Abstract Predicting the properties of proteins is an important procedure in protein engineering. It determines the subspace of mutations for protein modifications, which is critical to the success of the project, but heavily relies on the knowledge and experience of scientists. In this study, we propose a novel deep 3D-CNN model, Eq3DCNN, specifically designed for local environment-related tasks in protein engineering. Eq3DCNN uses basic atom descriptors and their coordinates as inputs, utilizing customized data augmentations to enhance its training efficiency. To make the Eq3DCNN extracted features with more generalization capability, we incorporated a rotation equivariant module to get rotation invariant features. Using cross-validations with different data splitting strategies and under the scenarios of zero-shot predictions, we demonstrate that Eq3DCNN outperformed other 3D-CNN models in stability predictions, and also well-preformed on other prediction tasks, such as the binding pocket and the secondary structure predictions. Our results also identified the key factors that contribute to the model’s accuracy and the scope of its applications. These findings may help scientists in designing better mutation experiments and increasing the success rate in protein engineering.