This study presents a data-driven approach for modeling subgrid-scale (SGS) stresses in turbulent channel flow using a fully connected neural network (FCNN), also known as a multilayer perceptron (MLP). Our data-driven closure model is based on localized learning, and the FCNN architecture has a point-to-point mapping framework. The MLP-based SGS model is trained using direct numerical simulation (DNS) data of a turbulent channel flow at the bulk Reynolds number Re=4400, corresponding to the friction Reynolds number 〖 Re〗_τ=180. Performance of the FCCN model is assessed using a priori analysis in comparison with the filtered DNS (fDNS) data of the turbulent channel flow. Resolved flow statistics, including filtered velocity gradients and wall distance, are employed as input parameters for training the FCNN. To evaluate the significance of the feature inputs, we utilize random forest regression, which reveals that y^+ is a critical factor in predicting the output, which can reduce the compactional costs. In an a priori test, the model achieved a correlation coefficient exceeding 95% for the components of the SGS stress tensor, which have non-zero mean values. Finally, potential strategies for improving prediction accuracy and overall performance are explored. The findings demonstrate that the proposed FCNN can accurately reproduce key turbulence characteristics, offering a promising step toward efficient, AI-based turbulence modeling.