Essential Protein Identification Using FCNN and Harris Hawks Optimization

This research addresses the critical task of computationally identifying essential proteins, vital for organismal survival, disease understanding, and drug discovery. Existing methods face limitations with diverse biological data, significant class imbalance, and complex model optimization. To overcome these, we developed a novel methodology for essential protein prediction utilizing a multi‐input deep neural network, specifically a hybrid fully connected neural network (FCNN). The proposed approach integrates diverse biological information: protein embeddings from PPI networks (node2vec), gene expression data capturing temporal and conditional dynamics, and subcellular localization features. We addressed the significant class imbalance using synthetic minority oversampling technique (SMOTE) to balance training data. Furthermore, to enhance network architecture and performance, we implemented layer‐specific hyperparameter tuning with the Harris Hawks Optimization (HHO) algorithm, adaptively adjusting dense layer neurons and dropout rates. The method was rigorously evaluated on the GSE7645 ( S. cerevisiae ) and E. coli datasets against state‐of‐the‐art techniques. Results demonstrated that the proposed multi‐input model, combined with HHO and SMOTE balancing, achieved superior performance for S. cerevisiae , with a significantly higher accuracy of 89 and a more balanced prediction: recall 82, precision 70, F 1‐score 75, compared to baselines. The proposed models achieved very good results for E. coli dataset with an accuracy of 91.89, a recall of 89.47, a specificity of 94.00, and an F 1‐score of 91.89, which are higher and balanced results in comparison with state‐of‐the‐art methods. These findings highlight the value of integrating diverse data sources and employing advanced optimization techniques to significantly advance computational essential protein prediction.