Abstract: Sugars are one of the reasons we feel better when eating. But eating too much high-sugar food can threaten body's health. Therefore, a heat issue is to create novel and safe non-nutritive sweeteners in recent years. However, the development and identification of sweeteners can also be a resource-consuming and long-term process. In this study, a feasible classification system of sweeteners was proposed using ligand and deep learning, in order to accurately sift sweetener candidates from numerous molecular libraries. Firstly, the SMILES strings in the database were imported after the dataset was extended, and then converted into two-dimensional graphics and extended-connectivity fingerprints (ECFP). A series of operations was selected to preprocess the generated two-dimensional images, including random brightness transformation, rotation, and flipping. Secondly, a deep learning model DenseNet was established for the molecular classification. The batch size of the training model was set to 64, and the initial learning rate was 0.005. A batch normalization (BN) was used in every layer of linear convolution to prevent overfitting and gradient extinction. A vector constantly attenuation function was also used for the learning rate, with the attenuation factor of 0.1 at every seven epochs. Thirdly, the attention mechanism was added to the back of each Dense Block module in the DenseNet, where a new layer was formed with the transition layer. After that, the generated ECFP matrix was fused with the extracted matrix. The classification task was finally completed to add two full-junction layers with different numbers of neurons and a Softmax classification layer. Two-dimensional data was the most suitable choice for rapidly sifting in the case of large datasets. Since some specific three-dimensional structures and theoretical difficulties were usually found in the sensory intensity of sweet taste in reality, it was necessary to use the feature fusion for better prediction performance. Finally, an independent test was conducted to verify the proposed model. The experimental results showed that the attention mechanism and feature fusion greatly improved the performance of the model. The model was obviously superior to most machine learning in all indicators, particularly for the tasks with big datasets. The highest accuracy rate was 0.934 (steady at around 0.934), while the fluctuation of loss value and accuracy was less than 0.001 and 0.005, respectively. The average accuracies of bitter, tasteless, weak sweet, and strong sweet classification were 0.96, 0.94, 0.91 and 0.92, respectively. In addition, cosine annealing was selected to adjust the learning rate for better performance of the model. As such, the learning rate was "jump out" of local optimum, and then found the direction of global optimum, because of the sudden increase of learning rate. A local optimal value was avoided because the learning rate generally remained to get smaller when the model was trained. The experiments showed that there was improved efficiency, indicating that the model was suitable for the sweet and bitter tastes. The finding can contribute to the high accuracy of sweetener prediction and the rapid development of low-calorie, even no-calorie sweeteners.