From sequences to images: Deep feature extraction for transposable element classification
Abstract
Transposable elements (TEs) play a fundamental role in genome evolution and contribute significantly to genetic diversity. However, their automatic classification remains challenging due to their high structural diversity and the lack of effective computational representations. In this study, we explored the potential of deep features extracted for TE sequence classification and introduced the use of general-purpose convolutional neural networks (ResNet50RS, InceptionV3, EfficientNetB0, and VGG16) to leverage their deep feature representations. To enable this integration, we proposed a two-dimensional image-based representation of TEs sequences designed to capture relevant structural and compositional patterns. These deep features were combined with classical machine learning algorithms to identify the most effective hybrid configurations, bridging traditional and modern deep learning approaches. We also evaluated end-to-end CNN models using the same representation, demonstrating that TE classification can be achieved without relying solely on specialized architectures. Experiments were conducted on five taxonomic datasets derived from multiple sources, including RepBase, covering both order- and superfamily-level classifications. The integration of deep features extracted by the state-of-the-art Transposable Elements Representation Learner (TERL) with a Random Forest classifier achieved the strongest overall performance, attaining a macro accuracy of 0.953, precision of 0.749, and F1-score of 0.717 on Dataset 3. Among general-purpose CNN architectures, ResNet50RS demonstrated the most robust and consistent behavior, surpassing TERL in superfamily-level classification tasks. Collectively, the results validate a generalizable 2D image-based representation that is fully compatible with modern CNNs, reducing dependence on domain-specific tools while revealing the untapped potential of transferable deep features. The proposed framework establishes a scalable foundation for future genomic classification pipelines, fostering feature reuse, transfer learning, and improved generalization in complex biological settings.