In the drug discovery process, unstable compounds in storage can lead to false positive or false negative bioassay conclusions. Prediction of the chemical stability of a compound by de novo methods is complex. Chemical instability prediction is commonly based on a model derived from empirical data. The COMDECOM (COMpound DECOMposition) project provides the empirical data for prediction of chemical stability. Models such as the extended-connectivity fingerprint and atom center fragments were built from the COMDECOM data and used for estimation of chemical stability, but deficits in the existing models remain. In this paper, we report DeepChemStable, a model employing an attention-based graph convolution network based on the COMDECOM data. The main advantage of this method is that DeepChemStable is an end-to-end model, which does not predefine structural fingerprint features, but instead, dynamically learns structural features and associates the features through the learning process of an attention-based graph convolution network. The previous ChemStable program relied on a rule-based method to reduce the false negatives. DeepChemStable, on the other hand, reduces the risk of false negatives without using a rule-based method. Because minimizing the rate of false negatives is a greater concern for instability prediction, this feature is a major improvement. This model achieves an AUC value of 84.7%, recall rate of 79.8%, and 10-fold stratified cross-validation accuracy of 79.1%.