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

Prompt and accurate identification of Listeria monocytogenes at the whole-genome level is essential for food safety surveillance and outbreak prevention, yet existing culture-based, PCR, and next-generation sequencing (NGS) workflows are either slow, labor-intensive, or rely on opaque machine learning models with limited interpretability. This study proposes an explainable genomic classification framework that couples transformer-based DNA embeddings with gradient boosting to distinguish Listeria from non-Listeria genomes. A curated dataset of 700 complete bacterial genomes (350 L. monocytogenes and 350 biologically related non-Listeria genomes) was assembled from the NCBI Assembly database and rigorously filtered to retain high-quality assemblies. Each genome was tokenized into overlapping 6-mers and encoded using the pretrained DNABERT model to obtain contextual genome-level embeddings, which were then classified with a LightGBM classifier. Model decisions were interpreted using SHapley Additive exPlanations (SHAP) to quantify the contribution of individual embedding dimensions and associated k-mer patterns to each prediction. After final verification of the confusion matrix, the proposed DNABERT + LightGBM + SHAP pipeline correctly classified 335/350 Listeria genomes and 330/350 non-Listeria genomes, corresponding to 665/700 correct classifications and a corrected accuracy of 95.00%. The same evaluation yielded precision of 94.37%, recall of 95.71%, F1-score of 95.03%, and an AUC of 0.9976. Comparative experiments show that the framework consistently outperforms conventional k-mer-based Random Forest, TF-IDF + SVM, CNN, XGBoost, and DNABERT + Logistic Regression baselines across all metrics. Beyond its strong predictive performance, SHAP-based analysis reveals discriminative sequence patterns that may correspond to putative genomic signatures of Listeria. Overall, the proposed approach provides a high-performance, interpretable tool for genome-scale Listeria identification and offers a transferable template for explainable pathogen genomics in food safety and public health applications.