, TF-IDF + SVM, CNN, XGBoost, and DNABERT + Logistic Regression.

Conclusions:

• The proposed framework offers a high-performance and interpretable solution for genome-scale Listeria identification.
• SHAP analysis identified discriminative sequence patterns potentially indicative of Listeria genomic signatures.
• This approach provides a transferable template for explainable pathogen genomics in food safety and public health.