Development of a novel aggregated deep learning framework for small biological datasets using overlapping subsequences

The development of deep learning models and techniques to harness the potential of data mining in biological sequences with limited data availability is a vital need. This necessity arises from unique biological characteristics and technical constraints that limit access to sufficient quantities of high-quality data across many biological and genetic fields. Building on a data augmentation strategy that generates overlapping augmented subsequences, an aggregated CNN-LSTM-Attention-Residual architecture was developed for analysis at the original full-length sequence level. The model was independently tested and validated on three sets of regulatory sequences from evolutionarily diverse organisms, chloroplasts sharing a common ancestor, and prokaryotic sequences, comprising 100, 50, and 100 sequences per group, respectively, within each dataset. Trained on augmented subsequences, the model effectively aggregated and transferred high-level features back to the original full-length sequences. It achieved high performance with approximately 96% accuracy, recall, precision, F1-score, and AUC at the subsequence level. The model correctly distinguished the sequences of the distinct groups at the full-length sequence level across all datasets. The strong performance confirmed that an ensemble of features learned from short, overlapping subsequences provide an informative signature for classifying full-length regulatory regions. The combination of first-level training on augmented subsequences with dual-level validation—evaluating both subsequences and their corresponding full-length sequences— provides a framework for advance model development in limited-data biological sequence analysis.