Beyond Benchmark Accuracy: Evaluation of Ab Initio MiRNA Detection Models with Genomic Screening Considerations

The detection of microRNA precursors in genomic sequences plays a critical role in understanding post-transcriptional gene regulation. Although many recent approaches report high accuracy on curated datasets, they often lack validation under realistic genomic conditions. In this study, we present a screening-based evaluation framework to benchmark the performance and computational efficiency of various microRNA precursor detection models, including architectures that combine Word2Vec embeddings with deep learning, long short-term memory networks, and support vector machines. A secondary structure-based state-of-the-art model was also included for comparison. Using DNA sequences from Ensembl and annotations from miRBase, we constructed a balanced dataset of over ninety thousand samples, enhanced with data augmentation. Models were evaluated using both isolated sequence classification and a sliding-window genomic screening strategy, enabling a more realistic assessment of performance across continuous genomic regions. Our results reveal that although multiple models achieve strong performance under traditional evaluation, their true positive rates decline in genomic screening settings, where challenges such as boundary detection and sequence context arise. The deep learning model, however, achieved the highest overall performance, with 97% accuracy and processing speeds exceeding thirteen thousand nucleotides per second. These findings underscore the importance of incorporating genomic sliding window evaluations in future studies to ensure models are robust and scalable for genome-wide applications, ultimately advancing miRNA discovery and gene regulation research.