Promoter Classification in Human Genome via DNA2Vec and UNK-Aware Deep Neural Networks

This study proposes a new hybrid model combining DNA2Vec-based embedded representations with UNK character support and a deep neural network (DNN) architecture for the classification of promoter and non-promoter DNA sequences belonging to the Homo sapiens genome. The model's objective is twofold: first, to minimize the loss of contextual information, and second, to enhance the generalization performance by representing unknown or low-confidence k-mer sequences with an UNK vector. The model, which was structured with a GELU activation function and an AdamW optimization algorithm, achieved strong and balanced results, including 85.03% accuracy, 0.9376 ROC-AUC, and 0.8444 F1 score, when evaluated using a stratified 5-fold cross-validation method. The findings indicate that the proposed structure provides a more straightforward yet effective approach in comparison to the more complex models documented in the extant literature. Furthermore, this architecture provides pragmatic and comprehensible solutions in bioinformatics applications, particularly since it facilitates motif-independent learning. Future work should address the generalization capacity be increased across species and that the integration with Transformer-based models be evaluated in future studies.