Application of 1D CNN based on KPLS weighting in ADMET classification

To solve the problem of low label sample size and redundant information between features in ADMET, this paper proposes a one-dimensional convolutional neural network based on kernel partial least squares (KPLS) weighting and adaptive increment (KPLS_CNN). Firstly, the training set is feature weighted using KPLS, and the weighted training set is inputted into a one-dimensional convolutional neural network (1D CNN) to obtain the initial weight values of the neurons in the network. Then, the weights are shared with the model of the test set and the classification results are calculated. Finally, by setting confidence levels, samples with higher confidence levels in the test set are added to the training set, and the training model is retrained to update the weights. After the updates, they are placed in the model of the test set. This model not only achieves the weighting of input data by KPLS, but also expands the training set when the sample size is insufficient, achieving weight sharing. We compared and analyzed the classification problem of ADMET using three models: KPLS_CNN, 1D CNN, and SVM. The experimental results showed that the KPLSWAI-CNN model had better final classification performance than other methods.