Dengue virus serotype classification from complete genomes using a Transformer with chunk-based sequence modeling and explainability analysis

Abstract

Artificial intelligence is increasingly used for genomic surveillance of infectious diseases, but predictive models must be evaluated under realistic generalization settings and interpreted cautiously. We present XAI-DENFormer, an explainability-oriented framework for dengue virus (DENV) serotype classification from complete viral genomes. A curated dataset of 29,091 publicly available DENV genomes was encoded as nucleotide-level one-hot sequences and processed with a chunk-based Transformer encoder designed for memory-efficient whole-genome analysis. Generalization was assessed using three complementary validation protocols: CD-HIT cluster-aware, geographical, and temporal validation. DENFormer variants were benchmarked against Logistic Regression, a feed-forward neural network, Longformer, and Performer under identical conditions. The mean-pooling DENFormer configuration maintained high performance across all three validation protocols, with accuracies of 98.68–99.91% across CD-HIT folds, 97.18–100.00% across geographical holdouts, and 96.35–99.86% across temporal holdouts. To inspect model behavior, we combined attention maps, Gradient × Input attributions, embedding visualizations, and regional statistical analyses. The resulting profiles supported distributed serotype-associated signals across structural, non-structural, and untranslated regions rather than reliance on a single genomic locus. XAI-DENFormer provides a robust and interpretable framework for large-scale DENV genome classification and model-guided genomic inspection.