SFOA-based optimized SeResNet-BiLSTM network with an attention mechanism for the detection of exons

Detection of exons in eukaryotic DNA sequences is a challenging research area in computational biology. In this work, a Starfish Optimization Algorithm (SFOA) based hybrid deep learning architecture is proposed for the efficient detection of exons, which is the amalgamation of squeeze and excitation residual network (SeResNet), bidirectional long short-term memory (BiLSTM), and a customized attention (CA) mechanism. Firstly, SeResNet retrieves the deeply hidden features by performing recalibration and interdependencies between the channels. After that, BiLSTM is utilized to extract temporal information for a better understanding of the input data. Later, a CA mechanism is employed to select the salient features by redistributing the feature weights obtained from the SeResNet-BiLSTM to enhance the detection accuracy. Finally, the SFOA is applied to automatically optimize the hyperparameters of the SeResNet-BiLSTM-CA framework to improve the feature extraction. Before applying the deep learning framework, the modified Gabor wavelet transform (MGWT) is used to extract the periodic feature information. The efficacy of the proposed model is validated using two benchmark datasets, namely the HMR195 and the Genscan training set. To verify the efficiency of the proposed method, its performance is compared with several pre-trained networks and various existing techniques using evaluation metrics. The simulation outcomes confirmed that the proposed SFOA-based hybrid optimized SeResNet-BiLSTM-CA model achieves superior performance than its counterparts.