The Anatomical Therapeutic Chemical (ATC) system assigns unique codes to drugs for tracking, aiding in prescription, public health policies, and research. Different classifiers with distinct properties have been used earlier to predict ATC drug classes. However, Limited research has been conducted on 17 molecular properties and classifier transparency, emphasising the need of categorising each feature to increase model explainability. In this paper,a Transparent-ATC-DNN (Transparent ATC drug class prediction using Deep Neural Network) deep neural network and SHAP based framework was proposed for ATC drug class prediction and feature categorisation based on their contribution towards prediction. The 17 molecular properties and ATC data were extracted from PubChem and SIDER databases and then mapped based on drug identifier. The resulting dataset contained information on 1117 drugs, including their molecular properties, and 14 ATC drug classes. The proposed framework was applied to this dataset utilizing MLSMOTE to manage multilabel class imbalance and compared with eight classifiers, namely the ETC, KNN, SVM, RF, DT, XGBoost, LSTM, and 1D-CNN. The results of multilabel DNN architecture demonstrated promising performance, achieving an average accuracy of 98.89 ± .002, a hamming loss of 1.50 ± .002, and an ROC-AUC of 98.47 ± .003. Features were evaluated and categorized into three distinct categories: highest, lowest, and none, showcasing the proposed framework's effectiveness and explanatory power in predicting ATC drug classes.