Analysis of immediate upstream gene regions in orthopoxvirus genomes through the use of DNA Duplex Stability and explainable machine learning

Abstract

Orthopoxviruses are double-stranded DNA viruses with a temporally regulated gene expression program divided into early, intermediate, and late stages. Although many temporal gene functions are known, the promoter features that distinguish expression program require more precise characterization. Here, we employed an explainable machine learning approach to classify upstream gene regions as either early or non-early using the DNA Duplex Stability (DDS) as a numerical representation. We trained a Random Forest model on NCBI orthopoxvirus genomes and evaluated it using a public transcriptomic dataset of the agent Orthopoxvirus monkeypox virus (MPXV). To determine how individual features contributed to model predictions, we employed SHAP explainable Artificial Intelligence tool. Our results show that the region from -10 to -1 relative to the start codon is highly informative for class distinction. We found that the TAAAT-like motif at position -3 is a strong predictor of the non-early class, even in sequences originally labeled as early, providing insights into its multi-stage functionality. These findings provide an interpretable framework for identifying CDS proximal regions features linked to the temporal regulation of orthopoxvirus genes. This approach may support regulatory region annotation in newly sequenced genomes and help prioritize promoter variants affecting viral gene expression.