Leveraging Architectural Diversity for DNA Sequence Classification

Source abstract

The rapid evolution of high-throughput sequencing technologies has effectively transformed DNA sequencers into advanced genomic biosensors. However, the accurate functional annotation and classification of raw DNA sequences directly from these biosensor outputs, without manual feature extraction, remain central challenges in computational genomics. While deep learning methods have shown remarkable success in specialized tasks, such as viral genome detection in complex sensor-derived biological samples, a comprehensive evaluation of how diverse architectural paradigms generalize across varied genomic sensing contexts warrants further investigation. In this study, we extend beyond domain-specific models by benchmarking a broad spectrum of deep learning architectures, ranging from pure Convolutional Neural Networks (CNNs) and self-attention mechanisms to hybrid CNN-Transformer frameworks and pre-trained genomic foundation models (DNABERT-2) augmented with motif-refinement modules; moreover, we propose a novel hybrid CNN-Transformer architecture. Furthermore, we develop a comprehensive ensemble strategy that integrates these diverse architectures while systematically varying key training hyperparameters, including data augmentation techniques and loss functions, to optimize predictive performance. We evaluate models across a diverse suite of genomic datasets, encompassing viral host classification, Escherichia coli promoter recognition, and human enhancer classification. Our comparative analysis highlights the relative trade-offs between convolutional local feature extraction and attention-based global context modeling when processing high-dimensional sequence data acquired by genomic sensors. The results demonstrate that ensemble and attention-guided models consistently achieve superior generalization and robustness across distinct genomic classification tasks, providing a systematic framework for robust computational analysis of sequencing derived DNA sequences.